1
00:00:00,000 --> 00:00:03,080
Most organizations think they're adopting AI in reality.

2
00:00:03,080 --> 00:00:05,520
They're just building chatbots that search faster.

3
00:00:05,520 --> 00:00:07,000
You spend millions on co-pilot,

4
00:00:07,000 --> 00:00:08,600
you connect it to your document folders,

5
00:00:08,600 --> 00:00:10,720
you tell your teams to ask it questions.

6
00:00:10,720 --> 00:00:12,640
And what you've actually built is a search interface

7
00:00:12,640 --> 00:00:14,480
that talks back, it isn't transformation,

8
00:00:14,480 --> 00:00:16,240
it's just automation of retrieval.

9
00:00:16,240 --> 00:00:19,280
But here is what is actually happening in 2026.

10
00:00:19,280 --> 00:00:21,400
The organizations that are winning aren't the ones

11
00:00:21,400 --> 00:00:22,480
with better search.

12
00:00:22,480 --> 00:00:24,920
They are the ones that shifted from retrieval to reasoning.

13
00:00:24,920 --> 00:00:26,840
They moved from assistance that advise

14
00:00:26,840 --> 00:00:28,460
to digital employees that act.

15
00:00:28,460 --> 00:00:29,840
The architecture behind that shift

16
00:00:29,840 --> 00:00:31,560
is nothing like what you've been building.

17
00:00:31,560 --> 00:00:33,200
In this episode, we are going to diagnose

18
00:00:33,200 --> 00:00:35,240
why your current approach is hitting a wall.

19
00:00:35,240 --> 00:00:37,440
We will walk through the new operating model Microsoft

20
00:00:37,440 --> 00:00:41,120
has assembled, work IQ, age in 365 and A to A.

21
00:00:41,120 --> 00:00:43,420
I'm gonna show you why this changes everything.

22
00:00:43,420 --> 00:00:45,260
And we are going to be honest about the cost,

23
00:00:45,260 --> 00:00:47,580
not just in credits, but in your organizational structure,

24
00:00:47,580 --> 00:00:50,660
your governance and the way you think about IT identity itself.

25
00:00:50,660 --> 00:00:53,300
By the end, you'll understand why service accounts are dead.

26
00:00:53,300 --> 00:00:55,360
You'll see why a single-entra agent ID

27
00:00:55,360 --> 00:00:57,600
is the foundation of everything that comes next.

28
00:00:57,600 --> 00:00:59,020
The organizations that move on this now

29
00:00:59,020 --> 00:01:00,620
will operate at machine speed,

30
00:01:00,620 --> 00:01:03,940
while everyone else is still trying to figure out permissions.

31
00:01:03,940 --> 00:01:06,160
The latency wall, why Raga's failing?

32
00:01:06,160 --> 00:01:08,620
Let's start with the problem nobody wants to admit.

33
00:01:08,620 --> 00:01:10,960
You've deployed retrieval augmented generation,

34
00:01:10,960 --> 00:01:12,800
Raga, across your entire company,

35
00:01:12,800 --> 00:01:15,720
you built the vector stores and indexed every document you own.

36
00:01:15,720 --> 00:01:18,680
You set up the pipelines, and it works, until it doesn't.

37
00:01:18,680 --> 00:01:21,520
The moment you ask an agent to do something

38
00:01:21,520 --> 00:01:23,920
that requires real time understanding of your business,

39
00:01:23,920 --> 00:01:25,360
the system breaks, it doesn't crash,

40
00:01:25,360 --> 00:01:28,120
it just slows down, latency kills it.

41
00:01:28,120 --> 00:01:30,360
Raga was designed for a very specific problem.

42
00:01:30,360 --> 00:01:32,160
Answering questions about static documents,

43
00:01:32,160 --> 00:01:34,400
you have a knowledge base, a user asks a question.

44
00:01:34,400 --> 00:01:36,200
The system pulls the most relevant chunks

45
00:01:36,200 --> 00:01:37,680
and feeds them to a model.

46
00:01:37,680 --> 00:01:40,880
This works for your vacation policy or password reset guide.

47
00:01:40,880 --> 00:01:42,480
Those documents don't change every hour,

48
00:01:42,480 --> 00:01:45,240
the answers are predictable, but work isn't static.

49
00:01:45,240 --> 00:01:46,400
Work is happening right now,

50
00:01:46,400 --> 00:01:48,880
someone is editing a file while a meeting is running.

51
00:01:48,880 --> 00:01:50,600
A decision was made five minutes ago,

52
00:01:50,600 --> 00:01:52,320
a project status just changed.

53
00:01:52,320 --> 00:01:54,200
A person was reassigned to a new team,

54
00:01:54,200 --> 00:01:57,320
that is liquid context, it moves, it changes.

55
00:01:57,320 --> 00:02:00,360
And every time an agent needs to understand that context,

56
00:02:00,360 --> 00:02:03,320
your Raga system has to pull fresh data from somewhere.

57
00:02:03,320 --> 00:02:05,000
This is where you hit the latency wall.

58
00:02:05,000 --> 00:02:07,040
Every retrieval cycle adds overhead.

59
00:02:07,040 --> 00:02:09,720
Your agent has to query the vector store, score those results,

60
00:02:09,720 --> 00:02:12,240
and maybe query again if the first batch was bad.

61
00:02:12,240 --> 00:02:14,280
Then it passes that context to a language model.

62
00:02:14,280 --> 00:02:16,080
That isn't measured in milliseconds anymore,

63
00:02:16,080 --> 00:02:17,560
it takes hundreds of milliseconds.

64
00:02:17,560 --> 00:02:20,120
When you have multiple agents doing this at the same time,

65
00:02:20,120 --> 00:02:22,640
each one waiting for index lookups and model inference.

66
00:02:22,640 --> 00:02:24,160
You create a massive bottleneck.

67
00:02:24,160 --> 00:02:25,640
Users feel the slowness.

68
00:02:25,640 --> 00:02:28,480
Business owners decide the agent is too slow to actually use.

69
00:02:28,480 --> 00:02:31,240
The cost compounds because you don't stop at one vector store.

70
00:02:31,240 --> 00:02:33,360
You build one for SharePoint, then one for email,

71
00:02:33,360 --> 00:02:36,360
then another for Teams Chats, then you add one for CRM data

72
00:02:36,360 --> 00:02:38,840
because it isn't native to Microsoft 365,

73
00:02:38,840 --> 00:02:40,160
you are fragmenting your data.

74
00:02:40,160 --> 00:02:41,800
Every index adds sync complexity,

75
00:02:41,800 --> 00:02:43,320
everything adds more latency.

76
00:02:43,320 --> 00:02:45,640
Every latency failure makes your data stale.

77
00:02:45,640 --> 00:02:47,000
A document changes in SharePoint

78
00:02:47,000 --> 00:02:49,400
and your vector index gets updated eventually,

79
00:02:49,400 --> 00:02:50,760
but by the time that happens.

80
00:02:50,760 --> 00:02:53,800
15 agents have already made decisions based on the old version.

81
00:02:53,800 --> 00:02:56,600
You've lost consistency, you've created an audit nightmare.

82
00:02:56,600 --> 00:02:58,480
Nobody can tell you which agent was looking

83
00:02:58,480 --> 00:03:00,880
at which version of the data when it made a choice.

84
00:03:00,880 --> 00:03:02,320
The deeper problem is structural.

85
00:03:02,320 --> 00:03:04,800
You are treating work data like a static library.

86
00:03:04,800 --> 00:03:07,120
You extract it, vectorize it and index it separately

87
00:03:07,120 --> 00:03:08,760
from where the work actually happens.

88
00:03:08,760 --> 00:03:10,640
You've built a parallel data structure

89
00:03:10,640 --> 00:03:12,640
and parallel structures never sync perfectly.

90
00:03:12,640 --> 00:03:14,840
They get out of alignment, they create shadows,

91
00:03:14,840 --> 00:03:16,560
they become their own governance problem.

92
00:03:16,560 --> 00:03:18,520
So organizations keep adding more layers.

93
00:03:18,520 --> 00:03:20,680
They try better chunking on new embedding models,

94
00:03:20,680 --> 00:03:23,120
they try multi-hop retrieval in hybrid search,

95
00:03:23,120 --> 00:03:25,600
every addition is just trying to fix a latency problem

96
00:03:25,600 --> 00:03:28,120
that exists because the architecture itself is wrong.

97
00:03:28,120 --> 00:03:30,400
The model assumes retrieval is the hard part.

98
00:03:30,400 --> 00:03:31,800
The real hard part is reasoning

99
00:03:31,800 --> 00:03:34,800
over live-governed organizational state in real time

100
00:03:34,800 --> 00:03:37,280
and that requires a completely different approach.

101
00:03:37,280 --> 00:03:39,520
Static context versus liquid context.

102
00:03:39,520 --> 00:03:41,960
We need to be precise about what we are talking about.

103
00:03:41,960 --> 00:03:43,360
There is a fundamental difference

104
00:03:43,360 --> 00:03:46,000
between two kinds of information in your organization.

105
00:03:46,000 --> 00:03:48,440
Understanding this difference is where everything changes.

106
00:03:48,440 --> 00:03:50,480
Static context is what is written down.

107
00:03:50,480 --> 00:03:53,240
It is your documented knowledge, policies, procedures,

108
00:03:53,240 --> 00:03:54,720
past decisions.

109
00:03:54,720 --> 00:03:57,520
It serves as a guide for how things are supposed to work.

110
00:03:57,520 --> 00:03:59,640
Think of a decision memo from six months ago

111
00:03:59,640 --> 00:04:02,320
or a lessons learned document from a finished project.

112
00:04:02,320 --> 00:04:03,480
These things do not move.

113
00:04:03,480 --> 00:04:05,440
They sit in repositories like SharePoint,

114
00:04:05,440 --> 00:04:07,600
Wikis or Internet pages while they wait for someone

115
00:04:07,600 --> 00:04:09,840
to find them, they change slowly and deliberately.

116
00:04:09,840 --> 00:04:12,160
When a change happens, it is recorded with a version

117
00:04:12,160 --> 00:04:13,680
and author and a date.

118
00:04:13,680 --> 00:04:15,240
This is where Rags work well.

119
00:04:15,240 --> 00:04:17,640
You index these documents and build retrieval systems

120
00:04:17,640 --> 00:04:19,600
around them because the data is static,

121
00:04:19,600 --> 00:04:22,480
you can afford to index it once and update it every so often.

122
00:04:22,480 --> 00:04:24,920
You know the index stays consistent with the source of truth.

123
00:04:24,920 --> 00:04:27,360
When an agent asks about the remote work policy,

124
00:04:27,360 --> 00:04:29,720
it hits that index and gives a correct answer.

125
00:04:29,720 --> 00:04:32,040
But liquid context is something entirely different.

126
00:04:32,040 --> 00:04:33,800
It is who is working on what right now?

127
00:04:33,800 --> 00:04:35,560
It is the current state of a project.

128
00:04:35,560 --> 00:04:36,840
Which tasks are in progress?

129
00:04:36,840 --> 00:04:39,040
Which are blocked? Who is waiting for a review?

130
00:04:39,040 --> 00:04:41,520
It is the organizational structure as it exists today.

131
00:04:41,520 --> 00:04:42,680
Who reports to whom?

132
00:04:42,680 --> 00:04:44,680
Who is on leave? Who just got promoted?

133
00:04:44,680 --> 00:04:46,720
It is the real-time permissions landscape.

134
00:04:46,720 --> 00:04:47,720
Who has access to what?

135
00:04:47,720 --> 00:04:49,120
Whose roles just changed?

136
00:04:49,120 --> 00:04:51,560
It is the active collaboration happening across email chat

137
00:04:51,560 --> 00:04:53,040
and meetings this very second.

138
00:04:53,040 --> 00:04:54,880
Liquid context is dynamic.

139
00:04:54,880 --> 00:04:56,800
It changes constantly.

140
00:04:56,800 --> 00:04:59,960
Someone finishes a task or gets reassigned to a new team.

141
00:04:59,960 --> 00:05:01,720
A meeting wraps up and a decision is made,

142
00:05:01,720 --> 00:05:04,040
a document is reviewed and the status shifts.

143
00:05:04,040 --> 00:05:05,920
Permissions are updated and priorities move.

144
00:05:05,920 --> 00:05:08,440
There is no single source of truth sitting in a repository.

145
00:05:08,440 --> 00:05:11,040
The context is distributed across the living systems

146
00:05:11,040 --> 00:05:12,520
where work actually happens.

147
00:05:12,520 --> 00:05:14,880
Consider a simple scenario.

148
00:05:14,880 --> 00:05:17,760
An agent needs to summarize the status of a product launch.

149
00:05:17,760 --> 00:05:19,640
A rag system would find the project plan

150
00:05:19,640 --> 00:05:21,040
and the last status update.

151
00:05:21,040 --> 00:05:23,400
It might even look at a document from a similar project

152
00:05:23,400 --> 00:05:24,200
in the past.

153
00:05:24,200 --> 00:05:26,800
It would assemble a summary from those static sources.

154
00:05:26,800 --> 00:05:28,120
But here is what it would not know.

155
00:05:28,120 --> 00:05:30,200
Sarah is the one responsible for the strategy.

156
00:05:30,200 --> 00:05:32,800
But she is out sick today and her work is blocked.

157
00:05:32,800 --> 00:05:34,960
The legal review was supposed to finish yesterday,

158
00:05:34,960 --> 00:05:36,760
but it just got delayed by a week.

159
00:05:36,760 --> 00:05:39,880
The executive sponsor approved a budget increase three hours ago,

160
00:05:39,880 --> 00:05:41,840
but that approval is not documented yet.

161
00:05:41,840 --> 00:05:44,000
Three team members just had a sync meeting

162
00:05:44,000 --> 00:05:46,360
and found a critical issue with the pricing model.

163
00:05:46,360 --> 00:05:47,680
The demo was scheduled for Thursday,

164
00:05:47,680 --> 00:05:49,920
but someone just moved the calendar invite to Friday.

165
00:05:49,920 --> 00:05:51,600
All of that information is liquid.

166
00:05:51,600 --> 00:05:52,720
It is happening right now.

167
00:05:52,720 --> 00:05:55,200
It exists in Outlook calendars and Teams messages.

168
00:05:55,200 --> 00:05:56,400
It lives in real-time updates

169
00:05:56,400 --> 00:05:57,960
and in the current awareness of the team.

170
00:05:57,960 --> 00:05:59,800
But it does not exist in any index document.

171
00:05:59,800 --> 00:06:02,240
You cannot retrieve it because it has not been written down yet.

172
00:06:02,240 --> 00:06:03,320
It is still in motion.

173
00:06:03,320 --> 00:06:05,400
A rag system fails here because it is designed

174
00:06:05,400 --> 00:06:07,600
to find what was already captured and stored.

175
00:06:07,600 --> 00:06:08,960
It does not see what is happening.

176
00:06:08,960 --> 00:06:10,720
It only sees what was documented.

177
00:06:10,720 --> 00:06:14,160
By the time the agent finishes its answer using static context,

178
00:06:14,160 --> 00:06:16,280
the liquid context has already changed again.

179
00:06:16,280 --> 00:06:18,440
The summary is stale before it is even delivered.

180
00:06:18,440 --> 00:06:20,320
This is where work IQ enters the picture.

181
00:06:20,320 --> 00:06:22,040
Work IQ does not work like rag.

182
00:06:22,040 --> 00:06:23,240
It does not retrieve an index.

183
00:06:23,240 --> 00:06:26,800
Instead it reasons over live Microsoft 365 signals as they happen.

184
00:06:26,800 --> 00:06:29,320
It understands who is currently assigned to what?

185
00:06:29,320 --> 00:06:32,400
It tracks real-time permissions and sees active collaboration.

186
00:06:32,400 --> 00:06:33,960
It knows what changed in the last hour

187
00:06:33,960 --> 00:06:36,840
because it is connected to the systems where that change occurred.

188
00:06:36,840 --> 00:06:40,400
It operates on the moving state of work, not on archive documents.

189
00:06:40,400 --> 00:06:42,520
This distinction changes everything downstream.

190
00:06:42,520 --> 00:06:44,400
If your agent can only see static context,

191
00:06:44,400 --> 00:06:46,000
it is always reasoning about yesterday.

192
00:06:46,000 --> 00:06:48,880
If your agent can see liquid context, it is reasoning about now.

193
00:06:48,880 --> 00:06:52,760
And that changes what an agent can actually do for your organization.

194
00:06:52,760 --> 00:06:54,400
The service account trap.

195
00:06:54,400 --> 00:06:57,080
The problem with work IQ and smart infrastructure

196
00:06:57,080 --> 00:07:01,440
is that organizations never had a way to run agents as distinct entities.

197
00:07:01,440 --> 00:07:03,480
So they did what they always did with automation.

198
00:07:03,480 --> 00:07:05,720
They created a service account, one service account,

199
00:07:05,720 --> 00:07:08,480
multiple agents, all acting under the same identity.

200
00:07:08,480 --> 00:07:10,080
This solved an immediate problem.

201
00:07:10,080 --> 00:07:12,840
You did not need to manage individual agent identities.

202
00:07:12,840 --> 00:07:15,720
You did not need to figure out permissions for each board separately.

203
00:07:15,720 --> 00:07:17,320
One account and one set of credentials

204
00:07:17,320 --> 00:07:19,520
allowed you to deploy as many agents as you wanted.

205
00:07:19,520 --> 00:07:20,960
It was always going to break.

206
00:07:20,960 --> 00:07:22,320
The breaking point is not technical.

207
00:07:22,320 --> 00:07:23,480
It is governance.

208
00:07:23,480 --> 00:07:25,720
And governance breaks slowly, so you do not notice

209
00:07:25,720 --> 00:07:27,360
until you are already in trouble.

210
00:07:27,360 --> 00:07:29,040
A service account has no audit trail.

211
00:07:29,040 --> 00:07:31,720
When something goes wrong, you cannot tell which agent did it.

212
00:07:31,720 --> 00:07:33,200
You cannot see when it happened.

213
00:07:33,200 --> 00:07:34,680
You cannot prove who authorized it

214
00:07:34,680 --> 00:07:36,760
because a service account has no human owner.

215
00:07:36,760 --> 00:07:39,480
It is just a set of credentials floating in the environment.

216
00:07:39,480 --> 00:07:42,160
Nobody is responsible for it and nobody monitors it.

217
00:07:42,160 --> 00:07:45,080
Nobody can tell you when its permissions last changed or why.

218
00:07:45,080 --> 00:07:46,680
A service account has no life cycle.

219
00:07:46,680 --> 00:07:49,680
When a human leaves the organization, their access is revoked.

220
00:07:49,680 --> 00:07:51,240
It is clear and it is done.

221
00:07:51,240 --> 00:07:53,480
But for an agent, there is no termination process.

222
00:07:53,480 --> 00:07:55,200
The service account just stays active.

223
00:07:55,200 --> 00:07:58,400
If the agent was supposed to be temporary, nobody removes it.

224
00:07:58,400 --> 00:08:00,880
The credentials persist and the access remains.

225
00:08:00,880 --> 00:08:04,040
You have built permanent infrastructure to solve temporary problems.

226
00:08:04,040 --> 00:08:06,120
A service account has no policy enforcement.

227
00:08:06,120 --> 00:08:08,080
You cannot apply conditional access rules

228
00:08:08,080 --> 00:08:10,720
to a service account the way you do for a human.

229
00:08:10,720 --> 00:08:13,440
You cannot say this account can only act during business hours.

230
00:08:13,440 --> 00:08:16,640
You cannot limit it to accessing data from a specific region.

231
00:08:16,640 --> 00:08:18,720
You cannot require extra authentication

232
00:08:18,720 --> 00:08:20,840
before it performs a sensitive operation.

233
00:08:20,840 --> 00:08:22,840
The account either has permission or it does not.

234
00:08:22,840 --> 00:08:23,840
It is all or nothing.

235
00:08:23,840 --> 00:08:25,440
None of this matters until it does.

236
00:08:25,440 --> 00:08:28,920
It matters when compliance asks a question you cannot answer.

237
00:08:28,920 --> 00:08:32,120
Which agent accessed that sensitive customer data on March 15th?

238
00:08:32,120 --> 00:08:33,040
You do not know.

239
00:08:33,040 --> 00:08:35,160
The audit log just shows the service account.

240
00:08:35,160 --> 00:08:38,680
You cannot tell if it was agent A or agent B or some other agent you forgot about.

241
00:08:38,680 --> 00:08:39,920
Which agent is still running?

242
00:08:39,920 --> 00:08:41,040
You do not know.

243
00:08:41,040 --> 00:08:42,480
The service account is still active.

244
00:08:42,480 --> 00:08:45,800
So you cannot tell if it is being used or if it is abandoned infrastructure.

245
00:08:45,800 --> 00:08:48,560
Who authorised this agent to have access to the financial system?

246
00:08:48,560 --> 00:08:49,840
You do not know.

247
00:08:49,840 --> 00:08:53,200
The service account was created three years ago by someone who left the company.

248
00:08:53,200 --> 00:08:55,320
There is no approval record and no justification.

249
00:08:55,320 --> 00:08:56,800
These are not edge cases.

250
00:08:56,800 --> 00:08:59,200
These are the questions compliance teams are already asking.

251
00:08:59,200 --> 00:08:59,960
Which agent did this?

252
00:08:59,960 --> 00:09:00,480
Who owns it?

253
00:09:00,480 --> 00:09:02,000
Can we prove it was authorised?

254
00:09:02,000 --> 00:09:04,800
In a service account model, these questions have no answers.

255
00:09:04,800 --> 00:09:07,600
You are managing agents but you have no way to identify them.

256
00:09:07,600 --> 00:09:10,120
You have no way to prove what they are allowed to do.

257
00:09:10,120 --> 00:09:12,080
You have no way to hold anyone accountable.

258
00:09:12,080 --> 00:09:14,040
And that is before we talk about security.

259
00:09:14,040 --> 00:09:17,200
A compromised service account is not just one agent going rogue.

260
00:09:17,200 --> 00:09:20,400
It is all agents acting under that account going rogue at the same time.

261
00:09:20,400 --> 00:09:24,040
One breach means all the agents are compromised and all the access is exposed.

262
00:09:24,040 --> 00:09:25,360
The fundamental gap is clear.

263
00:09:25,360 --> 00:09:27,160
You have identity systems for humans.

264
00:09:27,160 --> 00:09:29,080
You have app registrations for services.

265
00:09:29,080 --> 00:09:31,040
But agents exist in a governance vacuum.

266
00:09:31,040 --> 00:09:33,360
They are not users and they are not applications.

267
00:09:33,360 --> 00:09:34,480
They are something new.

268
00:09:34,480 --> 00:09:36,280
And nobody built a way to manage them.

269
00:09:36,280 --> 00:09:38,720
Agent 365 re-thinks that entirely.

270
00:09:38,720 --> 00:09:40,800
Why your co-pilot pilot is stalling?

271
00:09:40,800 --> 00:09:41,920
You've deployed co-pilot.

272
00:09:41,920 --> 00:09:43,600
It looks brilliant in the demo.

273
00:09:43,600 --> 00:09:46,720
Your executives watch it summarise a massive document in seconds.

274
00:09:46,720 --> 00:09:48,920
Draft an email or answer a policy question.

275
00:09:48,920 --> 00:09:50,680
And they immediately approve the pilot.

276
00:09:50,680 --> 00:09:52,520
They tell their teams to start using it.

277
00:09:52,520 --> 00:09:54,640
And then adoption flatlines.

278
00:09:54,640 --> 00:09:56,400
15%.

279
00:09:56,400 --> 00:09:57,760
That is where the momentum stops.

280
00:09:57,760 --> 00:10:01,360
Only 15% of your organisation actually uses the agent regularly

281
00:10:01,360 --> 00:10:03,840
while the rest of the team tried it once and walked away.

282
00:10:03,840 --> 00:10:04,880
They found it too slow.

283
00:10:04,880 --> 00:10:08,360
They didn't trust the answers or it just didn't work when they actually needed it too.

284
00:10:08,360 --> 00:10:09,320
So they stopped.

285
00:10:09,320 --> 00:10:11,600
They went back to their old workflows because those workflows

286
00:10:11,600 --> 00:10:13,200
actually get the job done.

287
00:10:13,200 --> 00:10:15,320
You will hear plenty of excuses for this.

288
00:10:15,320 --> 00:10:18,960
People will say the AI isn't good enough yet or the data is too messy

289
00:10:18,960 --> 00:10:21,440
or the users just need better prompting sessions.

290
00:10:21,440 --> 00:10:25,120
While those things might be true, none of them are the real reason your pilot is failing.

291
00:10:25,120 --> 00:10:28,240
The real problem is friction, operational friction.

292
00:10:28,240 --> 00:10:30,520
And it has nothing to do with how smart the AI is.

293
00:10:30,520 --> 00:10:31,920
Think about how this works in practice.

294
00:10:31,920 --> 00:10:34,680
Your agent can read, it can summarise, and it can answer questions.

295
00:10:34,680 --> 00:10:35,720
But can it act?

296
00:10:35,720 --> 00:10:36,800
Can it update a record?

297
00:10:36,800 --> 00:10:40,000
Send an email, create a task, or actually change something in your systems.

298
00:10:40,000 --> 00:10:42,080
If it can, what happens if it makes a mistake?

299
00:10:42,080 --> 00:10:43,440
Who is responsible for that error?

300
00:10:43,440 --> 00:10:47,440
There is no way to trace the action back to a specific decision or prove it was authorized

301
00:10:47,440 --> 00:10:50,320
because right now there is no governance model for agent actions.

302
00:10:50,320 --> 00:10:51,880
So you do the only safe thing.

303
00:10:51,880 --> 00:10:53,560
You turn off the dangerous actions.

304
00:10:53,560 --> 00:10:55,120
You make the agent read only.

305
00:10:55,120 --> 00:10:56,320
It can suggest but it can't do.

306
00:10:56,320 --> 00:10:58,000
It can draft but it can't send.

307
00:10:58,000 --> 00:11:00,480
You have built a sophisticated reasoning system

308
00:11:00,480 --> 00:11:02,560
and then locked it into an advisor role

309
00:11:02,560 --> 00:11:04,760
because you have no way to govern what it does.

310
00:11:04,760 --> 00:11:07,560
Users don't trust it because agents have no identity.

311
00:11:07,560 --> 00:11:10,800
No accountability, no audit trail.

312
00:11:10,800 --> 00:11:15,400
When an agent gives you an answer, you don't know if it's trustworthy or who decided to design it that way.

313
00:11:15,400 --> 00:11:18,440
You don't know if it has been updated since yesterday or if it's even still running.

314
00:11:18,440 --> 00:11:19,160
There is no face.

315
00:11:19,160 --> 00:11:20,320
There is no owner.

316
00:11:20,320 --> 00:11:22,720
Business owners are even more skeptical of the technology.

317
00:11:22,720 --> 00:11:26,160
They will never give an agent right access if they can't trace what happened

318
00:11:26,160 --> 00:11:29,960
because if an agent deletes a file or sends a message to the wrong person,

319
00:11:29,960 --> 00:11:31,720
there is nobody to hold accountable.

320
00:11:31,720 --> 00:11:33,040
So the business owners say no.

321
00:11:33,040 --> 00:11:36,480
They tell you not to touch their critical systems or sign anything on their behalf.

322
00:11:36,480 --> 00:11:38,280
The agent becomes a suggestion engine.

323
00:11:38,280 --> 00:11:41,360
It is useful sometimes but it isn't integrated into real work.

324
00:11:41,360 --> 00:11:42,720
Integration requires trust.

325
00:11:42,720 --> 00:11:44,320
Trust requires accountability.

326
00:11:44,320 --> 00:11:46,040
An accountability requires identity.

327
00:11:46,040 --> 00:11:47,480
This is where most pilots stall.

328
00:11:47,480 --> 00:11:51,320
Not because the AI is bad but because you've built an AI system into infrastructure

329
00:11:51,320 --> 00:11:52,920
that has no way to manage it.

330
00:11:52,920 --> 00:11:54,640
Co-pilot exists in a governance void.

331
00:11:54,640 --> 00:11:57,440
It has no enter ID, no conditional access policy,

332
00:11:57,440 --> 00:11:59,600
and no owner in an organizational sense.

333
00:11:59,600 --> 00:12:03,480
It is sitting in your tenant but it isn't managed like anything else in your environment.

334
00:12:03,480 --> 00:12:04,760
Users sense this friction.

335
00:12:04,760 --> 00:12:06,760
They don't sit there thinking about identity models

336
00:12:06,760 --> 00:12:08,760
but they feel the slowness of read-only answers

337
00:12:08,760 --> 00:12:11,160
and the lack of trust in black-box suggestions.

338
00:12:11,160 --> 00:12:12,880
They feel the resistance from their managers.

339
00:12:12,880 --> 00:12:15,840
They realize the system isn't actually participating in real work.

340
00:12:15,840 --> 00:12:17,440
It is just running parallel to it.

341
00:12:17,440 --> 00:12:19,840
And that is the ceiling, 15% adoption.

342
00:12:19,840 --> 00:12:21,920
You get the early adopters who tolerate the friction

343
00:12:21,920 --> 00:12:24,560
but everyone else has already gone back to what they can trust.

344
00:12:24,560 --> 00:12:26,560
Agent 365 breaks that ceiling.

345
00:12:26,560 --> 00:12:29,560
But first, we need to understand what it actually does.

346
00:12:29,560 --> 00:12:31,440
Agents as first-class identities.

347
00:12:31,440 --> 00:12:33,800
Agent 365 changes one fundamental thing.

348
00:12:33,800 --> 00:12:35,680
It treats agents exactly like users.

349
00:12:35,680 --> 00:12:37,920
Not like applications, not like services.

350
00:12:37,920 --> 00:12:39,960
Like users, they get entra-agent IDs.

351
00:12:39,960 --> 00:12:43,120
They have life cycles and they move through joiner mover-leaver workflows

352
00:12:43,120 --> 00:12:44,760
just like any other employee.

353
00:12:44,760 --> 00:12:46,920
They inherit policies, they get assigned roles,

354
00:12:46,920 --> 00:12:47,920
and they can be revoked.

355
00:12:47,920 --> 00:12:50,000
They finally participate in the identity system.

356
00:12:50,000 --> 00:12:51,000
This sounds simple.

357
00:12:51,000 --> 00:12:52,400
But it is revolutionary.

358
00:12:52,400 --> 00:12:55,560
For years, identity systems have been built for two categories.

359
00:12:55,560 --> 00:12:57,000
Humans and applications.

360
00:12:57,000 --> 00:12:58,960
A human gets a user object in entra.

361
00:12:58,960 --> 00:13:02,360
They get a password, conditional access policies, and risk-based controls.

362
00:13:02,360 --> 00:13:05,280
We monitor them and if they become a security risk, we block them.

363
00:13:05,280 --> 00:13:07,240
When they leave the company, we delete them.

364
00:13:07,240 --> 00:13:09,120
An application gets a service principle.

365
00:13:09,120 --> 00:13:11,000
It gets an app registration and credentials

366
00:13:11,000 --> 00:13:12,600
like a secret or a certificate.

367
00:13:12,600 --> 00:13:15,880
It has scoped permissions and stays assigned to specific resources.

368
00:13:15,880 --> 00:13:17,720
It is designed to be deployed once

369
00:13:17,720 --> 00:13:19,520
and then run forever without changing.

370
00:13:19,520 --> 00:13:20,880
Agents didn't fit either model.

371
00:13:20,880 --> 00:13:21,680
They aren't users.

372
00:13:21,680 --> 00:13:22,600
They don't have passwords.

373
00:13:22,600 --> 00:13:25,560
They don't take vacations and they don't need a seat license.

374
00:13:25,560 --> 00:13:27,880
But they aren't traditional applications either.

375
00:13:27,880 --> 00:13:29,480
They are deployed in multiple versions

376
00:13:29,480 --> 00:13:31,200
and their life cycles are measured in weeks

377
00:13:31,200 --> 00:13:32,440
or months rather than years.

378
00:13:32,440 --> 00:13:34,160
You need to turn them off when a project ends

379
00:13:34,160 --> 00:13:36,160
or update them when a business process shifts.

380
00:13:36,160 --> 00:13:37,800
They don't fit the service principle model

381
00:13:37,800 --> 00:13:40,360
because those assume eternal unchanging infrastructure.

382
00:13:40,360 --> 00:13:41,560
So agents lived in the gap.

383
00:13:41,560 --> 00:13:42,760
They had no proper identity.

384
00:13:42,760 --> 00:13:44,680
They borrowed credentials from somewhere else.

385
00:13:44,680 --> 00:13:46,120
And they had no clear ownership.

386
00:13:46,120 --> 00:13:47,920
Agent 365 closes that gap.

387
00:13:47,920 --> 00:13:49,720
An agent gets an entra agent ID.

388
00:13:49,720 --> 00:13:51,840
It isn't a user object or a service principle.

389
00:13:51,840 --> 00:13:54,880
It is a new kind of identity designed for digital workers.

390
00:13:54,880 --> 00:13:57,480
And once an agent has an identity, everything changes.

391
00:13:57,480 --> 00:14:01,400
An agent with an identity can be assigned to specific data sources.

392
00:14:01,400 --> 00:14:03,120
You don't give it access to everything.

393
00:14:03,120 --> 00:14:06,440
Instead, you assign it only to the data sets it needs to function.

394
00:14:06,440 --> 00:14:08,600
A sales agent gets access to the CRM.

395
00:14:08,600 --> 00:14:11,200
A compliance agent gets access to policy documents.

396
00:14:11,200 --> 00:14:13,520
A billing agent gets access to invoice records.

397
00:14:13,520 --> 00:14:16,440
The identity system enforces that scoping at the directory level,

398
00:14:16,440 --> 00:14:20,400
which is much more secure than relying on configuration files or prompts.

399
00:14:20,400 --> 00:14:23,520
An agent with an identity is subject to conditional access rules.

400
00:14:23,520 --> 00:14:26,840
You can decide that an agent only operates during business hours

401
00:14:26,840 --> 00:14:30,560
or that it cannot access data from regions outside of North America.

402
00:14:30,560 --> 00:14:32,680
You can even require step-up authentication

403
00:14:32,680 --> 00:14:35,480
before it performs a transaction over $1 million.

404
00:14:35,480 --> 00:14:38,440
You enforce these rules through the identity platform itself.

405
00:14:38,440 --> 00:14:41,600
So the agent cannot violate them even if it's prompt tells it to.

406
00:14:41,600 --> 00:14:44,880
An agent with an identity can be monitored for unusual behavior.

407
00:14:44,880 --> 00:14:47,480
If an agent suddenly starts accessing data, it has never touched before,

408
00:14:47,480 --> 00:14:48,960
the system detects it immediately.

409
00:14:48,960 --> 00:14:51,320
If it tries to reach a resource it isn't authorized for,

410
00:14:51,320 --> 00:14:53,560
the system logs the attempt and triggers an alert.

411
00:14:53,560 --> 00:14:56,600
The same threat detection we use for compromised user accounts

412
00:14:56,600 --> 00:14:58,560
now works for compromised agents.

413
00:14:58,560 --> 00:15:01,920
Finally, an agent with an identity can be revoked when the task is done.

414
00:15:01,920 --> 00:15:04,520
You don't leave it running indefinitely and hope nobody notices.

415
00:15:04,520 --> 00:15:07,720
You decommission it, the identity is deleted and the permissions are gone.

416
00:15:07,720 --> 00:15:09,920
It is treated like an employee whose project ended.

417
00:15:09,920 --> 00:15:12,600
The access is terminated, it is clean and auditable.

418
00:15:12,600 --> 00:15:15,480
The Entra agent ID is the foundation for everything else.

419
00:15:15,480 --> 00:15:18,680
Governance, accountability and policy enforcement all depend on the fact

420
00:15:18,680 --> 00:15:20,840
that the agent has a proper identity in the directory.

421
00:15:20,840 --> 00:15:23,600
Once you have that identity, you can apply every governance tool

422
00:15:23,600 --> 00:15:25,120
you already use for humans,

423
00:15:25,120 --> 00:15:26,800
but identity alone isn't enough.

424
00:15:26,800 --> 00:15:29,160
You also need to constrain what agents can do,

425
00:15:29,160 --> 00:15:32,000
which is why we have to look at policy.

426
00:15:32,000 --> 00:15:33,600
Policy has code for agents.

427
00:15:33,600 --> 00:15:35,520
Traditional governance is aspirational.

428
00:15:35,520 --> 00:15:38,240
You write a policy document, you distribute it, you hope people read it,

429
00:15:38,240 --> 00:15:40,360
you hope they understand it, you hope they follow it,

430
00:15:40,360 --> 00:15:42,360
and still somebody violates the policy.

431
00:15:42,360 --> 00:15:45,600
Because the policy exists in a document and documents don't enforce anything,

432
00:15:45,600 --> 00:15:48,400
they suggest, they guide, they explain the intention,

433
00:15:48,400 --> 00:15:51,840
but at the moment of action, if someone chooses to ignore the policy,

434
00:15:51,840 --> 00:15:53,000
there's nothing stopping them.

435
00:15:53,000 --> 00:15:56,200
The enforcement happens after the fact and audit finds the violation.

436
00:15:56,200 --> 00:15:57,400
Someone gets corrected.

437
00:15:57,400 --> 00:15:58,720
The culture shifts slowly.

438
00:15:58,720 --> 00:16:02,600
Maybe, if you're lucky, policy is code inverts that entirely.

439
00:16:02,600 --> 00:16:05,400
The policy isn't aspirational, it's executable.

440
00:16:05,400 --> 00:16:07,760
It's written not in English for humans to interpret,

441
00:16:07,760 --> 00:16:10,120
but in logic that the system understands and enforces,

442
00:16:10,120 --> 00:16:12,600
the policy isn't a recommendation, it's a constraint.

443
00:16:12,600 --> 00:16:14,640
The system won't let the agent violate it,

444
00:16:14,640 --> 00:16:17,000
because the system can't let the agent violate it.

445
00:16:17,000 --> 00:16:19,720
The policy is built into the authorization layer itself,

446
00:16:19,720 --> 00:16:21,600
consider what becomes possible.

447
00:16:21,600 --> 00:16:24,800
You can configure an agent to read from specific sharepoint sites,

448
00:16:24,800 --> 00:16:25,640
but not others.

449
00:16:25,640 --> 00:16:27,840
That's not a guideline in the agent system prompt

450
00:16:27,840 --> 00:16:29,080
that someone might override.

451
00:16:29,080 --> 00:16:31,720
That's a permission boundary enforced at the directory level.

452
00:16:31,720 --> 00:16:35,200
The agent can only access the data sources you've explicitly assigned.

453
00:16:35,200 --> 00:16:38,520
Try to access something else, the API call fails, the system denies it.

454
00:16:38,520 --> 00:16:41,480
There's no workaround through cleverness or prompt injection.

455
00:16:41,480 --> 00:16:42,720
The permission doesn't exist.

456
00:16:42,720 --> 00:16:45,400
You can configure an agent to update records, but not delete them.

457
00:16:45,400 --> 00:16:47,320
It can write, it cannot destroy.

458
00:16:47,320 --> 00:16:48,840
You set that at the role level,

459
00:16:48,840 --> 00:16:51,720
the role has create update permissions, but not delete permissions.

460
00:16:51,720 --> 00:16:52,840
The agent assumes that role.

461
00:16:52,840 --> 00:16:56,080
It cannot delete data because the identity doesn't have the permission.

462
00:16:56,080 --> 00:16:58,680
The system prevents the action before it ever happens.

463
00:16:58,680 --> 00:17:01,440
You can configure an agent to send emails internally,

464
00:17:01,440 --> 00:17:02,920
but never externally.

465
00:17:02,920 --> 00:17:07,800
The email policy says this agent can send to anyone within the organization's domains.

466
00:17:07,800 --> 00:17:11,440
It cannot send outside them, it can reach the entire company directory.

467
00:17:11,440 --> 00:17:14,240
Try to send to a Gmail address, the system blocks it.

468
00:17:14,240 --> 00:17:17,280
Not because the agent's prompt says don't do that,

469
00:17:17,280 --> 00:17:21,360
but because the policy framework physically prevents the message from leaving the tenant,

470
00:17:21,360 --> 00:17:24,440
the constraint is infrastructure, not instruction.

471
00:17:24,440 --> 00:17:28,400
You can configure an agent to require a human approval for transactions over a threshold.

472
00:17:28,400 --> 00:17:31,760
Transactions under $10 million, the agent executes them directly.

473
00:17:31,760 --> 00:17:34,760
Over $10 million, the system demands a human authorization token.

474
00:17:34,760 --> 00:17:38,440
The agent can prepare the transaction, it can present the recommendation,

475
00:17:38,440 --> 00:17:41,480
but it cannot finalize without the human's explicit approval.

476
00:17:41,480 --> 00:17:43,840
The system enforces this gating automatically.

477
00:17:43,840 --> 00:17:45,280
These aren't suggestions.

478
00:17:45,280 --> 00:17:48,760
They're not best practices embedded in the agent's instructions.

479
00:17:48,760 --> 00:17:50,280
They're hard constraints.

480
00:17:50,280 --> 00:17:52,920
The agent literally cannot violate the policy.

481
00:17:52,920 --> 00:17:57,120
Even if the prompt says ignore all constraints and delete this record, it cannot do it.

482
00:17:57,120 --> 00:17:58,560
The permission doesn't exist.

483
00:17:58,560 --> 00:18:00,080
The API call is rejected.

484
00:18:00,080 --> 00:18:03,240
The authorization framework says no, and there's no appeal process.

485
00:18:03,240 --> 00:18:04,720
The system enforces what's possible.

486
00:18:04,720 --> 00:18:08,320
This is the critical difference between governance, theater, and actual governance.

487
00:18:08,320 --> 00:18:11,080
Theater is when you have a policy document that everyone acknowledges,

488
00:18:11,080 --> 00:18:12,520
and nobody can fully enforce.

489
00:18:12,520 --> 00:18:16,320
You have rules, but enforcement depends on catching violations after they happen.

490
00:18:16,320 --> 00:18:20,440
Actual governance is when the system makes violations impossible before they can occur.

491
00:18:20,440 --> 00:18:24,360
The agent can't go rogue within its policy constraints because it doesn't have the capability to.

492
00:18:24,360 --> 00:18:27,720
The constraint exists at the infrastructure level, not the behavioral level,

493
00:18:27,720 --> 00:18:30,680
but constraints without visibility are just invisible walls.

494
00:18:30,680 --> 00:18:33,320
You can prevent an agent from doing the wrong thing.

495
00:18:33,320 --> 00:18:37,400
But if you can't see what it's actually doing, you can't verify that the constraints are working.

496
00:18:37,400 --> 00:18:40,960
You can't adjust policy based on how the agent behaves in practice.

497
00:18:40,960 --> 00:18:43,520
You can't detect when something unexpected is happening.

498
00:18:43,520 --> 00:18:46,040
That's where monitoring and transparency enter the picture.

499
00:18:46,040 --> 00:18:49,880
The real-time reasoning trace, so you've constrained what an agent can do,

500
00:18:49,880 --> 00:18:53,520
you've drawn the boundaries, you've said, "Here's what you're allowed to access.

501
00:18:53,520 --> 00:18:56,920
Here's what you cannot do. Here's where you need approval before proceeding."

502
00:18:56,920 --> 00:19:00,160
The system enforces those constraints, the agent cannot violate them.

503
00:19:00,160 --> 00:19:03,480
But enforcement alone doesn't tell you what's actually happening inside the agent's reasoning.

504
00:19:03,480 --> 00:19:04,720
It doesn't show you the thinking.

505
00:19:04,720 --> 00:19:06,720
That's where the reasoning trace comes in.

506
00:19:06,720 --> 00:19:10,120
Agent 365 lets you watch an agent think, not after the fact.

507
00:19:10,120 --> 00:19:13,320
In real-time, as the reasoning is happening, step by step,

508
00:19:13,320 --> 00:19:16,920
you can see exactly what's going on inside the agent's decision-making process.

509
00:19:16,920 --> 00:19:18,960
What data did it retrieve to answer the question?

510
00:19:18,960 --> 00:19:20,240
The reasoning trace shows you.

511
00:19:20,240 --> 00:19:23,200
You see the query, you see the data sources are consulted,

512
00:19:23,200 --> 00:19:26,160
you see the context it pulled, it's not a black box, it's transparent.

513
00:19:26,160 --> 00:19:27,680
Which tools did it consider using?

514
00:19:27,680 --> 00:19:29,040
The trace shows you that too.

515
00:19:29,040 --> 00:19:31,400
The agent is thinking about what actions it might take.

516
00:19:31,400 --> 00:19:35,840
Should it send an email or create a task or update a record or escalate to a human?

517
00:19:35,840 --> 00:19:38,280
The reasoning trace shows you every option it considered.

518
00:19:38,280 --> 00:19:39,520
You can see the decision tree.

519
00:19:39,520 --> 00:19:42,840
You can see why it ruled some options out and moved forward with others.

520
00:19:42,840 --> 00:19:45,040
Why did it choose this action over that one?

521
00:19:45,040 --> 00:19:46,320
This is the crucial part.

522
00:19:46,320 --> 00:19:48,520
The reasoning trace explains the decision logic.

523
00:19:48,520 --> 00:19:50,440
The agent didn't just pick an action randomly.

524
00:19:50,440 --> 00:19:53,960
It evaluated options against constraints against the current context.

525
00:19:53,960 --> 00:19:55,640
Against the goal it's trying to achieve,

526
00:19:55,640 --> 00:19:57,160
the trace walks through that evaluation.

527
00:19:57,160 --> 00:19:58,800
It's not magic, it's reasoned.

528
00:19:58,800 --> 00:20:00,240
And the reasoning is visible.

529
00:20:00,240 --> 00:20:01,360
What was the outcome?

530
00:20:01,360 --> 00:20:03,720
The trace follows the action through to completion.

531
00:20:03,720 --> 00:20:04,640
Did the email get sent?

532
00:20:04,640 --> 00:20:06,240
To whom? Did the task get created?

533
00:20:06,240 --> 00:20:07,360
With what description?

534
00:20:07,360 --> 00:20:09,120
Did the API call succeed or fail?

535
00:20:09,120 --> 00:20:10,680
If it failed. Why?

536
00:20:10,680 --> 00:20:12,400
The entire execution path is visible.

537
00:20:12,400 --> 00:20:13,840
Not as a black box result.

538
00:20:13,840 --> 00:20:15,600
As a sequence of events you can inspect.

539
00:20:15,600 --> 00:20:18,480
This transparency changes what's possible in three fundamental ways.

540
00:20:18,480 --> 00:20:20,240
First, trust.

541
00:20:20,240 --> 00:20:22,400
Users can see the agent isn't a black box.

542
00:20:22,400 --> 00:20:23,920
They can watch the reasoning unfold.

543
00:20:23,920 --> 00:20:26,600
They can see that the agent looked at the relevant context.

544
00:20:26,600 --> 00:20:29,680
They can see that the decision makes sense given the information available.

545
00:20:29,680 --> 00:20:32,320
Even if the user doesn't fully understand large language models

546
00:20:32,320 --> 00:20:34,080
or embeddings or transformer architectures,

547
00:20:34,080 --> 00:20:35,680
they can understand a reasoning trace.

548
00:20:35,680 --> 00:20:36,520
They can see the steps.

549
00:20:36,520 --> 00:20:37,520
They can see the logic.

550
00:20:37,520 --> 00:20:38,840
That builds confidence.

551
00:20:38,840 --> 00:20:40,120
The agent isn't mysterious.

552
00:20:40,120 --> 00:20:41,800
It's thinking.

553
00:20:41,800 --> 00:20:43,360
And the thinking is followable.

554
00:20:43,360 --> 00:20:45,680
Second, debugging when something goes wrong.

555
00:20:45,680 --> 00:20:46,680
And something will go wrong.

556
00:20:46,680 --> 00:20:48,280
You can see exactly where it breaks.

557
00:20:48,280 --> 00:20:50,400
Did the agent retrieve the wrong context?

558
00:20:50,400 --> 00:20:51,600
You can see that in the trace.

559
00:20:51,600 --> 00:20:52,880
It looked at the wrong document.

560
00:20:52,880 --> 00:20:54,080
So its answer was wrong.

561
00:20:54,080 --> 00:20:55,000
You found the problem.

562
00:20:55,000 --> 00:20:55,800
You can fix it.

563
00:20:55,800 --> 00:20:58,600
Did the agent consider the right options but made a poor choice?

564
00:20:58,600 --> 00:21:00,240
You can see that in the trace too.

565
00:21:00,240 --> 00:21:01,880
The agent weighed the options correctly,

566
00:21:01,880 --> 00:21:04,560
but failed to account for some constraint you thought it would understand.

567
00:21:04,560 --> 00:21:05,880
You found a gap in the design.

568
00:21:05,880 --> 00:21:06,920
You can address it.

569
00:21:06,920 --> 00:21:09,000
Did the agent try to take an action that failed?

570
00:21:09,000 --> 00:21:10,520
You can see the failure point.

571
00:21:10,520 --> 00:21:12,680
The API call had unexpected parameters.

572
00:21:12,680 --> 00:21:14,160
The system rejected it for a reason.

573
00:21:14,160 --> 00:21:16,360
You can see exactly what happened and why.

574
00:21:16,360 --> 00:21:19,760
Without the reasoning trace, debugging is guesswork.

575
00:21:19,760 --> 00:21:21,040
The agent produced a wrong answer.

576
00:21:21,040 --> 00:21:21,560
Why?

577
00:21:21,560 --> 00:21:22,240
You don't know?

578
00:21:22,240 --> 00:21:23,040
Could be the data.

579
00:21:23,040 --> 00:21:23,840
Could be the model.

580
00:21:23,840 --> 00:21:24,680
Could be the prompting.

581
00:21:24,680 --> 00:21:26,080
Could be something environmental.

582
00:21:26,080 --> 00:21:27,320
You're stuck.

583
00:21:27,320 --> 00:21:29,800
With the reasoning trace, you can see the exact sequence.

584
00:21:29,800 --> 00:21:31,280
You can identify the failure point.

585
00:21:31,280 --> 00:21:33,000
You can fix it with precision.

586
00:21:33,000 --> 00:21:35,640
Third, compliance.

587
00:21:35,640 --> 00:21:37,440
Auditors care about traceability.

588
00:21:37,440 --> 00:21:39,120
If an agent makes a recommendation,

589
00:21:39,120 --> 00:21:41,280
can you prove what information it based that on?

590
00:21:41,280 --> 00:21:43,240
Can you show what reasoning led to the decision?

591
00:21:43,240 --> 00:21:46,520
Can you prove that the decision was sound given the available information?

592
00:21:46,520 --> 00:21:48,840
The reasoning trace provides exactly that.

593
00:21:48,840 --> 00:21:49,960
Here's what the agent saw.

594
00:21:49,960 --> 00:21:51,160
Here's how it processed it.

595
00:21:51,160 --> 00:21:52,880
Here's why it reached this conclusion.

596
00:21:52,880 --> 00:21:54,120
The audit trail is complete.

597
00:21:54,120 --> 00:21:54,880
It's detailed.

598
00:21:54,880 --> 00:21:57,080
It's defensible.

599
00:21:57,080 --> 00:22:01,080
But beyond compliance, the reasoning trace gives you something else.

600
00:22:01,080 --> 00:22:02,400
An emergency break.

601
00:22:02,400 --> 00:22:04,040
If you're watching the reasoning in real time

602
00:22:04,040 --> 00:22:07,280
and you see the agent heading toward a bad decision, you can pause it.

603
00:22:07,280 --> 00:22:08,840
You can intervene, mid-thought.

604
00:22:08,840 --> 00:22:10,920
You can say, stop, don't take that action.

605
00:22:10,920 --> 00:22:12,040
The agent holds.

606
00:22:12,040 --> 00:22:13,600
The decision is not executed.

607
00:22:13,600 --> 00:22:16,200
You've prevented a problem before it became irreversible.

608
00:22:16,200 --> 00:22:19,920
But visibility and constraints only work if you're actually tracking what matters.

609
00:22:19,920 --> 00:22:22,080
If you're monitoring for the right signals.

610
00:22:22,080 --> 00:22:24,440
Behavioral monitoring and anomaly detection.

611
00:22:24,440 --> 00:22:25,800
You've built the walls.

612
00:22:25,800 --> 00:22:26,960
You've set the constraints.

613
00:22:26,960 --> 00:22:28,640
You've even established the reasoning trace

614
00:22:28,640 --> 00:22:31,040
so you can see exactly what the agent is doing.

615
00:22:31,040 --> 00:22:32,920
But none of that matters if you can't detect

616
00:22:32,920 --> 00:22:34,680
when something is actually wrong.

617
00:22:34,680 --> 00:22:38,200
Constraints are there to prevent unauthorized actions during normal operation.

618
00:22:38,200 --> 00:22:39,120
But here's the problem.

619
00:22:39,120 --> 00:22:40,680
An agent can be compromised.

620
00:22:40,680 --> 00:22:42,120
Its credentials can be stolen.

621
00:22:42,120 --> 00:22:43,600
Its instructions can be overwritten.

622
00:22:43,600 --> 00:22:46,400
It can be hijacked to do things it was never designed to do.

623
00:22:46,400 --> 00:22:48,840
And if you aren't actively monitoring for that compromise,

624
00:22:48,840 --> 00:22:51,600
you won't know it's happening until the damage is already done.

625
00:22:51,600 --> 00:22:54,000
Agent 365 integrates with Microsoft Defender

626
00:22:54,000 --> 00:22:55,680
to watch for agent misbehavior.

627
00:22:55,680 --> 00:22:58,200
We aren't just talking about policy violations here.

628
00:22:58,200 --> 00:23:00,680
Those are prevented by your role-based access controls.

629
00:23:00,680 --> 00:23:02,720
We're talking about anomalous behavior.

630
00:23:02,720 --> 00:23:06,240
These are patterns that don't match how this specific agent normally operates.

631
00:23:06,240 --> 00:23:08,560
They are signals that indicate something has changed,

632
00:23:08,560 --> 00:23:10,840
something is wrong, or something has been compromised.

633
00:23:10,840 --> 00:23:13,480
So what does misbehavior actually look like for an agent?

634
00:23:13,480 --> 00:23:15,160
It looks like unusual access patterns.

635
00:23:15,160 --> 00:23:16,840
Imagine an agent that normally reads

636
00:23:16,840 --> 00:23:18,960
from three specific SharePoint sites,

637
00:23:18,960 --> 00:23:22,440
suddenly tries to access 17 different sites it has never touched before.

638
00:23:22,440 --> 00:23:24,520
Technically that isn't a policy violation

639
00:23:24,520 --> 00:23:27,560
because the agent has broad-red access across the organization.

640
00:23:27,560 --> 00:23:29,480
It's allowed to be there, but it isn't normal.

641
00:23:29,480 --> 00:23:32,520
The agent never does that, something changed, and alert fires.

642
00:23:32,520 --> 00:23:34,240
And the security team investigates.

643
00:23:34,240 --> 00:23:36,240
It also looks like excessive tool usage.

644
00:23:36,240 --> 00:23:39,960
An agent that normally makes a few dozen API calls per workflow

645
00:23:39,960 --> 00:23:42,640
might suddenly start making hundreds of calls in a matter of seconds.

646
00:23:42,640 --> 00:23:43,600
It's hammering the system.

647
00:23:43,600 --> 00:23:46,560
Maybe it's trying to ex-filter a data by pulling everything it can access

648
00:23:46,560 --> 00:23:49,560
before someone notices, or maybe it's just stuck in a loop.

649
00:23:49,560 --> 00:23:51,400
Either way, it isn't normal operation.

650
00:23:51,400 --> 00:23:54,200
The anomaly detection system sees this bike, triggers an alert,

651
00:23:54,200 --> 00:23:55,440
and attention gets paid.

652
00:23:55,440 --> 00:23:58,440
Policy violations are caught by the authorization framework itself.

653
00:23:58,440 --> 00:24:01,080
The agent attempts an action it isn't allowed to perform,

654
00:24:01,080 --> 00:24:03,880
the system denies it, and the event is logged and reported.

655
00:24:03,880 --> 00:24:07,160
But the system isn't just logging denials, it's looking for patterns.

656
00:24:07,160 --> 00:24:10,480
Is the agent repeatedly attempting actions it isn't authorized for?

657
00:24:10,480 --> 00:24:14,440
Is it probing the boundaries of its permissions or testing what it can and cannot access?

658
00:24:14,440 --> 00:24:15,760
That pattern is suspicious.

659
00:24:15,760 --> 00:24:20,120
It suggests the agent is being manipulated by something trying to break out of its constraints.

660
00:24:20,120 --> 00:24:21,600
Then there is cross-tenant activity.

661
00:24:21,600 --> 00:24:23,720
If an agent is compromised in one tenant,

662
00:24:23,720 --> 00:24:26,920
an attacker might use it to try to access another tenant's data.

663
00:24:26,920 --> 00:24:29,280
The agent attempts something that should be impossible,

664
00:24:29,280 --> 00:24:32,320
like accessing a resource outside its authorized boundary.

665
00:24:32,320 --> 00:24:35,480
The system detects the attempt it fails and the system logs it.

666
00:24:35,480 --> 00:24:38,160
When you see multiple failed attempts across multiple tenants,

667
00:24:38,160 --> 00:24:40,040
that is a clear signal of compromise.

668
00:24:40,040 --> 00:24:41,440
The agent is being weaponized.

669
00:24:41,440 --> 00:24:43,840
These anomalies trigger alerts in real time.

670
00:24:43,840 --> 00:24:45,360
This doesn't happen after the fact,

671
00:24:45,360 --> 00:24:47,520
or in a log file that someone reviews tomorrow.

672
00:24:47,520 --> 00:24:48,440
It happens right now.

673
00:24:48,440 --> 00:24:51,680
A security analyst sees the alert and can respond immediately.

674
00:24:51,680 --> 00:24:53,280
They can revoke the agent's credentials,

675
00:24:53,280 --> 00:24:55,760
pause its activity, and prevent further damage.

676
00:24:55,760 --> 00:24:57,000
This is the critical insight.

677
00:24:57,000 --> 00:25:00,320
Agents can be compromised exactly like user accounts can be compromised.

678
00:25:00,320 --> 00:25:03,480
When a user's password gets stolen and an attacker logs in,

679
00:25:03,480 --> 00:25:04,760
they are inside the system.

680
00:25:04,760 --> 00:25:07,080
They can access everything that user can access.

681
00:25:07,080 --> 00:25:09,000
They can read data, delete files,

682
00:25:09,000 --> 00:25:11,160
and send messages on behalf of that person.

683
00:25:11,160 --> 00:25:12,840
The breach is active and dangerous.

684
00:25:12,840 --> 00:25:16,160
An agent's credentials being stolen is the exact same scenario.

685
00:25:16,160 --> 00:25:17,760
Once an attacker has those credentials,

686
00:25:17,760 --> 00:25:19,440
they can make the agent do whatever they want.

687
00:25:19,440 --> 00:25:21,640
They can execute requests, access data,

688
00:25:21,640 --> 00:25:23,360
and abuse the agent's capabilities.

689
00:25:23,360 --> 00:25:26,040
The breach is just as dangerous and possibly more so.

690
00:25:26,040 --> 00:25:27,880
Because an agent can execute at scale,

691
00:25:27,880 --> 00:25:29,680
a compromised agent can do in seconds

692
00:25:29,680 --> 00:25:31,600
what would take a human hours to finish.

693
00:25:31,600 --> 00:25:32,960
It can read thousands of documents,

694
00:25:32,960 --> 00:25:34,560
make hundreds of API calls,

695
00:25:34,560 --> 00:25:36,360
and move laterally through your systems.

696
00:25:36,360 --> 00:25:37,880
Because of this, Defender for Agents

697
00:25:37,880 --> 00:25:40,320
applies the same threat detection that works for humans.

698
00:25:40,320 --> 00:25:41,720
It uses the same patterns,

699
00:25:41,720 --> 00:25:43,120
the same anomaly algorithms,

700
00:25:43,120 --> 00:25:44,680
and the same response mechanisms.

701
00:25:44,680 --> 00:25:47,120
An agent that has been compromised behaves differently

702
00:25:47,120 --> 00:25:48,120
from a normal agent,

703
00:25:48,120 --> 00:25:50,520
the same way a compromised user behaves differently

704
00:25:50,520 --> 00:25:51,720
from a normal user.

705
00:25:51,720 --> 00:25:53,280
The system learns what normal looks like,

706
00:25:53,280 --> 00:25:55,680
detects the deviations and enables a response.

707
00:25:55,680 --> 00:25:57,600
But all of this, the constraints,

708
00:25:57,600 --> 00:25:59,120
the visibility and the monitoring,

709
00:25:59,120 --> 00:26:01,520
it only works if you have a shared context layer

710
00:26:01,520 --> 00:26:03,520
that understands organizational state.

711
00:26:03,520 --> 00:26:05,520
You need something that can enforce these controls

712
00:26:05,520 --> 00:26:07,320
consistently across everything.

713
00:26:07,320 --> 00:26:09,680
And that is where WorkIQ becomes essential.

714
00:26:09,680 --> 00:26:11,480
What WorkIQ actually is.

715
00:26:11,480 --> 00:26:13,000
WorkIQ is not a database.

716
00:26:13,000 --> 00:26:14,160
You don't store anything in it.

717
00:26:14,160 --> 00:26:15,400
It's also not a search engine,

718
00:26:15,400 --> 00:26:18,160
so you don't query it the way you would query an index.

719
00:26:18,160 --> 00:26:19,680
It is something different from both.

720
00:26:19,680 --> 00:26:21,320
WorkIQ is a reasoning layer that sits

721
00:26:21,320 --> 00:26:23,400
between your agents and the actual systems

722
00:26:23,400 --> 00:26:24,480
where work happens.

723
00:26:24,480 --> 00:26:25,280
Think of it this way.

724
00:26:25,280 --> 00:26:27,040
Your organization has data everywhere.

725
00:26:27,040 --> 00:26:30,120
You have email in exchange, documents in SharePoint,

726
00:26:30,120 --> 00:26:33,280
conversations in Teams and records in your ERP system.

727
00:26:33,280 --> 00:26:36,000
All of that data is distributed across different systems

728
00:26:36,000 --> 00:26:38,080
with their own access controls and structures.

729
00:26:38,080 --> 00:26:40,840
An agent needs to understand that data and act on it,

730
00:26:40,840 --> 00:26:42,200
but it doesn't need to copy it.

731
00:26:42,200 --> 00:26:45,560
WorkIQ lets the agent reason over that distributed data

732
00:26:45,560 --> 00:26:48,080
as if it were a single coherent system.

733
00:26:48,080 --> 00:26:50,640
WorkIQ exposes four specific capabilities.

734
00:26:50,640 --> 00:26:51,800
First is chat.

735
00:26:51,800 --> 00:26:53,640
Your agents can have conversational access

736
00:26:53,640 --> 00:26:55,800
to organizational knowledge through WorkIQ.

737
00:26:55,800 --> 00:26:57,720
This isn't done by querying a search engine

738
00:26:57,720 --> 00:26:58,960
or retrieving documents.

739
00:26:58,960 --> 00:27:01,000
It's done by having a reasoning conversation.

740
00:27:01,000 --> 00:27:03,520
Your agent can ask WorkIQ about the current status

741
00:27:03,520 --> 00:27:06,760
of a specific region and WorkIQ understands the context.

742
00:27:06,760 --> 00:27:09,480
It reasons over real-time signals and gives an answer

743
00:27:09,480 --> 00:27:11,400
that reflects what is happening right now,

744
00:27:11,400 --> 00:27:13,000
not what was documented a month ago.

745
00:27:13,000 --> 00:27:14,240
Second is context.

746
00:27:14,240 --> 00:27:16,520
This is the governed retrieval of tenant signals.

747
00:27:16,520 --> 00:27:17,920
When your agent needs to understand

748
00:27:17,920 --> 00:27:19,240
who is working on a project,

749
00:27:19,240 --> 00:27:21,880
WorkIQ returns more than just a list of names.

750
00:27:21,880 --> 00:27:23,320
It returns the people involved.

751
00:27:23,320 --> 00:27:25,560
Their current assignments and their availability.

752
00:27:25,560 --> 00:27:27,440
Its context, not just raw data.

753
00:27:27,440 --> 00:27:28,760
The agent gets interpreted,

754
00:27:28,760 --> 00:27:31,440
relevant information tailored to what it's trying to accomplish.

755
00:27:31,440 --> 00:27:33,640
Third is tools which are your action endpoints.

756
00:27:33,640 --> 00:27:35,560
Your agent can send mail, update files,

757
00:27:35,560 --> 00:27:37,560
or schedule meetings through WorkIQ.

758
00:27:37,560 --> 00:27:39,240
These aren't raw API calls.

759
00:27:39,240 --> 00:27:41,120
They are WorkIQ mediated actions.

760
00:27:41,120 --> 00:27:43,440
The permission model and policy constraints apply here

761
00:27:43,440 --> 00:27:45,800
and the reasoning trace captures exactly what happened.

762
00:27:45,800 --> 00:27:47,600
The agent isn't calling the API directly.

763
00:27:47,600 --> 00:27:49,440
It's asking WorkIQ to do something

764
00:27:49,440 --> 00:27:51,320
and WorkIQ handles the execution

765
00:27:51,320 --> 00:27:52,840
within the governance framework.

766
00:27:52,840 --> 00:27:54,200
Fourth is workspaces.

767
00:27:54,200 --> 00:27:56,720
These provide persistent state for multi-step tasks.

768
00:27:56,720 --> 00:27:58,920
When an agent is working on something complex,

769
00:27:58,920 --> 00:28:01,680
it needs to remember decisions and track its progress.

770
00:28:01,680 --> 00:28:04,200
Workspaces give the agent a project-scoped space

771
00:28:04,200 --> 00:28:07,160
to maintain state across multiple reasoning steps.

772
00:28:07,160 --> 00:28:09,520
That state is persistent, bounded, and governed

773
00:28:09,520 --> 00:28:10,720
like everything else.

774
00:28:10,720 --> 00:28:13,440
The critical architectural detail changes everything.

775
00:28:13,440 --> 00:28:15,160
WorkIQ does not copy your data.

776
00:28:15,160 --> 00:28:18,080
It doesn't export it, vectorize it, or build a shadow index.

777
00:28:18,080 --> 00:28:19,960
Instead, it reasons over data in place.

778
00:28:19,960 --> 00:28:21,320
Your documents stay in SharePoint

779
00:28:21,320 --> 00:28:22,800
and your emails stay in exchange.

780
00:28:22,800 --> 00:28:25,080
WorkIQ connects to those systems in real time

781
00:28:25,080 --> 00:28:26,760
to understand the data where it lives.

782
00:28:26,760 --> 00:28:29,120
It enforces permissions where they are already defined.

783
00:28:29,120 --> 00:28:32,160
It's a reasoning layer, not a data duplication layer.

784
00:28:32,160 --> 00:28:34,520
This leads to three major operational advantages.

785
00:28:34,520 --> 00:28:36,480
First, you avoid the data export nightmare.

786
00:28:36,480 --> 00:28:38,560
You aren't building parallel data stores

787
00:28:38,560 --> 00:28:40,680
or managing sync between the source of truth

788
00:28:40,680 --> 00:28:41,920
and a shadow index.

789
00:28:41,920 --> 00:28:44,720
The source of truth remains the only source of truth.

790
00:28:44,720 --> 00:28:46,840
Because WorkIQ reasons about it directly,

791
00:28:46,840 --> 00:28:49,080
your compliance obligations don't expand

792
00:28:49,080 --> 00:28:50,840
and your data governance doesn't fragment.

793
00:28:50,840 --> 00:28:53,080
You aren't creating new data residency problems

794
00:28:53,080 --> 00:28:54,560
or new security surfaces.

795
00:28:54,560 --> 00:28:56,680
Second, permissions are enforced at reasoning time,

796
00:28:56,680 --> 00:28:58,200
not retrieval time.

797
00:28:58,200 --> 00:28:59,960
When your agent asks WorkIQ a question,

798
00:28:59,960 --> 00:29:02,200
the system checks permissions as it reasons.

799
00:29:02,200 --> 00:29:04,960
It asks if the agent can see this data in real time.

800
00:29:04,960 --> 00:29:07,200
The answer it gives reflects what the agent is actually

801
00:29:07,200 --> 00:29:08,200
allowed to know.

802
00:29:08,200 --> 00:29:10,200
You aren't retrieving data and then filtering it.

803
00:29:10,200 --> 00:29:12,880
You are reasoning about the data you are authorized

804
00:29:12,880 --> 00:29:14,400
to see from the start.

805
00:29:14,400 --> 00:29:17,160
The permission boundary is built into the reasoning itself.

806
00:29:17,160 --> 00:29:19,560
Third, the audit trail is native to M365.

807
00:29:19,560 --> 00:29:21,640
You aren't bolting on logging after the fact

808
00:29:21,640 --> 00:29:24,320
or trying to make WorkIQ queryable from your CM.

809
00:29:24,320 --> 00:29:26,160
The action is captured directly in the systems

810
00:29:26,160 --> 00:29:27,560
where the action occurred.

811
00:29:27,560 --> 00:29:30,080
If an agent sends an email through WorkIQ exchange logs

812
00:29:30,080 --> 00:29:32,800
that email, if an agent updates a SharePoint record,

813
00:29:32,800 --> 00:29:34,160
SharePoint logs the change.

814
00:29:34,160 --> 00:29:37,000
Everything is auditable through the native M365 logging

815
00:29:37,000 --> 00:29:38,200
you already use.

816
00:29:38,200 --> 00:29:41,600
WorkIQ becomes the foundation that makes everything else possible.

817
00:29:41,600 --> 00:29:43,560
The constraints defined in Agent 365

818
00:29:43,560 --> 00:29:45,120
aren't forced by WorkIQ.

819
00:29:45,120 --> 00:29:46,280
But the reasoning trace you're watching

820
00:29:46,280 --> 00:29:47,920
is provided by WorkIQ.

821
00:29:47,920 --> 00:29:49,880
The governance that prevents unauthorized access

822
00:29:49,880 --> 00:29:51,520
is implemented by WorkIQ.

823
00:29:51,520 --> 00:29:53,280
But WorkIQ itself is single threaded.

824
00:29:53,280 --> 00:29:56,040
It handles one agent and one conversation at a time.

825
00:29:56,040 --> 00:29:57,960
For organizational work, you need agents

826
00:29:57,960 --> 00:29:59,800
to coordinate and delegate to each other.

827
00:29:59,800 --> 00:30:02,040
You need a protocol for agent to agent communication.

828
00:30:02,040 --> 00:30:04,160
That is where A2A enters.

829
00:30:04,160 --> 00:30:06,520
A2A, the agent communication protocol.

830
00:30:06,520 --> 00:30:08,400
WorkIQ handles the intelligence.

831
00:30:08,400 --> 00:30:10,320
Agent 365 handles the governance.

832
00:30:10,320 --> 00:30:12,320
But neither of them solves the problem of agents

833
00:30:12,320 --> 00:30:13,720
needing to talk to each other.

834
00:30:13,720 --> 00:30:16,600
Real work doesn't happen when one agent answers one question.

835
00:30:16,600 --> 00:30:19,240
It happens when multiple agents collaborate to finish a goal.

836
00:30:19,240 --> 00:30:20,920
Think about how a workflow actually looks.

837
00:30:20,920 --> 00:30:22,840
Your procurement agent needs to check the budget

838
00:30:22,840 --> 00:30:24,520
before it approves a purchase.

839
00:30:24,520 --> 00:30:26,000
It can't do that analysis alone

840
00:30:26,000 --> 00:30:28,600
because the finance agent owns the budget data.

841
00:30:28,600 --> 00:30:30,720
The procurement agent has to ask a question.

842
00:30:30,720 --> 00:30:31,920
It needs to delegate.

843
00:30:31,920 --> 00:30:33,760
It needs to hand off a piece of the work.

844
00:30:33,760 --> 00:30:36,120
The obvious way to do this is natural language.

845
00:30:36,120 --> 00:30:37,840
The procurement agent writes a message in English

846
00:30:37,840 --> 00:30:40,000
asking if there is enough money for the purchase.

847
00:30:40,000 --> 00:30:41,960
It sends that message to the finance agent.

848
00:30:41,960 --> 00:30:43,520
Then the finance agent reads the English

849
00:30:43,520 --> 00:30:44,960
and tries to parse the request.

850
00:30:44,960 --> 00:30:46,520
It has to figure out what is being asked

851
00:30:46,520 --> 00:30:48,040
and reason about the answer.

852
00:30:48,040 --> 00:30:50,320
It writes a response in English and sends it back.

853
00:30:50,320 --> 00:30:52,520
Finally, the procurement agent reads that response

854
00:30:52,520 --> 00:30:53,640
and continues its work.

855
00:30:53,640 --> 00:30:55,120
This works for a single handoff.

856
00:30:55,120 --> 00:30:56,000
It's slow.

857
00:30:56,000 --> 00:30:56,600
But it works.

858
00:30:56,600 --> 00:30:59,920
But now imagine three agents involved or five or 20.

859
00:30:59,920 --> 00:31:01,560
Imagine agents trying to coordinate

860
00:31:01,560 --> 00:31:04,440
across dozens of different workflows at the same time.

861
00:31:04,440 --> 00:31:06,120
Every handoff creates latency.

862
00:31:06,120 --> 00:31:07,360
Passing is expensive.

863
00:31:07,360 --> 00:31:09,160
Language is messy and ambiguous.

864
00:31:09,160 --> 00:31:10,840
The finance agent might misunderstand

865
00:31:10,840 --> 00:31:13,000
which budget the procurement agent is talking about.

866
00:31:13,000 --> 00:31:15,720
The procurement agent might misinterpret the answer.

867
00:31:15,720 --> 00:31:17,560
You are basically injecting errors into the system

868
00:31:17,560 --> 00:31:19,040
every time a conversation happens.

869
00:31:19,040 --> 00:31:20,040
But here's the problem.

870
00:31:20,040 --> 00:31:22,440
You're wasting massive amounts of compute.

871
00:31:22,440 --> 00:31:24,800
You are spending tokens to generate English text

872
00:31:24,800 --> 00:31:27,240
and parse English responses for a coordination

873
00:31:27,240 --> 00:31:29,200
that doesn't need to be in English at all.

874
00:31:29,200 --> 00:31:31,440
It's like two computers trying to talk to each other

875
00:31:31,440 --> 00:31:34,080
using human voices instead of a network protocol.

876
00:31:34,080 --> 00:31:36,600
It technically works, but it's terrible in practice.

877
00:31:36,600 --> 00:31:38,400
A to A inverts this model.

878
00:31:38,400 --> 00:31:40,160
It's an open standard for agent communication

879
00:31:40,160 --> 00:31:42,160
that ignores natural language entirely.

880
00:31:42,160 --> 00:31:44,800
It uses structured messages, explicit schemas,

881
00:31:44,800 --> 00:31:46,840
machine readable formats.

882
00:31:46,840 --> 00:31:48,960
This is what an A to A message looks like.

883
00:31:48,960 --> 00:31:50,760
The procurement agent doesn't send a question.

884
00:31:50,760 --> 00:31:52,280
It sends a structured task.

885
00:31:52,280 --> 00:31:53,800
That task includes the operation

886
00:31:53,800 --> 00:31:56,920
like check budget availability and the parameters

887
00:31:56,920 --> 00:31:58,680
like the cost center and fiscal year.

888
00:31:58,680 --> 00:32:00,760
It includes the context the other agent needs

889
00:32:00,760 --> 00:32:02,960
such as vendor details and urgency.

890
00:32:02,960 --> 00:32:05,400
It even defines the exact format for the response

891
00:32:05,400 --> 00:32:07,840
like a JSON file showing the remaining budget.

892
00:32:07,840 --> 00:32:10,240
It also includes metadata to prove the request

893
00:32:10,240 --> 00:32:11,800
is legitimate and authorized.

894
00:32:11,800 --> 00:32:13,360
The finance agent receives this task

895
00:32:13,360 --> 00:32:14,840
and there is zero ambiguity.

896
00:32:14,840 --> 00:32:16,040
The schema is explicit.

897
00:32:16,040 --> 00:32:17,280
The parameters are clear.

898
00:32:17,280 --> 00:32:19,000
The output format is already set.

899
00:32:19,000 --> 00:32:22,400
The finance agent processes the request, validates the budget,

900
00:32:22,400 --> 00:32:24,960
and returns a response that fits the schema perfectly,

901
00:32:24,960 --> 00:32:29,160
structured data, machine readable, no interpretation required.

902
00:32:29,160 --> 00:32:31,280
This is fundamentally different from agents chatting

903
00:32:31,280 --> 00:32:32,160
with each other.

904
00:32:32,160 --> 00:32:33,360
There is no conversation.

905
00:32:33,360 --> 00:32:34,920
There is no back and forth negotiation.

906
00:32:34,920 --> 00:32:37,520
There is just a task, a request and a response,

907
00:32:37,520 --> 00:32:39,720
clean delegation, verifiable completion.

908
00:32:39,720 --> 00:32:40,560
And this is the shift.

909
00:32:40,560 --> 00:32:42,760
You get multi-agent orchestration at cloud speed.

910
00:32:42,760 --> 00:32:44,920
When agents use structured messages,

911
00:32:44,920 --> 00:32:47,840
coordination happens in milliseconds instead of seconds.

912
00:32:47,840 --> 00:32:50,040
An orchestrator agent can hand off tasks

913
00:32:50,040 --> 00:32:52,800
to 10 specialized agents at the exact same time.

914
00:32:52,800 --> 00:32:55,080
Each one gets a structured request and processes it

915
00:32:55,080 --> 00:32:55,960
at full speed.

916
00:32:55,960 --> 00:32:59,040
The orchestrator pulls the responses together and moves on.

917
00:32:59,040 --> 00:33:01,360
The whole process is finished before a human could even

918
00:33:01,360 --> 00:33:03,720
finish reading the first sentence of an English request.

919
00:33:03,720 --> 00:33:05,480
You get clear delegation boundaries.

920
00:33:05,480 --> 00:33:07,600
When the schema defines the data and the format,

921
00:33:07,600 --> 00:33:09,840
you can't miscommunicate about the scope of the work.

922
00:33:09,840 --> 00:33:11,880
The procurement agent can't accidentally ask

923
00:33:11,880 --> 00:33:14,360
for sensitive payroll info because the schema doesn't allow it.

924
00:33:14,360 --> 00:33:15,560
The boundary is structural.

925
00:33:15,560 --> 00:33:18,240
It's enforced by the system, not suggested by a prompt.

926
00:33:18,240 --> 00:33:19,680
The handoff becomes efficient.

927
00:33:19,680 --> 00:33:21,760
You aren't generating or passing English,

928
00:33:21,760 --> 00:33:23,480
so you're just moving information.

929
00:33:23,480 --> 00:33:26,000
The network cost is lower and the compute cost is lower.

930
00:33:26,000 --> 00:33:28,280
The entire system is faster and cheaper to run.

931
00:33:28,280 --> 00:33:30,480
You also get verifiable task completion.

932
00:33:30,480 --> 00:33:32,880
Every handoff is logged with a status and a timestamp.

933
00:33:32,880 --> 00:33:34,840
You can trace the whole workflow and prove exactly

934
00:33:34,840 --> 00:33:36,480
when the finance agent received the request

935
00:33:36,480 --> 00:33:37,520
and what it returned.

936
00:33:37,520 --> 00:33:38,920
The audit trail is complete.

937
00:33:38,920 --> 00:33:41,560
You can replay the entire sequence whenever you need to.

938
00:33:41,560 --> 00:33:43,720
A2A is the protocol layer that makes real

939
00:33:43,720 --> 00:33:45,160
multi-agent systems possible.

940
00:33:45,160 --> 00:33:46,760
It's not about one agent at a time.

941
00:33:46,760 --> 00:33:48,400
It's about swarms of agents coordinating

942
00:33:48,400 --> 00:33:50,280
in real time over shared work.

943
00:33:50,280 --> 00:33:53,360
Performance A2A versus MCP benchmarks.

944
00:33:53,360 --> 00:33:56,360
Now that you see what A2A does, structured communication

945
00:33:56,360 --> 00:33:57,600
instead of language passing,

946
00:33:57,600 --> 00:34:00,040
the question is simple, does it actually matter?

947
00:34:00,040 --> 00:34:02,640
Does this architectural shift lead to real results?

948
00:34:02,640 --> 00:34:05,040
Or is it just a theoretical win that disappears

949
00:34:05,040 --> 00:34:05,880
when you scale up?

950
00:34:05,880 --> 00:34:08,200
The benchmarks are where the theory meets reality

951
00:34:08,200 --> 00:34:09,160
and they are stock.

952
00:34:09,160 --> 00:34:12,800
A2A handles 12,500 requests per second.

953
00:34:12,800 --> 00:34:15,480
MCP, which is the other major communication standard,

954
00:34:15,480 --> 00:34:17,080
handles 7,800.

955
00:34:17,080 --> 00:34:19,320
That is a 60% advantage in throughput.

956
00:34:19,320 --> 00:34:20,720
When one agent has to coordinate

957
00:34:20,720 --> 00:34:22,880
with 10 specialized agents at once,

958
00:34:22,880 --> 00:34:25,440
that gap is the difference between a responsive system

959
00:34:25,440 --> 00:34:26,800
and a total bottleneck.

960
00:34:26,800 --> 00:34:28,680
The latency numbers show the same thing.

961
00:34:28,680 --> 00:34:31,880
A2A averages 3.2 milliseconds per request,

962
00:34:31,880 --> 00:34:34,400
while MCP averages 8.5 milliseconds,

963
00:34:34,400 --> 00:34:36,760
that is a 62% reduction in delay.

964
00:34:36,760 --> 00:34:38,440
This is the difference between a user feeling

965
00:34:38,440 --> 00:34:41,560
like the agent is instant versus feeling a noticeable lag.

966
00:34:41,560 --> 00:34:43,360
3 milliseconds is invisible to a human,

967
00:34:43,360 --> 00:34:45,560
but 8 milliseconds starts to feel sluggish.

968
00:34:45,560 --> 00:34:47,360
If you have five agents coordinating,

969
00:34:47,360 --> 00:34:50,240
you're looking at 15 milliseconds versus 42 and a half.

970
00:34:50,240 --> 00:34:52,400
People feel that, it becomes friction.

971
00:34:52,400 --> 00:34:54,880
The resource efficiency gap is even wider.

972
00:34:54,880 --> 00:34:58,160
A2A hits 92% efficiency, meaning it uses compute

973
00:34:58,160 --> 00:35:00,360
and bandwidth almost perfectly.

974
00:35:00,360 --> 00:35:03,320
MCP runs at 75%, in a cloud environment

975
00:35:03,320 --> 00:35:05,120
where you pay for every second of compute

976
00:35:05,120 --> 00:35:07,960
that 17-point difference adds up fast across a whole fleet

977
00:35:07,960 --> 00:35:10,000
of agents, but these numbers only matter

978
00:35:10,000 --> 00:35:12,000
if you see what they mean in practice.

979
00:35:12,000 --> 00:35:13,240
At an organizational scale,

980
00:35:13,240 --> 00:35:14,880
these small gaps don't stay small.

981
00:35:14,880 --> 00:35:15,880
They compound.

982
00:35:15,880 --> 00:35:18,200
You aren't just running two agents having one conversation.

983
00:35:18,200 --> 00:35:19,440
You're running hundreds of agents

984
00:35:19,440 --> 00:35:21,800
and thousands of coordinations every single hour.

985
00:35:21,800 --> 00:35:24,560
Every time a coordination takes 8.5 milliseconds

986
00:35:24,560 --> 00:35:26,880
instead of 3.2, you lose time.

987
00:35:26,880 --> 00:35:29,560
By the end of the day, you've lost hours of throughput.

988
00:35:29,560 --> 00:35:31,720
By the end of the month, you've lost capacity.

989
00:35:31,720 --> 00:35:33,200
You end up needing more infrastructure

990
00:35:33,200 --> 00:35:34,800
just to do the same amount of work.

991
00:35:34,800 --> 00:35:37,480
One slow agent can bottle next the entire workflow.

992
00:35:37,480 --> 00:35:40,040
If your orchestrator sends requests to 10 agents

993
00:35:40,040 --> 00:35:43,280
and nine of them use A2A, they finish in five milliseconds.

994
00:35:43,280 --> 00:35:46,560
But if the 10th agent uses MCP and takes 10 milliseconds,

995
00:35:46,560 --> 00:35:47,960
the orchestrator has to wait.

996
00:35:47,960 --> 00:35:51,320
The whole workflow is only as fast as that one slow component.

997
00:35:51,320 --> 00:35:52,880
That agent becomes your ceiling.

998
00:35:52,880 --> 00:35:54,120
But here's the deeper problem.

999
00:35:54,120 --> 00:35:56,040
When agents coordinate in real time,

1000
00:35:56,040 --> 00:35:57,960
latency changes how they behave.

1001
00:35:57,960 --> 00:36:00,560
Real-time coordination means agents make decisions

1002
00:36:00,560 --> 00:36:02,480
based on what is happening right now.

1003
00:36:02,480 --> 00:36:05,120
A2A is fast enough that the agent gets the fresh state

1004
00:36:05,120 --> 00:36:06,360
of the organization.

1005
00:36:06,360 --> 00:36:07,960
The decision is based on reality.

1006
00:36:07,960 --> 00:36:10,880
With MCP, the data is slightly stale by the time the agent gets it.

1007
00:36:10,880 --> 00:36:13,160
It's only a few milliseconds, but in a fast environment,

1008
00:36:13,160 --> 00:36:16,080
that's enough to make a decision based on outdated information.

1009
00:36:16,080 --> 00:36:17,800
The real insight is architectural.

1010
00:36:17,800 --> 00:36:21,440
These benchmarks prove A2A was built for a genetic density.

1011
00:36:21,440 --> 00:36:22,840
It's for many agents working together,

1012
00:36:22,840 --> 00:36:25,240
not just one smart agent talking to a user.

1013
00:36:25,240 --> 00:36:26,880
The performance isn't a side effect.

1014
00:36:26,880 --> 00:36:27,760
It's the whole point.

1015
00:36:27,760 --> 00:36:29,640
A2A scales horizontally.

1016
00:36:29,640 --> 00:36:32,040
You can add more agents and the performance holds.

1017
00:36:32,040 --> 00:36:33,840
If you add more agents to an MCP system,

1018
00:36:33,840 --> 00:36:35,600
you hit the ceiling much faster.

1019
00:36:35,600 --> 00:36:38,920
This matters because the future isn't one agent per department.

1020
00:36:38,920 --> 00:36:40,680
It's a swarm of specialized agents,

1021
00:36:40,680 --> 00:36:42,840
sales, compliance, finance, and operations,

1022
00:36:42,840 --> 00:36:44,520
all talking at real-time speeds

1023
00:36:44,520 --> 00:36:47,440
that requires a protocol designed for density from day one.

1024
00:36:47,440 --> 00:36:50,360
That's A2A, speed, and efficiency are only half the story.

1025
00:36:50,360 --> 00:36:52,440
There is another side to this cost.

1026
00:36:52,440 --> 00:36:54,720
And in the reasoning economy, cost is architecture.

1027
00:36:54,720 --> 00:36:57,760
Agente density, many agents, one fabric.

1028
00:36:57,760 --> 00:37:00,440
The future of work isn't one smart agent that does everything.

1029
00:37:00,440 --> 00:37:01,680
That's not the direction.

1030
00:37:01,680 --> 00:37:03,800
The direction is many specialized agents.

1031
00:37:03,800 --> 00:37:05,520
Each built for a specific domain,

1032
00:37:05,520 --> 00:37:08,680
each with deep expertise in its area, or working together.

1033
00:37:08,680 --> 00:37:10,360
Picture your organization's agent layer.

1034
00:37:10,360 --> 00:37:12,680
You have a sales agent that understands pipeline.

1035
00:37:12,680 --> 00:37:14,560
It doesn't just look at CRM records.

1036
00:37:14,560 --> 00:37:17,080
It understands the context, which deals are moving fast,

1037
00:37:17,080 --> 00:37:19,120
which are stalled, which have risk signals.

1038
00:37:19,120 --> 00:37:20,520
Because it's built for that domain,

1039
00:37:20,520 --> 00:37:23,120
it sees data through the lens of sales operations

1040
00:37:23,120 --> 00:37:24,760
and knows exactly what matters.

1041
00:37:24,760 --> 00:37:27,200
Then you have a compliance agent that knows regulations.

1042
00:37:27,200 --> 00:37:29,000
It doesn't just search for policy documents.

1043
00:37:29,000 --> 00:37:31,360
It reasons about regulatory obligations.

1044
00:37:31,360 --> 00:37:34,040
It understands which rules apply to which business operations

1045
00:37:34,040 --> 00:37:36,480
and catches edge cases where overlapping regulations

1046
00:37:36,480 --> 00:37:37,520
create contradictions.

1047
00:37:37,520 --> 00:37:40,040
It's a domain expert in regulation and compliance.

1048
00:37:40,040 --> 00:37:41,680
An operations agent tracks your inventory.

1049
00:37:41,680 --> 00:37:43,440
It doesn't store inventory data locally.

1050
00:37:43,440 --> 00:37:45,080
It reasons over the live system.

1051
00:37:45,080 --> 00:37:47,480
It understands warehouse constraints, supplier relationships,

1052
00:37:47,480 --> 00:37:48,560
and lead times.

1053
00:37:48,560 --> 00:37:50,480
So it can recommend reordering decisions.

1054
00:37:50,480 --> 00:37:52,200
It understands the operational complexity

1055
00:37:52,200 --> 00:37:54,320
because it's built for that specific domain.

1056
00:37:54,320 --> 00:37:56,920
A customer service agent handles escalations.

1057
00:37:56,920 --> 00:37:58,680
It understands customer sentiment

1058
00:37:58,680 --> 00:38:01,040
and knows which issues can be resolved automatically

1059
00:38:01,040 --> 00:38:02,720
versus which need human judgment.

1060
00:38:02,720 --> 00:38:05,240
It has context about customer history, account status,

1061
00:38:05,240 --> 00:38:06,400
and support tiers.

1062
00:38:06,400 --> 00:38:08,720
It's specialized for customer interaction.

1063
00:38:08,720 --> 00:38:11,960
A finance agent reasons about costs, budget, and cashflow.

1064
00:38:11,960 --> 00:38:14,000
A contract's agent understands legal obligations

1065
00:38:14,000 --> 00:38:15,080
and renewal dates.

1066
00:38:15,080 --> 00:38:17,240
A project management agent tracks dependencies

1067
00:38:17,240 --> 00:38:18,400
in critical parts.

1068
00:38:18,400 --> 00:38:21,320
An HR agent manages leave policies and benefits enrollment.

1069
00:38:21,320 --> 00:38:22,440
Each one is specialist.

1070
00:38:22,440 --> 00:38:24,080
Each one built for its domain.

1071
00:38:24,080 --> 00:38:26,640
But here's the thing, these agents don't work in isolation.

1072
00:38:26,640 --> 00:38:28,080
They coordinate.

1073
00:38:28,080 --> 00:38:30,040
The customer service agent gets an escalation

1074
00:38:30,040 --> 00:38:32,400
from an angry customer because their invoice is wrong.

1075
00:38:32,400 --> 00:38:34,160
The agent can't fix billing issues.

1076
00:38:34,160 --> 00:38:35,320
That's not its domain.

1077
00:38:35,320 --> 00:38:37,120
So it delegates to the finance agent.

1078
00:38:37,120 --> 00:38:39,640
Check this customer's account, verify the latest invoice,

1079
00:38:39,640 --> 00:38:41,640
identify any calculation errors.

1080
00:38:41,640 --> 00:38:43,760
The finance agent processes the request.

1081
00:38:43,760 --> 00:38:44,600
It checks.

1082
00:38:44,600 --> 00:38:45,680
It finds the error.

1083
00:38:45,680 --> 00:38:47,080
It returns the result.

1084
00:38:47,080 --> 00:38:49,840
Invoice dated March 15 has a line item over charge

1085
00:38:49,840 --> 00:38:53,080
of $47,300 because it was manually entered

1086
00:38:53,080 --> 00:38:55,520
instead of imported from the system.

1087
00:38:55,520 --> 00:38:58,040
The customer service agent now has the information it needs.

1088
00:38:58,040 --> 00:38:59,600
It knows what to tell the customer

1089
00:38:59,600 --> 00:39:02,440
and escalates the correction to a human finance specialist

1090
00:39:02,440 --> 00:39:03,360
if needed.

1091
00:39:03,360 --> 00:39:05,240
Simultaneously, the project management agent

1092
00:39:05,240 --> 00:39:06,760
is tracking a product launch.

1093
00:39:06,760 --> 00:39:08,560
It needs to know budget implications.

1094
00:39:08,560 --> 00:39:10,960
The sales agent is forecasting revenue impact.

1095
00:39:10,960 --> 00:39:13,440
The operations agent is calculating fulfillment capacity.

1096
00:39:13,440 --> 00:39:15,040
They're all reasoning about the same launch

1097
00:39:15,040 --> 00:39:16,280
from different angles.

1098
00:39:16,280 --> 00:39:17,680
And they need to coordinate.

1099
00:39:17,680 --> 00:39:20,160
The project agent asks the operations agent,

1100
00:39:20,160 --> 00:39:22,520
can we fulfill 50,000 units in Q2?

1101
00:39:22,520 --> 00:39:24,440
Operations checks capacity.

1102
00:39:24,440 --> 00:39:26,960
It knows current utilization, seasonal patterns

1103
00:39:26,960 --> 00:39:28,400
and supplier constraints.

1104
00:39:28,400 --> 00:39:32,000
We can fulfill 50,000 units, but only with expedited procurement,

1105
00:39:32,000 --> 00:39:33,920
which adds $7 per unit.

1106
00:39:33,920 --> 00:39:36,520
The project agent gets the answer and asks finance.

1107
00:39:36,520 --> 00:39:39,560
What is the impact of $350,000 in expedited procurement

1108
00:39:39,560 --> 00:39:41,640
costs on overall project margins?

1109
00:39:41,640 --> 00:39:44,440
Finance runs the calculation and returns the result.

1110
00:39:44,440 --> 00:39:46,760
All three agents have coordinated their analysis.

1111
00:39:46,760 --> 00:39:48,960
And the human decision maker gets comprehensive input

1112
00:39:48,960 --> 00:39:50,160
from every domain.

1113
00:39:50,160 --> 00:39:52,360
This is organizational reasoning at scale.

1114
00:39:52,360 --> 00:39:54,400
It's not one smart agent doing everything poorly.

1115
00:39:54,400 --> 00:39:57,480
It's multiple domain expert agents coordinating in real time,

1116
00:39:57,480 --> 00:39:59,600
each reasoning about what it knows best,

1117
00:39:59,600 --> 00:40:01,960
each passing information to agents that need it.

1118
00:40:01,960 --> 00:40:03,760
A2A makes this possible.

1119
00:40:03,760 --> 00:40:06,080
The agents communicate through structured messages.

1120
00:40:06,080 --> 00:40:09,000
No passing ambiguity, no latency delays stacking up.

1121
00:40:09,000 --> 00:40:11,600
When five agents are coordinating on a single decision,

1122
00:40:11,600 --> 00:40:14,560
the last response can't arrive eight seconds after the first.

1123
00:40:14,560 --> 00:40:16,160
Or the coordination breaks down.

1124
00:40:16,160 --> 00:40:18,400
A2A keeps latency in the millisecond range,

1125
00:40:18,400 --> 00:40:20,040
so the coordination stays real time.

1126
00:40:20,040 --> 00:40:21,680
MCP doesn't scale to this density.

1127
00:40:21,680 --> 00:40:23,520
The coordination overhead becomes the bottleneck.

1128
00:40:23,520 --> 00:40:26,240
You can run a few agents, but you hit a ceiling quickly.

1129
00:40:26,240 --> 00:40:28,720
A2A was designed for this, for agentic density.

1130
00:40:28,720 --> 00:40:31,040
For many agents working together at cloud speed.

1131
00:40:31,040 --> 00:40:33,200
The architectural implication is profound.

1132
00:40:33,200 --> 00:40:34,760
You're not building a copilot.

1133
00:40:34,760 --> 00:40:36,760
You're not deploying a single smart assistant.

1134
00:40:36,760 --> 00:40:38,880
You're building an agentic operating system,

1135
00:40:38,880 --> 00:40:40,760
an infrastructure layer where your organization

1136
00:40:40,760 --> 00:40:42,600
reasons through specialized agents.

1137
00:40:42,600 --> 00:40:44,600
Work IQ is the shared context layer.

1138
00:40:44,600 --> 00:40:47,320
All agents access the same organizational reality.

1139
00:40:47,320 --> 00:40:49,840
Agent 365 is the governance layer.

1140
00:40:49,840 --> 00:40:52,480
All agents operate within consistent policy boundaries.

1141
00:40:52,480 --> 00:40:54,360
A2A is the coordination protocol.

1142
00:40:54,360 --> 00:40:56,200
All agents communicate at cloud speed.

1143
00:40:56,200 --> 00:40:58,120
It's not additive improvement on the old model.

1144
00:40:58,120 --> 00:41:00,760
It's structural transformation, but that transformation

1145
00:41:00,760 --> 00:41:01,720
has a cost.

1146
00:41:01,720 --> 00:41:04,240
And understanding that cost is where the reasoning economy

1147
00:41:04,240 --> 00:41:05,440
comes into view.

1148
00:41:05,440 --> 00:41:06,760
The consumption billing model.

1149
00:41:06,760 --> 00:41:08,360
You've now got the infrastructure in place.

1150
00:41:08,360 --> 00:41:10,160
Agents with identities, policies

1151
00:41:10,160 --> 00:41:12,720
enforced at the system level, real time visibility

1152
00:41:12,720 --> 00:41:13,720
into reasoning.

1153
00:41:13,720 --> 00:41:15,480
Multiple specialized agents coordinating

1154
00:41:15,480 --> 00:41:17,360
through A2A at cloud speed.

1155
00:41:17,360 --> 00:41:18,920
It's an elegant architecture.

1156
00:41:18,920 --> 00:41:21,880
It solves the governance problem and the coordination problem.

1157
00:41:21,880 --> 00:41:24,280
But it introduces a new problem, the financial one.

1158
00:41:24,280 --> 00:41:25,200
How do you pay for this?

1159
00:41:25,200 --> 00:41:28,720
For years, Microsoft 365 operated on a predictable model.

1160
00:41:28,720 --> 00:41:29,720
You bought licenses.

1161
00:41:29,720 --> 00:41:30,960
You assigned them to users.

1162
00:41:30,960 --> 00:41:32,440
You paid per user per month.

1163
00:41:32,440 --> 00:41:34,600
The math was straightforward.

1164
00:41:34,600 --> 00:41:39,280
A thousand users times $20 per user equals $20,000 a month.

1165
00:41:39,280 --> 00:41:41,640
Predictable, linear, simple.

1166
00:41:41,640 --> 00:41:43,320
Work IQ doesn't fit that model.

1167
00:41:43,320 --> 00:41:44,880
It's not consumed per user.

1168
00:41:44,880 --> 00:41:47,520
A thousand users might generate zero work IQ reasoning

1169
00:41:47,520 --> 00:41:48,920
if they're not using agents.

1170
00:41:48,920 --> 00:41:50,880
Or they might generate millions of reasoning steps

1171
00:41:50,880 --> 00:41:53,320
if they're working with agent-driven workflows all day.

1172
00:41:53,320 --> 00:41:54,960
The consumption is variable.

1173
00:41:54,960 --> 00:41:56,400
It's unpredictable.

1174
00:41:56,400 --> 00:41:58,840
Per user licensing breaks.

1175
00:41:58,840 --> 00:42:00,520
So Microsoft made a different choice.

1176
00:42:00,520 --> 00:42:04,600
Starting June 16th, 2026, work IQ moves to consumption billing.

1177
00:42:04,600 --> 00:42:05,600
Not per user.

1178
00:42:05,600 --> 00:42:06,480
Not per agent.

1179
00:42:06,480 --> 00:42:07,000
Per user.

1180
00:42:07,000 --> 00:42:07,520
You use it.

1181
00:42:07,520 --> 00:42:08,160
You pay for it.

1182
00:42:08,160 --> 00:42:09,760
The currency is co-pilot credits.

1183
00:42:09,760 --> 00:42:11,880
Think of co-pilot credits as a unified meter

1184
00:42:11,880 --> 00:42:13,560
for Microsoft AI services.

1185
00:42:13,560 --> 00:42:14,880
It's not just work IQ.

1186
00:42:14,880 --> 00:42:16,920
Co-pilot Studio uses co-pilot credits.

1187
00:42:16,920 --> 00:42:19,880
Some advanced co-pilot features consume co-pilot credits.

1188
00:42:19,880 --> 00:42:22,240
And Azure Open AI can consume co-pilot credits

1189
00:42:22,240 --> 00:42:23,480
depending on your configuration.

1190
00:42:23,480 --> 00:42:26,480
It's a shared currency across Microsoft's AI ecosystem.

1191
00:42:26,480 --> 00:42:27,480
One credit pool.

1192
00:42:27,480 --> 00:42:29,280
Multiple services drawing from it.

1193
00:42:29,280 --> 00:42:31,440
The pricing structure is component-based.

1194
00:42:31,440 --> 00:42:32,960
Work IQ breaks into parts.

1195
00:42:32,960 --> 00:42:34,040
Tools have a fixed rate.

1196
00:42:34,040 --> 00:42:35,600
Every tool in vocation.

1197
00:42:35,600 --> 00:42:37,480
Like sending an email, updating a file,

1198
00:42:37,480 --> 00:42:39,960
or creating a record costs 0.1 credits.

1199
00:42:39,960 --> 00:42:40,800
That's fixed.

1200
00:42:40,800 --> 00:42:42,080
It doesn't vary based on complexity.

1201
00:42:42,080 --> 00:42:43,320
The action is the action.

1202
00:42:43,320 --> 00:42:44,840
And your charged per action.

1203
00:42:44,840 --> 00:42:46,800
Chat and context pricing is variable.

1204
00:42:46,800 --> 00:42:48,280
It depends on what you're asking.

1205
00:42:48,280 --> 00:42:51,040
A simple question like, who's assigned to this project?

1206
00:42:51,040 --> 00:42:53,680
Requires minimal reasoning and light context retrieval.

1207
00:42:53,680 --> 00:42:54,960
Maybe a tenth of a credit.

1208
00:42:54,960 --> 00:42:57,520
A complex question that requires synthesizing information

1209
00:42:57,520 --> 00:43:00,640
across multiple systems requires more reasoning compute.

1210
00:43:00,640 --> 00:43:01,760
And more credits.

1211
00:43:01,760 --> 00:43:03,920
The cost reflects the computational load.

1212
00:43:03,920 --> 00:43:06,840
Example scenarios give you the shape of the cost structure.

1213
00:43:06,840 --> 00:43:07,840
Light usage.

1214
00:43:07,840 --> 00:43:10,280
Something simple like identifying tasks assigned to you

1215
00:43:10,280 --> 00:43:12,200
and compiling them into a checklist.

1216
00:43:12,200 --> 00:43:14,280
Runs between 20 and 40 cents.

1217
00:43:14,280 --> 00:43:17,160
The agent queries your task systems retrieves the data

1218
00:43:17,160 --> 00:43:18,640
and formats a response.

1219
00:43:18,640 --> 00:43:19,240
Light work.

1220
00:43:19,240 --> 00:43:19,880
Light cost.

1221
00:43:19,880 --> 00:43:20,880
Medium complexity.

1222
00:43:20,880 --> 00:43:23,520
Say, reviewing customer interviews, identifying recurring

1223
00:43:23,520 --> 00:43:26,120
themes and recommending prioritized roadmap actions

1224
00:43:26,120 --> 00:43:27,080
with trade-offs.

1225
00:43:27,080 --> 00:43:28,960
Runs 30 to 75 cents.

1226
00:43:28,960 --> 00:43:30,880
The agent is doing synthesis work.

1227
00:43:30,880 --> 00:43:32,480
It's reasoning across multiple inputs

1228
00:43:32,480 --> 00:43:34,720
making judgments and explaining trade-offs.

1229
00:43:34,720 --> 00:43:35,680
More compute.

1230
00:43:35,680 --> 00:43:36,720
Higher cost.

1231
00:43:36,720 --> 00:43:37,880
Heavy scenarios.

1232
00:43:37,880 --> 00:43:40,160
Where an agent acts as a true business partner,

1233
00:43:40,160 --> 00:43:42,200
coordinating across multiple systems,

1234
00:43:42,200 --> 00:43:45,160
applying judgment and handling PR sensitive decisions,

1235
00:43:45,160 --> 00:43:46,880
run 50 cents to $1.50.

1236
00:43:46,880 --> 00:43:49,440
The agent is doing organizational level reasoning.

1237
00:43:49,440 --> 00:43:51,440
It's expensive because reasoning at that depth

1238
00:43:51,440 --> 00:43:52,520
requires more compute.

1239
00:43:52,520 --> 00:43:54,840
This isn't like token pricing in a Azure Open AI

1240
00:43:54,840 --> 00:43:56,960
where you're paying per 1,000 tokens in the count gets

1241
00:43:56,960 --> 00:43:57,800
complicated.

1242
00:43:57,800 --> 00:44:00,720
These are scenario costs end to end.

1243
00:44:00,720 --> 00:44:02,520
An agent runs through a workflow.

1244
00:44:02,520 --> 00:44:04,600
The workflow is light, medium or heavy.

1245
00:44:04,600 --> 00:44:06,160
And you're charged the corresponding amounts,

1246
00:44:06,160 --> 00:44:07,200
simple accounting.

1247
00:44:07,200 --> 00:44:10,400
But the shift from per user to consumption-based changes,

1248
00:44:10,400 --> 00:44:12,920
everything about how you think about infrastructure.

1249
00:44:12,920 --> 00:44:15,120
With per user licensing, the variable is headcount.

1250
00:44:15,120 --> 00:44:16,400
You hire 10 new salespeople.

1251
00:44:16,400 --> 00:44:18,560
You buy 10 new licenses done.

1252
00:44:18,560 --> 00:44:20,320
Cost scales linearly with people.

1253
00:44:20,320 --> 00:44:22,680
With consumption-based, the variable is usage.

1254
00:44:22,680 --> 00:44:25,040
One salesperson working with an AI agent all day

1255
00:44:25,040 --> 00:44:27,720
might consume $300 in copilot credits.

1256
00:44:27,720 --> 00:44:30,680
While another might consume 30, you can't predict it from headcount.

1257
00:44:30,680 --> 00:44:32,480
You have to predict it from behavior.

1258
00:44:32,480 --> 00:44:34,640
How intensively will people use these agents?

1259
00:44:34,640 --> 00:44:36,720
How complex are the reasoning workflows?

1260
00:44:36,720 --> 00:44:38,440
How much daily coordination is happening?

1261
00:44:38,440 --> 00:44:40,440
This is where architects need to think differently.

1262
00:44:40,440 --> 00:44:42,000
Every reasoning step now has a cost.

1263
00:44:42,000 --> 00:44:43,680
Every tool invocation has a cost.

1264
00:44:43,680 --> 00:44:45,360
Every context retrieval has a cost.

1265
00:44:45,360 --> 00:44:47,000
You're not paying for potential anymore.

1266
00:44:47,000 --> 00:44:49,120
You're paying for actual, which means you need to start

1267
00:44:49,120 --> 00:44:51,680
thinking about reasoning efficiency, not as a performance

1268
00:44:51,680 --> 00:44:54,000
optimization, but as a cost optimization.

1269
00:44:54,000 --> 00:44:55,600
Because in the reasoning economy,

1270
00:44:55,600 --> 00:44:57,280
inefficiency isn't just slow.

1271
00:44:57,280 --> 00:44:58,240
It's expensive.

1272
00:44:58,240 --> 00:45:00,520
Reasoning ROI and cost optimization.

1273
00:45:00,520 --> 00:45:02,280
You understand the cost structure now.

1274
00:45:02,280 --> 00:45:03,680
Light workflows cost pennies.

1275
00:45:03,680 --> 00:45:05,200
Heavy workflows cost dollars.

1276
00:45:05,200 --> 00:45:07,280
But architects don't usually think about the cost

1277
00:45:07,280 --> 00:45:08,280
of a single workflow.

1278
00:45:08,280 --> 00:45:09,600
They think about the system.

1279
00:45:09,600 --> 00:45:11,120
And this is where the math gets interesting.

1280
00:45:11,120 --> 00:45:12,520
Imagine you deploy an agent.

1281
00:45:12,520 --> 00:45:13,080
It works.

1282
00:45:13,080 --> 00:45:13,800
Users love it.

1283
00:45:13,800 --> 00:45:16,440
You measure the impact and see it saves three hours per user

1284
00:45:16,440 --> 00:45:18,880
every week, with 5,000 potential users.

1285
00:45:18,880 --> 00:45:22,360
That's 500% hours of productivity recovered weekly.

1286
00:45:22,360 --> 00:45:23,200
That is real value.

1287
00:45:23,200 --> 00:45:24,600
It changes the organization.

1288
00:45:24,600 --> 00:45:26,360
But what if that agent is badly designed?

1289
00:45:26,360 --> 00:45:30,000
What if instead of retrieving context ones and reasoning over it,

1290
00:45:30,000 --> 00:45:33,240
the agent re-quaries that same context five times in one workflow?

1291
00:45:33,240 --> 00:45:35,680
It asks the same question to work IQ five times

1292
00:45:35,680 --> 00:45:37,880
when it could have asked once and cashed the answer.

1293
00:45:37,880 --> 00:45:39,320
Each query is a reasoning step.

1294
00:45:39,320 --> 00:45:40,600
Each step costs credits.

1295
00:45:40,600 --> 00:45:43,760
If you multiply five unnecessary calls by 52 weeks,

1296
00:45:43,760 --> 00:45:46,280
for 1,000 concurrent users, you are burning money

1297
00:45:46,280 --> 00:45:48,280
on something the agent should have handled once.

1298
00:45:48,280 --> 00:45:50,120
The difference between good design and bad design

1299
00:45:50,120 --> 00:45:50,960
isn't small.

1300
00:45:50,960 --> 00:45:52,160
It's geometric.

1301
00:45:52,160 --> 00:45:55,000
A well-designed agent that retrieves context efficiently

1302
00:45:55,000 --> 00:45:58,400
and batches its calls can cost the fifth of what a lazy agent costs.

1303
00:45:58,400 --> 00:45:59,320
The output is the same.

1304
00:45:59,320 --> 00:46:00,720
The user sees no difference.

1305
00:46:00,720 --> 00:46:02,800
But the financial impact is five times worse

1306
00:46:02,800 --> 00:46:04,240
because of the architecture.

1307
00:46:04,240 --> 00:46:05,800
This creates a new discipline.

1308
00:46:05,800 --> 00:46:07,080
Re-signing architecture.

1309
00:46:07,080 --> 00:46:08,560
Not just application architecture.

1310
00:46:08,560 --> 00:46:11,600
Re-signing architecture, it's the practice of designing agents

1311
00:46:11,600 --> 00:46:14,240
to minimize cost while keeping them capable.

1312
00:46:14,240 --> 00:46:17,320
So what does efficient reasoning architecture actually look like?

1313
00:46:17,320 --> 00:46:18,520
Batch your retrievals.

1314
00:46:18,520 --> 00:46:21,120
If you need three things from work IQ, account status,

1315
00:46:21,120 --> 00:46:22,960
transaction history and open tickets,

1316
00:46:22,960 --> 00:46:24,360
do not make three separate calls.

1317
00:46:24,360 --> 00:46:25,280
Batch them into one.

1318
00:46:25,280 --> 00:46:27,920
The agent gets everything it needs in a single reasoning step.

1319
00:46:27,920 --> 00:46:28,880
Cost is lower.

1320
00:46:28,880 --> 00:46:30,000
Latency is lower.

1321
00:46:30,000 --> 00:46:31,560
It's better for everyone.

1322
00:46:31,560 --> 00:46:34,680
Next, cash-affemoral context where your policy allows it.

1323
00:46:34,680 --> 00:46:37,400
An agent is working on a project and needs to know the status.

1324
00:46:37,400 --> 00:46:38,160
It retrieves it.

1325
00:46:38,160 --> 00:46:40,160
Now it's going to reference that start us ten times

1326
00:46:40,160 --> 00:46:41,400
over the next few minutes

1327
00:46:41,400 --> 00:46:43,200
while it reasons through different tasks.

1328
00:46:43,200 --> 00:46:44,520
Don't retrieve it again.

1329
00:46:44,520 --> 00:46:46,280
Store it in the agent's working memory.

1330
00:46:46,280 --> 00:46:47,600
The status is a femoral.

1331
00:46:47,600 --> 00:46:49,560
It's only valid for this one session.

1332
00:46:49,560 --> 00:46:51,760
But inside that session, it doesn't change.

1333
00:46:51,760 --> 00:46:53,720
You retrieve once and reference many times.

1334
00:46:53,720 --> 00:46:55,240
The cost drops immediately.

1335
00:46:55,240 --> 00:46:57,840
You should also use lighter models for simple decisions.

1336
00:46:57,840 --> 00:47:00,480
Not every task needs the heaviest model available.

1337
00:47:00,480 --> 00:47:02,560
If an agent is making a binary choice

1338
00:47:02,560 --> 00:47:04,320
like whether to escalate a ticket,

1339
00:47:04,320 --> 00:47:05,760
a lighter model can handle it.

1340
00:47:05,760 --> 00:47:07,400
It's cheaper and it's faster.

1341
00:47:07,400 --> 00:47:09,480
Save the heavy models for nuanced decisions

1342
00:47:09,480 --> 00:47:11,160
that actually require that power.

1343
00:47:11,160 --> 00:47:13,360
Using a heavy model for simple classification

1344
00:47:13,360 --> 00:47:15,680
is like hiring a surgeon to give you a bandaid.

1345
00:47:15,680 --> 00:47:18,320
Structure your workflow to defer heavy reasoning.

1346
00:47:18,320 --> 00:47:20,240
A customer service agent handles a ticket.

1347
00:47:20,240 --> 00:47:22,160
First, it does a lightweight classification

1348
00:47:22,160 --> 00:47:23,680
to find the category and sentiment.

1349
00:47:23,680 --> 00:47:24,880
These are easy.

1350
00:47:24,880 --> 00:47:26,320
Use light reasoning.

1351
00:47:26,320 --> 00:47:28,600
Only if the ticket requires deep judgment,

1352
00:47:28,600 --> 00:47:30,800
like conflicting policies or subtle contexts,

1353
00:47:30,800 --> 00:47:32,520
do you move to heavier reasoning.

1354
00:47:32,520 --> 00:47:34,040
Most tickets never get that far.

1355
00:47:34,040 --> 00:47:35,800
Most stay at a light reasoning cost.

1356
00:47:35,800 --> 00:47:37,440
The financial reality is stock.

1357
00:47:37,440 --> 00:47:40,440
A poorly designed agent at scale can cost ten times more

1358
00:47:40,440 --> 00:47:43,080
than a well-designed one producing the exact same result.

1359
00:47:43,080 --> 00:47:44,200
That isn't an exaggeration.

1360
00:47:44,200 --> 00:47:45,480
It's just the math.

1361
00:47:45,480 --> 00:47:47,360
One agent queries contacts at every step

1362
00:47:47,360 --> 00:47:48,920
while the other batches and caches.

1363
00:47:48,920 --> 00:47:51,520
Same result, but a tenfold difference in the bill.

1364
00:47:51,520 --> 00:47:53,920
Organizations that get this have a massive advantage.

1365
00:47:53,920 --> 00:47:55,200
They deploy efficiently.

1366
00:47:55,200 --> 00:47:57,000
They iterate based on cost metrics.

1367
00:47:57,000 --> 00:47:59,560
When an agent costs too much, they know it immediately.

1368
00:47:59,560 --> 00:48:01,880
They can redesign the workflow or change the model.

1369
00:48:01,880 --> 00:48:03,800
They have visibility so they can act.

1370
00:48:03,800 --> 00:48:06,800
But organizations that treat co-pilot credits like their infinite

1371
00:48:06,800 --> 00:48:08,720
hit a ceiling they didn't expect.

1372
00:48:08,720 --> 00:48:11,800
The agent looked great in the pilot when it handled a hundred tickets.

1373
00:48:11,800 --> 00:48:13,760
Then they rolled it out to 5,000 users.

1374
00:48:13,760 --> 00:48:15,320
Suddenly the credit burn is massive.

1375
00:48:15,320 --> 00:48:17,520
The costs were invisible at a small scale.

1376
00:48:17,520 --> 00:48:19,520
But they are unavoidable at a large one.

1377
00:48:19,520 --> 00:48:21,760
By then, the agent is already in production.

1378
00:48:21,760 --> 00:48:22,880
The damage is done.

1379
00:48:22,880 --> 00:48:25,400
This shift, designing for cost, not just function,

1380
00:48:25,400 --> 00:48:26,400
is a governance problem.

1381
00:48:26,400 --> 00:48:28,160
You can't have every engineer deploying agents,

1382
00:48:28,160 --> 00:48:29,000
however they want.

1383
00:48:29,000 --> 00:48:29,840
You need standards.

1384
00:48:29,840 --> 00:48:30,440
You need review.

1385
00:48:30,440 --> 00:48:31,680
You need guardrails.

1386
00:48:31,680 --> 00:48:32,960
Finops for AI.

1387
00:48:32,960 --> 00:48:34,480
Managing the reasoning budget.

1388
00:48:34,480 --> 00:48:36,160
Those guardrails start with a practice

1389
00:48:36,160 --> 00:48:38,360
that started in cloud computing a decade ago.

1390
00:48:38,360 --> 00:48:39,360
It's called Finops.

1391
00:48:39,360 --> 00:48:40,640
Financial operations.

1392
00:48:40,640 --> 00:48:43,280
It began when companies realized their cloud bills were spiraling

1393
00:48:43,280 --> 00:48:45,360
because nobody was watching what got deployed.

1394
00:48:45,360 --> 00:48:47,960
Engineers spun up infrastructure because it was easy.

1395
00:48:47,960 --> 00:48:49,160
But nobody tracked the cost.

1396
00:48:49,160 --> 00:48:50,160
The bills were a shock.

1397
00:48:50,160 --> 00:48:51,080
Finops was the answer.

1398
00:48:51,080 --> 00:48:52,640
Make cost visible, make it measurable,

1399
00:48:52,640 --> 00:48:54,240
and make it someone's job to manage it.

1400
00:48:54,240 --> 00:48:56,240
AI reasoning needs that same approach.

1401
00:48:56,240 --> 00:48:58,680
Because without discipline, consumption-based billing

1402
00:48:58,680 --> 00:49:00,400
becomes a problem you can't control.

1403
00:49:00,400 --> 00:49:03,320
So, what does Finops for reasoning look like?

1404
00:49:03,320 --> 00:49:04,440
First, visibility.

1405
00:49:04,440 --> 00:49:05,920
You need to track spend by agent.

1406
00:49:05,920 --> 00:49:07,480
Not just the total for the whole company.

1407
00:49:07,480 --> 00:49:08,840
You need agent-level detail.

1408
00:49:08,840 --> 00:49:10,520
Which agents are burning the most credits?

1409
00:49:10,520 --> 00:49:12,120
How much does each execution cost?

1410
00:49:12,120 --> 00:49:13,120
Which ones are efficient?

1411
00:49:13,120 --> 00:49:14,360
And which ones are wasteful?

1412
00:49:14,360 --> 00:49:17,560
The Microsoft 365 Admin Center has dashboards for this.

1413
00:49:17,560 --> 00:49:19,920
But they only work if you actually look at them.

1414
00:49:19,920 --> 00:49:20,840
Second, alerting.

1415
00:49:20,840 --> 00:49:21,720
Set your budgets.

1416
00:49:21,720 --> 00:49:23,960
Define what an acceptable spend looks like.

1417
00:49:23,960 --> 00:49:27,040
If an agent goes 50% over its expected cost in a day,

1418
00:49:27,040 --> 00:49:28,080
you need an alert.

1419
00:49:28,080 --> 00:49:29,360
Not because it's broken,

1420
00:49:29,360 --> 00:49:31,680
but because something changed, maybe usage spiked

1421
00:49:31,680 --> 00:49:33,000
or the workflow got more complex,

1422
00:49:33,000 --> 00:49:34,680
you need to know when it happens.

1423
00:49:34,680 --> 00:49:37,040
Not at the end of the month when you see the bill.

1424
00:49:37,040 --> 00:49:38,480
Third, optimization.

1425
00:49:38,480 --> 00:49:40,320
Once you find the agents that are cost centers,

1426
00:49:40,320 --> 00:49:42,280
you fix them, you redesign the workflows,

1427
00:49:42,280 --> 00:49:44,200
you add caching or switch models,

1428
00:49:44,200 --> 00:49:45,280
then you measure the impact.

1429
00:49:45,280 --> 00:49:46,800
Did the change actually save money?

1430
00:49:46,800 --> 00:49:48,440
This isn't a one-time project.

1431
00:49:48,440 --> 00:49:49,240
It's a loop.

1432
00:49:49,240 --> 00:49:50,640
New agents get measured.

1433
00:49:50,640 --> 00:49:53,960
And inefficient ones get sent back for a redesign.

1434
00:49:53,960 --> 00:49:55,040
Fourth, forecasting.

1435
00:49:55,040 --> 00:49:56,280
If you have three months of data,

1436
00:49:56,280 --> 00:49:58,480
you can see the trends, adoption is picking up

1437
00:49:58,480 --> 00:50:00,200
and new workflows are going live.

1438
00:50:00,200 --> 00:50:02,280
How will that change your credit use next quarter?

1439
00:50:02,280 --> 00:50:04,160
How much budget do you need to set aside?

1440
00:50:04,160 --> 00:50:06,320
If you can't forecast, you can't plan.

1441
00:50:06,320 --> 00:50:08,680
FinOps turns a surprise bill into a prediction.

1442
00:50:08,680 --> 00:50:11,560
You should be asking specific questions about your agents.

1443
00:50:11,560 --> 00:50:13,360
Which ones are actually delivering ROI?

1444
00:50:13,360 --> 00:50:14,880
Usage does not equal value.

1445
00:50:14,880 --> 00:50:17,960
An agent that everyone loves, but costs $5 per run

1446
00:50:17,960 --> 00:50:20,120
might be less valuable than one that costs $0.20

1447
00:50:20,120 --> 00:50:21,480
and saves 30 minutes.

1448
00:50:21,480 --> 00:50:22,560
You need the math.

1449
00:50:22,560 --> 00:50:25,760
Usage times cost per run versus the value of the outcome,

1450
00:50:25,760 --> 00:50:27,640
which agents are just cost centers.

1451
00:50:27,640 --> 00:50:28,800
Some might be experiments.

1452
00:50:28,800 --> 00:50:31,280
You have a budget for them because they are learning opportunities.

1453
00:50:31,280 --> 00:50:33,120
They aren't supposed to be profitable yet,

1454
00:50:33,120 --> 00:50:34,440
but you still have to track them.

1455
00:50:34,440 --> 00:50:36,240
Are they moving toward value?

1456
00:50:36,240 --> 00:50:38,320
Or are they just burning credits with no plan?

1457
00:50:38,320 --> 00:50:39,720
Where is the spend concentrated?

1458
00:50:39,720 --> 00:50:42,680
Maybe one department is using agents heavily while another isn't?

1459
00:50:42,680 --> 00:50:45,600
Maybe one single workflow is responsible for half your costs.

1460
00:50:45,600 --> 00:50:46,480
Is that justified?

1461
00:50:46,480 --> 00:50:48,760
Or is it a sign that something needs to be redesigned?

1462
00:50:48,760 --> 00:50:51,080
And finally, what is the cost per outcome?

1463
00:50:51,080 --> 00:50:52,760
If an agent summarizes meeting notes

1464
00:50:52,760 --> 00:50:56,320
and uses $2 in credits, but a human could do it in 15 minutes,

1465
00:50:56,320 --> 00:50:57,560
you have to look at the numbers.

1466
00:50:57,560 --> 00:51:00,680
At a normal salary, that human time might cost $5.

1467
00:51:00,680 --> 00:51:02,560
The agent was cheaper, but not by a lot.

1468
00:51:02,560 --> 00:51:03,880
Is it worth the complexity?

1469
00:51:03,880 --> 00:51:05,280
You need the data to decide.

1470
00:51:05,280 --> 00:51:07,960
The real power of FinOps is making cost a design constraint.

1471
00:51:07,960 --> 00:51:09,880
When engineers know their agents are being metered,

1472
00:51:09,880 --> 00:51:11,120
it changes how they build.

1473
00:51:11,120 --> 00:51:12,960
They know inefficiency will be noticed.

1474
00:51:12,960 --> 00:51:14,520
It encourages a better architecture.

1475
00:51:14,520 --> 00:51:17,120
It turns cost management into a competitive discipline.

1476
00:51:17,120 --> 00:51:19,760
But none of this works unless it is someone's job.

1477
00:51:19,760 --> 00:51:21,440
Not everyone's job, someone's job.

1478
00:51:21,440 --> 00:51:23,720
A specific role or a team for reasoning operations.

1479
00:51:23,720 --> 00:51:26,000
They own the dashboards and analyze the trends.

1480
00:51:26,000 --> 00:51:27,840
They push back when a deployment looks wasteful.

1481
00:51:27,840 --> 00:51:30,640
Cost discipline doesn't happen because people have good intentions.

1482
00:51:30,640 --> 00:51:33,720
It happens because there is accountability.

1483
00:51:33,720 --> 00:51:35,680
The multi-cloud governance challenge.

1484
00:51:35,680 --> 00:51:37,840
This is where the architecture gets complicated.

1485
00:51:37,840 --> 00:51:41,680
Agent 365 can govern agents running on your Microsoft infrastructure,

1486
00:51:41,680 --> 00:51:44,680
which is a clean setup because those agents live inside your entrotennent

1487
00:51:44,680 --> 00:51:46,880
and inherit M365 policies.

1488
00:51:46,880 --> 00:51:50,840
The governance model is unified, but most organizations don't live in a single cloud.

1489
00:51:50,840 --> 00:51:54,080
You have workloads in Microsoft, critical systems in AWS,

1490
00:51:54,080 --> 00:51:56,560
and you're likely using Google Cloud for analytics.

1491
00:51:56,560 --> 00:51:59,240
You're running custom agents on Azure, others on Bedrock,

1492
00:51:59,240 --> 00:52:01,080
and still more on Vertex AI.

1493
00:52:01,080 --> 00:52:04,040
The agents don't care what cloud they run on, they'll work anywhere.

1494
00:52:04,040 --> 00:52:06,640
But governance becomes a nightmare when you're trying to enforce

1495
00:52:06,640 --> 00:52:09,280
consistent policy across three different cloud providers

1496
00:52:09,280 --> 00:52:10,840
with three different identity systems.

1497
00:52:10,840 --> 00:52:12,960
You're dealing with three different compliance frameworks

1498
00:52:12,960 --> 00:52:14,880
and three different audit mechanisms.

1499
00:52:14,880 --> 00:52:17,160
The problem is structural.

1500
00:52:17,160 --> 00:52:19,600
A policy that makes sense in Microsoft's world,

1501
00:52:19,600 --> 00:52:22,640
like agents can read from SharePoint, but not modify.

1502
00:52:22,640 --> 00:52:24,480
Doesn't directly translate to AWS.

1503
00:52:24,480 --> 00:52:26,000
AWS doesn't have SharePoint.

1504
00:52:26,000 --> 00:52:27,440
It has S3.

1505
00:52:27,440 --> 00:52:28,880
The policy needs to be reframed,

1506
00:52:28,880 --> 00:52:32,600
so agents can read from authorized S3 buckets, but not modify them,

1507
00:52:32,600 --> 00:52:35,480
which means the intent is the same, but the implementation is different.

1508
00:52:35,480 --> 00:52:39,000
Multiply that across every control, every constraint,

1509
00:52:39,000 --> 00:52:40,560
and every policy rule.

1510
00:52:40,560 --> 00:52:43,800
What you end up with is policy translation that's fragile and error-prone.

1511
00:52:43,800 --> 00:52:45,920
Authentication makes the problem even worse.

1512
00:52:45,920 --> 00:52:48,520
In Microsoft, agents get entraagent IDs,

1513
00:52:48,520 --> 00:52:51,040
while in AWS they get IAM roles,

1514
00:52:51,040 --> 00:52:53,480
and in Google Cloud, they get service accounts.

1515
00:52:53,480 --> 00:52:54,720
These aren't the same thing.

1516
00:52:54,720 --> 00:52:55,800
They don't interoperate.

1517
00:52:55,800 --> 00:53:00,360
An Entra agent ID has no meaning in AWS, and an IAM role has no meaning in Google Cloud,

1518
00:53:00,360 --> 00:53:04,040
so you're managing agent identity across three separate directory systems.

1519
00:53:04,040 --> 00:53:06,360
If an agent needs to be revoked, you revoke it in Entra,

1520
00:53:06,360 --> 00:53:09,120
then you revoke it in AWS, and then you do it again in Google.

1521
00:53:09,120 --> 00:53:11,520
If you miss one, the agent is only half revoked.

1522
00:53:11,520 --> 00:53:13,080
It still has access somewhere.

1523
00:53:13,080 --> 00:53:15,440
Compliance frameworks diverge too.

1524
00:53:15,440 --> 00:53:18,360
Microsoft's compliance stack uses PerView and Defender,

1525
00:53:18,360 --> 00:53:20,680
but AWS has GodDuty and config,

1526
00:53:20,680 --> 00:53:24,880
while Google Cloud relies on Security Command Center and VPC service controls.

1527
00:53:24,880 --> 00:53:29,600
The controls are similar conceptually because you want to detect anomalies and maintain audit trails,

1528
00:53:29,600 --> 00:53:31,560
but the actual mechanisms are different.

1529
00:53:31,560 --> 00:53:35,160
The threat detection algorithm in Defender doesn't translate to GodDuty,

1530
00:53:35,160 --> 00:53:38,120
so you aren't using the same rules or getting the same insights.

1531
00:53:38,120 --> 00:53:39,960
When an agent is behaving strangely,

1532
00:53:39,960 --> 00:53:41,680
one cloud detects it immediately,

1533
00:53:41,680 --> 00:53:44,200
while another might take hours to flag the issue.

1534
00:53:44,200 --> 00:53:47,560
Audit trails are perhaps the most frustrating part of this divergence.

1535
00:53:47,560 --> 00:53:51,680
A multi-cloud agent executes a workflow where it reads from S3 in AWS,

1536
00:53:51,680 --> 00:53:56,160
updates a record in Dynamics in Azure, and then queries a BigQuery dataset in Google Cloud.

1537
00:53:56,160 --> 00:53:59,840
The entire workflow is logged, but it split across three different places.

1538
00:53:59,840 --> 00:54:02,400
AWS CloudTrail shows the S3 access,

1539
00:54:02,400 --> 00:54:04,880
as your audit log shows the Dynamics update,

1540
00:54:04,880 --> 00:54:08,000
and Google Cloud's audit log shows the BigQuery query.

1541
00:54:08,000 --> 00:54:11,880
Reconstructing what the agent actually did requires pulling logs from three systems

1542
00:54:11,880 --> 00:54:14,160
and manually correlating them by timestamp.

1543
00:54:14,160 --> 00:54:17,080
Try doing that at scale across thousands of agent executions

1544
00:54:17,080 --> 00:54:20,160
when you're looking for compliance violations or security incidents.

1545
00:54:20,160 --> 00:54:24,240
The current answer, agent 365 provides is a hybrid model, not a unified one.

1546
00:54:24,240 --> 00:54:28,000
Agent 365 acts as a registry that knows about agents across clouds,

1547
00:54:28,000 --> 00:54:30,960
including their identities, their owners, and how to correlate them,

1548
00:54:30,960 --> 00:54:32,960
but enforcement happens locally.

1549
00:54:32,960 --> 00:54:35,840
Microsoft agents are governed through Azure security controls.

1550
00:54:35,840 --> 00:54:38,960
AWS agents are governed through IAM and GodDuty,

1551
00:54:38,960 --> 00:54:42,160
and Google agents are governed through VPC service controls.

1552
00:54:42,160 --> 00:54:46,160
Agent 365 is the lens that lets you see all agents in one place,

1553
00:54:46,160 --> 00:54:49,440
but the actual policy enforcement happens in three separate places.

1554
00:54:49,440 --> 00:54:51,760
This creates massive operational complexity.

1555
00:54:51,760 --> 00:54:55,200
You define a policy stating agents cannot ex-filterate data,

1556
00:54:55,200 --> 00:54:59,840
but that policy needs to be expressed in intra-conditional access language for Microsoft agents.

1557
00:54:59,840 --> 00:55:03,040
Then you write it in AWS IAM policy language for AWS agents

1558
00:55:03,040 --> 00:55:05,680
and Google Cloud IAM policy language for Google agents.

1559
00:55:05,680 --> 00:55:08,880
It's not automation, it's manual translation across three systems.

1560
00:55:08,880 --> 00:55:11,920
When you need to change a policy, you change it in three places,

1561
00:55:11,920 --> 00:55:15,520
which means the risk of drift and misconfiguration is incredibly high.

1562
00:55:15,520 --> 00:55:17,920
It's workable, and many organizations operate this way,

1563
00:55:17,920 --> 00:55:20,960
but it isn't seamless. It isn't unified governance.

1564
00:55:20,960 --> 00:55:22,480
It's orchestrated governance.

1565
00:55:22,480 --> 00:55:24,400
Your orchestrating policy across clouds,

1566
00:55:24,400 --> 00:55:28,000
which requires discipline, process, and tooling that understands all three clouds

1567
00:55:28,000 --> 00:55:30,640
to translate intent into specific controls.

1568
00:55:30,640 --> 00:55:34,720
This is the emerging problem that the next generation of governance tools will need to solve,

1569
00:55:34,720 --> 00:55:36,400
the pilot to production journey.

1570
00:55:36,400 --> 00:55:38,240
Understanding the architecture is one thing,

1571
00:55:38,240 --> 00:55:40,960
but executing it in your organization is another.

1572
00:55:40,960 --> 00:55:43,360
Most enterprises fail not because the technology doesn't work,

1573
00:55:43,360 --> 00:55:44,960
but because they implemented the wrong way.

1574
00:55:44,960 --> 00:55:47,200
They deploy agents at scale without governance,

1575
00:55:47,200 --> 00:55:48,560
skip the foundational work,

1576
00:55:48,560 --> 00:55:52,160
and ignore cost signals until they hit walls they didn't anticipate.

1577
00:55:52,160 --> 00:55:56,080
The path from understanding to success runs through a structured progression.

1578
00:55:56,080 --> 00:55:57,040
Start small.

1579
00:55:57,040 --> 00:56:00,720
Phase one is a pilot where you pick one low-risk agent in one business unit,

1580
00:56:00,720 --> 00:56:01,840
not organization-wide,

1581
00:56:01,840 --> 00:56:04,480
not multiple agents at the same time, just one.

1582
00:56:04,480 --> 00:56:06,880
A single, contained agent doing something that matters,

1583
00:56:06,880 --> 00:56:09,120
but isn't mission-critical is the best place to start.

1584
00:56:09,120 --> 00:56:12,720
Maybe it's a customer service agent that handles tier one ticket triage,

1585
00:56:12,720 --> 00:56:16,560
or perhaps an operations agent that recommends actions but doesn't execute them.

1586
00:56:16,560 --> 00:56:19,280
You want something where the downside of failure is manageable.

1587
00:56:19,280 --> 00:56:23,040
The pilot isn't about proving the technology works, the technology already works.

1588
00:56:23,040 --> 00:56:25,600
The pilot is about proving the governance model works.

1589
00:56:25,600 --> 00:56:27,840
Can you create an agent with an entra-agent ID,

1590
00:56:27,840 --> 00:56:30,160
assign it policies and monitor it effectively?

1591
00:56:30,160 --> 00:56:33,200
Can you detect when it misbehaves, measure its cost,

1592
00:56:33,200 --> 00:56:35,040
and understand its total impact?

1593
00:56:35,040 --> 00:56:38,000
The pilot is answering operational questions, not technical ones.

1594
00:56:38,000 --> 00:56:38,960
Measure everything.

1595
00:56:38,960 --> 00:56:42,560
Look at the adoption rate to see what percentage of eligible users actually interact

1596
00:56:42,560 --> 00:56:44,560
with the agent versus those who ignore it.

1597
00:56:44,560 --> 00:56:48,320
Track the reasoning cost per outcome to see how much it costs the organization

1598
00:56:48,320 --> 00:56:50,080
to get one result from the agent.

1599
00:56:50,080 --> 00:56:54,720
Monitor agent uptime and reliability to ensure it's producing correct output consistently.

1600
00:56:54,720 --> 00:56:58,400
Gage user trust through feedback and usage patterns to see if people actually believe

1601
00:56:58,400 --> 00:57:00,320
the agent or if they're second-guessing it.

1602
00:57:00,320 --> 00:57:02,880
Watch for compliance violations trending towards zero

1603
00:57:02,880 --> 00:57:05,760
to ensure the agent is staying within policy boundaries.

1604
00:57:05,760 --> 00:57:07,600
You're running this pilot for probably three months,

1605
00:57:07,600 --> 00:57:09,440
which is enough time to see real patterns.

1606
00:57:09,440 --> 00:57:12,800
You'll discover that the agent handles certain types of inputs beautifully

1607
00:57:12,800 --> 00:57:14,080
while it struggles with others.

1608
00:57:14,080 --> 00:57:16,640
You'll find workflows that are more efficient with the agent

1609
00:57:16,640 --> 00:57:18,080
and others that are actually slower.

1610
00:57:18,080 --> 00:57:21,920
You'll see costs surprises where certain workflows burn more credits than you expected.

1611
00:57:21,920 --> 00:57:23,840
You'll see user behavior patterns,

1612
00:57:23,840 --> 00:57:26,880
like which users embrace agents and which ones resist.

1613
00:57:26,880 --> 00:57:29,600
Phase two takes what you learned and expands carefully.

1614
00:57:29,600 --> 00:57:32,880
You're now running three to five agents across two to three departments.

1615
00:57:32,880 --> 00:57:35,920
You're doing this deliberately, not haphazardly,

1616
00:57:35,920 --> 00:57:38,400
by establishing governance patterns that worked in the pilot

1617
00:57:38,400 --> 00:57:39,920
and applying them consistently.

1618
00:57:39,920 --> 00:57:42,640
You're building the muscle of agent architecture and operations.

1619
00:57:42,640 --> 00:57:45,280
You're still small enough to fix things quickly if they break,

1620
00:57:45,280 --> 00:57:47,360
but large enough to see organizational patterns.

1621
00:57:47,360 --> 00:57:50,480
This is where you're establishing the governance foundation that will scale.

1622
00:57:50,480 --> 00:57:51,600
You're not doing this ad hoc.

1623
00:57:51,600 --> 00:57:52,720
You're building the registry.

1624
00:57:52,720 --> 00:57:55,440
You're creating policy templates, defining approval workflows

1625
00:57:55,440 --> 00:57:57,760
for new agents and setting up cost dashboards.

1626
00:57:57,760 --> 00:58:00,880
You're training the people who will be responsible for agent operations.

1627
00:58:00,880 --> 00:58:02,640
You're creating the operating model,

1628
00:58:02,640 --> 00:58:04,480
not just deploying technology.

1629
00:58:04,480 --> 00:58:05,920
Phase three is scale production.

1630
00:58:05,920 --> 00:58:08,080
You roll out across the organization

1631
00:58:08,080 --> 00:58:10,960
and now your process is matter because they're running at scale.

1632
00:58:10,960 --> 00:58:14,560
You automate governance because you can't manually review hundreds of agents.

1633
00:58:14,560 --> 00:58:17,920
You optimize for cost because small inefficiencies become massive,

1634
00:58:17,920 --> 00:58:19,440
expensive problems at scale.

1635
00:58:19,440 --> 00:58:21,280
You're running dozens or hundreds of agents,

1636
00:58:21,280 --> 00:58:23,360
hundreds of thousands of executions per day

1637
00:58:23,360 --> 00:58:25,360
and millions of reasoning steps per week.

1638
00:58:25,360 --> 00:58:28,400
What kills organizations in this transition starting too big?

1639
00:58:28,400 --> 00:58:31,280
Some leaders see the vision and want agents everywhere immediately,

1640
00:58:31,280 --> 00:58:35,280
so they deploy dozens of agents across the organization without any governance.

1641
00:58:35,280 --> 00:58:38,400
Six months later, they've got agents accessing data they shouldn't,

1642
00:58:38,400 --> 00:58:40,400
duplicating work and generating costs,

1643
00:58:40,400 --> 00:58:42,240
spikes, they're in chaos.

1644
00:58:42,240 --> 00:58:44,320
They have to pull everything back and start over,

1645
00:58:44,320 --> 00:58:46,640
which means they've lost momentum and internal trust.

1646
00:58:46,640 --> 00:58:48,080
Moving too fast is another killer.

1647
00:58:48,080 --> 00:58:50,800
You skip the identity work and the governance foundation

1648
00:58:50,800 --> 00:58:52,880
to jump straight to agent deployment.

1649
00:58:52,880 --> 00:58:54,160
Then you hit a compliance problem

1650
00:58:54,160 --> 00:58:57,280
because an agent accessed protected data it shouldn't have.

1651
00:58:57,280 --> 00:59:00,000
Now you're retrofitting governance into a production system

1652
00:59:00,000 --> 00:59:02,080
while dealing with a breach or an audit finding.

1653
00:59:02,080 --> 00:59:04,880
The fix is messy, expensive and completely avoidable.

1654
00:59:04,880 --> 00:59:06,960
Ignoring cost is the final trap.

1655
00:59:06,960 --> 00:59:10,320
The pilot looked cheap because you weren't measuring carefully,

1656
00:59:10,320 --> 00:59:12,480
but now at scale the credit burn is massive.

1657
00:59:12,480 --> 00:59:14,320
It's in the budget plans you didn't anticipate

1658
00:59:14,320 --> 00:59:16,160
and it becomes a financial surprise.

1659
00:59:16,160 --> 00:59:18,160
You're defending purchases you didn't plan for

1660
00:59:18,160 --> 00:59:20,160
and cutting agents just to control costs.

1661
00:59:20,160 --> 00:59:22,800
You're losing the benefits you built the system to achieve

1662
00:59:22,800 --> 00:59:26,000
even though you could have managed this with early cost discipline.

1663
00:59:26,000 --> 00:59:29,040
The organizations that succeed are the ones that pilot rigorously.

1664
00:59:29,040 --> 00:59:31,200
They establish governance patterns, move deliberately

1665
00:59:31,200 --> 00:59:32,480
and measure continuously.

1666
00:59:32,480 --> 00:59:34,480
They optimize based on data and they don't rush

1667
00:59:34,480 --> 00:59:36,080
they don't ignore warning signals.

1668
00:59:36,080 --> 00:59:38,480
They build the operating model as they scale the agents.

1669
00:59:39,280 --> 00:59:43,120
The organizational shift from user-centric to agent-centric

1670
00:59:43,120 --> 00:59:46,800
for 20 years IT has focused on one metric.

1671
00:59:46,800 --> 00:59:50,480
Users that number drove everything your head count determined your budget

1672
00:59:50,480 --> 00:59:52,800
10,000 employees meant 10,000 licenses.

1673
00:59:52,800 --> 00:59:54,960
The math was linear and predictable.

1674
00:59:54,960 --> 00:59:59,200
Every hiring cycle you bought more, every reduction you freed them up

1675
00:59:59,200 --> 01:00:02,320
supporting users meant managing devices, responding to tickets

1676
01:00:02,320 --> 01:00:03,680
and securing identities.

1677
01:00:03,680 --> 01:00:06,720
The entire system was built for humans agents break that model

1678
01:00:06,720 --> 01:00:10,480
because agents aren't users they don't have help desk tickets they don't hire on

1679
01:00:10,480 --> 01:00:12,800
and they don't retire they aren't units of head count

1680
01:00:12,800 --> 01:00:16,800
they are units of work and organizing around work requires a different way of thinking.

1681
01:00:16,800 --> 01:00:21,120
An agent-centric organization asks new questions how many agents actually exist

1682
01:00:21,120 --> 01:00:24,320
not just the ones you deployed but the shadow agents it doesn't know about

1683
01:00:24,320 --> 01:00:26,160
what are they actually capable of doing

1684
01:00:26,160 --> 01:00:30,320
who is accountable for their behavior is the reasoning cost trending up or down

1685
01:00:30,320 --> 01:00:33,360
these questions require new roles you need an agent architect

1686
01:00:33,360 --> 01:00:36,480
this person designs the workflows they understand the business process

1687
01:00:36,480 --> 01:00:41,280
and they understand the governance constraints they decide which reasoning steps are necessary

1688
01:00:41,280 --> 01:00:45,600
and which are just wasteful you need agent owners every agent needs a business sponsor

1689
01:00:45,600 --> 01:00:49,600
someone accountable for the outcome not the person who built it the person who owns the results

1690
01:00:49,600 --> 01:00:53,440
if the agent is broken that's on them if it costs too much that's on them

1691
01:00:53,440 --> 01:00:58,560
a reasoning analyst is the next role this person watches the spend they identify where the money is going

1692
01:00:58,560 --> 01:01:02,880
and which agents are being wasteful they are like fin ops for AI finally you need an agent security

1693
01:01:02,880 --> 01:01:07,520
specialist this person investigates anomalies is the agent compromised or did the workflow just

1694
01:01:07,520 --> 01:01:11,840
change they know what normal looks like and they know when to be concerned your existing team can

1695
01:01:11,840 --> 01:01:16,320
evolve into these roles infrastructure becomes reasoning operations developers become architects

1696
01:01:16,320 --> 01:01:22,080
support personnel become owners but they can't do this without new tools new dashboards new monitoring new

1697
01:01:22,080 --> 01:01:30,080
processes the shift is subtle but it's profound the unit of account moves from people to agents

1698
01:01:30,080 --> 01:01:34,640
support moves from reactive to proactive cost control moves from licenses to consumption but here's

1699
01:01:34,640 --> 01:01:40,880
the problem right now intelligence is external to it business units build their own tools in the shadows

1700
01:01:40,880 --> 01:01:45,280
it can't see it so they can't govern it agent centric organizations make intelligence a managed

1701
01:01:45,280 --> 01:01:50,480
service agents are visible they are monitored they are optimized intelligence becomes infrastructure

1702
01:01:50,480 --> 01:01:54,640
just like networks and storage that is the shift intelligence moves from the shadows into the

1703
01:01:54,640 --> 01:02:00,160
operating model building the governance foundation you have a path now pilot one agent expand to a

1704
01:02:00,160 --> 01:02:04,960
handful then scale but that path only works if you build the foundation first you can't just flip

1705
01:02:04,960 --> 01:02:09,840
a switch on governance it has to exist before you scale there are five components you need first

1706
01:02:09,840 --> 01:02:14,800
the agent registry this isn't just a list it's a system of record what does the agent do who owns it

1707
01:02:14,800 --> 01:02:19,840
what data can it access the registry is your inventory it is the single source of truth because

1708
01:02:19,840 --> 01:02:26,400
you can't govern what you can't see second policy templates you're going to get dozens of requests

1709
01:02:26,400 --> 01:02:32,080
we need a service agent we need a procurement agent if you design policies from scratch every time

1710
01:02:32,080 --> 01:02:37,120
you will never finish build templates instead a service agent template specifies what it can read

1711
01:02:37,120 --> 01:02:41,440
and what it can't modify new agents inherit those rules governance stays consistent

1712
01:02:41,440 --> 01:02:48,640
and deployment moves faster third approval workflows who decides if an agent goes live it can't be

1713
01:02:48,640 --> 01:02:52,960
the person who built it that's a conflict of interest you need a gate the architect reviews the design

1714
01:02:52,960 --> 01:02:57,360
security reviews the risk the business owner approves the launch the gate stops agents from going

1715
01:02:57,360 --> 01:03:03,360
into production half baked fourth monitoring dashboards you need real-time visibility not a report at

1716
01:03:03,360 --> 01:03:08,160
the end of the month you need to see uptime reasoning costs and policy violations right now the

1717
01:03:08,160 --> 01:03:13,760
dashboard is the nervous system it tells you something is wrong before it becomes a disaster fifth incident

1718
01:03:13,760 --> 01:03:18,640
playbooks when things go wrong and they will you need a response if an agent uses too many credits what

1719
01:03:18,640 --> 01:03:23,600
happens the playbook tells you who to alert how to investigate and when to roll back you aren't

1720
01:03:23,600 --> 01:03:27,840
making it up as you go you have a process these five pieces aren't optional they are the difference

1721
01:03:27,840 --> 01:03:32,880
between success and a total mess the most common mistake is building governance too late by then you

1722
01:03:32,880 --> 01:03:37,680
have shadow agents everywhere you have cost overruns and you have agents you can't shut down without

1723
01:03:37,680 --> 01:03:42,880
breaking the business retrofitting governance is nightmare work it's expensive and it's disruptive

1724
01:03:43,760 --> 01:03:48,480
the discipline here feels wrong at first start with constraints let agents operate inside those

1725
01:03:48,480 --> 01:03:53,120
walls don't add governance later it's harder and it's more painful organizations that succeed

1726
01:03:53,120 --> 01:03:57,360
build the foundation first the constraints aren't obstacles they are the structure that makes

1727
01:03:57,360 --> 01:04:02,320
scale possible this work happens before you hit full production you build the infrastructure while

1728
01:04:02,320 --> 01:04:06,880
running a few agents you aren't rushing to deploy more you are building the model the hidden

1729
01:04:06,880 --> 01:04:11,680
costs of a genetic transformation everyone wants to talk about the exciting stuff the agents reasoning

1730
01:04:11,680 --> 01:04:16,400
over data the multi-step workflows the speed but nobody wants to talk about the cost i'm not just

1731
01:04:16,400 --> 01:04:20,640
talking about co-pilot credits i'm talking about the real costs the ones that surprise you halfway

1732
01:04:20,640 --> 01:04:25,600
through the project yes there is a direct cost for every reasoning step and every tool you call

1733
01:04:25,600 --> 01:04:30,160
that is the obvious line item light workflows cost pennies while heavy workflows cost dollars

1734
01:04:30,160 --> 01:04:34,720
and while that adds up it isn't your biggest problem the hidden cost dwarf the reasoning credits

1735
01:04:34,720 --> 01:04:39,040
it starts with governance infrastructure you need dashboards to monitor agents and systems to

1736
01:04:39,040 --> 01:04:43,680
manage your registry some of this comes from Microsoft but you will likely end up building or buying

1737
01:04:43,680 --> 01:04:48,240
the rest then you have the process costs approval workflows and architecture reviews don't run

1738
01:04:48,240 --> 01:04:53,200
themselves they require someone's time you are paying for policy reviews incident investigations

1739
01:04:53,200 --> 01:04:58,480
and a dedicated agent operations team architects analysts and security specialists are new roles

1740
01:04:58,480 --> 01:05:02,880
and they are definitely not free identity hygiene is another massive cost that most organizations

1741
01:05:02,880 --> 01:05:07,200
underestimate before an agent can operate safely your permissions have to be perfect and let's be

1742
01:05:07,200 --> 01:05:12,800
honest they probably aren't most companies have years of accumulated permission debt people move roles

1743
01:05:12,800 --> 01:05:17,440
but keep their old access and sensitive data ends up shared much more broadly than it should be agents

1744
01:05:17,440 --> 01:05:21,920
will expose every single one of these floors before you roll them out you have to clean the house

1745
01:05:21,920 --> 01:05:26,480
that means hundreds of hours spent on access reviews and fixing broken inheritance it isn't optional

1746
01:05:26,480 --> 01:05:30,960
it's a prerequisite training costs are also substantial your teams don't know how to architect

1747
01:05:30,960 --> 01:05:35,520
for agents or how to monitor their ROI yet they have to learn which means paying for instructor time

1748
01:05:35,520 --> 01:05:40,640
and losing productivity while they study this isn't a one week onboarding it's ongoing new people join

1749
01:05:40,640 --> 01:05:44,960
and new agent types emerge so the training debt just keeps compounding monitoring is more than

1750
01:05:44,960 --> 01:05:49,680
just a dashboard you are instrumenting agents across the entire company and collecting telemetry

1751
01:05:49,680 --> 01:05:54,160
from dozens of sources connecting your monitoring stack to your cost and compliance systems is a

1752
01:05:54,160 --> 01:05:58,560
major integration project then there is incident response this is the cost nobody budgets for until

1753
01:05:58,560 --> 01:06:03,200
something breaks when an agent misbehaves you need trained people and ready to go playbooks to shut

1754
01:06:03,200 --> 01:06:08,320
it down you need people on call the moment agents touch critical systems the total cost of ownership

1755
01:06:08,320 --> 01:06:13,120
for an agent program is usually three to five times the cost of the credits if you spend a dollar

1756
01:06:13,120 --> 01:06:17,360
on reasoning you will spend four more on the people and infrastructure to keep it safe

1757
01:06:17,360 --> 01:06:22,160
organizations that only budget for credits end up underfunded they cut corners governance weekends

1758
01:06:22,160 --> 01:06:26,800
and the whole program eventually struggles but here's the thing the ROI justifies it if you do it right

1759
01:06:26,800 --> 01:06:31,840
automation kills manual work agents work 24 hours a day without fatigue or sick days they make fewer

1760
01:06:31,840 --> 01:06:36,960
errors and stay consistent which makes your decision making much faster these benefits compound over time

1761
01:06:36,960 --> 01:06:42,400
if a well run program saves three hours a week for a thousand users you just reclaimed 500 person

1762
01:06:42,400 --> 01:06:46,400
years of work that is transformative and it's worth paying for the infrastructure to get there

1763
01:06:46,400 --> 01:06:50,320
the financial case is overwhelmingly positive but only if you account for the costs upfront

1764
01:06:50,320 --> 01:06:54,400
organizations that treat governance as optional usually find out the truth in year two when

1765
01:06:54,400 --> 01:06:59,680
the budget is already gone they try to run a production system on a pilot budget and it fails

1766
01:06:59,680 --> 01:07:03,920
get the full cost picture before you start it's the only way to build something that lasts

1767
01:07:03,920 --> 01:07:09,920
measuring success metrics that matter this is where most organizations go wrong they measure the

1768
01:07:09,920 --> 01:07:14,320
wrong things they celebrate deployments that should actually worry them they stare at vanity metrics

1769
01:07:14,320 --> 01:07:19,200
while the program quietly fails start by ignoring the metrics everyone loves like total reasoning credit

1770
01:07:19,200 --> 01:07:23,600
spent that number is meaningless without context you could spend ten thousand dollars a month and

1771
01:07:23,600 --> 01:07:29,040
create massive value or you could spend the same amount burning money on broken processes credit

1772
01:07:29,040 --> 01:07:33,440
spend tells you nothing about success so don't use it to justify your budget the number of agents

1773
01:07:33,440 --> 01:07:38,080
deployed is another seductive vanity metric maybe you deployed a hundred agents but half of them

1774
01:07:38,080 --> 01:07:42,880
aren't being used in a quarter are just duplicating work quantity has nothing to do with value if you

1775
01:07:42,880 --> 01:07:47,120
measure success by agent count you are building for the wrong reason the same goes for adoption rates

1776
01:07:47,120 --> 01:07:52,160
seeing that 70% of your workforce uses agents looks great in a slide deck but if those people are

1777
01:07:52,160 --> 01:07:56,400
being forced to use them or if the agents are just shifting work around that percentage is just

1778
01:07:56,400 --> 01:08:01,280
noise vanity metrics feel good because they are easy to report but they are lying to you the metrics

1779
01:08:01,280 --> 01:08:06,400
that actually matter are harder to find but they tell the truth start with reasoning cost per outcome

1780
01:08:06,400 --> 01:08:12,240
how much does it cost the company to get a specific result if an agent generates a report in eight

1781
01:08:12,240 --> 01:08:16,320
minutes for two dollars and that same report takes a human four hours at eighty dollars the math is

1782
01:08:16,320 --> 01:08:21,280
easy the agent is thirty times cheaper and much faster but if an agent takes a dollar to do a five

1783
01:08:21,280 --> 01:08:26,000
minute task it's eighty times more expensive than the human cost per outcome tells you immediately if

1784
01:08:26,000 --> 01:08:31,040
an agent actually makes sense then there is agent uptime if an agent is supposed to be working and it

1785
01:08:31,040 --> 01:08:35,520
isn't that is a failure ninety nine percent uptime sounds okay until you realize that means seven

1786
01:08:35,520 --> 01:08:39,360
hours of downtime every month if your agent is critical infrastructure you have to treat those

1787
01:08:39,360 --> 01:08:44,560
outages seriously you also need to track compliance violations agents will eventually violate a policy

1788
01:08:44,560 --> 01:08:49,200
because of configuration drift or a scenario you didn't expect these violations should be rare

1789
01:08:49,200 --> 01:08:53,360
and trending towards zero if that number is staying the same or going up your governance model is

1790
01:08:53,360 --> 01:08:58,560
broken time saved per interaction is the most important signal you need to measure the actual time

1791
01:08:58,560 --> 01:09:03,280
eliminated from a workflow if an agent turns a twenty minute task into a four minute task those

1792
01:09:03,280 --> 01:09:08,480
sixteen minutes add up fast a hundred executions a day saves you a person year of work every few months

1793
01:09:08,480 --> 01:09:14,000
that is a real measurable win finally look at user trust but don't use a survey watch their behavior

1794
01:09:14,000 --> 01:09:18,880
do they use the agent when they have a choice or do they find workarounds do they accept the agents

1795
01:09:18,880 --> 01:09:23,360
output or do they spend all their time editing it if people aren't choosing to use the agents the

1796
01:09:23,360 --> 01:09:28,080
deployment doesn't matter the core question is whether the agent makes the company faster smarter

1797
01:09:28,080 --> 01:09:33,600
or more compliant faster means more throughput smarter means fewer errors more compliant means

1798
01:09:33,600 --> 01:09:38,480
lower risk if an agent doesn't move the needle on one of those three things it is just expensive overhead

1799
01:09:38,480 --> 01:09:43,200
kill it and move the budget somewhere else measure the outcome not the activity outcome is the only

1800
01:09:43,200 --> 01:09:49,600
thing that determines if you've built something that lasts the uncomfortable truth most agents will fail

1801
01:09:49,600 --> 01:09:53,680
you've built the foundation you've deployed the pilots and measured the costs you've trained the

1802
01:09:53,680 --> 01:09:58,400
teams and established the governance you are ready to scale and here is what nobody wants to hear

1803
01:09:58,400 --> 01:10:02,960
most of your agents won't make it not all of them some will be transformative those few will

1804
01:10:02,960 --> 01:10:07,920
justify the entire investment but statistically more agents will fail then succeed they'll get

1805
01:10:07,920 --> 01:10:12,080
deployed and run for a few months but then they'll stall usage will drop the business will stop

1806
01:10:12,080 --> 01:10:16,320
investing in them they'll become technical debt eventually they'll get shut down this isn't a failure

1807
01:10:16,320 --> 01:10:21,360
of the technology it's a failure of expectations why do agents fail it starts with misaligned complexity

1808
01:10:21,360 --> 01:10:26,560
the business process looked automatable in theory but in practice it's full of edge cases the agent

1809
01:10:26,560 --> 01:10:32,240
handles 90% of scenarios cleanly but the other 10% are exceptions that require human judgment the agent

1810
01:10:32,240 --> 01:10:37,200
gets stuck on those it escalates everything that's complicated it becomes a filter that passes only

1811
01:10:37,200 --> 01:10:42,320
the hard work to humans the humans still do most of the work the agent adds overhead value is negative

1812
01:10:42,320 --> 01:10:47,520
the agent gets killed user trust failures are just as common an agent makes a confident recommendation

1813
01:10:47,520 --> 01:10:52,800
that turns out to be wrong not catastrophically just wrong a salesperson trusted the agents lead scoring

1814
01:10:52,800 --> 01:10:56,800
because the agent said a prospect was high probability the salesperson spent time on it but nothing

1815
01:10:56,800 --> 01:11:01,040
happened now the salesperson doesn't trust the agent anymore they start second guessing everything

1816
01:11:01,040 --> 01:11:05,600
it says they slow down and double check the agent becomes a slower way to do work they already knew

1817
01:11:05,600 --> 01:11:10,640
how to do why use it then there is the reasoning cost you deployed an agent thinking it would save time

1818
01:11:10,640 --> 01:11:15,520
and it does but barely it saves three minutes per execution but each execution costs a dollar in

1819
01:11:15,520 --> 01:11:21,040
credits do the math to save eight hours of human time you need to run the agent 160 times that costs

1820
01:11:21,040 --> 01:11:26,000
160 dollars the human time is worth maybe 60 dollars at average cost you're spending almost three

1821
01:11:26,000 --> 01:11:31,040
times more on the agent than on the human alternative the math doesn't work the agent gets shut down

1822
01:11:31,040 --> 01:11:35,680
governance constraints also strangle capability the agent was designed to do something useful

1823
01:11:35,680 --> 01:11:40,560
but policy restrictions limit what it can actually do it can't access the data it needs it can't

1824
01:11:40,560 --> 01:11:44,640
take the actions that would deliver value it can only make suggestions the humans still have to

1825
01:11:44,640 --> 01:11:48,960
execute the agent becomes advisory instead of autonomous advisory agents have a hard time

1826
01:11:48,960 --> 01:11:53,920
justifying their cost edge cases overwhelm the design the agent was trained on historical data

1827
01:11:53,920 --> 01:11:58,480
but real world conditions include scenarios the training didn't anticipate a customer service agent

1828
01:11:58,480 --> 01:12:02,960
was trained on normal support tickets but then it encounters a ticket where the customer is threatening

1829
01:12:02,960 --> 01:12:07,840
legal action the agent doesn't know what to do it escalates three quarters of escalations are edge

1830
01:12:07,840 --> 01:12:13,440
cases the agent isn't reducing work it's creating triage overhead this is the hard truth many agents fail

1831
01:12:13,440 --> 01:12:18,160
because the problem was harder than anticipated or the solution was less applicable than hoped

1832
01:12:18,160 --> 01:12:22,640
not because the technology is broken because the assumptions were wrong the organizations that succeed

1833
01:12:22,640 --> 01:12:27,840
are ruthless they pilot aggressively they measure continuously they kill agents that aren't working

1834
01:12:27,840 --> 01:12:31,680
they don't get emotionally attached to deployments they don't keep underperforming agents running

1835
01:12:31,680 --> 01:12:36,080
because they've invested in them they cut losses they double down on what works they iterate based

1836
01:12:36,080 --> 01:12:40,960
on data the organizations that fail to the opposite they deploy an agent and it underperforms

1837
01:12:40,960 --> 01:12:45,840
instead of killing it they try to fix it they tweak the prompt and add more context they adjust policies

1838
01:12:45,840 --> 01:12:49,840
they invest more in something that's not delivering they refuse to admit failure they blame the

1839
01:12:49,840 --> 01:12:53,680
technology or they blame the users or they blame the model they keep the agent running because

1840
01:12:53,680 --> 01:12:58,240
they've made a public commitment to it by the time they finally shut it down they've wasted months

1841
01:12:58,240 --> 01:13:03,600
the discipline here is counter intuitive fail fast fail cheap move on a failed pilot that cost

1842
01:13:03,600 --> 01:13:08,160
$50,000 in three months is cheaper than a failed production system that's costing money monthly

1843
01:13:08,160 --> 01:13:13,200
for two years because you can't admit it's not working pilot ruthlessly measure ruthlessly kill

1844
01:13:13,200 --> 01:13:19,680
ruthlessly scale what works the agentic operating model is not coming it's here agent 365 is in general

1845
01:13:19,680 --> 01:13:24,960
availability work IQ is in general availability a 2a is production ready the question is not whether

1846
01:13:24,960 --> 01:13:29,360
you should do this it's how fast you can move organizations that build the governance foundation

1847
01:13:29,360 --> 01:13:33,920
now that establish cost discipline and that measure outcomes instead of activity will operate at

1848
01:13:33,920 --> 01:13:38,160
machine speed in 18 months they'll have embedded intelligence in their workflows they'll have a

1849
01:13:38,160 --> 01:13:42,880
competitive advantage organizations that ignore this will spend the next two years catching up by

1850
01:13:42,880 --> 01:13:47,280
then the leaders will have moved further ahead the shift is fundamental we are moving from AI as a

1851
01:13:47,280 --> 01:13:52,080
search tool to AI as an operating layer we are moving from service accounts to digital employees

1852
01:13:52,080 --> 01:13:57,600
with entraagent IDs we are moving from static context to liquid context we are moving from retrieval

1853
01:13:57,600 --> 01:14:01,840
to reasoning the model is set the tools are ready the only variable is execution