1
00:00:00,040 --> 00:00:03,319
Welcome to the M365 FM podcast,
where we explore the people,

2
00:00:03,319 --> 00:00:05,759
ideas and technologies shaping
the future of work.

3
00:00:05,759 --> 00:00:07,639
Why Copilot Studio feels
complicated.

4
00:00:07,639 --> 00:00:10,439
Before you start, you can build
a working AI agent without

5
00:00:10,439 --> 00:00:11,759
writing a single line of code.

6
00:00:11,759 --> 00:00:13,220
Copilot studio is Microsoft's

7
00:00:13,220 --> 00:00:14,640
workshop for that, a place where

8
00:00:14,640 --> 00:00:16,039
you describe the assistant you

9
00:00:16,039 --> 00:00:17,239
want and connect it to the

10
00:00:17,239 --> 00:00:18,600
documents and systems it should

11
00:00:18,600 --> 00:00:19,399
know about.

12
00:00:19,399 --> 00:00:20,920
It is not the copilot already

13
00:00:20,920 --> 00:00:21,920
sitting inside your word and

14
00:00:21,920 --> 00:00:22,559
outlook.

15
00:00:22,559 --> 00:00:24,879
And that mix up is where most of
the confusion starts.

16
00:00:24,879 --> 00:00:28,239
Welcome to the M365 FM podcast,
where we explore the people,

17
00:00:28,239 --> 00:00:30,679
ideas, and technologies shaping
the future of work.

18
00:00:30,679 --> 00:00:33,240
Today we are looking at Copilot
Studio the way a beginner

19
00:00:33,240 --> 00:00:36,119
actually meets it, and by the
end you will know what it is,

20
00:00:36,119 --> 00:00:39,359
what it costs, and the design
mistake that sinks more first

21
00:00:39,359 --> 00:00:40,920
agents than anything else.

22
00:00:40,920 --> 00:00:43,579
Most people hear the name and
picture a developer hunched over

23
00:00:43,579 --> 00:00:45,399
code, and that picture is wrong.

24
00:00:45,399 --> 00:00:46,799
It is also the reason plenty of

25
00:00:46,799 --> 00:00:48,320
good people never open the tool

26
00:00:48,320 --> 00:00:49,079
twice.

27
00:00:49,079 --> 00:00:50,039
You click a link.

28
00:00:50,039 --> 00:00:52,920
You land on a page full of terms
you have never met, and you

29
00:00:52,920 --> 00:00:55,159
decide the whole thing sits
above your pay grade.

30
00:00:55,159 --> 00:00:55,640
It doesn't.

31
00:00:55,640 --> 00:00:58,280
And that gap between how it
looks and how it works is the

32
00:00:58,280 --> 00:00:59,439
thing I want to close for you.

33
00:00:59,439 --> 00:01:00,679
So stay with me for the next

34
00:01:00,679 --> 00:01:02,439
twenty minutes or so before we

35
00:01:02,439 --> 00:01:03,479
finish, you'll understand what

36
00:01:03,479 --> 00:01:05,439
an agent is, how one gets built,

37
00:01:05,439 --> 00:01:06,840
where it fits inside Microsoft's

38
00:01:06,840 --> 00:01:08,280
world, and the honest limits

39
00:01:08,280 --> 00:01:09,319
that never make it into the

40
00:01:09,319 --> 00:01:10,040
sales pitch.

41
00:01:10,040 --> 00:01:11,760
We'll start with the plain
version.

42
00:01:11,760 --> 00:01:13,840
What Copilot Studio actually is.

43
00:01:13,840 --> 00:01:15,719
Copilot studio is a low code

44
00:01:15,719 --> 00:01:17,239
studio from Microsoft for

45
00:01:17,239 --> 00:01:18,920
building and managing AI agents

46
00:01:18,920 --> 00:01:19,799
and workflows.

47
00:01:19,799 --> 00:01:21,239
Low code means you build with

48
00:01:21,239 --> 00:01:22,519
your mouse in plain language

49
00:01:22,519 --> 00:01:23,480
rather than a programming

50
00:01:23,480 --> 00:01:25,040
language, and the studio carries

51
00:01:25,040 --> 00:01:26,000
the technical machinery

52
00:01:26,000 --> 00:01:26,760
underneath.

53
00:01:26,760 --> 00:01:28,280
You drag things into place.

54
00:01:28,280 --> 00:01:31,159
You type what you want in
ordinary sentences, and you test

55
00:01:31,159 --> 00:01:33,200
the result as you go.

56
00:01:33,200 --> 00:01:35,519
The tool isn't brand new, which
explains a lot of the older

57
00:01:35,519 --> 00:01:37,319
videos you may have seen.

58
00:01:37,319 --> 00:01:40,120
Copilot studio grew out of
something called Power Virtual

59
00:01:40,120 --> 00:01:43,359
Agents, and Microsoft folded
that old bot builder into

60
00:01:43,359 --> 00:01:46,719
Copilot Studio before adding
generative AI on top.

61
00:01:46,719 --> 00:01:49,939
If you watched a Power Virtual
Agents tutorial a few years ago,

62
00:01:49,939 --> 00:01:52,519
you were looking at the same
idea with a different engine.

63
00:01:52,519 --> 00:01:55,000
Inside the studio, you can build
three kinds of things.

64
00:01:55,000 --> 00:01:57,920
An agent is an AI assistant that
holds a conversation and

65
00:01:57,920 --> 00:01:59,760
completes tasks on your behalf.

66
00:01:59,760 --> 00:02:02,430
A workflow is an automation you
assemble with a drag and drop

67
00:02:02,430 --> 00:02:05,189
designer with human checkpoints
built in so a person can approve

68
00:02:05,189 --> 00:02:06,189
the risky steps.

69
00:02:06,189 --> 00:02:08,990
An agent flow is the older
automation format close to Power

70
00:02:08,990 --> 00:02:11,590
Automate that now runs inside
Copilot Studio as well.

71
00:02:11,590 --> 00:02:12,909
Those pieces can call each other

72
00:02:12,909 --> 00:02:14,409
so a workflow can hand a job to

73
00:02:14,409 --> 00:02:15,669
an agent and wait for the

74
00:02:15,669 --> 00:02:16,150
answer.

75
00:02:16,150 --> 00:02:18,110
A small example makes this
concrete.

76
00:02:18,110 --> 00:02:19,629
Employees keep asking how many

77
00:02:19,629 --> 00:02:21,030
vacation days they have left,

78
00:02:21,030 --> 00:02:22,310
and the same person in human

79
00:02:22,310 --> 00:02:23,830
resources answers it all day

80
00:02:23,830 --> 00:02:24,389
long.

81
00:02:24,389 --> 00:02:27,229
An agent can read the policy and
the person's record and reply on

82
00:02:27,229 --> 00:02:30,469
its own at any hour without a
human retyping a word.

83
00:02:30,469 --> 00:02:31,469
What separates this from the

84
00:02:31,469 --> 00:02:32,789
older generation of bots is the

85
00:02:32,789 --> 00:02:33,750
AI underneath.

86
00:02:33,750 --> 00:02:35,430
An old bot matched keywords and

87
00:02:35,430 --> 00:02:36,750
fell apart the moment you phrase

88
00:02:36,750 --> 00:02:37,789
things differently.

89
00:02:37,789 --> 00:02:40,310
A modern agent understands the
intent behind the request,

90
00:02:40,310 --> 00:02:43,210
reasons about it, and pulls an
answer from the knowledge you

91
00:02:43,210 --> 00:02:45,789
connected, which is why it feels
less like a phone menu and more

92
00:02:45,789 --> 00:02:47,229
like a conversation.

93
00:02:47,229 --> 00:02:48,229
That is the surface of it,

94
00:02:48,229 --> 00:02:49,270
though, and the useful question

95
00:02:49,270 --> 00:02:50,629
is what actually sits inside an

96
00:02:50,629 --> 00:02:51,229
agent?

97
00:02:51,229 --> 00:02:53,550
The answer is simpler than the
name suggests, and it's where

98
00:02:53,550 --> 00:02:55,629
the next few minutes pay off.

99
00:02:55,629 --> 00:02:56,870
The pieces that make an agent

100
00:02:56,870 --> 00:02:58,669
work, every agent, however

101
00:02:58,669 --> 00:03:00,210
clever it looks on the outside,

102
00:03:00,210 --> 00:03:01,590
is built from a small set of

103
00:03:01,590 --> 00:03:02,229
parts.

104
00:03:02,229 --> 00:03:03,349
Learn these four and copilot

105
00:03:03,349 --> 00:03:04,469
studio stops feeling like a

106
00:03:04,469 --> 00:03:05,310
foreign language.

107
00:03:05,310 --> 00:03:06,389
Start with instructions.

108
00:03:06,389 --> 00:03:09,030
This is the job description you
write for the agent in ordinary

109
00:03:09,030 --> 00:03:11,629
sentences, the same way you
would brief a new colleague on

110
00:03:11,629 --> 00:03:12,710
their first morning.

111
00:03:12,710 --> 00:03:15,590
You say who the agent helps, how
it should sound, and what it

112
00:03:15,590 --> 00:03:16,870
must never do.

113
00:03:16,870 --> 00:03:18,150
Beginners either rush this part

114
00:03:18,150 --> 00:03:19,509
or panic over it and neither is

115
00:03:19,509 --> 00:03:19,990
needed.

116
00:03:19,990 --> 00:03:21,310
You don't want a perfect essay.

117
00:03:21,310 --> 00:03:24,110
You want a clear, specific
brief, the kind you would give a

118
00:03:24,110 --> 00:03:26,550
bright colleague who has never
worked in your department.

119
00:03:26,550 --> 00:03:27,430
Next comes knowledge.

120
00:03:27,430 --> 00:03:28,389
This is where the agent is

121
00:03:28,389 --> 00:03:29,830
allowed to read from, and you

122
00:03:29,830 --> 00:03:31,270
can point it at documents, a

123
00:03:31,270 --> 00:03:32,909
website, a SharePoint site, or a

124
00:03:32,909 --> 00:03:33,909
business system.

125
00:03:33,909 --> 00:03:35,389
The agent answers from those
sources.

126
00:03:35,389 --> 00:03:37,669
So the quality of its answer
depends on the quality of what

127
00:03:37,669 --> 00:03:38,430
you feed it.

128
00:03:38,430 --> 00:03:39,909
Thin or outdated documents

129
00:03:39,909 --> 00:03:41,710
produce thin or outdated answers

130
00:03:41,710 --> 00:03:42,550
every time.

131
00:03:42,550 --> 00:03:43,669
Then there are topics.

132
00:03:43,669 --> 00:03:46,150
A topic is one scripted
conversation you design in

133
00:03:46,150 --> 00:03:49,469
advance with steps, questions,
and conditions built into it.

134
00:03:49,469 --> 00:03:51,830
You reach for topics when you
can predict the question,

135
00:03:51,830 --> 00:03:54,830
because a topic gives the same
answer every single time.

136
00:03:54,830 --> 00:03:57,990
When a request falls outside
your topics, the agent can still

137
00:03:57,990 --> 00:04:01,030
generate a reply from the
sources you connected, so it's

138
00:04:01,030 --> 00:04:02,150
never truly stuck.

139
00:04:02,150 --> 00:04:04,870
The last piece is actions, which
people also call tools.

140
00:04:04,870 --> 00:04:07,150
This is what the agent can
actually do beyond talking.

141
00:04:07,150 --> 00:04:08,629
It can create a support ticket,

142
00:04:08,629 --> 00:04:10,110
send an email, look up an order,

143
00:04:10,110 --> 00:04:11,270
or update a record in your

144
00:04:11,270 --> 00:04:12,469
company's systems.

145
00:04:12,469 --> 00:04:15,870
An agent that only talks is a
chatbot, and an agent that acts

146
00:04:15,870 --> 00:04:17,949
is the one people actually want.

147
00:04:17,949 --> 00:04:20,670
Then there's the channel, which
is simply where the agent lives.

148
00:04:20,670 --> 00:04:23,829
You can publish it into teams
Microsoft copilot, SharePoint, a

149
00:04:23,829 --> 00:04:26,310
mobile app, a company website,
or even WhatsApp.

150
00:04:26,310 --> 00:04:28,269
So people meet it where they
already are.

151
00:04:28,269 --> 00:04:30,430
Picture a new hire on their
first day at work.

152
00:04:30,430 --> 00:04:32,949
You hand them a job description,
a stack of documents, a script

153
00:04:32,949 --> 00:04:35,870
for the common questions, and
permission to use a few systems.

154
00:04:35,870 --> 00:04:38,230
Everything they do afterward
comes from those four things,

155
00:04:38,230 --> 00:04:39,709
and nothing comes from magic.

156
00:04:39,709 --> 00:04:42,449
The parts are easy to supply,
which is exactly why the way you

157
00:04:42,449 --> 00:04:44,990
wire them together is where
beginners get into trouble.

158
00:04:44,990 --> 00:04:47,149
I'll show you that spot in a few
minutes.

159
00:04:47,149 --> 00:04:49,589
Building your first agent
without writing code.

160
00:04:49,589 --> 00:04:52,069
Building one is less dramatic
than people expect.

161
00:04:52,069 --> 00:04:54,990
You sign in at Copilot Studio,
Microsoft dot com, and you

162
00:04:54,990 --> 00:04:57,389
either start from a template or
describe the agent you want.

163
00:04:57,389 --> 00:04:59,709
In plain language, copilot
studio proposes the

164
00:04:59,709 --> 00:05:02,509
instructions, the knowledge, and
the tools for you, and then you

165
00:05:02,509 --> 00:05:04,779
refine the result by answering a
few follow up questions.

166
00:05:04,779 --> 00:05:06,420
From there, you connect the
knowledge.

167
00:05:06,420 --> 00:05:09,019
You aim the agent at the
documents or the site it should

168
00:05:09,019 --> 00:05:12,139
read, and you wait while
Microsoft indexes that content,

169
00:05:12,139 --> 00:05:14,500
which can take a little while on
a large library.

170
00:05:14,500 --> 00:05:17,120
While it works, you add a topic
or two for the questions you

171
00:05:17,120 --> 00:05:18,139
know are coming.

172
00:05:18,139 --> 00:05:21,060
You write the trigger phrases a
real person would type, so a

173
00:05:21,060 --> 00:05:23,579
password reset topic might
listen for I forgot my password,

174
00:05:23,579 --> 00:05:26,220
I'm locked out or my account
won't let me in.

175
00:05:26,220 --> 00:05:29,060
Then you lay out the response
step by step before you publish.

176
00:05:29,060 --> 00:05:29,620
You test.

177
00:05:29,620 --> 00:05:31,939
Every agent has a test chat
panel where you talk to it the

178
00:05:31,939 --> 00:05:34,740
way a user would, and you can
watch which topic fired and

179
00:05:34,740 --> 00:05:36,060
which knowledge it leaned on.

180
00:05:36,060 --> 00:05:37,259
Beginners skip this part, and

181
00:05:37,259 --> 00:05:38,540
it's the part that saves you

182
00:05:38,540 --> 00:05:39,860
because the flaws show up in the

183
00:05:39,860 --> 00:05:41,019
test long before a customer

184
00:05:41,019 --> 00:05:41,660
finds them.

185
00:05:41,660 --> 00:05:43,220
One detail catches people out.

186
00:05:43,220 --> 00:05:45,660
A trial license lets you build
and test, but it won't let you

187
00:05:45,660 --> 00:05:48,379
publish, so a weekend of work
can stall at the very last.

188
00:05:48,379 --> 00:05:50,620
Click check what your license
actually covers before you

189
00:05:50,620 --> 00:05:52,500
promise your team a live agent.

190
00:05:52,500 --> 00:05:55,300
Underneath all of this sits what
Microsoft calls a harness the

191
00:05:55,300 --> 00:05:57,259
engine that decides how much
your agent can reason.

192
00:05:57,259 --> 00:05:58,420
There are a few to choose from,

193
00:05:58,420 --> 00:05:59,579
running from simple and scripted

194
00:05:59,579 --> 00:06:00,620
to heavy and reasoning driven,

195
00:06:00,620 --> 00:06:01,860
and the choice changes both what

196
00:06:01,860 --> 00:06:03,000
your agent can do and what it

197
00:06:03,000 --> 00:06:03,620
costs.

198
00:06:03,620 --> 00:06:06,459
Microsoft is also rolling out a
way to build by describing the

199
00:06:06,459 --> 00:06:08,819
business goal and letting the
studio choose the pieces for

200
00:06:08,819 --> 00:06:11,339
you, though that part is still
marked as preview.

201
00:06:11,339 --> 00:06:13,139
A concrete example pulls this
together.

202
00:06:13,139 --> 00:06:15,980
An IT team builds an agent that
resets passwords and files

203
00:06:15,980 --> 00:06:18,740
tickets for the requests that
repeat, and it quietly takes the

204
00:06:18,740 --> 00:06:20,220
easy half of the queue.

205
00:06:20,220 --> 00:06:21,779
The humans keep the hard half

206
00:06:21,779 --> 00:06:22,939
and the day gets calmer for

207
00:06:22,939 --> 00:06:23,420
everyone.

208
00:06:23,420 --> 00:06:27,500
So what people actually use it
for the scenarios people build

209
00:06:27,500 --> 00:06:28,899
are surprisingly ordinary.

210
00:06:28,899 --> 00:06:29,939
And that's the point.

211
00:06:29,939 --> 00:06:33,180
An agent does its best work on a
task that repeats a thousand

212
00:06:33,180 --> 00:06:36,060
times and rarely needs a human
to think hard.

213
00:06:36,060 --> 00:06:38,459
Internal IT support is the
classic case.

214
00:06:38,459 --> 00:06:39,519
The agent takes the ticket,

215
00:06:39,519 --> 00:06:41,019
answers the password question,

216
00:06:41,019 --> 00:06:42,300
and routes anything messy to a

217
00:06:42,300 --> 00:06:42,819
person.

218
00:06:42,819 --> 00:06:45,860
Human resources uses it the same
way for the policy questions

219
00:06:45,860 --> 00:06:48,699
that would otherwise land in one
inbox every single day.

220
00:06:48,699 --> 00:06:50,060
Customer service teams put an

221
00:06:50,060 --> 00:06:51,379
agent on the website so people

222
00:06:51,379 --> 00:06:52,420
get an answer at two in the

223
00:06:52,420 --> 00:06:53,660
morning when no one is staffing

224
00:06:53,660 --> 00:06:54,379
the phone.

225
00:06:54,379 --> 00:06:56,620
Ordinary doesn't mean small,
though, because these agents

226
00:06:56,620 --> 00:06:58,740
work around the clock and never
take a lunch break.

227
00:06:58,740 --> 00:07:00,060
Onboarding is another popular
one.

228
00:07:00,060 --> 00:07:03,459
Someone who just started can ask
about benefits, tools, and the

229
00:07:03,459 --> 00:07:06,459
first week checklist at eleven
at night and hear back without

230
00:07:06,459 --> 00:07:09,060
waiting on whichever colleague
happens to be online.

231
00:07:09,060 --> 00:07:10,220
Then there are the bigger bills.

232
00:07:10,220 --> 00:07:13,259
Twenty four Hour Fitness created
an agent called flex to support

233
00:07:13,259 --> 00:07:16,259
two million members, letting it
handle the routine questions so

234
00:07:16,259 --> 00:07:17,740
staff can take the harder ones.

235
00:07:17,740 --> 00:07:20,740
Assyrian uses agents to push
invoice processing forward with

236
00:07:20,740 --> 00:07:21,939
far less human handling.

237
00:07:21,939 --> 00:07:25,060
Microsoft lists agents for
balance sheet reconciliation,

238
00:07:25,060 --> 00:07:27,579
recruitment, screening and
contract review, which are the

239
00:07:27,579 --> 00:07:30,579
document heavy jobs that used to
eat entire afternoons.

240
00:07:30,579 --> 00:07:31,939
The pattern never really
changes.

241
00:07:31,939 --> 00:07:34,060
You find the question or the
task that repeats.

242
00:07:34,060 --> 00:07:36,740
You check that it has a clear
answer, and you hand that slice

243
00:07:36,740 --> 00:07:37,540
to an agent.

244
00:07:37,540 --> 00:07:40,379
You keep the judgment and give
away the repetition.

245
00:07:40,379 --> 00:07:42,660
You don't need a grand project
to start either.

246
00:07:42,660 --> 00:07:45,540
The strongest first agent is
usually a narrow one built for a

247
00:07:45,540 --> 00:07:48,579
single team on a single topic,
where you can see within a week

248
00:07:48,579 --> 00:07:49,980
whether it actually helped.

249
00:07:49,980 --> 00:07:51,300
Microsoft says ninety percent of

250
00:07:51,300 --> 00:07:52,459
the fortune five hundred has

251
00:07:52,459 --> 00:07:54,220
used Copilot Studio, and that

252
00:07:54,220 --> 00:07:55,420
figure comes from Microsoft

253
00:07:55,420 --> 00:07:55,740
itself.

254
00:07:55,740 --> 00:07:59,060
So treat it as marketing rather
than an audited number.

255
00:07:59,060 --> 00:08:01,620
The number that matters to you
is smaller and closer to home.

256
00:08:01,620 --> 00:08:03,019
If your week holds one task you

257
00:08:03,019 --> 00:08:05,149
do over and over, that task is a

258
00:08:05,149 --> 00:08:06,089
candidate.

259
00:08:06,089 --> 00:08:07,930
How Copilot studio fits with
Microsoft.

260
00:08:07,930 --> 00:08:09,350
Three sixty five and Azure

261
00:08:09,350 --> 00:08:10,930
Copilot Studio lives inside the

262
00:08:10,930 --> 00:08:11,769
power platform.

263
00:08:11,769 --> 00:08:13,089
The same Microsoft family as

264
00:08:13,089 --> 00:08:15,050
Power Apps, Power Automate, and

265
00:08:15,050 --> 00:08:16,009
Power BI.

266
00:08:16,009 --> 00:08:17,170
That placement matters because

267
00:08:17,170 --> 00:08:18,329
your agent can reach the same

268
00:08:18,329 --> 00:08:19,689
data and the same business

269
00:08:19,689 --> 00:08:21,129
systems those tools already

270
00:08:21,129 --> 00:08:21,689
touch.

271
00:08:21,689 --> 00:08:24,009
When you publish an agent, it
can show up inside Microsoft

272
00:08:24,009 --> 00:08:24,889
three hundred sixty five.

273
00:08:24,889 --> 00:08:26,370
It can appear in teams in

274
00:08:26,370 --> 00:08:28,050
SharePoint or inside Microsoft

275
00:08:28,050 --> 00:08:29,689
Copilot chat, so it meets people

276
00:08:29,689 --> 00:08:31,050
inside the tools they already

277
00:08:31,050 --> 00:08:31,410
use.

278
00:08:31,410 --> 00:08:32,889
The Azure side is mostly about

279
00:08:32,889 --> 00:08:34,529
billing and brains pay as you

280
00:08:34,529 --> 00:08:34,730
go.

281
00:08:34,730 --> 00:08:37,789
Usage runs through an Azure
subscription, and your agent can

282
00:08:37,789 --> 00:08:40,850
borrow models from Azure AI
foundry when the default model

283
00:08:40,850 --> 00:08:42,450
isn't the right fit for the job.

284
00:08:42,450 --> 00:08:45,210
Copilot studio also reaches
outward through connectors, and

285
00:08:45,210 --> 00:08:48,110
Microsoft lists more than a
thousand pre-built ones, plus

286
00:08:48,110 --> 00:08:50,730
MCP servers, which are a
standard way to plug an agent

287
00:08:50,730 --> 00:08:51,929
into another system.

288
00:08:51,929 --> 00:08:53,370
That is how a simple chat window

289
00:08:53,370 --> 00:08:54,870
ends up able to talk to a real

290
00:08:54,870 --> 00:08:56,090
business application.

291
00:08:56,090 --> 00:08:58,409
Now the comparison that clears
up the most confusion.

292
00:08:58,409 --> 00:08:59,049
Microsoft three.

293
00:08:59,049 --> 00:09:00,250
Sixty five copilot, which

294
00:09:00,250 --> 00:09:01,289
Microsoft now also calls

295
00:09:01,289 --> 00:09:03,210
Microsoft Copilot, is a finished

296
00:09:03,210 --> 00:09:03,649
assistant.

297
00:09:03,649 --> 00:09:05,610
You use Inside Word, Outlook,
and Teams.

298
00:09:05,610 --> 00:09:06,929
Copilot studio is the workshop

299
00:09:06,929 --> 00:09:08,090
where you build assistance of

300
00:09:08,090 --> 00:09:08,730
your own.

301
00:09:08,730 --> 00:09:10,649
One is the meal and the other is
the kitchen.

302
00:09:10,649 --> 00:09:12,049
Microsoft's own rule of thumb is

303
00:09:12,049 --> 00:09:13,450
worth remembering if your agent

304
00:09:13,450 --> 00:09:14,850
faces customers, spans

305
00:09:14,850 --> 00:09:16,409
departments, runs a complex

306
00:09:16,409 --> 00:09:17,730
process, or needs governance and

307
00:09:17,730 --> 00:09:18,970
audit trails, you build it in

308
00:09:18,970 --> 00:09:20,009
Copilot Studio.

309
00:09:20,009 --> 00:09:21,490
If it's a small internal helper

310
00:09:21,490 --> 00:09:22,850
that answers questions, the

311
00:09:22,850 --> 00:09:24,169
lighter Agent Builder inside

312
00:09:24,169 --> 00:09:25,210
Microsoft three hundred sixty

313
00:09:25,210 --> 00:09:26,490
five copilot may be enough, and

314
00:09:26,490 --> 00:09:27,610
you can move it into Copilot

315
00:09:27,610 --> 00:09:29,169
Studio later if it outgrows

316
00:09:29,169 --> 00:09:29,929
that.

317
00:09:29,929 --> 00:09:32,289
A second comparison matters just
as much.

318
00:09:32,289 --> 00:09:33,690
Power automate runs a fixed set

319
00:09:33,690 --> 00:09:34,809
of steps the same way every

320
00:09:34,809 --> 00:09:36,289
time, like clockwork, while a

321
00:09:36,289 --> 00:09:37,769
copilot studio agent decides

322
00:09:37,769 --> 00:09:38,649
what to do based on the

323
00:09:38,649 --> 00:09:40,370
conversation, most teams end up

324
00:09:40,370 --> 00:09:41,929
using both with the agent making

325
00:09:41,929 --> 00:09:43,129
the call and the flow doing the

326
00:09:43,129 --> 00:09:44,169
reliable work.

327
00:09:44,169 --> 00:09:46,210
There's also a governance layer,
and it matters more.

328
00:09:46,210 --> 00:09:49,009
As agents multiply,
administrators can inventory

329
00:09:49,009 --> 00:09:52,289
every agent, set usage limits,
and keep an experimental agent

330
00:09:52,289 --> 00:09:55,490
away from a production system or
from one admin center.

331
00:09:55,490 --> 00:09:58,490
That's the unglamorous half of
the story, and it's the half

332
00:09:58,490 --> 00:10:01,210
that keeps a clever idea from
becoming a quiet risk.

333
00:10:01,210 --> 00:10:01,570
For you.

334
00:10:01,570 --> 00:10:02,970
It means Copilot Studio is a

335
00:10:02,970 --> 00:10:04,129
working skill, not a side

336
00:10:04,129 --> 00:10:04,450
project.

337
00:10:04,450 --> 00:10:05,639
You learn for fun.

338
00:10:05,639 --> 00:10:06,759
It plugs straight into the

339
00:10:06,759 --> 00:10:07,919
systems your company already

340
00:10:07,919 --> 00:10:09,360
runs.

341
00:10:09,360 --> 00:10:11,360
The mistakes beginners keep
making.

342
00:10:11,360 --> 00:10:13,480
Here is the design mistake I
promised you.

343
00:10:13,480 --> 00:10:16,039
Teams who audit copilot studio
agents keep finding the same

344
00:10:16,039 --> 00:10:18,840
problems in roughly the same
order, and the first one is by

345
00:10:18,840 --> 00:10:20,080
far the most common.

346
00:10:20,080 --> 00:10:23,320
Topics overlap when two topics
share similar trigger phrases.

347
00:10:23,320 --> 00:10:24,519
The agent can't tell which one

348
00:10:24,519 --> 00:10:25,879
should answer, and the result is

349
00:10:25,879 --> 00:10:27,399
an agent that answers twice,

350
00:10:27,399 --> 00:10:28,860
fires the wrong topic, or keeps

351
00:10:28,860 --> 00:10:30,240
asking you to rephrase before it

352
00:10:30,240 --> 00:10:31,200
will commit.

353
00:10:31,200 --> 00:10:34,080
The fix is fewer and clearer
topics, each with a handful of

354
00:10:34,080 --> 00:10:37,000
distinct phrases rather than
twenty near-identical ones.

355
00:10:37,000 --> 00:10:39,240
Mistake two is a fallback that
goes nowhere.

356
00:10:39,240 --> 00:10:40,879
Copilot studio ships with a

357
00:10:40,879 --> 00:10:42,299
topic that asks the user to

358
00:10:42,299 --> 00:10:44,279
rephrase, and then a path called

359
00:10:44,279 --> 00:10:45,639
escalate, but escalate only

360
00:10:45,639 --> 00:10:47,320
reaches a human once you point

361
00:10:47,320 --> 00:10:48,720
it at a real queue or a

362
00:10:48,720 --> 00:10:49,840
ticketing system.

363
00:10:49,840 --> 00:10:53,080
Left untouched, it simply tells
people that nobody is available.

364
00:10:53,080 --> 00:10:54,679
The third trap is dumping in
knowledge.

365
00:10:54,679 --> 00:10:57,200
The instinct is to connect an
entire SharePoint site or a

366
00:10:57,200 --> 00:10:59,919
shared drive and hope for the
best, which produces an agent

367
00:10:59,919 --> 00:11:01,440
that sites the wrong document.

368
00:11:01,440 --> 00:11:04,519
It can also fail silently on any
file marked confidential or

369
00:11:04,519 --> 00:11:07,519
password protected, because
those files show as ready and

370
00:11:07,519 --> 00:11:08,519
then never answer.

371
00:11:08,519 --> 00:11:10,759
Then there's deciding
authentication late, whether

372
00:11:10,759 --> 00:11:12,000
people sign in and how.

373
00:11:12,000 --> 00:11:14,840
Decides which channels your
agent can even reach, and

374
00:11:14,840 --> 00:11:17,720
changing that decision after the
topics are built breaks them.

375
00:11:17,720 --> 00:11:19,080
Decide at first, then build.

376
00:11:19,080 --> 00:11:21,039
The fifth is never reading the
numbers.

377
00:11:21,039 --> 00:11:22,799
Copilot studio tracks resolution

378
00:11:22,799 --> 00:11:24,039
rate, escalation rate and

379
00:11:24,039 --> 00:11:25,700
abandon rate, and most agents

380
00:11:25,700 --> 00:11:26,980
never have those numbers opened

381
00:11:26,980 --> 00:11:28,000
after launch.

382
00:11:28,000 --> 00:11:29,240
A monthly look tells you which

383
00:11:29,240 --> 00:11:30,559
topic needs work and which one

384
00:11:30,559 --> 00:11:31,639
needs retiring.

385
00:11:31,639 --> 00:11:33,120
The last one is having no owner.

386
00:11:33,120 --> 00:11:34,559
When nobody is accountable, the

387
00:11:34,559 --> 00:11:35,879
agent runs on stale answers

388
00:11:35,879 --> 00:11:37,480
until someone complains, and by

389
00:11:37,480 --> 00:11:38,440
then people have stopped

390
00:11:38,440 --> 00:11:39,200
trusting it.

391
00:11:39,200 --> 00:11:40,720
Name one person before you ship

392
00:11:40,720 --> 00:11:42,159
it with the job of keeping it

393
00:11:42,159 --> 00:11:42,759
honest.

394
00:11:42,759 --> 00:11:44,080
None of these are code problems.

395
00:11:44,080 --> 00:11:46,720
Every single one is a design
problem, which means everyone is

396
00:11:46,720 --> 00:11:49,000
something a beginner can get
right on purpose.

397
00:11:49,000 --> 00:11:50,919
The limits you should know
before you start.

398
00:11:50,919 --> 00:11:54,240
Mistakes you can fix, but limits
you have to design around.

399
00:11:54,240 --> 00:11:57,000
And a few of them matter a great
deal for a beginner.

400
00:11:57,000 --> 00:11:58,519
As you plan your first agent,

401
00:11:58,519 --> 00:11:59,840
check which of these edges your

402
00:11:59,840 --> 00:12:01,279
idea touches because the answer

403
00:12:01,279 --> 00:12:01,720
changes.

404
00:12:01,720 --> 00:12:03,080
The whole design cost is the

405
00:12:03,080 --> 00:12:04,259
first one, and it catches people

406
00:12:04,259 --> 00:12:04,879
off guard.

407
00:12:04,879 --> 00:12:07,960
Copilot studio runs on copilot
credits and different actions

408
00:12:07,960 --> 00:12:09,200
burn different amounts.

409
00:12:09,200 --> 00:12:11,639
A simple scripted answer costs
one credit.

410
00:12:11,639 --> 00:12:14,960
A generated answer costs two,
and a heavy reasoning response

411
00:12:14,960 --> 00:12:16,480
can cost one hundred or more.

412
00:12:16,480 --> 00:12:19,399
The same agent can be cheap or
expensive, depending only on how

413
00:12:19,399 --> 00:12:20,159
you built it.

414
00:12:20,159 --> 00:12:23,360
The cheapest way to start is the
path many companies already own.

415
00:12:23,360 --> 00:12:25,559
If your organization pays for
Microsoft three hundred sixty

416
00:12:25,559 --> 00:12:28,559
five copilot, internal answers
for licensed employees often

417
00:12:28,559 --> 00:12:31,480
don't burn credits at all, which
makes an internal helper agent

418
00:12:31,480 --> 00:12:32,759
close to free to run.

419
00:12:32,759 --> 00:12:35,600
Start there and save the paid
credits for the agents that face

420
00:12:35,600 --> 00:12:36,600
the outside world.

421
00:12:36,600 --> 00:12:38,639
Watch the capacity rules as
well.

422
00:12:38,639 --> 00:12:40,960
Unused credits don't roll over
into the next month, and if you

423
00:12:40,960 --> 00:12:43,539
run past your purchased
capacity, the agent can switch

424
00:12:43,539 --> 00:12:46,240
off until someone adds more for
a customer facing agent.

425
00:12:46,240 --> 00:12:48,799
That's the difference between a
busy day and a dark one.

426
00:12:48,799 --> 00:12:50,679
Then come the hard technical
limits.

427
00:12:50,679 --> 00:12:52,720
A single model response has a
five minute cap.

428
00:12:52,720 --> 00:12:55,679
Confidential or password
protected documents get skipped

429
00:12:55,679 --> 00:12:56,840
without an error.

430
00:12:56,840 --> 00:12:58,039
Agents can't write and run their

431
00:12:58,039 --> 00:13:00,120
own code, so heavy analysis over

432
00:13:00,120 --> 00:13:01,220
a big spreadsheet may not go

433
00:13:01,220 --> 00:13:01,519
well.

434
00:13:01,519 --> 00:13:04,759
Microsoft also states plainly
that Copilot Studio isn't a

435
00:13:04,759 --> 00:13:06,674
medical device, that it
shouldn't be used for

436
00:13:06,674 --> 00:13:09,830
emergencies, and that any agent
you build is your own product

437
00:13:09,830 --> 00:13:11,309
with your name on the result.

438
00:13:11,309 --> 00:13:12,990
A final limit is easy to miss.

439
00:13:12,990 --> 00:13:15,909
Copilot studio supports a long
list of languages from English

440
00:13:15,909 --> 00:13:18,950
and Spanish to Japanese and
Hindi, so one agent can serve a

441
00:13:18,950 --> 00:13:21,590
multilingual team, but it will
only be as current as the

442
00:13:21,590 --> 00:13:23,230
content you keep feeding it.

443
00:13:23,230 --> 00:13:24,950
An agent left alone drifts.

444
00:13:24,950 --> 00:13:26,230
None of that should scare you
off.

445
00:13:26,230 --> 00:13:28,909
It should tell you where the
edges are so your first agent

446
00:13:28,909 --> 00:13:30,830
lives safely inside them.

447
00:13:30,830 --> 00:13:32,990
A short recap and one question
for you.

448
00:13:32,990 --> 00:13:34,909
So the short version goes like
this.

449
00:13:34,909 --> 00:13:38,289
Copilot Studio is Microsoft's
low code workshop for building

450
00:13:38,289 --> 00:13:40,750
AI agents that answer questions
and take action.

451
00:13:40,750 --> 00:13:42,149
It grew out of power virtual
agents.

452
00:13:42,149 --> 00:13:44,590
It sits inside the power
platform, and it publishes into

453
00:13:44,590 --> 00:13:47,269
teams, websites and the tools
your team already uses.

454
00:13:47,269 --> 00:13:50,070
It isn't the copilot inside your
word documents, and the real

455
00:13:50,070 --> 00:13:52,190
skill is design, not code.

456
00:13:52,190 --> 00:13:54,230
Now, one question for you, and I
want you to answer it in your

457
00:13:54,230 --> 00:13:55,750
head before you click away.

458
00:13:55,750 --> 00:13:59,230
What is the one task in your job
that you repeat every week and

459
00:13:59,230 --> 00:14:02,230
would happily hand to an agent
first, if this made the topic

460
00:14:02,230 --> 00:14:03,429
feel less intimidating?

461
00:14:03,429 --> 00:14:05,149
Subscribe to M365 FM.

462
00:14:05,149 --> 00:14:06,110
To master the tools and the

463
00:14:06,110 --> 00:14:07,629
mindset shaping the future of

464
00:14:07,629 --> 00:14:08,230
work.