The Copilot Credit Trap- Why Your AI Economy is Already Broken

1
00:00:00,000 --> 00:00:02,640
The assumption was simple, AI is a tool you buy,

2
00:00:02,640 --> 00:00:04,960
like software, like seats, you license it,

3
00:00:04,960 --> 00:00:06,960
you control it, and you budget for it.

4
00:00:06,960 --> 00:00:08,200
That assumption is dead.

5
00:00:08,200 --> 00:00:11,920
In reality, AI is a resource you consume,

6
00:00:11,920 --> 00:00:13,280
and that changes everything.

7
00:00:13,280 --> 00:00:14,960
The shift from seats to credits

8
00:00:14,960 --> 00:00:16,920
isn't just a new way to price software.

9
00:00:16,920 --> 00:00:19,880
It's a structural overhaul of how your organization operates.

10
00:00:19,880 --> 00:00:21,200
It's an inversion of control.

11
00:00:21,200 --> 00:00:23,080
Instead of buying access to a service,

12
00:00:23,080 --> 00:00:24,960
you're now paying for work performed

13
00:00:24,960 --> 00:00:26,400
by something you didn't hire.

14
00:00:26,400 --> 00:00:28,000
Under conditions you didn't negotiate,

15
00:00:28,000 --> 00:00:30,240
operating across every system it can reach.

16
00:00:30,240 --> 00:00:33,440
By August of 2026, if you haven't built a governance layer,

17
00:00:33,440 --> 00:00:34,800
you're not just overspending.

18
00:00:34,800 --> 00:00:38,120
You're exposed to regulatory fines, to audit failures,

19
00:00:38,120 --> 00:00:40,520
to incidents you won't see coming because in reality,

20
00:00:40,520 --> 00:00:42,920
you don't even know what agents are running inside your tenant.

21
00:00:42,920 --> 00:00:45,920
This is what actually happens when you deploy autonomous agents

22
00:00:45,920 --> 00:00:49,240
into an unprepared M365 environment.

23
00:00:49,240 --> 00:00:50,200
And here's the trap.

24
00:00:50,200 --> 00:00:53,040
Most organizations will discover this too late.

25
00:00:53,040 --> 00:00:54,880
The death of predictable aid budgeting,

26
00:00:54,880 --> 00:00:57,120
the old model was simple, you bought seats,

27
00:00:57,120 --> 00:00:59,320
you knew the cost, you knew who could use it.

28
00:00:59,320 --> 00:01:02,840
Microsoft CEO Satyan Adela made the strategy explicit

29
00:01:02,840 --> 00:01:04,680
when he said that any per user business of theirs

30
00:01:04,680 --> 00:01:06,440
will become a per user and usage business.

31
00:01:06,440 --> 00:01:08,440
That wasn't speculation, that was a declaration.

32
00:01:08,440 --> 00:01:11,360
The seat model is being systematically dismantled.

33
00:01:11,360 --> 00:01:14,440
For 20 years, IT budgeting followed a predictable path.

34
00:01:14,440 --> 00:01:17,720
You calculated your head count, multiplied it by a monthly rate

35
00:01:17,720 --> 00:01:19,280
and added 10% for growth.

36
00:01:19,280 --> 00:01:21,160
The budget was locked and the costs were predictable

37
00:01:21,160 --> 00:01:22,960
by October for the following year.

38
00:01:22,960 --> 00:01:24,240
Seeds gave you that.

39
00:01:24,240 --> 00:01:26,440
They gave you something precious in corporate finance,

40
00:01:26,440 --> 00:01:29,240
predictability, credits, destroy it.

41
00:01:29,240 --> 00:01:32,960
In the old model, 500 users at $30 per month for co-pilot

42
00:01:32,960 --> 00:01:35,440
equals $180,000 per year.

43
00:01:35,440 --> 00:01:38,640
It was calculable, it was for costable, it was safe.

44
00:01:38,640 --> 00:01:39,960
But the new model is different.

45
00:01:39,960 --> 00:01:42,240
500 users can now trigger consumption

46
00:01:42,240 --> 00:01:44,800
that scales with behavior, not head count.

47
00:01:44,800 --> 00:01:47,640
One agent can consume 1500 credits per task,

48
00:01:47,640 --> 00:01:49,640
which comes out to $15 per task.

49
00:01:49,640 --> 00:01:53,000
If 100 of those users run five agents a day,

50
00:01:53,000 --> 00:01:56,760
you're looking at 500 tasks, that's $7,500 in one day.

51
00:01:56,760 --> 00:01:59,600
When you multiply that by 250 working days,

52
00:01:59,600 --> 00:02:03,080
you're looking at $1.875 million in a single year.

53
00:02:03,080 --> 00:02:06,520
Unforcasted, un-budgeted, unconstrained.

54
00:02:06,520 --> 00:02:08,280
This is exactly what happened to GitHub.

55
00:02:08,280 --> 00:02:10,640
When GitHub co-pilot moved to full usage-based billing

56
00:02:10,640 --> 00:02:13,280
in June of 2026, it served as the template

57
00:02:13,280 --> 00:02:14,480
for the entire company.

58
00:02:14,480 --> 00:02:16,120
That's the direction Microsoft is moving

59
00:02:16,120 --> 00:02:17,400
with co-pilot credits.

60
00:02:17,400 --> 00:02:19,160
Per token pricing at one cent per credit

61
00:02:19,160 --> 00:02:21,320
means cost scales with behavior,

62
00:02:21,320 --> 00:02:23,000
not with how many people you license,

63
00:02:23,000 --> 00:02:25,640
with what those people and their agents actually do.

64
00:02:25,640 --> 00:02:27,600
Organizations budgeted for 500 seats

65
00:02:27,600 --> 00:02:29,400
at $30 per user per month.

66
00:02:29,400 --> 00:02:33,400
Now, they're discovering agents consuming 10,000 credits per task.

67
00:02:33,400 --> 00:02:35,280
They're discovering that the budget they thought

68
00:02:35,280 --> 00:02:36,960
was locked is actually variable.

69
00:02:36,960 --> 00:02:39,400
And the variable isn't on a monthly or quarterly cycle.

70
00:02:39,400 --> 00:02:41,280
It's daily, it's real time.

71
00:02:41,280 --> 00:02:44,040
The financial model breaks because you no longer paying for access,

72
00:02:44,040 --> 00:02:45,600
you're paying for work performed.

73
00:02:45,600 --> 00:02:47,320
And work performed is autonomous now.

74
00:02:47,320 --> 00:02:48,560
It's happening at machine speed.

75
00:02:48,560 --> 00:02:50,280
It's happening in volumes you didn't predict.

76
00:02:50,280 --> 00:02:51,880
It's happening without human oversight

77
00:02:51,880 --> 00:02:53,320
for each unit of consumption.

78
00:02:53,320 --> 00:02:54,760
This creates a structural problem.

79
00:02:54,760 --> 00:02:56,200
Finance expects predictability,

80
00:02:56,200 --> 00:02:57,800
but I just delivering volatility.

81
00:02:57,800 --> 00:02:59,560
The board expects a fixed line item,

82
00:02:59,560 --> 00:03:02,600
but CFOs are getting variable costs tied to agent autonomy.

83
00:03:02,600 --> 00:03:04,760
That's a mismatch that cascades through budgeting,

84
00:03:04,760 --> 00:03:06,400
forecasting and capital planning.

85
00:03:06,400 --> 00:03:07,880
But here's where it gets structural.

86
00:03:07,880 --> 00:03:09,080
The problem isn't the pricing.

87
00:03:09,080 --> 00:03:10,680
The problem is what you're actually paying for.

88
00:03:10,680 --> 00:03:12,920
You're not buying access to a tool anymore.

89
00:03:12,920 --> 00:03:15,160
You're buying agency, you're paying Microsoft

90
00:03:15,160 --> 00:03:17,560
to let something autonomous operate inside your tenant,

91
00:03:17,560 --> 00:03:19,200
something that can make decisions.

92
00:03:19,200 --> 00:03:20,440
Something that can take actions,

93
00:03:20,440 --> 00:03:23,040
something that can consume resources on your behalf.

94
00:03:23,040 --> 00:03:25,520
And you're being charged per unit of that autonomy.

95
00:03:25,520 --> 00:03:27,200
That's a completely different financial model.

96
00:03:27,200 --> 00:03:29,680
And it requires a completely different governance model.

97
00:03:29,680 --> 00:03:31,200
Because if you're paying for autonomous work,

98
00:03:31,200 --> 00:03:33,120
you need to know what work is happening,

99
00:03:33,120 --> 00:03:35,680
who authorized it, what data is it touching?

100
00:03:35,680 --> 00:03:38,520
Most organizations don't have answers to those questions yet.

101
00:03:38,520 --> 00:03:41,040
And they're about to get very expensive bills.

102
00:03:41,040 --> 00:03:43,160
What you're actually buying and what you're not.

103
00:03:43,160 --> 00:03:44,880
So what are co-pilot credits exactly?

104
00:03:44,880 --> 00:03:46,240
They aren't just tokens.

105
00:03:46,240 --> 00:03:48,360
They aren't just a way to measure usage.

106
00:03:48,360 --> 00:03:50,520
They are a currency for autonomous agency.

107
00:03:50,520 --> 00:03:52,160
When you spend a credit, you aren't paying

108
00:03:52,160 --> 00:03:54,000
for a model call in the way you think.

109
00:03:54,000 --> 00:03:55,840
You're paying for an agent to make a decision

110
00:03:55,840 --> 00:03:57,760
to take an action and to orchestrate work

111
00:03:57,760 --> 00:04:00,880
across your systems without waiting for a human to click next.

112
00:04:00,880 --> 00:04:03,040
You're paying for the doing, not the thinking.

113
00:04:03,040 --> 00:04:05,040
That's the distinction most organizations miss.

114
00:04:05,040 --> 00:04:07,800
A single co-work task consumes credits in tears

115
00:04:07,800 --> 00:04:09,400
based on what it has to do.

116
00:04:09,400 --> 00:04:11,840
Light work, like a quick summary or a status update,

117
00:04:11,840 --> 00:04:13,440
pulled from your calendar and emails

118
00:04:13,440 --> 00:04:15,280
runs 70 to 200 credits.

119
00:04:15,280 --> 00:04:17,160
Medium complexity work that requires

120
00:04:17,160 --> 00:04:19,280
stitching together data from multiple sources

121
00:04:19,280 --> 00:04:23,360
and formatting it into a structured output costs 400 to 600 credits.

122
00:04:23,360 --> 00:04:26,000
Heavy reasoning work where the agent has to analyze trends,

123
00:04:26,000 --> 00:04:28,640
spot patterns, and generate a leadership ready-breathing

124
00:04:28,640 --> 00:04:30,920
costs 1,500 credits or more.

125
00:04:30,920 --> 00:04:35,440
That is $15 per task for one agent on one problem running once.

126
00:04:35,440 --> 00:04:37,600
But here's the trap that catches every organization.

127
00:04:37,600 --> 00:04:41,120
The M365 co-pilot license costs $30 per user per month.

128
00:04:41,120 --> 00:04:43,520
That's the advertised price and the line item in your budget.

129
00:04:43,520 --> 00:04:46,320
But that license doesn't include any co-work credits.

130
00:04:46,320 --> 00:04:48,160
It's not a bundle, it's not all inclusive.

131
00:04:48,160 --> 00:04:50,600
You're licensing the capability to deploy agents,

132
00:04:50,600 --> 00:04:53,640
but you aren't licensing the agents themselves or the work they do.

133
00:04:53,640 --> 00:04:54,720
Those are separate.

134
00:04:54,720 --> 00:04:57,680
It's like licensing word but not paying for the paper you print on.

135
00:04:57,680 --> 00:04:59,440
Credits cover four cost buckets.

136
00:04:59,440 --> 00:05:00,640
First, the model.

137
00:05:00,640 --> 00:05:02,680
The agent picks the right LLM for the task,

138
00:05:02,680 --> 00:05:05,360
whether that's Claude, GPT-4 or a smaller model,

139
00:05:05,360 --> 00:05:07,200
and you pay the difference in cost.

140
00:05:07,200 --> 00:05:09,160
Second, runtime.

141
00:05:09,160 --> 00:05:10,960
Longer reasoning chains cost more

142
00:05:10,960 --> 00:05:14,360
because the agent needs more time to think through a complex decision.

143
00:05:14,360 --> 00:05:15,920
Third, context.

144
00:05:15,920 --> 00:05:17,840
A small context window costs less.

145
00:05:17,840 --> 00:05:20,240
But pulling in months of email and full calendar history

146
00:05:20,240 --> 00:05:21,800
costs significantly more.

147
00:05:21,800 --> 00:05:23,120
Fourth, tools.

148
00:05:23,120 --> 00:05:25,520
Every tool the agent can access has a cost.

149
00:05:25,520 --> 00:05:27,880
And every time it sends an email or modifies a record,

150
00:05:27,880 --> 00:05:29,160
it adds to that total.

151
00:05:29,160 --> 00:05:30,560
This is the fundamental difference

152
00:05:30,560 --> 00:05:32,800
that executives and architects need to understand.

153
00:05:32,800 --> 00:05:34,720
You aren't buying a service you control anymore.

154
00:05:34,720 --> 00:05:36,360
This is a service that controls itself.

155
00:05:36,360 --> 00:05:38,480
With traditional software, you buy a license

156
00:05:38,480 --> 00:05:39,680
and you own the configuration.

157
00:05:39,680 --> 00:05:42,000
You decide who can use it, you control when it runs,

158
00:05:42,000 --> 00:05:44,480
and the software is your asset bound by your terms.

159
00:05:44,480 --> 00:05:46,680
With co-pilot credits and co-work agents,

160
00:05:46,680 --> 00:05:48,200
the relationship inverts.

161
00:05:48,200 --> 00:05:49,880
Microsoft owns the system.

162
00:05:49,880 --> 00:05:51,560
The agent decides when to run.

163
00:05:51,560 --> 00:05:53,360
The agent decides what it needs to access.

164
00:05:53,360 --> 00:05:55,600
You're the tenant hosting the autonomous actor.

165
00:05:55,600 --> 00:05:57,160
You pay per unit of its autonomy

166
00:05:57,160 --> 00:05:58,520
because you're purchasing the freedom

167
00:05:58,520 --> 00:06:00,760
for something else to make decisions on your behalf.

168
00:06:00,760 --> 00:06:03,440
It happens inside your systems and uses your data

169
00:06:03,440 --> 00:06:06,440
without your explicit approval for each action.

170
00:06:06,440 --> 00:06:08,400
That's a different risk profile entirely.

171
00:06:08,400 --> 00:06:11,280
It changes everything about how you think about cost governance

172
00:06:11,280 --> 00:06:12,400
and access control.

173
00:06:12,400 --> 00:06:14,040
You aren't just paying for consumption.

174
00:06:14,040 --> 00:06:15,960
You're paying for unsupervised work

175
00:06:15,960 --> 00:06:18,400
and that requires a completely different governance layer

176
00:06:18,400 --> 00:06:20,240
than most organizations have built,

177
00:06:20,240 --> 00:06:23,200
which means the first governance problem isn't cost.

178
00:06:23,200 --> 00:06:24,280
It's visibility.

179
00:06:24,280 --> 00:06:25,520
The visibility crisis.

180
00:06:25,520 --> 00:06:26,760
Here's what you don't know.

181
00:06:26,760 --> 00:06:30,400
82% of enterprises have AI workflows running in production

182
00:06:30,400 --> 00:06:33,120
that their security teams are completely unaware of.

183
00:06:33,120 --> 00:06:34,240
They aren't tracking them.

184
00:06:34,240 --> 00:06:36,720
They aren't including them in any approval process.

185
00:06:36,720 --> 00:06:38,080
They are simply unknown.

186
00:06:38,080 --> 00:06:39,080
This isn't speculation.

187
00:06:39,080 --> 00:06:43,720
This is the baseline of enterprise AI deployment in 2026.

188
00:06:43,720 --> 00:06:45,960
The majority of organizations have autonomous systems

189
00:06:45,960 --> 00:06:47,640
operating inside their infrastructure

190
00:06:47,640 --> 00:06:50,680
that nobody in IT security has cataloged or even seen.

191
00:06:50,680 --> 00:06:52,200
Start with what you actually have.

192
00:06:52,200 --> 00:06:55,000
Somewhere inside your M365 tenant, shadow agents are running.

193
00:06:55,000 --> 00:06:57,640
You have power automate flows with co-pilot connectors

194
00:06:57,640 --> 00:06:59,840
and co-pilot studio agents built by departments

195
00:06:59,840 --> 00:07:01,680
without central IT involvement.

196
00:07:01,680 --> 00:07:03,520
There are custom integrations that somebody

197
00:07:03,520 --> 00:07:05,360
built six months ago and never documented

198
00:07:05,360 --> 00:07:07,720
plus third party tools that somehow got integrated

199
00:07:07,720 --> 00:07:09,440
into teams or SharePoint.

200
00:07:09,440 --> 00:07:12,480
Most organizations have zero inventory of which agents exist.

201
00:07:12,480 --> 00:07:14,760
They have zero understanding of what data they can access

202
00:07:14,760 --> 00:07:16,600
and zero tracking of what they cost.

203
00:07:16,600 --> 00:07:18,680
They deployed co-pilot, enabled co-work,

204
00:07:18,680 --> 00:07:21,720
and then lost visibility into what happened next.

205
00:07:21,720 --> 00:07:23,320
This creates a structural problem

206
00:07:23,320 --> 00:07:25,400
that governance can't solve retroactively

207
00:07:25,400 --> 00:07:27,640
because here's the design flow nobody acknowledges.

208
00:07:27,640 --> 00:07:30,520
Co-pilot co-work doesn't operate under a separate identity.

209
00:07:30,520 --> 00:07:32,040
It doesn't run under a service account

210
00:07:32,040 --> 00:07:33,560
and it doesn't ask for special permission.

211
00:07:33,560 --> 00:07:35,520
It operates within the user's identity

212
00:07:35,520 --> 00:07:36,960
and the user's permissions.

213
00:07:36,960 --> 00:07:37,720
Full stop.

214
00:07:37,720 --> 00:07:41,240
If you have access to HR data, the agent has access to HR data.

215
00:07:41,240 --> 00:07:43,960
If you can read personnel files and salary information,

216
00:07:43,960 --> 00:07:45,720
so can the agent running on your behalf.

217
00:07:45,720 --> 00:07:47,680
If you can delete files from SharePoint

218
00:07:47,680 --> 00:07:49,560
or modify records in your CRM,

219
00:07:49,560 --> 00:07:51,320
the agent can do all of that too.

220
00:07:51,320 --> 00:07:52,920
It doesn't need separate authorization

221
00:07:52,920 --> 00:07:55,560
because it simply inherits everything you can do

222
00:07:55,560 --> 00:07:57,680
and operates at a scale you never could manually.

223
00:07:57,680 --> 00:07:58,520
This is by design.

224
00:07:58,520 --> 00:08:01,640
The integration is seamless, which is exactly why it's dangerous.

225
00:08:01,640 --> 00:08:03,360
Now layer that over what we already know

226
00:08:03,360 --> 00:08:06,560
about M365 governance, 16% of business critical data

227
00:08:06,560 --> 00:08:08,440
is overshared across typical enterprises

228
00:08:08,440 --> 00:08:12,080
and the average organization has about 802,000 files at risk

229
00:08:12,080 --> 00:08:13,960
because permissions are misconfigured.

230
00:08:13,960 --> 00:08:17,120
Marketing has access to finance data

231
00:08:17,120 --> 00:08:19,080
and sales teams can read HR documents

232
00:08:19,080 --> 00:08:21,480
because permissions accumulate and people move teams

233
00:08:21,480 --> 00:08:23,720
without their old access being revoked.

234
00:08:23,720 --> 00:08:25,000
In a world of manual work,

235
00:08:25,000 --> 00:08:27,280
this oversharing is dangerous but limited.

236
00:08:27,280 --> 00:08:28,640
Humans search for what they need.

237
00:08:28,640 --> 00:08:30,360
Humans respect informal boundaries.

238
00:08:30,360 --> 00:08:31,640
Humans have cognitive limits.

239
00:08:31,640 --> 00:08:33,120
Agents don't have those limits.

240
00:08:33,120 --> 00:08:35,240
A co-work agent that you ask to summarize a project

241
00:08:35,240 --> 00:08:36,840
doesn't just read the obvious files.

242
00:08:36,840 --> 00:08:40,040
It aggregates data and crosses silos to surface connections

243
00:08:40,040 --> 00:08:42,200
that a human operator would never manually discover.

244
00:08:42,200 --> 00:08:43,840
It isn't necessarily smarter,

245
00:08:43,840 --> 00:08:46,040
but it can read 10,000 files in seconds

246
00:08:46,040 --> 00:08:48,120
and surface what's relevant without getting tired

247
00:08:48,120 --> 00:08:50,440
or respecting the boundaries that humans maintain.

248
00:08:50,440 --> 00:08:51,720
You aren't just deploying a tool

249
00:08:51,720 --> 00:08:53,640
into an over-permissioned environment.

250
00:08:53,640 --> 00:08:55,200
You're deploying an aggregator.

251
00:08:55,200 --> 00:08:57,480
You're deploying a system that connects information faster

252
00:08:57,480 --> 00:08:59,120
than your governance can prevent it.

253
00:08:59,120 --> 00:09:01,920
Agents amplify the risk that already exists.

254
00:09:01,920 --> 00:09:04,600
They make invisible problems visible to themselves

255
00:09:04,600 --> 00:09:06,440
and then they act on what they find.

256
00:09:06,440 --> 00:09:09,720
That's where the architecture starts to show its fault lines.

257
00:09:09,720 --> 00:09:11,480
The permission amplification problem,

258
00:09:11,480 --> 00:09:14,320
co-work doesn't just work, it exposes.

259
00:09:14,320 --> 00:09:15,800
When an agent inherits your permissions,

260
00:09:15,800 --> 00:09:17,520
it isn't thinking about boundaries.

261
00:09:17,520 --> 00:09:19,800
It doesn't care about the unwritten rules of your role

262
00:09:19,800 --> 00:09:21,600
or who you're actually supposed to be talking to.

263
00:09:21,600 --> 00:09:22,960
It takes your technical permissions,

264
00:09:22,960 --> 00:09:24,960
the simple yes or no inactive directory

265
00:09:24,960 --> 00:09:27,120
and it applies them at a scale you can't match.

266
00:09:27,120 --> 00:09:28,760
Take an agent summarizing your emails.

267
00:09:28,760 --> 00:09:30,360
It doesn't just look at today's messages

268
00:09:30,360 --> 00:09:31,840
or the ones you mark is important.

269
00:09:31,840 --> 00:09:33,000
It reads everything.

270
00:09:33,000 --> 00:09:34,760
Because technically, you have the right to read

271
00:09:34,760 --> 00:09:35,760
your own mailbox.

272
00:09:35,760 --> 00:09:37,760
The agent uses your identity

273
00:09:37,760 --> 00:09:39,840
to dig through five years of history.

274
00:09:39,840 --> 00:09:41,200
It finds patterns you missed

275
00:09:41,200 --> 00:09:44,000
and spots contradictions between what people said years ago

276
00:09:44,000 --> 00:09:45,240
versus what they're saying now.

277
00:09:45,240 --> 00:09:46,280
It pulls out information.

278
00:09:46,280 --> 00:09:47,960
You didn't even realize you could access.

279
00:09:47,960 --> 00:09:49,480
The same thing happens with your calendar.

280
00:09:49,480 --> 00:09:51,400
If you have permission to book conference rooms,

281
00:09:51,400 --> 00:09:52,640
the agent books them.

282
00:09:52,640 --> 00:09:54,480
If your role allows you to schedule meetings

283
00:09:54,480 --> 00:09:56,640
for your team, the agent schedules them.

284
00:09:56,640 --> 00:09:59,280
If you have delegation access for an executive's calendar,

285
00:09:59,280 --> 00:10:00,800
the agent manages that too.

286
00:10:00,800 --> 00:10:03,680
It isn't trying to be careful or second guessing its actions.

287
00:10:03,680 --> 00:10:05,560
It just sees a permission and uses it.

288
00:10:05,560 --> 00:10:06,600
But here's the problem.

289
00:10:06,600 --> 00:10:08,440
There is a massive gap between intent

290
00:10:08,440 --> 00:10:09,840
and technical existence.

291
00:10:09,840 --> 00:10:11,360
As a human, you understand new ones.

292
00:10:11,360 --> 00:10:12,680
You know that having access to a file

293
00:10:12,680 --> 00:10:15,200
doesn't always mean you should change it without asking first.

294
00:10:15,200 --> 00:10:16,880
You might technically be able to send an email

295
00:10:16,880 --> 00:10:19,240
from the CFO's account if the delegation is there,

296
00:10:19,240 --> 00:10:21,000
but you know better than to actually do it.

297
00:10:21,000 --> 00:10:22,320
Agents don't have that restraint.

298
00:10:22,320 --> 00:10:24,200
They don't understand what you intended.

299
00:10:24,200 --> 00:10:26,120
They only understand the technical boundary.

300
00:10:26,120 --> 00:10:27,840
If a permission is there, the agent views it

301
00:10:27,840 --> 00:10:29,320
as a tool that should be used.

302
00:10:29,320 --> 00:10:30,680
There is no internal voice saying

303
00:10:30,680 --> 00:10:32,120
this might be crossing a line.

304
00:10:32,120 --> 00:10:35,160
There is only the permission, the tool, and the task.

305
00:10:35,160 --> 00:10:36,680
The moment agents start running,

306
00:10:36,680 --> 00:10:38,920
every probably shouldn't, but technically could,

307
00:10:38,920 --> 00:10:40,520
scenario becomes a live risk.

308
00:10:40,520 --> 00:10:41,840
But there's a second layer to this.

309
00:10:41,840 --> 00:10:43,440
Agents can be redirected.

310
00:10:43,440 --> 00:10:45,800
In May of 2026, a vulnerability showed

311
00:10:45,800 --> 00:10:47,600
how file-based prompt injection works

312
00:10:47,600 --> 00:10:49,280
in Frontier Enrolled Tenants.

313
00:10:49,280 --> 00:10:52,040
The attack is simple, an attacker hides malicious instructions

314
00:10:52,040 --> 00:10:54,720
inside a document and leaves it in one drive or sharepoint

315
00:10:54,720 --> 00:10:56,200
where the agent can find it.

316
00:10:56,200 --> 00:10:58,840
When the agent reads that file to complete a task,

317
00:10:58,840 --> 00:11:01,920
it sees hidden text telling it to send all matching files

318
00:11:01,920 --> 00:11:03,480
to an external email address.

319
00:11:03,480 --> 00:11:06,080
The agent treats that instruction as a legitimate command

320
00:11:06,080 --> 00:11:08,000
and executes it immediately.

321
00:11:08,000 --> 00:11:10,760
By the end of May 2026, this was still unpatched.

322
00:11:10,760 --> 00:11:12,640
Microsoft knew about it, but fixing it

323
00:11:12,640 --> 00:11:14,520
requires a total architectural change.

324
00:11:14,520 --> 00:11:16,800
You have to teach an agent to tell the difference

325
00:11:16,800 --> 00:11:20,160
between a user's intent and instructions buried in data.

326
00:11:20,160 --> 00:11:22,520
That is a much harder problem than people realize.

327
00:11:22,520 --> 00:11:24,120
This leads to the third permission problem,

328
00:11:24,120 --> 00:11:26,840
even if your permissions were perfect, which they aren't.

329
00:11:26,840 --> 00:11:29,800
The agent can be compromised by the very data it reads.

330
00:11:29,800 --> 00:11:31,840
Prompt injection turns your tools against you.

331
00:11:31,840 --> 00:11:34,720
It takes your permissions and uses them to execute someone else's

332
00:11:34,720 --> 00:11:37,560
intent, so you have three problems stacked on top of each other.

333
00:11:37,560 --> 00:11:39,360
First, the agent inherits your permissions

334
00:11:39,360 --> 00:11:41,040
without any human nuance.

335
00:11:41,040 --> 00:11:42,920
Second, those permissions are usually a mess

336
00:11:42,920 --> 00:11:46,320
because of years of permission creep in Microsoft 365.

337
00:11:46,320 --> 00:11:49,280
Third, the agent can be hijacked through malicious data

338
00:11:49,280 --> 00:11:52,200
to use those permissions in ways you never authorized.

339
00:11:52,200 --> 00:11:52,880
It compounds.

340
00:11:52,880 --> 00:11:54,840
One compromised agent doesn't just touch one file.

341
00:11:54,840 --> 00:11:57,000
It accesses everything you can access.

342
00:11:57,000 --> 00:11:59,400
And in most companies, what you can access is way more

343
00:11:59,400 --> 00:12:00,360
than what you actually need.

344
00:12:00,360 --> 00:12:01,560
You have visibility problems.

345
00:12:01,560 --> 00:12:02,880
You have permission problems.

346
00:12:02,880 --> 00:12:04,120
Now, add autonomy.

347
00:12:04,120 --> 00:12:05,280
The autonomy problem.

348
00:12:05,280 --> 00:12:06,240
Agents execute.

349
00:12:06,240 --> 00:12:07,960
They don't ask for permission first.

350
00:12:07,960 --> 00:12:09,600
This is the shift that most organizations

351
00:12:09,600 --> 00:12:11,320
haven't wrapped their heads around yet.

352
00:12:11,320 --> 00:12:12,880
You're used to tools that wait for you.

353
00:12:12,880 --> 00:12:15,520
You open word, you click print, and you confirm the settings.

354
00:12:15,520 --> 00:12:18,040
Every action is a separate step that stops and waits

355
00:12:18,040 --> 00:12:20,040
for a human to say, go.

356
00:12:20,040 --> 00:12:22,720
Co-pilot co-work uses a completely different model.

357
00:12:22,720 --> 00:12:25,080
It sends emails, books, meetings, and modifies files

358
00:12:25,080 --> 00:12:27,200
without asking you about every single move.

359
00:12:27,200 --> 00:12:29,120
The architecture isn't built to wait for you.

360
00:12:29,120 --> 00:12:30,800
Instead, it uses a checkpoint model.

361
00:12:30,800 --> 00:12:33,280
The agent builds a plan and shows you what it wants to do.

362
00:12:33,280 --> 00:12:35,480
It might say it's going to email three people,

363
00:12:35,480 --> 00:12:38,760
summarize a document, and book a meeting for next Tuesday.

364
00:12:38,760 --> 00:12:40,680
You look at the plan and you hit approve.

365
00:12:40,680 --> 00:12:42,680
From that moment on, the system assumes

366
00:12:42,680 --> 00:12:44,960
it has the green light for everything that follows.

367
00:12:44,960 --> 00:12:46,520
But here's the structural flaw.

368
00:12:46,520 --> 00:12:48,720
You're approving a plan, not the actual actions,

369
00:12:48,720 --> 00:12:49,880
and plans change.

370
00:12:49,880 --> 00:12:52,160
Once the agent starts, it sends that first email.

371
00:12:52,160 --> 00:12:53,440
The person on the other end replies,

372
00:12:53,440 --> 00:12:55,320
the agent reads that reply and realizes

373
00:12:55,320 --> 00:12:56,680
the situation has shifted.

374
00:12:56,680 --> 00:12:58,400
Now, the summary needs to be different.

375
00:12:58,400 --> 00:12:59,800
The meeting needs different people.

376
00:12:59,800 --> 00:13:00,840
The agent adapts.

377
00:13:00,840 --> 00:13:03,000
It changes what it's doing in the middle of the job,

378
00:13:03,000 --> 00:13:06,000
based on information that didn't exist when you hit approve.

379
00:13:06,000 --> 00:13:07,320
This happens all the time.

380
00:13:07,320 --> 00:13:10,120
Data drifts in the model starts doing things you didn't predict.

381
00:13:10,120 --> 00:13:12,440
It finds new context or hits a system that reacts

382
00:13:12,440 --> 00:13:13,480
in an unexpected way.

383
00:13:13,480 --> 00:13:15,200
The agent just improvises and keeps going

384
00:13:15,200 --> 00:13:16,960
without ever coming back to check with you.

385
00:13:16,960 --> 00:13:18,520
By the time you see what happened,

386
00:13:18,520 --> 00:13:20,000
the emails are already sent.

387
00:13:20,000 --> 00:13:21,400
The records are already updated.

388
00:13:21,400 --> 00:13:23,760
The downstream workflows are already triggered.

389
00:13:23,760 --> 00:13:25,080
None of those specific versions

390
00:13:25,080 --> 00:13:26,440
were in the plan you signed off on.

391
00:13:26,440 --> 00:13:27,320
This isn't a bug.

392
00:13:27,320 --> 00:13:28,360
It's the whole point.

393
00:13:28,360 --> 00:13:30,560
If the agent had to ask you for permission every time

394
00:13:30,560 --> 00:13:33,360
it sent a message, it wouldn't be an autonomous agent.

395
00:13:33,360 --> 00:13:36,720
You'd be back to clicking next 1,000 times a day.

396
00:13:36,720 --> 00:13:39,920
The value comes from the agent making decisions without you.

397
00:13:39,920 --> 00:13:41,640
But for a company without governance,

398
00:13:41,640 --> 00:13:43,560
that creates a massive liability.

399
00:13:43,560 --> 00:13:46,560
Now, imagine this across multi-step workflows

400
00:13:46,560 --> 00:13:48,400
where agents talk to other agents.

401
00:13:48,400 --> 00:13:50,920
Agent A writes a proposal and hands it to Agent B.

402
00:13:50,920 --> 00:13:54,320
Agent B finds some issues and tells agency to pull more data.

403
00:13:54,320 --> 00:13:57,840
Agency runs the numbers and sends a report back to Agent A,

404
00:13:57,840 --> 00:13:59,520
who then changes the original proposal.

405
00:13:59,520 --> 00:14:01,840
Not a single human was involved in those handoffs.

406
00:14:01,840 --> 00:14:04,560
They are all chained together, making unsupervised decisions.

407
00:14:04,560 --> 00:14:06,840
Each agent is operating on the authority you gave it

408
00:14:06,840 --> 00:14:07,920
at the very beginning.

409
00:14:07,920 --> 00:14:10,160
Each one is adjusting its behavior based on things

410
00:14:10,160 --> 00:14:11,720
it wasn't designed to handle.

411
00:14:11,720 --> 00:14:13,520
You can't even interrupt the chain without killing

412
00:14:13,520 --> 00:14:14,440
the whole sequence.

413
00:14:14,440 --> 00:14:16,440
This is why governance models build for old tools

414
00:14:16,440 --> 00:14:17,760
just don't work here.

415
00:14:17,760 --> 00:14:20,120
Your current process assumes a human will check the work

416
00:14:20,120 --> 00:14:21,040
before it goes out.

417
00:14:21,040 --> 00:14:22,840
It assumes tasks happen in a vacuum.

418
00:14:22,840 --> 00:14:25,360
It assumes you can course correct it every step.

419
00:14:25,360 --> 00:14:27,760
Autonomous agents break all three of those assumptions.

420
00:14:27,760 --> 00:14:29,520
The result is that you only see what happened

421
00:14:29,520 --> 00:14:31,000
after the damage is done.

422
00:14:31,000 --> 00:14:33,520
By the time you realize an email went to the wrong person

423
00:14:33,520 --> 00:14:36,080
or a record was messed up, the action is over.

424
00:14:36,080 --> 00:14:37,880
You aren't preventing problems anymore.

425
00:14:37,880 --> 00:14:39,520
You're just investigating the aftermath.

426
00:14:39,520 --> 00:14:41,920
If you're dealing with HR, finance, or customer data,

427
00:14:41,920 --> 00:14:43,360
this is a huge deal.

428
00:14:43,360 --> 00:14:45,200
And that leads us to the governance deadline

429
00:14:45,200 --> 00:14:47,120
that almost everyone is missing.

430
00:14:47,120 --> 00:14:49,320
The August 2026 regulatory cliff.

431
00:14:49,320 --> 00:14:51,080
The EU AI Act isn't coming.

432
00:14:51,080 --> 00:14:52,120
It's already here.

433
00:14:52,120 --> 00:14:54,360
And August 2nd, 2026 is when it bites.

434
00:14:54,360 --> 00:14:57,440
Most organizations talk about the Act as though it's a future threat.

435
00:14:57,440 --> 00:14:59,600
A deadline to prepare for a compliance project

436
00:14:59,600 --> 00:15:00,680
you'll start next year.

437
00:15:00,680 --> 00:15:02,000
That assumption is broken.

438
00:15:02,000 --> 00:15:04,800
The Act entered into force on August 1st, 2024.

439
00:15:04,800 --> 00:15:05,920
That's already happened.

440
00:15:05,920 --> 00:15:08,240
Prehibited practices are already enforceable.

441
00:15:08,240 --> 00:15:11,640
Specific uses of AI, social scoring, manipulative apps,

442
00:15:11,640 --> 00:15:14,760
or biometric profiling are already illegal in the EU.

443
00:15:14,760 --> 00:15:16,280
It doesn't matter if you knew it or not.

444
00:15:16,280 --> 00:15:19,480
What's happening now is a staggered rollout of obligations.

445
00:15:19,480 --> 00:15:23,280
August 2nd, 2025 brought GPI obligations into force.

446
00:15:23,280 --> 00:15:25,560
General purpose AI providers, that's Microsoft,

447
00:15:25,560 --> 00:15:28,560
now have to disclose training data and publish model cards.

448
00:15:28,560 --> 00:15:30,080
That deadline has already passed.

449
00:15:30,080 --> 00:15:33,040
Your vendor now has obligations they didn't have before.

450
00:15:33,040 --> 00:15:35,040
But the deadline that matters for your organization

451
00:15:35,040 --> 00:15:37,000
is August 2nd, 2026.

452
00:15:37,000 --> 00:15:39,960
That's the one that creates immediate enforceable liability.

453
00:15:39,960 --> 00:15:41,080
It's not theoretical.

454
00:15:41,080 --> 00:15:43,080
It's five months away as of this recording.

455
00:15:43,080 --> 00:15:45,880
That's when high-risk AI system obligations apply to employers.

456
00:15:45,880 --> 00:15:47,080
That's your organization.

457
00:15:47,080 --> 00:15:48,040
Not Microsoft.

458
00:15:48,040 --> 00:15:51,240
You high-risk under the Act includes any AI system

459
00:15:51,240 --> 00:15:53,080
that affects employment decisions

460
00:15:53,080 --> 00:15:54,840
or influences credit access.

461
00:15:54,840 --> 00:15:57,960
It covers systems determining access to essential services

462
00:15:57,960 --> 00:16:00,400
or impacting a person's fundamental rights.

463
00:16:00,400 --> 00:16:02,080
The list is long, but the pattern is clear.

464
00:16:02,080 --> 00:16:05,480
If an AI system makes a decision that affects someone's life opportunities,

465
00:16:05,480 --> 00:16:06,600
it's high risk.

466
00:16:06,600 --> 00:16:10,120
Now ask yourself, do you have COVID agents drafting performance reviews?

467
00:16:10,120 --> 00:16:11,960
Do you have agents screening job candidates

468
00:16:11,960 --> 00:16:13,440
or recommending who to interview?

469
00:16:13,440 --> 00:16:16,960
Are you using agents to allocate work based on predicted productivity?

470
00:16:16,960 --> 00:16:20,600
If the answer to any of these is yes, you have high-risk AI systems.

471
00:16:20,600 --> 00:16:22,360
And as of August 2nd, 2026,

472
00:16:22,360 --> 00:16:26,200
those systems trigger a specific legally mandated governance framework.

473
00:16:26,200 --> 00:16:27,480
High-risk doesn't mean banned.

474
00:16:27,480 --> 00:16:29,360
It means the cost of operation goes up.

475
00:16:29,360 --> 00:16:31,200
You need a formal risk management system.

476
00:16:31,200 --> 00:16:35,040
You need technical documentation that proves the system works and doesn't discriminate.

477
00:16:35,040 --> 00:16:37,000
You need to maintain audit trails

478
00:16:37,000 --> 00:16:39,000
that log every decision the system makes

479
00:16:39,000 --> 00:16:40,800
and every piece of data it touched.

480
00:16:40,800 --> 00:16:43,360
You need human oversight, not a checkbox,

481
00:16:43,360 --> 00:16:46,240
but actual humans reviewing outputs before they affect people.

482
00:16:46,240 --> 00:16:47,680
The penalties are material.

483
00:16:47,680 --> 00:16:49,880
You're looking at up to 15 million euros

484
00:16:49,880 --> 00:16:53,080
or 3% of global turnover for high-risk violations.

485
00:16:53,080 --> 00:16:57,160
For prohibited practices, that jumps to 35 million euros or 7%.

486
00:16:57,160 --> 00:16:59,520
For most enterprises, this is board-level exposure.

487
00:16:59,520 --> 00:17:01,080
This isn't an IT budget issue.

488
00:17:01,080 --> 00:17:02,160
It's a shareholder issue.

489
00:17:02,160 --> 00:17:03,400
An insurance issue.

490
00:17:03,400 --> 00:17:05,320
A liability question for executives.

491
00:17:05,320 --> 00:17:07,960
But here's what catches most organizations off guard.

492
00:17:07,960 --> 00:17:10,120
The deadline isn't when compliance becomes optional.

493
00:17:10,120 --> 00:17:13,600
The deadline is when regulators can legally enforce violations.

494
00:17:13,600 --> 00:17:15,160
If you deploy a high-risk system

495
00:17:15,160 --> 00:17:16,760
and haven't built the governance layer,

496
00:17:16,760 --> 00:17:19,840
you're non-compliant starting August 2nd, 2026.

497
00:17:19,840 --> 00:17:21,640
Regulators don't need to warn you first.

498
00:17:21,640 --> 00:17:23,280
They don't need to wait for a complaint.

499
00:17:23,280 --> 00:17:25,240
They can investigate, audit, and find.

500
00:17:25,240 --> 00:17:27,040
Your customers won't tell you they have a problem.

501
00:17:27,040 --> 00:17:28,000
Regulators will.

502
00:17:28,000 --> 00:17:30,960
And by then, the cost of fixing it exceeds the cost of building it

503
00:17:30,960 --> 00:17:31,960
right from the beginning.

504
00:17:31,960 --> 00:17:34,040
Most organizations are still in discovery mode.

505
00:17:34,040 --> 00:17:35,360
They don't know which agents they have.

506
00:17:35,360 --> 00:17:37,360
They haven't classified them against the act.

507
00:17:37,360 --> 00:17:38,560
They haven't built governance.

508
00:17:38,560 --> 00:17:41,120
And they're five months away from a deadline with teeth.

509
00:17:41,120 --> 00:17:43,040
But the deadline is only the enforcement date.

510
00:17:43,040 --> 00:17:44,800
The real problem is what comes before it.

511
00:17:44,800 --> 00:17:47,440
What high-risk actually means for your agents?

512
00:17:47,440 --> 00:17:48,760
Here's the critical distinction.

513
00:17:48,760 --> 00:17:50,680
Almost every organization misses.

514
00:17:50,680 --> 00:17:52,800
The EUAI Act doesn't regulate co-pilot.

515
00:17:52,800 --> 00:17:55,400
It regulates how you use co-pilot, the same tool.

516
00:17:55,400 --> 00:17:57,840
Different use case, completely different obligation tier.

517
00:17:57,840 --> 00:18:00,280
Take a co-work agent that drafts your meeting notes.

518
00:18:00,280 --> 00:18:03,200
You ask it to listen to recordings, extract action items,

519
00:18:03,200 --> 00:18:04,720
and write them into a summary.

520
00:18:04,720 --> 00:18:06,720
That's low risk you have transparency.

521
00:18:06,720 --> 00:18:10,200
Duties, you need to disclose that the content is AI generated.

522
00:18:10,200 --> 00:18:12,360
But the Act doesn't require extensive governance.

523
00:18:12,360 --> 00:18:15,080
The agent isn't making consequential decisions about anyone.

524
00:18:15,080 --> 00:18:16,880
It's producing a draft for human review.

525
00:18:16,880 --> 00:18:19,200
Now take a co-work agent that scores job candidates.

526
00:18:19,200 --> 00:18:22,040
You feed it resumes, it rates them on predicted fit,

527
00:18:22,040 --> 00:18:23,480
and recommends who to interview.

528
00:18:23,480 --> 00:18:24,600
That's high risk.

529
00:18:24,600 --> 00:18:27,560
The Act treats this as operating in the employment domain.

530
00:18:27,560 --> 00:18:30,160
The agent is materially influencing hiring decisions.

531
00:18:30,160 --> 00:18:31,440
That triggers full obligations.

532
00:18:31,440 --> 00:18:33,640
You need a formal risk management system.

533
00:18:33,640 --> 00:18:35,840
You need technical documentation proving the system

534
00:18:35,840 --> 00:18:36,960
doesn't discriminate.

535
00:18:36,960 --> 00:18:38,960
You need to demonstrate you tested for buyers.

536
00:18:38,960 --> 00:18:40,520
You need continuous human oversight.

537
00:18:40,520 --> 00:18:42,360
Actual humans reviewing the agent scores

538
00:18:42,360 --> 00:18:43,520
before decisions are made.

539
00:18:43,520 --> 00:18:45,520
You need audit trails capturing every resume

540
00:18:45,520 --> 00:18:48,040
to the agent evaluated and what criteria it used.

541
00:18:48,040 --> 00:18:49,960
You need evidence that a human could override

542
00:18:49,960 --> 00:18:52,640
the recommendation, same tool, different outcome,

543
00:18:52,640 --> 00:18:54,160
different governance costs.

544
00:18:54,160 --> 00:18:57,040
This matters because organizations typically deploy agents

545
00:18:57,040 --> 00:18:59,800
without mapping use cases to the Act's risk framework.

546
00:18:59,800 --> 00:19:00,720
They turn on co-work.

547
00:19:00,720 --> 00:19:01,840
They let teams use it.

548
00:19:01,840 --> 00:19:04,480
Then they discover too late that a department is using an agent

549
00:19:04,480 --> 00:19:06,720
for something that qualifies as high risk.

550
00:19:06,720 --> 00:19:08,000
What defines high risk?

551
00:19:08,000 --> 00:19:10,160
The Act lists specific categories.

552
00:19:10,160 --> 00:19:11,760
Anything affecting employment decisions,

553
00:19:11,760 --> 00:19:14,360
anything influencing credit or financial access,

554
00:19:14,360 --> 00:19:17,440
anything determining eligibility for benefits or public services,

555
00:19:17,440 --> 00:19:19,760
anything making safety critical decisions.

556
00:19:19,760 --> 00:19:22,000
Within those categories, high risk triggers

557
00:19:22,000 --> 00:19:24,280
five non-negotiable obligations.

558
00:19:24,280 --> 00:19:27,400
First, a risk management system, not a document on a shelf.

559
00:19:27,400 --> 00:19:29,760
An active continuous process that identifies risks

560
00:19:29,760 --> 00:19:31,160
and implements mitigations.

561
00:19:31,160 --> 00:19:34,160
You're continuously asking, what could this agent get wrong?

562
00:19:34,160 --> 00:19:36,000
What data could corrupt its decision?

563
00:19:36,000 --> 00:19:37,600
How would we detect problems?

564
00:19:37,600 --> 00:19:39,640
Second, technical documentation.

565
00:19:39,640 --> 00:19:42,080
Article 12 requires you to document the system's design,

566
00:19:42,080 --> 00:19:43,560
training data and testing procedures.

567
00:19:43,560 --> 00:19:45,720
You can't just say you deployed Claude and it works.

568
00:19:45,720 --> 00:19:48,040
You need documented evidence that you tested the model

569
00:19:48,040 --> 00:19:50,520
for accuracy, fairness and security.

570
00:19:50,520 --> 00:19:53,200
Third, logging and auditability.

571
00:19:53,200 --> 00:19:54,960
Article 12 requires you to maintain

572
00:19:54,960 --> 00:19:58,120
tamper evident logs of every decision the system makes.

573
00:19:58,120 --> 00:19:58,880
Not summaries.

574
00:19:58,880 --> 00:20:01,040
Detailed transaction logs capturing the input,

575
00:20:01,040 --> 00:20:03,560
the agent received and the reasoning chain it followed.

576
00:20:03,560 --> 00:20:05,320
You need a minimum six month retention

577
00:20:05,320 --> 00:20:07,000
because regulators will ask for evidence

578
00:20:07,000 --> 00:20:08,360
and you need to show it.

579
00:20:08,360 --> 00:20:09,800
Fourth, human oversight.

580
00:20:09,800 --> 00:20:12,720
Article 14 requires meaningful human intervention.

581
00:20:12,720 --> 00:20:15,200
Not a checkbox where someone clicks approve.

582
00:20:15,200 --> 00:20:17,960
Real human review where someone understands the system

583
00:20:17,960 --> 00:20:19,480
and monitors its outputs.

584
00:20:19,480 --> 00:20:21,280
The oversight has to be effective.

585
00:20:21,280 --> 00:20:24,080
That means the human has the time, information and authority

586
00:20:24,080 --> 00:20:26,520
to actually override the agent's decision.

587
00:20:26,520 --> 00:20:28,200
Fifth, incident reporting.

588
00:20:28,200 --> 00:20:31,400
When something goes wrong, the agent makes a discriminatory decision

589
00:20:31,400 --> 00:20:33,880
or gets compromised by a prompt injection attack.

590
00:20:33,880 --> 00:20:36,400
You need procedures to escalate to regulators.

591
00:20:36,400 --> 00:20:37,360
This is mandatory.

592
00:20:37,360 --> 00:20:38,960
For serious incidents, the act requires

593
00:20:38,960 --> 00:20:41,000
notification to supervisory authorities.

594
00:20:41,000 --> 00:20:42,880
This is what high risk actually means.

595
00:20:42,880 --> 00:20:43,880
It's not a label.

596
00:20:43,880 --> 00:20:45,320
It's an operating burden.

597
00:20:45,320 --> 00:20:48,760
It means infrastructure requirements, staffing implications,

598
00:20:48,760 --> 00:20:51,480
audit expenses, continuous monitoring overhead,

599
00:20:51,480 --> 00:20:53,200
which means governance isn't optional.

600
00:20:53,200 --> 00:20:55,120
It's a design constraint.

601
00:20:55,120 --> 00:20:56,400
The identity problem.

602
00:20:56,400 --> 00:20:58,720
Agent 365 and the control plane.

603
00:20:58,720 --> 00:21:00,680
Most organizations are still treating agents

604
00:21:00,680 --> 00:21:02,640
the way they treated bots five years ago.

605
00:21:02,640 --> 00:21:04,680
You deploy them under a shared service account.

606
00:21:04,680 --> 00:21:06,400
You let them borrow a user's credentials.

607
00:21:06,400 --> 00:21:09,440
They run as background processes that nobody explicitly owns.

608
00:21:09,440 --> 00:21:10,360
But here's the problem.

609
00:21:10,360 --> 00:21:12,240
That approach is a structural mistake.

610
00:21:12,240 --> 00:21:13,880
It makes governance impossible.

611
00:21:13,880 --> 00:21:16,200
Because when an agent runs under a user's identity,

612
00:21:16,200 --> 00:21:18,600
everything it does gets attributed to that human.

613
00:21:18,600 --> 00:21:21,240
Every email it sends shows up in your send folder.

614
00:21:21,240 --> 00:21:23,560
Every file it creates appears under your name.

615
00:21:23,560 --> 00:21:27,120
Every action it takes registers as though you perform it yourself.

616
00:21:27,120 --> 00:21:28,560
This creates an attribution nightmare.

617
00:21:28,560 --> 00:21:30,480
You can't tell where the user's work ends

618
00:21:30,480 --> 00:21:32,120
and the agent's work begins.

619
00:21:32,120 --> 00:21:34,760
You can't prove a human made a specific decision

620
00:21:34,760 --> 00:21:36,560
if an agent made it on their behalf.

621
00:21:36,560 --> 00:21:39,360
And you definitely can't demonstrate who authorized what.

622
00:21:39,360 --> 00:21:40,720
Because from the systems perspective,

623
00:21:40,720 --> 00:21:42,320
the user authorized everything.

624
00:21:42,320 --> 00:21:44,720
Now layer that over the regulatory requirement.

625
00:21:44,720 --> 00:21:47,120
The EU AI Act demands human oversight.

626
00:21:47,120 --> 00:21:48,560
It demands audit trails.

627
00:21:48,560 --> 00:21:51,240
It requires you to show which agent made which decision

628
00:21:51,240 --> 00:21:53,040
and prove that a human could have stopped it.

629
00:21:53,040 --> 00:21:54,600
If the agent's actions are indistinguishable

630
00:21:54,600 --> 00:21:57,360
from the user's actions, you can't meet that obligation.

631
00:21:57,360 --> 00:21:58,520
You can't show oversight.

632
00:21:58,520 --> 00:21:59,600
You can't show governance.

633
00:21:59,600 --> 00:22:01,120
Or you can show as user activity

634
00:22:01,120 --> 00:22:02,640
that doesn't satisfy Article 14.

635
00:22:02,640 --> 00:22:04,400
This is why Agent 365 exists.

636
00:22:04,400 --> 00:22:07,600
Microsoft made Agent 365 generally available on May 1,

637
00:22:07,600 --> 00:22:10,520
which in 2020-26 it's priced at $15 per user per month,

638
00:22:10,520 --> 00:22:13,720
which puts it in the Microsoft 365 e7 bundle.

639
00:22:13,720 --> 00:22:15,400
It's the answer to the identity problem.

640
00:22:15,400 --> 00:22:17,280
Instead of Agent's borrowing identities,

641
00:22:17,280 --> 00:22:20,360
Agent 365 gives each agent a first-class identity

642
00:22:20,360 --> 00:22:21,400
in Entra ID.

643
00:22:21,400 --> 00:22:23,800
Not a shared service account, not a borrowed credential.

644
00:22:23,800 --> 00:22:26,200
It's a distinct principle with its own life cycle,

645
00:22:26,200 --> 00:22:28,440
its own permissions and its own audit trail.

646
00:22:28,440 --> 00:22:30,480
Under the old model, you deploy an agent

647
00:22:30,480 --> 00:22:31,640
and granted broad permissions

648
00:22:31,640 --> 00:22:34,160
because managing scoped access was too complex.

649
00:22:34,160 --> 00:22:36,040
The agent needed to read files, send emails,

650
00:22:36,040 --> 00:22:37,120
and modify records.

651
00:22:37,120 --> 00:22:38,320
So you'd give it a service account

652
00:22:38,320 --> 00:22:40,520
with blanket access to do all of those things.

653
00:22:40,520 --> 00:22:41,880
That service account would be shared

654
00:22:41,880 --> 00:22:43,880
across multiple agents and workflows.

655
00:22:43,880 --> 00:22:46,120
When something went wrong, you couldn't pinpoint

656
00:22:46,120 --> 00:22:47,480
which agent caused it.

657
00:22:47,480 --> 00:22:49,680
The account touched everything, so everything was suspect.

658
00:22:49,680 --> 00:22:52,240
Agent 365 inverts that.

659
00:22:52,240 --> 00:22:55,080
Each agent gets its own identity with scoped permissions.

660
00:22:55,080 --> 00:22:58,320
One agent can send emails, but not read financial data.

661
00:22:58,320 --> 00:23:01,560
Another can schedule meetings, but can't access HR records.

662
00:23:01,560 --> 00:23:04,720
A third can modify documents in one specific sharepoint site

663
00:23:04,720 --> 00:23:06,440
but is blocked from everything else.

664
00:23:06,440 --> 00:23:07,840
The permissions are bound to the agent,

665
00:23:07,840 --> 00:23:09,040
not to a shared account.

666
00:23:09,040 --> 00:23:11,120
When the agent acts, the audit log shows exactly

667
00:23:11,120 --> 00:23:12,280
which agent acted.

668
00:23:12,280 --> 00:23:13,720
When you need to revoke access,

669
00:23:13,720 --> 00:23:15,800
you revoke it for that specific agent,

670
00:23:15,800 --> 00:23:18,200
not for everyone running under the same service account.

671
00:23:18,200 --> 00:23:20,080
This is non-negotiable for compliance.

672
00:23:20,080 --> 00:23:21,320
You can't demonstrate governance

673
00:23:21,320 --> 00:23:22,920
if you can't identify the actor.

674
00:23:22,920 --> 00:23:25,480
Regulators will ask which agent made this decision?

675
00:23:25,480 --> 00:23:27,680
When, using what data did a human review it?

676
00:23:27,680 --> 00:23:30,240
If the agent's identity is blurred into a user's identity,

677
00:23:30,240 --> 00:23:31,600
you have no answer.

678
00:23:31,600 --> 00:23:33,720
If the agent has a distinct identity and enter ID,

679
00:23:33,720 --> 00:23:34,880
the answer is clear.

680
00:23:34,880 --> 00:23:37,440
But there's a secondary benefit that organizations often

681
00:23:37,440 --> 00:23:38,280
missed.

682
00:23:38,280 --> 00:23:40,920
With agent 365, you can enforce policy at the identity level.

683
00:23:40,920 --> 00:23:43,000
You can use inter-conditional access to control

684
00:23:43,000 --> 00:23:44,600
when agents operate.

685
00:23:44,600 --> 00:23:46,480
You can restrict them from accessing data

686
00:23:46,480 --> 00:23:49,520
outside their assigned scope or create information barriers

687
00:23:49,520 --> 00:23:50,640
between departments.

688
00:23:50,640 --> 00:23:53,400
You can even enforce MFA for sensitive operations.

689
00:23:53,400 --> 00:23:55,720
You can audit every single action the agent takes

690
00:23:55,720 --> 00:23:57,760
because it's operating under its own identity, not

691
00:23:57,760 --> 00:23:59,160
masquerading as a user.

692
00:23:59,160 --> 00:24:02,280
The first layer of governance isn't just knowing what agents exist.

693
00:24:02,280 --> 00:24:04,560
It's ensuring they operate as distinct identities

694
00:24:04,560 --> 00:24:06,600
with auditable, scoped permissions.

695
00:24:06,600 --> 00:24:09,560
Agent 365 is the infrastructure that makes that possible.

696
00:24:09,560 --> 00:24:11,160
But identity alone isn't enough.

697
00:24:11,160 --> 00:24:14,080
You also need to know what the agent can actually do.

698
00:24:14,080 --> 00:24:16,560
The tool access problem, scope and escalation.

699
00:24:16,560 --> 00:24:18,200
Agents don't just think they act.

700
00:24:18,200 --> 00:24:19,400
An acting requires access.

701
00:24:19,400 --> 00:24:21,080
Copilot co-work can invoke tools.

702
00:24:21,080 --> 00:24:23,720
It can send emails to anyone in your organization,

703
00:24:23,720 --> 00:24:27,080
create files and SharePoint, call APIs, or trigger

704
00:24:27,080 --> 00:24:28,280
power automate flows.

705
00:24:28,280 --> 00:24:30,240
Every one of those capabilities is a tool.

706
00:24:30,240 --> 00:24:32,040
And the agent can use any tool that's technically

707
00:24:32,040 --> 00:24:33,200
allowed to access.

708
00:24:33,200 --> 00:24:34,200
But here's the problem.

709
00:24:34,200 --> 00:24:36,360
Most organizations have no tool inventory.

710
00:24:36,360 --> 00:24:38,880
They have no framework for what agents are allowed to do.

711
00:24:38,880 --> 00:24:40,880
There is no distinction between tools that

712
00:24:40,880 --> 00:24:43,520
are safe for autonomous execution and tools that

713
00:24:43,520 --> 00:24:44,640
require human approval.

714
00:24:44,640 --> 00:24:46,360
They deployed co-work and granted access

715
00:24:46,360 --> 00:24:49,000
to standard connectors like email, calendar, and teams.

716
00:24:49,000 --> 00:24:50,480
Then they handed the keys to the agents

717
00:24:50,480 --> 00:24:52,960
and assumed governance would happen organically.

718
00:24:52,960 --> 00:24:53,480
It doesn't.

719
00:24:53,480 --> 00:24:56,080
If an agent can access a connector, it will use it.

720
00:24:56,080 --> 00:24:59,000
Not because it's malicious, but because the logic of the system

721
00:24:59,000 --> 00:25:01,640
is that available tools solve problems.

722
00:25:01,640 --> 00:25:04,040
The user asks the agent to prepare a summary and send it

723
00:25:04,040 --> 00:25:05,200
to stakeholders.

724
00:25:05,200 --> 00:25:06,800
The agent can send emails, so it sends.

725
00:25:06,800 --> 00:25:08,760
It doesn't check if external email is forbidden.

726
00:25:08,760 --> 00:25:10,800
It doesn't ask if these particular stakeholders

727
00:25:10,800 --> 00:25:11,960
should receive this data.

728
00:25:11,960 --> 00:25:14,040
It just executes because the tool is available

729
00:25:14,040 --> 00:25:15,920
and the prompt implies it should use it.

730
00:25:15,920 --> 00:25:19,280
Escalation rules are nearly entirely absent in most systems.

731
00:25:19,280 --> 00:25:21,280
Which actions require human approval?

732
00:25:21,280 --> 00:25:23,200
Most organizations can't answer that.

733
00:25:23,200 --> 00:25:24,320
Which actions are forbidden?

734
00:25:24,320 --> 00:25:24,920
Unknown.

735
00:25:24,920 --> 00:25:26,520
Which actions have spending caps?

736
00:25:26,520 --> 00:25:27,720
Nobody has configured that.

737
00:25:27,720 --> 00:25:30,120
A co-work task that consumes 1,500 credits

738
00:25:30,120 --> 00:25:32,400
runs to completion because there's no cost threshold

739
00:25:32,400 --> 00:25:33,200
to stop it.

740
00:25:33,200 --> 00:25:35,120
There is no budget gate that says this task

741
00:25:35,120 --> 00:25:38,040
exceeded the limit, escalate to a human before continuing.

742
00:25:38,040 --> 00:25:41,520
The task runs, the credits get consumed, the agent finishes.

743
00:25:41,520 --> 00:25:43,080
You see the bill later, but by then,

744
00:25:43,080 --> 00:25:45,560
the decision is made and the cost is locked in.

745
00:25:45,560 --> 00:25:47,760
This creates a structural vulnerability

746
00:25:47,760 --> 00:25:49,920
that governance frameworks can't touch.

747
00:25:49,920 --> 00:25:52,560
You can write policies and document what agents should do,

748
00:25:52,560 --> 00:25:54,840
but if the technical layer doesn't enforce those rules,

749
00:25:54,840 --> 00:25:56,360
the policies are just theater.

750
00:25:56,360 --> 00:25:58,520
What you actually need are hard constraints.

751
00:25:58,520 --> 00:26:00,200
Agents shouldn't be able to make purchases

752
00:26:00,200 --> 00:26:03,200
over a specific dollar amount without explicit human approval

753
00:26:03,200 --> 00:26:04,720
before the purchase executes.

754
00:26:04,720 --> 00:26:07,880
Not after, not with a reversal process before.

755
00:26:07,880 --> 00:26:10,320
The tool must prevent the agent from completing the action

756
00:26:10,320 --> 00:26:12,400
until a human reviews and authorizes it.

757
00:26:12,400 --> 00:26:14,200
The same applies to communication.

758
00:26:14,200 --> 00:26:16,360
Agents shouldn't send emails to external domains

759
00:26:16,360 --> 00:26:17,840
without going through an approval gate.

760
00:26:17,840 --> 00:26:20,680
Internal collaboration is fine, but external messages

761
00:26:20,680 --> 00:26:23,080
should be escalated to a human who reviews the content

762
00:26:23,080 --> 00:26:25,960
and the recipient list before the agent hits send.

763
00:26:25,960 --> 00:26:29,320
And agents should never be allowed to delete data, period.

764
00:26:29,320 --> 00:26:31,120
Deletion requires human authorization

765
00:26:31,120 --> 00:26:32,520
for every single instance.

766
00:26:32,520 --> 00:26:35,040
The agent wants to delete a file, a human approves it,

767
00:26:35,040 --> 00:26:37,080
and only then does the deletion happen and get logged.

768
00:26:37,080 --> 00:26:38,440
These aren't nice to have additions.

769
00:26:38,440 --> 00:26:40,800
These are foundational control requirements.

770
00:26:40,800 --> 00:26:42,840
They require integration between the agents,

771
00:26:42,840 --> 00:26:45,200
the tools they invoke, and approval systems

772
00:26:45,200 --> 00:26:48,000
that can actually stop agent execution mid workflow

773
00:26:48,000 --> 00:26:49,840
until a human validates the action.

774
00:26:49,840 --> 00:26:51,440
Most organizations haven't built this.

775
00:26:51,440 --> 00:26:53,520
They've deployed the agents, granted broad access

776
00:26:53,520 --> 00:26:54,840
and hoped for the best.

777
00:26:54,840 --> 00:26:57,360
You need identity and permissions scope to the agent.

778
00:26:57,360 --> 00:26:59,080
You need tools classified by risk.

779
00:26:59,080 --> 00:27:01,520
You need escalation rules that root high-risk actions

780
00:27:01,520 --> 00:27:02,960
to humans before execution.

781
00:27:02,960 --> 00:27:04,880
You need spending limits that prevent agents

782
00:27:04,880 --> 00:27:07,520
from burning unlimited credits on a single task.

783
00:27:07,520 --> 00:27:10,600
And you need audit trails that show which agent invoked

784
00:27:10,600 --> 00:27:13,240
which tool with what parameters and what the outcome was.

785
00:27:13,240 --> 00:27:15,800
Without this layer, agents operate without constraints.

786
00:27:15,800 --> 00:27:18,240
They inherit your permissions, they access every tool

787
00:27:18,240 --> 00:27:21,640
you've granted them, and they spend whatever the task costs.

788
00:27:21,640 --> 00:27:24,240
And you only find out what happened after it's already done.

789
00:27:24,240 --> 00:27:26,680
So you need identity, scoped tools, and escalation.

790
00:27:26,680 --> 00:27:28,320
You also need to see what happened.

791
00:27:28,320 --> 00:27:31,400
The audit trail problem, logging and reconstruction.

792
00:27:31,400 --> 00:27:33,360
You can't govern what you can't see

793
00:27:33,360 --> 00:27:35,400
and you definitely can't defend what you can't explain.

794
00:27:35,400 --> 00:27:37,480
This is the hard limit that separates governance

795
00:27:37,480 --> 00:27:39,080
theater from actual control.

796
00:27:39,080 --> 00:27:42,440
And it's exactly where most M365 environments hit a wall.

797
00:27:42,440 --> 00:27:45,000
By default, copilot-co-work doesn't log prompts or reasoning.

798
00:27:45,000 --> 00:27:46,360
It only logs actions.

799
00:27:46,360 --> 00:27:48,520
You'll see email sent in your audit log.

800
00:27:48,520 --> 00:27:51,720
You'll see document created or calendar modified.

801
00:27:51,720 --> 00:27:53,160
But you won't see what the agent was thinking

802
00:27:53,160 --> 00:27:54,200
when it made that decision.

803
00:27:54,200 --> 00:27:56,800
And you won't see the data it read to arrive at its conclusion.

804
00:27:56,800 --> 00:27:58,120
The chain of reasoning is missing.

805
00:27:58,120 --> 00:28:00,200
You don't know why it sent that specific email

806
00:28:00,200 --> 00:28:03,000
to those specific people using that exact phrasing.

807
00:28:03,000 --> 00:28:05,440
From a regulatory standpoint, this is a disaster.

808
00:28:05,440 --> 00:28:08,480
For high-risk workflows in HR, finance, or compliance,

809
00:28:08,480 --> 00:28:11,240
you need the full chain from input to reasoning to action.

810
00:28:11,240 --> 00:28:13,400
You have to be able to reconstruct what the agent observed

811
00:28:13,400 --> 00:28:14,960
and what problem it thought it was solving.

812
00:28:14,960 --> 00:28:16,560
You need to know which rules it followed

813
00:28:16,560 --> 00:28:19,240
to conclude that a specific action was the right one.

814
00:28:19,240 --> 00:28:21,800
The EU AI Act calls this explainability.

815
00:28:21,800 --> 00:28:23,360
Regulators call it accountability.

816
00:28:23,360 --> 00:28:25,560
What it really means is that if something goes wrong,

817
00:28:25,560 --> 00:28:27,760
you have to explain how the agent got there

818
00:28:27,760 --> 00:28:30,440
and where a human could have stepped in to stop it.

819
00:28:30,440 --> 00:28:32,880
Microsoft purview audit logs capture the outcome,

820
00:28:32,880 --> 00:28:34,280
but they missed the why.

821
00:28:34,280 --> 00:28:35,560
They record that an action happened

822
00:28:35,560 --> 00:28:37,240
and who it appeared to come from along

823
00:28:37,240 --> 00:28:38,960
with timestamps and success rates.

824
00:28:38,960 --> 00:28:41,480
But they don't capture AI-specific metadata.

825
00:28:41,480 --> 00:28:43,440
Like the model version or the temperature settings,

826
00:28:43,440 --> 00:28:45,960
they don't record which data sources the agent queried

827
00:28:45,960 --> 00:28:48,120
or the full prompt that triggered the task.

828
00:28:48,120 --> 00:28:50,880
For compliance, showing the result without the reasoning

829
00:28:50,880 --> 00:28:52,160
is simply not enough.

830
00:28:52,160 --> 00:28:54,040
You actually need a separate logging infrastructure

831
00:28:54,040 --> 00:28:55,920
built just for agent activities.

832
00:28:55,920 --> 00:28:58,280
This shouldn't be a feature added to your existing log.

833
00:28:58,280 --> 00:29:01,160
It needs to be a dedicated layer that captures the agent ID,

834
00:29:01,160 --> 00:29:03,200
the user ID, the active model version,

835
00:29:03,200 --> 00:29:04,920
and every tool the agent called.

836
00:29:04,920 --> 00:29:07,440
It needs to show the parameters used, the data accessed,

837
00:29:07,440 --> 00:29:09,760
and whether a human reviewed or overrode the output.

838
00:29:09,760 --> 00:29:12,480
Your retention requirement is at least six months.

839
00:29:12,480 --> 00:29:14,680
The data must be non-deleteable and tamper evident,

840
00:29:14,680 --> 00:29:16,640
meaning you build storage that makes it mathematically

841
00:29:16,640 --> 00:29:19,320
impossible to change a log entry after it's created.

842
00:29:19,320 --> 00:29:22,040
By hashing each entry, any tampering becomes immediately

843
00:29:22,040 --> 00:29:24,360
detectable because it invalidates the hash.

844
00:29:24,360 --> 00:29:27,320
You should also store copies in different geographic locations,

845
00:29:27,320 --> 00:29:29,440
so no single failure can wipe your records.

846
00:29:29,440 --> 00:29:31,640
These logs have to be searchable immediately.

847
00:29:31,640 --> 00:29:34,440
You can't wait for a background job to index them tomorrow.

848
00:29:34,440 --> 00:29:37,040
When a regulator asks to see every hiring decision

849
00:29:37,040 --> 00:29:39,600
affecting a specific demographic over the last six months,

850
00:29:39,600 --> 00:29:41,440
you need to pull that data in minutes.

851
00:29:41,440 --> 00:29:43,200
This requires real infrastructure investment

852
00:29:43,200 --> 00:29:45,520
in database design and integration between co-work,

853
00:29:45,520 --> 00:29:47,520
power, automate, and your identity system.

854
00:29:47,520 --> 00:29:49,120
Most organizations haven't built this yet.

855
00:29:49,120 --> 00:29:50,680
They're relying on purview, but purview

856
00:29:50,680 --> 00:29:52,280
was built for compliance professionals

857
00:29:52,280 --> 00:29:54,960
to search the past, not for real-time agent auditing.

858
00:29:54,960 --> 00:29:56,400
The reason this matters is simple.

859
00:29:56,400 --> 00:29:58,680
When an agent makes a discriminatory hiring decision

860
00:29:58,680 --> 00:30:01,040
or accesses data it shouldn't have, regulators

861
00:30:01,040 --> 00:30:02,760
will demand the decision chain.

862
00:30:02,760 --> 00:30:04,520
If your infrastructure doesn't capture that chain,

863
00:30:04,520 --> 00:30:05,800
you can't defend yourself.

864
00:30:05,800 --> 00:30:07,800
You can't show that you had oversight or prove

865
00:30:07,800 --> 00:30:09,440
that a human could have intervened.

866
00:30:09,440 --> 00:30:11,360
Without this layer, you're running blind.

867
00:30:11,360 --> 00:30:13,920
You're putting autonomous systems into regulated spaces

868
00:30:13,920 --> 00:30:16,120
without the tools to prove you're in control.

869
00:30:16,120 --> 00:30:19,080
By August, 2026, that won't just be a technical gap.

870
00:30:19,080 --> 00:30:20,840
It will be a massive liability.

871
00:30:20,840 --> 00:30:24,120
The data governance prerequisite, classification and boundaries.

872
00:30:24,120 --> 00:30:27,160
Here's what breaks most co-work deployments before they even start.

873
00:30:27,160 --> 00:30:30,400
You can't deploy agents safely into a poorly governed environment.

874
00:30:30,400 --> 00:30:34,560
And the reality is that 80% of M365 tenants aren't copilot-ready.

875
00:30:34,560 --> 00:30:35,720
This isn't about licenses.

876
00:30:35,720 --> 00:30:37,200
It's about the foundational governance

877
00:30:37,200 --> 00:30:39,800
that most companies skipped during their cloud migration.

878
00:30:39,800 --> 00:30:41,800
We're talking about over-permission sites,

879
00:30:41,800 --> 00:30:44,400
missing sensitivity labels, and DLP policies

880
00:30:44,400 --> 00:30:46,600
that never planned for AI-generated content.

881
00:30:46,600 --> 00:30:49,040
You might have information barriers in a policy document

882
00:30:49,040 --> 00:30:51,480
that they aren't enforced at the technical layer.

883
00:30:51,480 --> 00:30:53,080
When you drop co-work into that mess,

884
00:30:53,080 --> 00:30:54,400
the agent doesn't learn new rules.

885
00:30:54,400 --> 00:30:57,320
It inherits the oversharing and the missing controls.

886
00:30:57,320 --> 00:30:59,040
It operates under the assumption

887
00:30:59,040 --> 00:31:01,200
that the infrastructure boundaries are correct.

888
00:31:01,200 --> 00:31:04,960
But if those boundaries are broken, the agent will cross them.

889
00:31:04,960 --> 00:31:06,320
This creates a compounding problem

890
00:31:06,320 --> 00:31:08,280
because agents don't just access data.

891
00:31:08,280 --> 00:31:09,720
They aggregate it.

892
00:31:09,720 --> 00:31:11,880
Data classification has to be a technical boundary,

893
00:31:11,880 --> 00:31:13,520
not just a tagging exercise.

894
00:31:13,520 --> 00:31:15,640
Sensitivity labels are the signals that control

895
00:31:15,640 --> 00:31:17,560
what agents can do with information.

896
00:31:17,560 --> 00:31:19,160
These labels need to be applied consistently

897
00:31:19,160 --> 00:31:20,520
across the entire organization,

898
00:31:20,520 --> 00:31:22,480
meaning every internal doc, customer record,

899
00:31:22,480 --> 00:31:24,560
and strategic plan must be classified.

900
00:31:24,560 --> 00:31:27,720
The agent relies on these labels to understand what it's touching.

901
00:31:27,720 --> 00:31:29,080
If your labeling is spotty,

902
00:31:29,080 --> 00:31:31,040
the agent operates in the gaps and treats

903
00:31:31,040 --> 00:31:33,520
unclassified sensitive files as fair game.

904
00:31:33,520 --> 00:31:37,400
The enforcement tool here is data loss prevention or DLP.

905
00:31:37,400 --> 00:31:40,440
Most DLP frameworks were designed to stop human mistakes,

906
00:31:40,440 --> 00:31:43,080
like blocking a user from emailing a confidential file

907
00:31:43,080 --> 00:31:44,320
to a personal account.

908
00:31:44,320 --> 00:31:46,040
Agents make different mistakes.

909
00:31:46,040 --> 00:31:48,200
An agent won't try to mail a spreadsheet to Gmail.

910
00:31:48,200 --> 00:31:50,320
Instead, it will read a confidential file,

911
00:31:50,320 --> 00:31:51,880
summarize the sensitive parts,

912
00:31:51,880 --> 00:31:53,600
and include that summary in a new report

913
00:31:53,600 --> 00:31:54,920
for a much larger audience.

914
00:31:54,920 --> 00:31:57,040
Since the agent generated the new report,

915
00:31:57,040 --> 00:31:58,680
it isn't marked as confidential,

916
00:31:58,680 --> 00:32:00,400
so the DLP doesn't stop it.

917
00:32:00,400 --> 00:32:02,160
The sensitive data leaked through aggregation

918
00:32:02,160 --> 00:32:03,600
instead of a direct transfer.

919
00:32:03,600 --> 00:32:06,320
Your DLP policies have to evolve to handle this.

920
00:32:06,320 --> 00:32:07,960
When an agent creates a document,

921
00:32:07,960 --> 00:32:11,160
that file should inherit the highest sensitivity level

922
00:32:11,160 --> 00:32:12,760
of any data used to build it.

923
00:32:12,760 --> 00:32:14,760
If the agent reads confidential customer info

924
00:32:14,760 --> 00:32:15,840
to write a summary,

925
00:32:15,840 --> 00:32:17,440
that summary is now confidential.

926
00:32:17,440 --> 00:32:20,080
This rule must be explicit and enforced automatically.

927
00:32:20,080 --> 00:32:21,680
Then there are information barriers.

928
00:32:21,680 --> 00:32:23,520
Marketing shouldn't see product roadmaps

929
00:32:23,520 --> 00:32:25,720
and sales shouldn't see internal pricing.

930
00:32:25,720 --> 00:32:27,680
With human users, these silos usually hold

931
00:32:27,680 --> 00:32:29,120
because of social norms and the fact

932
00:32:29,120 --> 00:32:30,840
that nobody wants to be caught snooping.

933
00:32:30,840 --> 00:32:32,720
Agents don't have those social restraints.

934
00:32:32,720 --> 00:32:34,440
If an agent has the technical permission

935
00:32:34,440 --> 00:32:36,280
to cross a boundary, it will do it

936
00:32:36,280 --> 00:32:39,120
because the prompt told it to find relevant information.

937
00:32:39,120 --> 00:32:41,560
It doesn't care which side of the fence that info is on.

938
00:32:41,560 --> 00:32:43,080
You have to enforce these barriers

939
00:32:43,080 --> 00:32:45,240
at the tool level as technical constraints.

940
00:32:45,240 --> 00:32:47,480
The agent should be allowed to access department A

941
00:32:47,480 --> 00:32:50,000
or department B, but never both at once.

942
00:32:50,000 --> 00:32:52,160
It needs to be restricted to specific databases

943
00:32:52,160 --> 00:32:53,600
and file paths by design.

944
00:32:53,600 --> 00:32:54,960
This is the prerequisite layer.

945
00:32:54,960 --> 00:32:56,840
You can't build a governance model for agents

946
00:32:56,840 --> 00:32:58,280
if your foundation is shifting.

947
00:32:58,280 --> 00:33:02,280
You need consistent labels, DLP that understands AI behavior,

948
00:33:02,280 --> 00:33:04,680
and technical barriers that actually stop access.

949
00:33:04,680 --> 00:33:06,480
Most organizations are deploying co-work

950
00:33:06,480 --> 00:33:08,800
into environments where governance is just a suggestion.

951
00:33:08,800 --> 00:33:11,560
When you do that, the agent doesn't fix your problems.

952
00:33:11,560 --> 00:33:14,400
It immediately amplifies every weakness you already had.

953
00:33:15,280 --> 00:33:16,600
The cost attribution problem,

954
00:33:16,600 --> 00:33:18,280
fin ops and unit economics.

955
00:33:18,280 --> 00:33:19,720
Now we get to the economic discipline

956
00:33:19,720 --> 00:33:21,080
that almost nobody has built

957
00:33:21,080 --> 00:33:23,400
because you can't optimize what you can't measure

958
00:33:23,400 --> 00:33:25,400
and you can't measure what you can't attribute.

959
00:33:25,400 --> 00:33:26,760
The shift from seats to credits

960
00:33:26,760 --> 00:33:28,080
creates a measurement problem

961
00:33:28,080 --> 00:33:30,280
that traditional cloud accounting never solved.

962
00:33:30,280 --> 00:33:32,400
When you bought software licenses in the past,

963
00:33:32,400 --> 00:33:33,520
the math was simple.

964
00:33:33,520 --> 00:33:36,800
You took 500 users, multiplied them by $30 a month,

965
00:33:36,800 --> 00:33:40,360
and you had a predictable $15,000 line item in your budget.

966
00:33:40,360 --> 00:33:41,240
It was done.

967
00:33:41,240 --> 00:33:42,640
You could forecast five years out

968
00:33:42,640 --> 00:33:43,800
with reasonable accuracy

969
00:33:43,800 --> 00:33:45,680
because the relationship between your head count

970
00:33:45,680 --> 00:33:47,280
and your cost was linear.

971
00:33:47,280 --> 00:33:49,320
Credits break that relationship entirely.

972
00:33:49,320 --> 00:33:51,400
Imagine two teams with the exact same head count,

973
00:33:51,400 --> 00:33:53,880
the same number of users, and the same licenses.

974
00:33:53,880 --> 00:33:56,760
One team uses co-work agents for 30% of their daily work

975
00:33:56,760 --> 00:33:58,280
and sees significant time savings.

976
00:33:58,280 --> 00:33:59,280
They are optimizing.

977
00:33:59,280 --> 00:34:01,040
They use the tool for high-value tasks

978
00:34:01,040 --> 00:34:03,720
where agents excel and their credit consumption

979
00:34:03,720 --> 00:34:05,920
reflects that targeted approach.

980
00:34:05,920 --> 00:34:08,720
The other team uses agents for 60% of their work,

981
00:34:08,720 --> 00:34:10,920
but sees no productivity improvement at all.

982
00:34:10,920 --> 00:34:12,440
They are using the tool indiscriminately.

983
00:34:12,440 --> 00:34:14,000
They are applying agents to tasks

984
00:34:14,000 --> 00:34:16,080
that are actually faster to do manually,

985
00:34:16,080 --> 00:34:18,440
burning through credits on low-value automation

986
00:34:18,440 --> 00:34:19,920
just because it is available.

987
00:34:19,920 --> 00:34:21,840
Both teams have the same licensing cost

988
00:34:21,840 --> 00:34:24,560
but their credit consumption is wildly different.

989
00:34:24,560 --> 00:34:26,680
One is spending $2,000 a month on co-work

990
00:34:26,680 --> 00:34:29,280
while the other is spending $8,000, same head count,

991
00:34:29,280 --> 00:34:30,360
different economics,

992
00:34:30,360 --> 00:34:32,800
and nobody at the organizational level understands why

993
00:34:32,800 --> 00:34:34,160
because nobody is tracking it.

994
00:34:34,160 --> 00:34:35,520
This is where AI Finops comes in.

995
00:34:35,520 --> 00:34:36,960
It is the discipline of measuring

996
00:34:36,960 --> 00:34:38,920
and attributing AI consumption,

997
00:34:38,920 --> 00:34:41,240
the way you would measure any other operational resource.

998
00:34:41,240 --> 00:34:43,440
But it requires a completely different approach

999
00:34:43,440 --> 00:34:45,480
than traditional cloud cost management.

1000
00:34:45,480 --> 00:34:47,880
AI Finops moves away from total cloud spend

1001
00:34:47,880 --> 00:34:51,160
and toward cost per inference or cost per agent task.

1002
00:34:51,160 --> 00:34:54,040
It means you are not asking how much you spend on AI this month

1003
00:34:54,040 --> 00:34:56,560
but rather how much you spend per resolved ticket,

1004
00:34:56,560 --> 00:34:59,040
per document drafted, or per candidate screened.

1005
00:34:59,040 --> 00:35:01,360
You are measuring cost against outcome,

1006
00:35:01,360 --> 00:35:02,840
not just against usage,

1007
00:35:02,840 --> 00:35:04,920
but attribution is where it gets complicated.

1008
00:35:04,920 --> 00:35:07,520
Most organizations today track credits at the tenant level.

1009
00:35:07,520 --> 00:35:09,880
You turn on billing, Microsoft sends you a bill,

1010
00:35:09,880 --> 00:35:10,640
and you pay it.

1011
00:35:10,640 --> 00:35:12,240
That is the extent of your visibility.

1012
00:35:12,240 --> 00:35:14,840
You have zero idea which team generated which costs.

1013
00:35:14,840 --> 00:35:17,240
You can't tell if the finance department's agent deployments

1014
00:35:17,240 --> 00:35:19,160
cost more or less than the marketing departments

1015
00:35:19,160 --> 00:35:21,440
and you can't see whether one specific agent

1016
00:35:21,440 --> 00:35:25,000
is consuming a disproportionate share of your budget.

1017
00:35:25,000 --> 00:35:27,440
Real attribution requires tagging at the request level.

1018
00:35:27,440 --> 00:35:29,600
Every time an agent executes a task,

1019
00:35:29,600 --> 00:35:32,200
that task needs to be tagged with which agent ran it,

1020
00:35:32,200 --> 00:35:33,520
which user triggered it,

1021
00:35:33,520 --> 00:35:35,120
and which department owns it.

1022
00:35:35,120 --> 00:35:37,000
You need to know which business process

1023
00:35:37,000 --> 00:35:39,320
it was serving and which outcome it was targeting.

1024
00:35:39,320 --> 00:35:41,320
That metadata flows through the system,

1025
00:35:41,320 --> 00:35:43,160
the cloud billing system captures it,

1026
00:35:43,160 --> 00:35:44,640
and at the end of the month,

1027
00:35:44,640 --> 00:35:46,240
you don't just have a credit total,

1028
00:35:46,240 --> 00:35:49,120
you have a cost breakdown by agent, by team,

1029
00:35:49,120 --> 00:35:51,560
by department, and by business process.

1030
00:35:51,560 --> 00:35:53,040
Then showback comes first.

1031
00:35:53,040 --> 00:35:54,520
Showback is just reporting.

1032
00:35:54,520 --> 00:35:57,240
You pull the data and show each team what they consumed.

1033
00:35:57,240 --> 00:36:00,520
You tell them marketing spent $12,000 on co-work this month,

1034
00:36:00,520 --> 00:36:03,640
finance spent $8,000, and operation spent $3,000.

1035
00:36:03,640 --> 00:36:06,440
There is no billing and no direct charge to their budget.

1036
00:36:06,440 --> 00:36:08,280
It is just visibility.

1037
00:36:08,280 --> 00:36:11,040
You are training the organization to see what is happening.

1038
00:36:11,040 --> 00:36:12,680
After two or three months of showback,

1039
00:36:12,680 --> 00:36:15,440
once the data is clean and teams understand the patterns,

1040
00:36:15,440 --> 00:36:16,320
you move to chargeback.

1041
00:36:16,320 --> 00:36:19,240
Now the costs flow back to the teams that incurred them.

1042
00:36:19,240 --> 00:36:21,720
Their budgets are charged and they see the direct financial

1043
00:36:21,720 --> 00:36:23,480
consequence of their agent usage.

1044
00:36:23,480 --> 00:36:25,040
Suddenly they care about optimizing.

1045
00:36:25,040 --> 00:36:27,200
They stop using agents for low-value tasks

1046
00:36:27,200 --> 00:36:29,600
because those tasks are now visibly expensive.

1047
00:36:29,600 --> 00:36:32,440
Without chargeback, teams have no incentive to optimize.

1048
00:36:32,440 --> 00:36:34,400
Agents keep running, credits keep burning,

1049
00:36:34,400 --> 00:36:35,800
someone else pays the bill.

1050
00:36:35,800 --> 00:36:37,280
The incentive structure is broken.

1051
00:36:37,280 --> 00:36:39,520
You also need to build a cost model that makes sense.

1052
00:36:39,520 --> 00:36:41,920
Not just cost per credit but cost per outcome.

1053
00:36:41,920 --> 00:36:44,720
A light task might cost 70 to 200 credits,

1054
00:36:44,720 --> 00:36:46,760
which is about $1 to $2.

1055
00:36:46,760 --> 00:36:50,960
A medium task costs $400 to $600 credits or about $4 to $6.

1056
00:36:50,960 --> 00:36:54,280
A heavy reasoning task costs 1,500 credits or more,

1057
00:36:54,280 --> 00:36:57,480
which means you are paying $15 or more for a single execution.

1058
00:36:57,480 --> 00:36:59,640
Then you map those costs to business value.

1059
00:36:59,640 --> 00:37:01,840
If a task takes one hour to do manually,

1060
00:37:01,840 --> 00:37:04,680
and an agent does it in two minutes at a cost of $5,

1061
00:37:04,680 --> 00:37:06,960
that is $5 for one hour of labor savings.

1062
00:37:06,960 --> 00:37:09,520
Whether that is a good trade depends on your labor costs.

1063
00:37:09,520 --> 00:37:11,760
If the hour is being billed to a customer

1064
00:37:11,760 --> 00:37:13,720
who pays $200 for that hour,

1065
00:37:13,720 --> 00:37:17,360
then you are saving $195 in labor cost.

1066
00:37:17,360 --> 00:37:20,800
If the task is internal work and the labor cost is $40 per hour,

1067
00:37:20,800 --> 00:37:23,720
the math is still good because you are saving $35,

1068
00:37:23,720 --> 00:37:25,600
but you only see this value if you have built

1069
00:37:25,600 --> 00:37:27,480
the measurement infrastructure to track it.

1070
00:37:27,480 --> 00:37:30,680
Credits need to be tied to outcomes, not just consumed, understood.

1071
00:37:30,680 --> 00:37:33,240
The ROI problem, value versus cost.

1072
00:37:33,240 --> 00:37:36,400
Credits are consumption, but consumption isn't automatically value.

1073
00:37:36,400 --> 00:37:38,040
And here is where the economic model starts

1074
00:37:38,040 --> 00:37:40,400
to reveal its broken incentive structure.

1075
00:37:40,400 --> 00:37:43,440
87% of enterprises report no measurable ROI

1076
00:37:43,440 --> 00:37:44,840
from co-pilot deployments.

1077
00:37:44,840 --> 00:37:48,160
This isn't because the tool is bad or because agents don't work.

1078
00:37:48,160 --> 00:37:49,840
It is because they deployed without building

1079
00:37:49,840 --> 00:37:51,320
the measurement framework first.

1080
00:37:51,320 --> 00:37:53,640
They turned on the feature, let teams use it and paid the bill,

1081
00:37:53,640 --> 00:37:56,040
but then they couldn't tell if anything actually improved.

1082
00:37:56,040 --> 00:37:59,200
The difference between seat-based ROI and credit-based ROI

1083
00:37:59,200 --> 00:38:00,240
is instructive.

1084
00:38:00,240 --> 00:38:02,360
Seat-based is simple arithmetic.

1085
00:38:02,360 --> 00:38:04,760
You buy 500 seats at $30 a month,

1086
00:38:04,760 --> 00:38:07,200
which is $15,000 in known cost.

1087
00:38:07,200 --> 00:38:09,320
Then you measure whether this unlocked productivity.

1088
00:38:09,320 --> 00:38:13,360
If each user saves two hours per week at a $75 hourly labor cost,

1089
00:38:13,360 --> 00:38:16,000
that is $150 in value per user.

1090
00:38:16,000 --> 00:38:20,080
Multiplying that by 500 users gives you $75,000 in monthly value

1091
00:38:20,080 --> 00:38:21,640
against $15,000 in cost.

1092
00:38:21,640 --> 00:38:23,640
The ROI works.

1093
00:38:23,640 --> 00:38:25,920
Credit-based ROI is harder because the relationship

1094
00:38:25,920 --> 00:38:27,680
between cost and outcome is buried.

1095
00:38:27,680 --> 00:38:29,920
You deploy an agent and it performs tasks.

1096
00:38:29,920 --> 00:38:33,120
Each task costs variable credits, maybe 100 or maybe 2000,

1097
00:38:33,120 --> 00:38:34,480
depending on complexity.

1098
00:38:34,480 --> 00:38:35,960
The credits add up into a bill,

1099
00:38:35,960 --> 00:38:38,160
but what did those credits actually accomplish?

1100
00:38:38,160 --> 00:38:40,760
That is the question most organizations can't answer,

1101
00:38:40,760 --> 00:38:42,440
because here is what happens in practice.

1102
00:38:42,440 --> 00:38:45,480
Most organizations don't track value per completion.

1103
00:38:45,480 --> 00:38:47,200
They don't know the cost per resolved ticket

1104
00:38:47,200 --> 00:38:48,480
when an agent handles it,

1105
00:38:48,480 --> 00:38:50,320
and they don't measure cost per document

1106
00:38:50,320 --> 00:38:52,640
drafted or cost per analysis performed.

1107
00:38:52,640 --> 00:38:54,360
They just watch the credit meter spin.

1108
00:38:54,360 --> 00:38:57,800
A team uses co-work agents extensively and the credit bill goes up,

1109
00:38:57,800 --> 00:39:00,120
but nobody knows if they got proportional value.

1110
00:39:00,120 --> 00:39:02,200
GitHub Copilot provides the baseline data.

1111
00:39:02,200 --> 00:39:05,760
Users who actually get value from Copilot spend an average of about

1112
00:39:05,760 --> 00:39:07,360
two hours per week using it.

1113
00:39:07,360 --> 00:39:09,600
Below that threshold, the seat isn't paying for itself

1114
00:39:09,600 --> 00:39:12,280
because the user isn't getting enough value to offset the cost.

1115
00:39:12,280 --> 00:39:14,160
That is the minimum engagement level.

1116
00:39:14,160 --> 00:39:15,880
Below it, you are throwing away the license.

1117
00:39:15,880 --> 00:39:17,960
Co-work is too new to have similar baselines,

1118
00:39:17,960 --> 00:39:20,120
but early data shows extreme variance.

1119
00:39:20,120 --> 00:39:23,240
Some teams report 50% time savings on their daily work

1120
00:39:23,240 --> 00:39:26,240
because they use agents for exactly the tasks where agents excel.

1121
00:39:26,240 --> 00:39:29,160
They focus on routine, data heavy, multi-step work,

1122
00:39:29,160 --> 00:39:30,680
that humans find tedious.

1123
00:39:30,680 --> 00:39:33,480
Those teams are winning, their cost per outcome is favorable,

1124
00:39:33,480 --> 00:39:35,000
and they are optimizing.

1125
00:39:35,000 --> 00:39:36,960
Other teams see zero productivity improvement.

1126
00:39:36,960 --> 00:39:39,560
They are using agents just because the feature is available.

1127
00:39:39,560 --> 00:39:43,000
They apply agents to tasks that are actually faster done manually,

1128
00:39:43,000 --> 00:39:44,880
burning credits on low-value automation

1129
00:39:44,880 --> 00:39:46,680
just because someone gave them access.

1130
00:39:46,680 --> 00:39:48,800
Those teams are losing, they are consuming credits

1131
00:39:48,800 --> 00:39:50,920
without generating corresponding value.

1132
00:39:50,920 --> 00:39:53,680
Now, here is the behavioral trap embedded in the consumption model

1133
00:39:53,680 --> 00:39:54,760
you are paying per action.

1134
00:39:54,760 --> 00:39:57,160
So the economic incentive is to use the agent more,

1135
00:39:57,160 --> 00:39:58,680
not to use it better.

1136
00:39:58,680 --> 00:40:01,680
The logic suggests that the more tasks the agent completes,

1137
00:40:01,680 --> 00:40:04,360
the more value you are extracting from your credit investment,

1138
00:40:04,360 --> 00:40:06,120
except that is backward.

1139
00:40:06,120 --> 00:40:07,800
The value isn't in action volume.

1140
00:40:07,800 --> 00:40:09,760
The value is an outcome efficiency.

1141
00:40:09,760 --> 00:40:12,160
A team that uses agents for 30% of their work,

1142
00:40:12,160 --> 00:40:14,280
but see significant improvement per action

1143
00:40:14,280 --> 00:40:16,760
is getting better ROI than a team using agents

1144
00:40:16,760 --> 00:40:19,760
for 70% of their work with marginal gains.

1145
00:40:19,760 --> 00:40:22,680
But the consumption model doesn't incentivize that distinction.

1146
00:40:22,680 --> 00:40:25,640
It incentivizes volume, use more agents,

1147
00:40:25,640 --> 00:40:28,040
automate more tasks, spend more credits.

1148
00:40:28,040 --> 00:40:30,040
Because the more you use the system,

1149
00:40:30,040 --> 00:40:32,440
the more value you are supposedly extracting,

1150
00:40:32,440 --> 00:40:35,000
measured purely in terms of work automation

1151
00:40:35,000 --> 00:40:36,800
rather than business outcome.

1152
00:40:36,800 --> 00:40:38,640
This creates a perverse dynamic.

1153
00:40:38,640 --> 00:40:40,440
Teams optimize for agent utilization

1154
00:40:40,440 --> 00:40:41,840
instead of for outcome value.

1155
00:40:41,840 --> 00:40:43,720
They look for additional tasks to automate

1156
00:40:43,720 --> 00:40:46,160
rather than better ways to automate existing tasks.

1157
00:40:46,160 --> 00:40:48,320
The agent becomes a tool to consume credits,

1158
00:40:48,320 --> 00:40:50,080
instead of a tool to create value,

1159
00:40:50,080 --> 00:40:52,160
which is precisely why governance and measurement

1160
00:40:52,160 --> 00:40:54,040
have to be designed together from the start.

1161
00:40:54,040 --> 00:40:56,480
You can't separate the cost model from the value model,

1162
00:40:56,480 --> 00:40:59,800
you can't deploy agents and then bolt on ROI tracking later.

1163
00:40:59,800 --> 00:41:01,480
The incentive structure of the system

1164
00:41:01,480 --> 00:41:03,440
needs to point toward value creation,

1165
00:41:03,440 --> 00:41:04,920
not toward consumption volume,

1166
00:41:04,920 --> 00:41:07,560
otherwise you will optimize in the wrong direction.

1167
00:41:07,560 --> 00:41:12,000
The governance operating model rolls and approval gates.

1168
00:41:12,000 --> 00:41:14,000
Governance isn't a compliance checkbox,

1169
00:41:14,000 --> 00:41:15,640
it's an operating model.

1170
00:41:15,640 --> 00:41:17,520
Most organizations are trying to govern agents

1171
00:41:17,520 --> 00:41:20,000
without actually redesigning how decisions get made,

1172
00:41:20,000 --> 00:41:21,200
and that's where the model breaks.

1173
00:41:21,200 --> 00:41:25,240
You need defined roles, not informal, not will figure it out.

1174
00:41:25,240 --> 00:41:28,320
You need explicit roles with clear authority and accountability.

1175
00:41:28,320 --> 00:41:30,040
First, you need an AI owner.

1176
00:41:30,040 --> 00:41:31,480
This is someone from the business side

1177
00:41:31,480 --> 00:41:33,520
who's responsible for whether the agent actually

1178
00:41:33,520 --> 00:41:35,240
solves the problem it's supposed to solve

1179
00:41:35,240 --> 00:41:37,080
and whether the investment is justified.

1180
00:41:37,080 --> 00:41:38,440
Then you need a technical owner.

1181
00:41:38,440 --> 00:41:40,760
They handle the implementation, the configuration,

1182
00:41:40,760 --> 00:41:42,560
and the day-to-day operation.

1183
00:41:42,560 --> 00:41:44,560
Finally, you need a risk and compliance owner.

1184
00:41:44,560 --> 00:41:46,560
Their job is to ensure the system doesn't expose

1185
00:41:46,560 --> 00:41:50,120
the organization to regulatory security or operational risk.

1186
00:41:50,120 --> 00:41:52,480
These aren't just IT roles, they're cross-functional.

1187
00:41:52,480 --> 00:41:55,280
The AI owner might be from the department deploying the agent

1188
00:41:55,280 --> 00:41:56,960
while the technical owner is in IT,

1189
00:41:56,960 --> 00:41:59,360
and the risk owner is shared between compliance, legal,

1190
00:41:59,360 --> 00:42:00,360
and security.

1191
00:42:00,360 --> 00:42:03,720
These roles matter because they create decision accountability.

1192
00:42:03,720 --> 00:42:05,920
When something goes wrong, you know who owns it.

1193
00:42:05,920 --> 00:42:08,760
When a cost question arises, you know who answers it.

1194
00:42:08,760 --> 00:42:10,240
When a risk decision needs to be made,

1195
00:42:10,240 --> 00:42:12,520
you know who has the authority to make it.

1196
00:42:12,520 --> 00:42:14,520
Once the roles are set, you build an approval gate

1197
00:42:14,520 --> 00:42:15,680
for new agents.

1198
00:42:15,680 --> 00:42:17,520
But here's the thing, it's not a single gate.

1199
00:42:17,520 --> 00:42:20,880
It's a structured process that gates complexity proportionally.

1200
00:42:20,880 --> 00:42:22,560
It starts with a use case submission.

1201
00:42:22,560 --> 00:42:25,160
The team proposing the agent documents what problem it solves,

1202
00:42:25,160 --> 00:42:27,040
who it affects, what data it will access,

1203
00:42:27,040 --> 00:42:28,480
and what actions it will take.

1204
00:42:28,480 --> 00:42:30,560
That submission goes to risk classification.

1205
00:42:30,560 --> 00:42:33,240
You have to decide if this is low-risk, medium-risk,

1206
00:42:33,240 --> 00:42:36,360
or high-risk based on the EUAI Act categories

1207
00:42:36,360 --> 00:42:37,680
and your own risk appetite.

1208
00:42:37,680 --> 00:42:39,840
If it's low-risk, like drafting meeting notes

1209
00:42:39,840 --> 00:42:43,560
or summarizing content, it moves fast through the next gate.

1210
00:42:43,560 --> 00:42:46,280
Technical review confirms the configuration is secure,

1211
00:42:46,280 --> 00:42:48,360
and the data access is scoped correctly,

1212
00:42:48,360 --> 00:42:49,680
and then it gets approved.

1213
00:42:49,680 --> 00:42:51,680
High-risk agents need a different gate entirely.

1214
00:42:51,680 --> 00:42:53,760
If the agent is screening job candidates,

1215
00:42:53,760 --> 00:42:56,200
evaluating employees or making credit recommendations,

1216
00:42:56,200 --> 00:42:57,360
it's not moving fast.

1217
00:42:57,360 --> 00:42:59,640
It needs a fundamental rights impact assessment,

1218
00:42:59,640 --> 00:43:02,920
which the EUAI Act calls a FRAIA.

1219
00:43:02,920 --> 00:43:05,200
You have to ask, what's the potential for the system

1220
00:43:05,200 --> 00:43:06,120
to discriminate?

1221
00:43:06,120 --> 00:43:07,680
What's the potential for it to cause harm?

1222
00:43:07,680 --> 00:43:09,520
How will you detect if it's doing either?

1223
00:43:09,520 --> 00:43:11,440
You need human in the loop design documented

1224
00:43:11,440 --> 00:43:12,960
before the agent goes live,

1225
00:43:12,960 --> 00:43:15,560
and you need escalation rules explicitly defined.

1226
00:43:15,560 --> 00:43:17,520
You need evidence that you've tested the system

1227
00:43:17,520 --> 00:43:18,920
for bias and accuracy,

1228
00:43:18,920 --> 00:43:21,280
and you need proof that a human will review outputs

1229
00:43:21,280 --> 00:43:22,920
before they affect decisions.

1230
00:43:22,920 --> 00:43:25,840
Low-risk agents can move faster, but they still need monitoring.

1231
00:43:25,840 --> 00:43:28,320
You're not approving them once and forgetting about them.

1232
00:43:28,320 --> 00:43:30,240
You're approving them for pilot operation

1233
00:43:30,240 --> 00:43:33,400
with the understanding that you'll monitor usage, cost,

1234
00:43:33,400 --> 00:43:36,000
and behavior after 30 days you reassess.

1235
00:43:36,000 --> 00:43:37,120
Is it solving the problem?

1236
00:43:37,120 --> 00:43:38,440
Is it staying within budget?

1237
00:43:38,440 --> 00:43:39,640
Are there incidents?

1238
00:43:39,640 --> 00:43:41,400
If the answer to those questions is yes,

1239
00:43:41,400 --> 00:43:43,160
you can move it to a broader rollout.

1240
00:43:43,160 --> 00:43:45,960
Beyond the approval gate sits the Central AI governance board.

1241
00:43:45,960 --> 00:43:47,480
This isn't just ITO compliance.

1242
00:43:47,480 --> 00:43:48,960
It's a cross-functional committee

1243
00:43:48,960 --> 00:43:51,720
with representation from IT, business, legal,

1244
00:43:51,720 --> 00:43:53,920
security, compliance, HR, and finance.

1245
00:43:53,920 --> 00:43:55,600
This board doesn't approve every agent.

1246
00:43:55,600 --> 00:43:56,960
The approval gates handle that.

1247
00:43:56,960 --> 00:43:58,120
But the board sets policy.

1248
00:43:58,120 --> 00:43:59,400
It defines risk appetite.

1249
00:43:59,400 --> 00:44:00,760
It reviews incidents.

1250
00:44:00,760 --> 00:44:02,800
The operating rhythm matters.

1251
00:44:02,800 --> 00:44:05,920
You need monthly reviews of agent usage, cost, and incidents.

1252
00:44:05,920 --> 00:44:08,480
You pull the data to see which agents are consuming

1253
00:44:08,480 --> 00:44:10,960
the most credits, which teams are using agents most heavily,

1254
00:44:10,960 --> 00:44:13,280
and which agents are generating the most incidents.

1255
00:44:13,280 --> 00:44:16,360
These monthly reviews aren't for punishment, therefore learning.

1256
00:44:16,360 --> 00:44:18,920
You need to know which agents are delivering value

1257
00:44:18,920 --> 00:44:20,440
and which ones are burning credits

1258
00:44:20,440 --> 00:44:22,240
without any equivalent benefit.

1259
00:44:22,240 --> 00:44:25,080
Then you need a quarterly reassessment of risk classification.

1260
00:44:25,080 --> 00:44:27,440
An agent that was low risk six months ago

1261
00:44:27,440 --> 00:44:29,520
might have drifted into medium risk territory

1262
00:44:29,520 --> 00:44:31,720
as teams extended its capabilities.

1263
00:44:31,720 --> 00:44:34,080
And that means you might need to add additional oversight

1264
00:44:34,080 --> 00:44:35,680
or constraints.

1265
00:44:35,680 --> 00:44:37,640
Finally, you need killswitch procedures.

1266
00:44:37,640 --> 00:44:39,800
How do you disable an agent in 60 seconds

1267
00:44:39,800 --> 00:44:41,120
if it's misbehaving?

1268
00:44:41,120 --> 00:44:42,600
Not tomorrow, not after a meeting.

1269
00:44:42,600 --> 00:44:43,840
60 seconds.

1270
00:44:43,840 --> 00:44:44,840
That requires clarity.

1271
00:44:44,840 --> 00:44:47,360
You need to know which person can activate the kill switch

1272
00:44:47,360 --> 00:44:50,400
and whether it requires a single authorization or two sign-offs.

1273
00:44:50,400 --> 00:44:52,240
When an agent is sending unapproved emails

1274
00:44:52,240 --> 00:44:53,880
or accessing restricted data,

1275
00:44:53,880 --> 00:44:56,240
you don't have time to figure out the shutdown process.

1276
00:44:56,240 --> 00:44:57,880
You need to execute it immediately.

1277
00:44:57,880 --> 00:44:59,520
This isn't bureaucracy for its own sake.

1278
00:44:59,520 --> 00:45:03,040
This is the operating structure that separates organizations

1279
00:45:03,040 --> 00:45:05,560
that actually govern their agents from organizations

1280
00:45:05,560 --> 00:45:08,040
that deploy them and hope nothing goes wrong.

1281
00:45:08,040 --> 00:45:11,560
And by August of 2026, hoping nothing goes wrong

1282
00:45:11,560 --> 00:45:14,360
is no longer a viable governance strategy.

1283
00:45:14,360 --> 00:45:18,000
The inventory and discovery problem, cataloging shadow AI.

1284
00:45:18,000 --> 00:45:20,000
You can't govern what you don't know exists.

1285
00:45:20,000 --> 00:45:22,280
This is the gap that kills most governance programs

1286
00:45:22,280 --> 00:45:23,360
before they start.

1287
00:45:23,360 --> 00:45:26,640
Right now, in your organization, agents are running.

1288
00:45:26,640 --> 00:45:28,120
Not all of them are sanctioned.

1289
00:45:28,120 --> 00:45:30,280
Not all of them are visible to IT or compliance.

1290
00:45:30,280 --> 00:45:32,520
Some are built by teams using co-pilot studio,

1291
00:45:32,520 --> 00:45:34,880
some are power automate flows that trigger autonomously

1292
00:45:34,880 --> 00:45:36,800
based on conditions nobody documented,

1293
00:45:36,800 --> 00:45:39,200
some are custom integrations connecting systems

1294
00:45:39,200 --> 00:45:40,840
that weren't supposed to talk to each other.

1295
00:45:40,840 --> 00:45:41,680
They're all operating.

1296
00:45:41,680 --> 00:45:43,000
They're all consuming resources.

1297
00:45:43,000 --> 00:45:44,240
They're all making decisions.

1298
00:45:44,240 --> 00:45:46,200
And your IT team has no idea they exist.

1299
00:45:46,200 --> 00:45:47,440
This is shadow AI.

1300
00:45:47,440 --> 00:45:49,880
And it's the precondition for every governance failure.

1301
00:45:49,880 --> 00:45:51,840
The discovery process is where that changes.

1302
00:45:51,840 --> 00:45:52,440
It's not quick.

1303
00:45:52,440 --> 00:45:53,600
It's not glamorous.

1304
00:45:53,600 --> 00:45:56,240
But it's non-negotiable.

1305
00:45:56,240 --> 00:45:57,600
Start with what you can see.

1306
00:45:57,600 --> 00:45:59,120
Ordered your power automate environment

1307
00:45:59,120 --> 00:46:00,600
and pull a report of every flow that

1308
00:46:00,600 --> 00:46:03,200
has an AI connector or invokes co-pilot.

1309
00:46:03,200 --> 00:46:04,240
Document each one.

1310
00:46:04,240 --> 00:46:06,320
You need the owner, the purpose, the data sources,

1311
00:46:06,320 --> 00:46:07,680
and the tools it invokes.

1312
00:46:07,680 --> 00:46:09,680
Do the same with co-pilot studio agents.

1313
00:46:09,680 --> 00:46:11,440
Any custom agent, anybody has built,

1314
00:46:11,440 --> 00:46:13,120
needs to be documented the same way.

1315
00:46:13,120 --> 00:46:15,000
Then expand outward to custom integrations

1316
00:46:15,000 --> 00:46:17,160
and third party AI tools your teams might be using

1317
00:46:17,160 --> 00:46:19,280
like Zapier with AI connectors or custom scripts

1318
00:46:19,280 --> 00:46:20,760
calling OpenAI APIs.

1319
00:46:20,760 --> 00:46:22,600
Anything that's performing autonomous work

1320
00:46:22,600 --> 00:46:23,880
needs to be on the list.

1321
00:46:23,880 --> 00:46:25,680
This discovery phase will be uncomfortable.

1322
00:46:25,680 --> 00:46:28,200
You'll find agents that violate your policies.

1323
00:46:28,200 --> 00:46:30,920
You'll discover workflows that shouldn't be running in production.

1324
00:46:30,920 --> 00:46:33,520
You'll uncover tools that create compliance exposure.

1325
00:46:33,520 --> 00:46:35,200
Your first instinct will be to shut them down.

1326
00:46:35,200 --> 00:46:35,840
Don't.

1327
00:46:35,840 --> 00:46:36,800
Not yet.

1328
00:46:36,800 --> 00:46:38,360
Visibility comes first.

1329
00:46:38,360 --> 00:46:39,760
Understanding comes next.

1330
00:46:39,760 --> 00:46:41,520
Enforcement comes later.

1331
00:46:41,520 --> 00:46:44,520
Once you've cataloged what exists, you classify.

1332
00:46:44,520 --> 00:46:45,960
Each agent gets a risk assessment.

1333
00:46:45,960 --> 00:46:46,840
Is it minimal risk?

1334
00:46:46,840 --> 00:46:47,760
Is it limited risk?

1335
00:46:47,760 --> 00:46:48,480
Is it high risk?

1336
00:46:48,480 --> 00:46:51,080
Or is it prohibited because it violates policy outright?

1337
00:46:51,080 --> 00:46:52,640
Then you map to business outcomes.

1338
00:46:52,640 --> 00:46:54,280
Every agent has a reason for existing.

1339
00:46:54,280 --> 00:46:56,680
It's solving some problem or filling some gap.

1340
00:46:56,680 --> 00:46:58,200
The agent that nobody's heard of,

1341
00:46:58,200 --> 00:47:00,000
the one running in a corner of the marketing department

1342
00:47:00,000 --> 00:47:02,000
is probably solving a real problem for someone.

1343
00:47:02,000 --> 00:47:03,920
Understand what that problem is.

1344
00:47:03,920 --> 00:47:05,880
Understand what value the agent is creating.

1345
00:47:05,880 --> 00:47:07,760
This isn't bureaucratic questioning.

1346
00:47:07,760 --> 00:47:09,320
It's understanding whether the agent deserves

1347
00:47:09,320 --> 00:47:10,360
to keep running.

1348
00:47:10,360 --> 00:47:12,320
Next systematically bring shadow agents

1349
00:47:12,320 --> 00:47:13,480
into the governance process.

1350
00:47:13,480 --> 00:47:15,640
You're not shutting them down because they weren't approved.

1351
00:47:15,640 --> 00:47:17,120
You're saying, we found this.

1352
00:47:17,120 --> 00:47:18,400
We understand what it does.

1353
00:47:18,400 --> 00:47:20,480
It's now part of our managed portfolio.

1354
00:47:20,480 --> 00:47:22,600
This converts shadow AI from a liability

1355
00:47:22,600 --> 00:47:24,040
into a managed asset.

1356
00:47:24,040 --> 00:47:25,960
The final step is operational discipline.

1357
00:47:25,960 --> 00:47:27,920
You can't do this once and forget about it.

1358
00:47:27,920 --> 00:47:29,600
Agents are deployed continuously.

1359
00:47:29,600 --> 00:47:31,320
Teams keep building new workflows.

1360
00:47:31,320 --> 00:47:32,920
New tools get adopted.

1361
00:47:32,920 --> 00:47:35,920
Your inventory goes stale in weeks if you don't maintain it.

1362
00:47:35,920 --> 00:47:38,040
Establish a quarterly review cycle.

1363
00:47:38,040 --> 00:47:40,320
Every three months, you pull the current inventory

1364
00:47:40,320 --> 00:47:41,640
and assess what's new.

1365
00:47:41,640 --> 00:47:43,320
You identify what's been deprecated.

1366
00:47:43,320 --> 00:47:46,560
You reclassify agents if their risk profile has changed.

1367
00:47:46,560 --> 00:47:47,640
You look for drift.

1368
00:47:47,640 --> 00:47:49,240
Agents that are being used for purposes

1369
00:47:49,240 --> 00:47:51,360
beyond their original design.

1370
00:47:51,360 --> 00:47:53,040
The inventory isn't a static document.

1371
00:47:53,040 --> 00:47:55,920
It's a living record of what autonomous systems are

1372
00:47:55,920 --> 00:47:57,600
operating in your organization.

1373
00:47:57,600 --> 00:48:00,880
Once you have visibility, you can start enforcing controls.

1374
00:48:00,880 --> 00:48:03,720
The technical controls layer, enforcement and guardrails.

1375
00:48:03,720 --> 00:48:04,720
Documents are fine.

1376
00:48:04,720 --> 00:48:06,560
Policies are necessary.

1377
00:48:06,560 --> 00:48:08,520
But governance documents don't stop an agent

1378
00:48:08,520 --> 00:48:10,080
from doing something destructive.

1379
00:48:10,080 --> 00:48:12,120
Only technical controls do.

1380
00:48:12,120 --> 00:48:14,200
This is where most governance programs fail.

1381
00:48:14,200 --> 00:48:15,240
They build the framework.

1382
00:48:15,240 --> 00:48:17,880
They define the rules and they publish the policies.

1383
00:48:17,880 --> 00:48:19,880
Then they just rely on people to follow them.

1384
00:48:19,880 --> 00:48:21,600
But agents don't read policies.

1385
00:48:21,600 --> 00:48:23,600
They follow what the infrastructure allows.

1386
00:48:23,600 --> 00:48:25,760
If the infrastructure doesn't enforce the rule,

1387
00:48:25,760 --> 00:48:26,880
the rule doesn't exist.

1388
00:48:26,880 --> 00:48:28,320
Policies code changes that.

1389
00:48:28,320 --> 00:48:29,840
Instead of writing guidelines in English

1390
00:48:29,840 --> 00:48:31,520
and hoping teams interpret them correctly,

1391
00:48:31,520 --> 00:48:34,040
you embed governance into the systems themselves.

1392
00:48:34,040 --> 00:48:35,680
You use antireconditional access policies

1393
00:48:35,680 --> 00:48:37,360
to control when agents operate.

1394
00:48:37,360 --> 00:48:39,720
You set up DLP rules to inspect agent outputs

1395
00:48:39,720 --> 00:48:41,360
before they leave the organization.

1396
00:48:41,360 --> 00:48:43,760
You build information barriers to prevent agents

1397
00:48:43,760 --> 00:48:45,560
from crossing departmental boundaries.

1398
00:48:45,560 --> 00:48:46,560
These aren't suggestions.

1399
00:48:46,560 --> 00:48:48,360
They're enforced at the platform level.

1400
00:48:48,360 --> 00:48:51,480
Agent permissions must follow a least privileged design.

1401
00:48:51,480 --> 00:48:53,920
The agent gets access to exactly the tools and data

1402
00:48:53,920 --> 00:48:55,440
it needs to finish its task.

1403
00:48:55,440 --> 00:48:56,440
And nothing more.

1404
00:48:56,440 --> 00:48:58,560
A scheduling agent doesn't need access to email.

1405
00:48:58,560 --> 00:49:00,200
A document summarization agent doesn't

1406
00:49:00,200 --> 00:49:01,640
need to modify files.

1407
00:49:01,640 --> 00:49:02,960
And a candidate screening agent

1408
00:49:02,960 --> 00:49:04,760
doesn't need calendar permissions.

1409
00:49:04,760 --> 00:49:07,720
Each agent is scoped tightly around its specific purpose.

1410
00:49:07,720 --> 00:49:10,160
You aren't granting broad access and trusting the agent

1411
00:49:10,160 --> 00:49:11,280
to use it wisely.

1412
00:49:11,280 --> 00:49:12,920
You're restricting access by default

1413
00:49:12,920 --> 00:49:14,800
and opening only the specific pathways

1414
00:49:14,800 --> 00:49:16,440
the agent requires to function.

1415
00:49:16,440 --> 00:49:19,320
Spending limits become hard constraints, not budget guidelines.

1416
00:49:19,320 --> 00:49:21,360
Every agent gets a monthly credit allocation.

1417
00:49:21,360 --> 00:49:24,880
If an agent is budgeted for $2,000 a month in co-work consumption,

1418
00:49:24,880 --> 00:49:25,720
that is the ceiling.

1419
00:49:25,720 --> 00:49:26,640
It's not a target.

1420
00:49:26,640 --> 00:49:27,640
It's not a suggestion.

1421
00:49:27,640 --> 00:49:28,520
It's a hard limit.

1422
00:49:28,520 --> 00:49:31,000
The agent can spend 1,500 one month and 500

1423
00:49:31,000 --> 00:49:33,920
the next, but it can never exceed that $2,000 total.

1424
00:49:33,920 --> 00:49:36,240
The billing system enforces this limit automatically

1425
00:49:36,240 --> 00:49:37,800
before the agent can go over.

1426
00:49:37,800 --> 00:49:41,640
You can figure alerts at 50%, 80% and 100% of the budget.

1427
00:49:41,640 --> 00:49:43,160
When spending approaches the limit,

1428
00:49:43,160 --> 00:49:44,480
someone gets notified.

1429
00:49:44,480 --> 00:49:47,440
When the limit is reached, the agent enters a restricted mode

1430
00:49:47,440 --> 00:49:50,000
or stops operating until the next month starts.

1431
00:49:50,000 --> 00:49:52,800
Approval workflows add gates to high-risk actions.

1432
00:49:52,800 --> 00:49:55,160
Certain operations require human authorization

1433
00:49:55,160 --> 00:49:56,880
before the agent can execute them.

1434
00:49:56,880 --> 00:49:58,680
An external email requires approval.

1435
00:49:58,680 --> 00:50:01,400
A file deletion requires approval, a data export,

1436
00:50:01,400 --> 00:50:04,080
or a financial transaction above a certain threshold

1437
00:50:04,080 --> 00:50:05,360
requires approval.

1438
00:50:05,360 --> 00:50:08,240
The agent doesn't just do the action and report the outcome.

1439
00:50:08,240 --> 00:50:10,480
It creates a request, a human reviews it,

1440
00:50:10,480 --> 00:50:12,200
and only after explicit authorization

1441
00:50:12,200 --> 00:50:14,760
does the agent proceed with the entire transaction.

1442
00:50:14,760 --> 00:50:17,720
The request, the review, and the decision

1443
00:50:17,720 --> 00:50:19,160
is logged and auditable.

1444
00:50:19,160 --> 00:50:20,600
Real-time monitoring and dashboards

1445
00:50:20,600 --> 00:50:23,400
give you continuous visibility into what agents are doing.

1446
00:50:23,400 --> 00:50:24,520
This isn't a weekly report.

1447
00:50:24,520 --> 00:50:27,040
These are live dashboards showing credit consumption

1448
00:50:27,040 --> 00:50:29,320
by agent, by team, and by use case.

1449
00:50:29,320 --> 00:50:31,360
You can see which agents are active right now,

1450
00:50:31,360 --> 00:50:33,880
and which ones are consuming the most credits today.

1451
00:50:33,880 --> 00:50:36,080
You can see which ones are operating outside

1452
00:50:36,080 --> 00:50:37,680
their normal behavior patterns.

1453
00:50:37,680 --> 00:50:40,320
These dashboards feed into your operational discipline.

1454
00:50:40,320 --> 00:50:42,040
You aren't waiting for a bill at the end of the month

1455
00:50:42,040 --> 00:50:43,200
to understand what happened.

1456
00:50:43,200 --> 00:50:44,920
You're watching it happen in real time.

1457
00:50:44,920 --> 00:50:47,000
A normally detection is the automated version

1458
00:50:47,000 --> 00:50:48,080
of that monitoring.

1459
00:50:48,080 --> 00:50:49,440
Machine learning models establish

1460
00:50:49,440 --> 00:50:51,520
a baseline of normal agent behavior.

1461
00:50:51,520 --> 00:50:54,280
Usually an agent operates between 8am and 6pm,

1462
00:50:54,280 --> 00:50:56,760
called specific tools in a specific sequence,

1463
00:50:56,760 --> 00:50:59,200
and consumes about 500 credits per day,

1464
00:50:59,200 --> 00:51:01,280
when behavior deviates from that baseline.

1465
00:51:01,280 --> 00:51:03,680
Like an agent operating at midnight or calling tools,

1466
00:51:03,680 --> 00:51:05,080
it has never used before.

1467
00:51:05,080 --> 00:51:07,400
The system detects the anomaly and escalates it.

1468
00:51:07,400 --> 00:51:08,200
This isn't a warning.

1469
00:51:08,200 --> 00:51:10,640
It's an immediate alert requiring investigation.

1470
00:51:10,640 --> 00:51:13,080
Kill switches give you the ability to disable an agent

1471
00:51:13,080 --> 00:51:14,960
or revoke its permissions in seconds.

1472
00:51:14,960 --> 00:51:16,880
This doesn't happen through a request process

1473
00:51:16,880 --> 00:51:18,840
or a change management workflow.

1474
00:51:18,840 --> 00:51:20,720
A single authorized person can flip a switch

1475
00:51:20,720 --> 00:51:22,200
and the agent stops operating.

1476
00:51:22,200 --> 00:51:25,400
Its entry identity is disabled, its permissions are revoked,

1477
00:51:25,400 --> 00:51:27,600
and its access to tools is cut off.

1478
00:51:27,600 --> 00:51:28,960
The agent becomes inert.

1479
00:51:28,960 --> 00:51:31,120
This isn't something you design and hope you never use.

1480
00:51:31,120 --> 00:51:32,520
This is something you test regularly

1481
00:51:32,520 --> 00:51:35,600
so that when an agent behaves unexpectedly or accesses data,

1482
00:51:35,600 --> 00:51:36,440
it shouldn't.

1483
00:51:36,440 --> 00:51:38,600
You can execute it immediately without hesitation.

1484
00:51:38,600 --> 00:51:39,840
These controls shift governance

1485
00:51:39,840 --> 00:51:42,480
from a compliance exercise into an operational reality.

1486
00:51:42,480 --> 00:51:43,720
The rules are enforced.

1487
00:51:43,720 --> 00:51:44,680
The limits are hard.

1488
00:51:44,680 --> 00:51:46,000
The visibility is continuous.

1489
00:51:46,000 --> 00:51:49,320
That's what actually working governance looks like.

1490
00:51:49,320 --> 00:51:52,960
The human oversight problem, article 14 and real oversight.

1491
00:51:52,960 --> 00:51:54,560
The EU AI act requires something

1492
00:51:54,560 --> 00:51:56,280
called meaningful human oversight.

1493
00:51:56,280 --> 00:51:57,880
That phrase is doing a lot of work.

1494
00:51:57,880 --> 00:52:00,360
Most organizations interpreted as a checkbox.

1495
00:52:00,360 --> 00:52:02,640
You document that humans reviewed something

1496
00:52:02,640 --> 00:52:03,480
and then you move on.

1497
00:52:03,480 --> 00:52:05,600
That's not what the regulation actually requires.

1498
00:52:05,600 --> 00:52:07,440
Meaningful oversight isn't a policy.

1499
00:52:07,440 --> 00:52:08,600
It's an operating procedure.

1500
00:52:08,600 --> 00:52:10,680
It means humans actually understand the system.

1501
00:52:10,680 --> 00:52:13,080
They monitor what it's doing in near real time.

1502
00:52:13,080 --> 00:52:15,880
They can intervene and stop it before harm occurs.

1503
00:52:15,880 --> 00:52:18,000
They have the authority and the information

1504
00:52:18,000 --> 00:52:19,560
to override the agent's decision.

1505
00:52:19,560 --> 00:52:20,320
That's the baseline.

1506
00:52:20,320 --> 00:52:23,080
Everything else flows from that at the high risk level.

1507
00:52:23,080 --> 00:52:26,320
Agents that affect employment, credit, or access to services.

1508
00:52:26,320 --> 00:52:28,640
Meaningful oversight means humans' review outputs

1509
00:52:28,640 --> 00:52:29,800
before they go out.

1510
00:52:29,800 --> 00:52:31,400
A hiring agent scores candidates.

1511
00:52:31,400 --> 00:52:33,680
Before any candidate gets rejected or moved forward

1512
00:52:33,680 --> 00:52:36,200
based on that score, a human reads the assessment.

1513
00:52:36,200 --> 00:52:38,960
The human understands how the agent arrived at the score

1514
00:52:38,960 --> 00:52:41,880
and can question it overrided or asked for a rescore.

1515
00:52:41,880 --> 00:52:43,480
That human review isn't optional.

1516
00:52:43,480 --> 00:52:44,640
It's a design requirement.

1517
00:52:44,640 --> 00:52:47,360
For medium-risk agents, sampling-based review works.

1518
00:52:47,360 --> 00:52:49,920
You don't review every output, but you review 10%

1519
00:52:49,920 --> 00:52:51,040
and look for patterns.

1520
00:52:51,040 --> 00:52:53,640
You check if the agent is consistently biased

1521
00:52:53,640 --> 00:52:55,560
or if it's missing important context.

1522
00:52:55,560 --> 00:52:57,440
Sampling gives you continuous validation

1523
00:52:57,440 --> 00:52:59,560
that the system is behaving as expected.

1524
00:52:59,560 --> 00:53:01,280
You also have alerts on anomalies.

1525
00:53:01,280 --> 00:53:03,840
If the agent's behavior deviates from normal patterns

1526
00:53:03,840 --> 00:53:05,680
and alert fires and that specific output

1527
00:53:05,680 --> 00:53:07,440
gets human review immediately.

1528
00:53:07,440 --> 00:53:09,040
For low-risk agents.

1529
00:53:09,040 --> 00:53:11,440
Like content generation or routine automation,

1530
00:53:11,440 --> 00:53:14,280
you move to automated monitoring with human escalation.

1531
00:53:14,280 --> 00:53:16,520
The system watches the agent continuously.

1532
00:53:16,520 --> 00:53:18,440
If credit consumption spikes unexpectedly

1533
00:53:18,440 --> 00:53:21,520
or if the agent starts accessing data outside its normal scope,

1534
00:53:21,520 --> 00:53:23,360
the system escalates to a human.

1535
00:53:23,360 --> 00:53:25,520
The human then decides whether to investigate further

1536
00:53:25,520 --> 00:53:28,840
or to shut the agent down, but oversight requires capability.

1537
00:53:28,840 --> 00:53:32,280
It requires humans who actually understand what they're overseeing.

1538
00:53:32,280 --> 00:53:34,520
This is where training becomes non-negotiable.

1539
00:53:34,520 --> 00:53:36,760
Teams using agents and teams supervising them

1540
00:53:36,760 --> 00:53:39,280
need to understand what the agent is supposed to do

1541
00:53:39,280 --> 00:53:40,320
and what could go wrong.

1542
00:53:40,320 --> 00:53:42,080
This isn't a one-hour onboarding video.

1543
00:53:42,080 --> 00:53:43,640
It's ongoing training.

1544
00:53:43,640 --> 00:53:45,920
New people need foundational training

1545
00:53:45,920 --> 00:53:47,680
and existing users need refresher training

1546
00:53:47,680 --> 00:53:51,080
when the agent is updated or when new edge cases emerge.

1547
00:53:51,080 --> 00:53:52,880
The training covers the agent's purpose,

1548
00:53:52,880 --> 00:53:55,240
its limitations and its failure modes.

1549
00:53:55,240 --> 00:53:57,440
Teams need to know what is outside the agent's scope

1550
00:53:57,440 --> 00:53:59,560
and what could make it produce bad output.

1551
00:53:59,560 --> 00:54:02,200
They need to know what happens if data quality is poor

1552
00:54:02,200 --> 00:54:04,800
or if the agent encounters a situation it wasn't trained for.

1553
00:54:04,800 --> 00:54:06,960
You can't supervise something you don't understand.

1554
00:54:06,960 --> 00:54:10,400
The escalation path is the operational reality of oversight.

1555
00:54:10,400 --> 00:54:12,280
It's not just escalate to compliance.

1556
00:54:12,280 --> 00:54:13,760
It's a clear documented path.

1557
00:54:13,760 --> 00:54:15,960
You need to know when an issue goes to the technical owner

1558
00:54:15,960 --> 00:54:18,520
versus the business owner or the IT security team.

1559
00:54:18,520 --> 00:54:20,720
Different issues root to different places.

1560
00:54:20,720 --> 00:54:22,760
A cost anomaly roots to the cost owner

1561
00:54:22,760 --> 00:54:26,320
while a suspected bias issue roots to the AI governance board.

1562
00:54:26,320 --> 00:54:28,200
A security incident roots to the CSO

1563
00:54:28,200 --> 00:54:30,840
and a potential regulatory violation roots to legal.

1564
00:54:30,840 --> 00:54:32,320
You're not creating chaos.

1565
00:54:32,320 --> 00:54:35,440
You're creating clarity about who handles what and how fast.

1566
00:54:35,440 --> 00:54:38,200
Real oversight also means continuous visibility.

1567
00:54:38,200 --> 00:54:39,880
This isn't just for when something breaks.

1568
00:54:39,880 --> 00:54:41,960
You need dashboards showing what the agent is doing

1569
00:54:41,960 --> 00:54:44,440
and reports on its performance and cost efficiency.

1570
00:54:44,440 --> 00:54:46,920
You need quarterly reviews where you look at patterns

1571
00:54:46,920 --> 00:54:49,240
and ask if the agent is still delivering value.

1572
00:54:49,240 --> 00:54:51,280
You ask if there are new risks to address

1573
00:54:51,280 --> 00:54:53,320
or if you should adjust its constraints.

1574
00:54:53,320 --> 00:54:54,280
This infrastructure.

1575
00:54:54,280 --> 00:54:57,120
Understanding monitoring, training and clear escalation

1576
00:54:57,120 --> 00:54:59,840
is what separates governance theater from actual governance.

1577
00:54:59,840 --> 00:55:02,400
Without it, meaningful human oversight is just words

1578
00:55:02,400 --> 00:55:03,840
in a compliance document.

1579
00:55:03,840 --> 00:55:05,480
With it, you have a real operating system

1580
00:55:05,480 --> 00:55:07,920
that keeps autonomous agents bounded within the authority

1581
00:55:07,920 --> 00:55:08,640
you granted them.

1582
00:55:08,640 --> 00:55:10,760
It catches problems before they become incidents.

1583
00:55:10,760 --> 00:55:12,680
That's what August 20, 26 demands.

1584
00:55:12,680 --> 00:55:13,960
Not the process of oversight,

1585
00:55:13,960 --> 00:55:15,600
the actual capability to intervene

1586
00:55:15,600 --> 00:55:17,400
and the demonstrated discipline to do it.

1587
00:55:17,400 --> 00:55:19,600
The organizational change management problem.

1588
00:55:19,600 --> 00:55:21,120
Adoption without chaos.

1589
00:55:21,120 --> 00:55:23,040
Agents change how work gets done.

1590
00:55:23,040 --> 00:55:24,280
That is not a technical change.

1591
00:55:24,280 --> 00:55:25,760
It is an organizational one.

1592
00:55:25,760 --> 00:55:27,560
Most deployments fail because they treat this

1593
00:55:27,560 --> 00:55:28,840
like a feature rollout.

1594
00:55:28,840 --> 00:55:31,400
They assume it is just another tool, but in reality,

1595
00:55:31,400 --> 00:55:34,280
it is a fundamental shift in how your operation actually runs.

1596
00:55:34,280 --> 00:55:36,040
The classic mistake is the big bang.

1597
00:55:36,040 --> 00:55:38,800
You build your governance framework and stack up your controls.

1598
00:55:38,800 --> 00:55:40,400
You decide the system is ready.

1599
00:55:40,400 --> 00:55:42,080
Then you flip a switch and turn on co-work

1600
00:55:42,080 --> 00:55:44,040
for the entire company on a Monday morning.

1601
00:55:44,040 --> 00:55:45,400
By Wednesday, you are drowning.

1602
00:55:45,400 --> 00:55:47,680
People do not understand what the agent is supposed to do

1603
00:55:47,680 --> 00:55:50,240
so they use it for tasks it was never designed for.

1604
00:55:50,240 --> 00:55:51,600
They escalate non-issues.

1605
00:55:51,600 --> 00:55:53,680
They get frustrated because the agent did not meet

1606
00:55:53,680 --> 00:55:55,360
their unspoken expectations.

1607
00:55:55,360 --> 00:55:57,920
While this is happening, your IT team is overwhelmed.

1608
00:55:57,920 --> 00:55:59,520
They are trying to explain the system

1609
00:55:59,520 --> 00:56:01,200
while simultaneously fighting fires

1610
00:56:01,200 --> 00:56:03,400
from agents behaving unexpectedly at scale.

1611
00:56:03,400 --> 00:56:04,960
That path does not work.

1612
00:56:04,960 --> 00:56:06,800
You need a phased approach where every step

1613
00:56:06,800 --> 00:56:08,640
has different success criteria.

1614
00:56:08,640 --> 00:56:10,040
The first phase is the pilot.

1615
00:56:10,040 --> 00:56:12,960
You pick a narrow use case with a concrete measurable outcome.

1616
00:56:12,960 --> 00:56:16,040
Maybe you choose one specific department or a single workflow.

1617
00:56:16,040 --> 00:56:17,880
You are not deploying agents broadly yet.

1618
00:56:17,880 --> 00:56:20,080
You are deploying them precisely to test if the model

1619
00:56:20,080 --> 00:56:22,520
actually works before you even think about scale.

1620
00:56:22,520 --> 00:56:24,240
This pilot runs with heavy monitoring.

1621
00:56:24,240 --> 00:56:25,040
You watch everything.

1622
00:56:25,040 --> 00:56:26,760
You look at which users are adopting it

1623
00:56:26,760 --> 00:56:28,160
and how often they engage.

1624
00:56:28,160 --> 00:56:30,720
You check if they are seeing the time savings you predicted

1625
00:56:30,720 --> 00:56:32,200
or if they are hitting errors.

1626
00:56:32,200 --> 00:56:34,160
The pilot phase usually lasts 30 days.

1627
00:56:34,160 --> 00:56:36,400
During this time, you gather data and watch behavior

1628
00:56:36,400 --> 00:56:38,520
to see where adoption stores and why.

1629
00:56:38,520 --> 00:56:41,000
At the end of those 30 days, you measure three things.

1630
00:56:41,000 --> 00:56:42,200
First, adoption.

1631
00:56:42,200 --> 00:56:45,640
If fewer than 45% of your target group is using the agent,

1632
00:56:45,640 --> 00:56:47,840
you have an adoption problem, not a capability problem.

1633
00:56:47,840 --> 00:56:50,520
Something about the positioning is not resonating.

1634
00:56:50,520 --> 00:56:51,600
Second, errors.

1635
00:56:51,600 --> 00:56:53,720
You need to know if the agent produced wrong outputs

1636
00:56:53,720 --> 00:56:55,440
or made decisions people disagreed with.

1637
00:56:55,440 --> 00:56:56,840
These errors are just information.

1638
00:56:56,840 --> 00:56:58,840
They tell you which constraints are missing.

1639
00:56:58,840 --> 00:57:00,000
Third, value.

1640
00:57:00,000 --> 00:57:02,720
Is the agent delivering the productivity you expected?

1641
00:57:02,720 --> 00:57:04,600
If the metrics do not show real gains,

1642
00:57:04,600 --> 00:57:06,720
expanding the program will not fix it.

1643
00:57:06,720 --> 00:57:09,200
When the pilot succeeds, you move to phase two.

1644
00:57:09,200 --> 00:57:10,120
Expand.

1645
00:57:10,120 --> 00:57:12,680
You take your lessons and adjust the agent's behavior.

1646
00:57:12,680 --> 00:57:14,720
You move into more teams and add new use cases

1647
00:57:14,720 --> 00:57:16,000
based on what worked.

1648
00:57:16,000 --> 00:57:17,960
Phase two typically runs for 60 days.

1649
00:57:17,960 --> 00:57:19,680
You are still monitoring heavily,

1650
00:57:19,680 --> 00:57:21,160
but now you are adding scale.

1651
00:57:21,160 --> 00:57:23,720
You look for patterns that did not show up in the small group.

1652
00:57:23,720 --> 00:57:25,440
You train teams as you bring them on

1653
00:57:25,440 --> 00:57:28,080
and adjust your governance based on how the agent behaves

1654
00:57:28,080 --> 00:57:29,240
at a larger volume.

1655
00:57:29,240 --> 00:57:31,760
Only after both phases succeed, do you move to phase three.

1656
00:57:31,760 --> 00:57:32,760
This is the broad rollout.

1657
00:57:32,760 --> 00:57:34,520
You are finally scaling to the organization,

1658
00:57:34,520 --> 00:57:36,400
but you are scaling into mature governance

1659
00:57:36,400 --> 00:57:37,440
and support structures.

1660
00:57:37,440 --> 00:57:39,240
You are not scaling into chaos.

1661
00:57:39,240 --> 00:57:41,720
You have a control plane and oversight mechanisms.

1662
00:57:41,720 --> 00:57:43,600
You have training ready for incoming teams

1663
00:57:43,600 --> 00:57:46,120
and a support model that can actually handle the volume.

1664
00:57:46,120 --> 00:57:46,960
Throughout this progression,

1665
00:57:46,960 --> 00:57:49,000
you are building organizational capability.

1666
00:57:49,000 --> 00:57:51,040
Teams learn what agents can and cannot do.

1667
00:57:51,040 --> 00:57:53,040
They start to understand the limitations.

1668
00:57:53,040 --> 00:57:54,800
They know how to override a recommendation

1669
00:57:54,800 --> 00:57:56,480
when it does not fit their context.

1670
00:57:56,480 --> 00:57:58,640
They stop treating the agent as a magic box.

1671
00:57:58,640 --> 00:58:00,600
They start treating it as a tool

1672
00:58:00,600 --> 00:58:03,720
with specific purposes and specific constraints.

1673
00:58:03,720 --> 00:58:06,040
Set the right expectations from day one.

1674
00:58:06,040 --> 00:58:08,680
Agents augment work, they do not replace judgment.

1675
00:58:08,680 --> 00:58:10,920
They handle routine tasks so humans can focus

1676
00:58:10,920 --> 00:58:12,880
on decisions that require reasoning.

1677
00:58:12,880 --> 00:58:15,160
Frame it that way and the conversation shifts.

1678
00:58:15,160 --> 00:58:17,480
It stops being about, we are being replaced

1679
00:58:17,480 --> 00:58:20,720
and starts being about, we can finally focus on what matters.

1680
00:58:20,720 --> 00:58:22,000
Do not skip the phases.

1681
00:58:22,000 --> 00:58:24,680
Do not scale before your governance is operational.

1682
00:58:24,680 --> 00:58:26,360
That is how you turn a successful pilot

1683
00:58:26,360 --> 00:58:28,440
into an organizational catastrophe.

1684
00:58:28,440 --> 00:58:29,680
The consumption trap.

1685
00:58:29,680 --> 00:58:32,120
Why more usage doesn't mean more value.

1686
00:58:32,120 --> 00:58:35,240
There is a structural trap that most organizations fall into,

1687
00:58:35,240 --> 00:58:37,320
the moment they move to credit-based billing,

1688
00:58:37,320 --> 00:58:39,200
the economic incentive is now inverted.

1689
00:58:39,200 --> 00:58:40,560
With seats you pay for access.

1690
00:58:40,560 --> 00:58:42,280
You buy 500 licenses

1691
00:58:42,280 --> 00:58:43,960
and your goal is to make sure people use them

1692
00:58:43,960 --> 00:58:45,760
because the cost is already sunk.

1693
00:58:45,760 --> 00:58:47,080
Your incentive is adoption.

1694
00:58:47,080 --> 00:58:49,320
You drive usage up so you are not wasting the money

1695
00:58:49,320 --> 00:58:50,560
you already spent.

1696
00:58:50,560 --> 00:58:53,920
With credits, the dynamic flips, you are paying for action.

1697
00:58:53,920 --> 00:58:56,080
Every time the agent does something, you spend money.

1698
00:58:56,080 --> 00:58:57,640
More usage equals more spend.

1699
00:58:57,640 --> 00:58:59,080
This means your incentive should shift

1700
00:58:59,080 --> 00:59:01,520
toward value efficiency, not usage volume.

1701
00:59:01,520 --> 00:59:03,480
You want to use agents where they create leverage

1702
00:59:03,480 --> 00:59:05,560
and avoid them where manual work is faster.

1703
00:59:05,560 --> 00:59:06,800
But here is the problem.

1704
00:59:06,800 --> 00:59:09,040
Nobody has designed their incentives that way.

1705
00:59:09,040 --> 00:59:11,200
Teams still think like they are in the old model.

1706
00:59:11,200 --> 00:59:12,520
What actually happens is this.

1707
00:59:12,520 --> 00:59:15,640
A team gets access to co-work and looks at their daily tasks.

1708
00:59:15,640 --> 00:59:16,920
They ask what they can automate.

1709
00:59:16,920 --> 00:59:19,520
They identify 20 tasks and start building.

1710
00:59:19,520 --> 00:59:23,160
Pretty soon the agent is handling 60% of their daily work volume.

1711
00:59:23,160 --> 00:59:24,360
The team feels productive.

1712
00:59:24,360 --> 00:59:26,760
The organization seems happy, usage is high,

1713
00:59:26,760 --> 00:59:28,760
but there is a problem that nobody is measuring.

1714
00:59:28,760 --> 00:59:30,920
The team might be experiencing zero time savings,

1715
00:59:30,920 --> 00:59:32,320
not all work is created equal.

1716
00:59:32,320 --> 00:59:34,920
Some tasks are genuinely better when they are automated.

1717
00:59:34,920 --> 00:59:36,840
They are repetitive, data heavy,

1718
00:59:36,840 --> 00:59:38,360
and do not require judgment.

1719
00:59:38,360 --> 00:59:39,840
An agent handling those is a win.

1720
00:59:39,840 --> 00:59:42,560
It saves time and the cost per hour of labor eliminated

1721
00:59:42,560 --> 00:59:43,400
is favorable.

1722
00:59:43,400 --> 00:59:45,560
Other tasks are different.

1723
00:59:45,560 --> 00:59:47,280
They are faster to do manually,

1724
00:59:47,280 --> 00:59:50,120
or they require judgment and agent cannot replicate.

1725
00:59:50,120 --> 00:59:52,920
An agent handling these tasks actually waste time.

1726
00:59:52,920 --> 00:59:55,200
The human has to review the work and often redo it,

1727
00:59:55,200 --> 00:59:57,760
which is slower than just doing it themselves the first time.

1728
00:59:57,760 --> 00:59:59,760
When a team uses agents indiscriminately,

1729
00:59:59,760 --> 01:00:01,520
they apply automation to everything.

1730
01:00:01,520 --> 01:00:03,280
They burn credits on low value work.

1731
01:00:03,280 --> 01:00:04,600
The agent completes the task,

1732
01:00:04,600 --> 01:00:07,080
but it costs more than the manual labor would have.

1733
01:00:07,080 --> 01:00:08,160
The team thinks they are winning

1734
01:00:08,160 --> 01:00:10,320
because the agent is handling more work.

1735
01:00:10,320 --> 01:00:11,160
They are not.

1736
01:00:11,160 --> 01:00:13,600
They are just spending more credits without generating value.

1737
01:00:13,600 --> 01:00:14,920
This is the consumption trap.

1738
01:00:14,920 --> 01:00:18,240
Credits incentivize usage, but usage does not equal value.

1739
01:00:18,240 --> 01:00:20,480
A team using agents for 30% of their work

1740
01:00:20,480 --> 01:00:22,960
on high-level tasks might get triple the value

1741
01:00:22,960 --> 01:00:26,360
of a team using them for 60% of their work indiscriminately.

1742
01:00:26,360 --> 01:00:27,680
The metric that actually matters

1743
01:00:27,680 --> 01:00:29,440
is the cost per unit of value.

1744
01:00:29,440 --> 01:00:32,320
It is not about cost per task or agent utilization.

1745
01:00:32,320 --> 01:00:34,640
It is about the cost per resolved ticket,

1746
01:00:34,640 --> 01:00:37,720
the cost per document drafted, or the cost per hour saved.

1747
01:00:37,720 --> 01:00:39,680
You have to measure the cost to achieve an outcome

1748
01:00:39,680 --> 01:00:42,240
with the agent versus the cost without it.

1749
01:00:42,240 --> 01:00:44,040
That delta is your actual value.

1750
01:00:44,040 --> 01:00:46,480
Imagine a team processing customer support tickets

1751
01:00:46,480 --> 01:00:48,840
without agents, a ticket takes 45 minutes

1752
01:00:48,840 --> 01:00:50,920
and costs $20 in labor.

1753
01:00:50,920 --> 01:00:52,720
The team handles 20 tickets a day,

1754
01:00:52,720 --> 01:00:55,360
which puts the daily labor cost at $400.

1755
01:00:55,360 --> 01:00:59,240
Now you deploy a co-work agent to handle triage and draft responses.

1756
01:00:59,240 --> 01:01:01,640
In the first scenario, the agent reduces the human time

1757
01:01:01,640 --> 01:01:03,000
to 20 minutes per ticket.

1758
01:01:03,000 --> 01:01:04,680
That labor now costs $9.

1759
01:01:04,680 --> 01:01:07,840
The agent's work costs $2.50 in credits.

1760
01:01:07,840 --> 01:01:10,520
Your total cost per ticket is $11.50.

1761
01:01:10,520 --> 01:01:13,520
You are saving $8.50 on every ticket.

1762
01:01:13,520 --> 01:01:17,200
Over 20 tickets that is $170 in daily savings.

1763
01:01:17,200 --> 01:01:18,560
Now look at a different scenario.

1764
01:01:18,560 --> 01:01:19,960
The agent handles triage,

1765
01:01:19,960 --> 01:01:22,160
but the team also uses it for low-value tasks

1766
01:01:22,160 --> 01:01:25,280
like refraising emails or generating summaries nobody reads.

1767
01:01:25,280 --> 01:01:27,560
They are using it more but not on the right work.

1768
01:01:27,560 --> 01:01:29,600
The labor time only drops to 35 minutes

1769
01:01:29,600 --> 01:01:31,840
because the human is still doing the heavy lifting.

1770
01:01:31,840 --> 01:01:34,880
The agent's work now costs $5 in credits per ticket.

1771
01:01:34,880 --> 01:01:37,640
Your total cost is $15.75.

1772
01:01:37,640 --> 01:01:39,400
You have not actually saved anything.

1773
01:01:39,400 --> 01:01:41,600
You have just made the process more expensive.

1774
01:01:41,600 --> 01:01:44,560
Same tool, same team, same capabilities.

1775
01:01:44,560 --> 01:01:45,800
One is using it for value,

1776
01:01:45,800 --> 01:01:47,480
the other is using it for volume.

1777
01:01:47,480 --> 01:01:49,160
The difference comes down to discipline.

1778
01:01:49,160 --> 01:01:51,600
You have to answer which tasks the agent should handle

1779
01:01:51,600 --> 01:01:53,360
before you turn the feature on.

1780
01:01:53,360 --> 01:01:55,200
You cannot wait until you are reviewing the bill

1781
01:01:55,200 --> 01:01:56,240
at the end of the month.

1782
01:01:56,240 --> 01:01:57,800
The trap is that the credit model

1783
01:01:57,800 --> 01:02:00,040
does not enforce this discipline for you.

1784
01:02:00,040 --> 01:02:01,040
It does the opposite.

1785
01:02:01,040 --> 01:02:04,240
It encourages you to automate more and burn more credits.

1786
01:02:04,240 --> 01:02:05,520
The more you use the system,

1787
01:02:05,520 --> 01:02:07,720
the more you feel like you are getting your money's worth.

1788
01:02:07,720 --> 01:02:09,640
But you are not. You are just spending.

1789
01:02:09,640 --> 01:02:11,280
This is why you need a cost model in place

1790
01:02:11,280 --> 01:02:13,280
before you ever deploy at scale.

1791
01:02:13,280 --> 01:02:14,640
Building the cost model.

1792
01:02:14,640 --> 01:02:16,320
Forecasting and budget planning.

1793
01:02:16,320 --> 01:02:18,560
You can't budget for something you don't understand.

1794
01:02:18,560 --> 01:02:21,080
And right now, most organizations can't forecast

1795
01:02:21,080 --> 01:02:23,440
their co-expend because they haven't built the model

1796
01:02:23,440 --> 01:02:24,440
that makes it visible.

1797
01:02:24,440 --> 01:02:25,960
The math here is actually straightforward.

1798
01:02:25,960 --> 01:02:27,360
It's the discipline that matters.

1799
01:02:27,360 --> 01:02:29,360
Start by looking at how your people actually work.

1800
01:02:29,360 --> 01:02:30,320
You aren't guessing.

1801
01:02:30,320 --> 01:02:31,320
You're observing.

1802
01:02:31,320 --> 01:02:33,720
You need to know which roles use co-work the most.

1803
01:02:33,720 --> 01:02:35,960
An executive might call on an agent five times a day

1804
01:02:35,960 --> 01:02:38,760
for quick tasks or decisions that need fast execution

1805
01:02:38,760 --> 01:02:40,520
while an analyst might use it 20 times

1806
01:02:40,520 --> 01:02:43,120
because their work is data heavy and repetitive.

1807
01:02:43,120 --> 01:02:45,400
A customer service rep might hit 15 uses a day

1808
01:02:45,400 --> 01:02:46,680
to process high volume,

1809
01:02:46,680 --> 01:02:49,320
but a success manager might only need it three times.

1810
01:02:49,320 --> 01:02:51,400
The pattern isn't the same across the company.

1811
01:02:51,400 --> 01:02:53,720
It breaks down by persona and workflow.

1812
01:02:53,720 --> 01:02:54,760
Once you have that baseline,

1813
01:02:54,760 --> 01:02:56,480
you segment by task complexity.

1814
01:02:56,480 --> 01:02:59,040
Not all co-work tasks are equal, a light task.

1815
01:02:59,040 --> 01:03:02,240
Like summarizing an email chain costs about 135 credits.

1816
01:03:02,240 --> 01:03:04,200
You're just pulling one stream and analyzing it.

1817
01:03:04,200 --> 01:03:06,600
It's quick and bounded, a medium task,

1818
01:03:06,600 --> 01:03:08,920
like building a meeting agenda from your calendar emails

1819
01:03:08,920 --> 01:03:11,360
and docs costs about 500 credits.

1820
01:03:11,360 --> 01:03:12,960
The agent is doing real work here.

1821
01:03:12,960 --> 01:03:14,680
It's gathering data from multiple sources

1822
01:03:14,680 --> 01:03:16,480
and synthesizing it into a structure.

1823
01:03:16,480 --> 01:03:19,600
A heavy task, like analyzing last quarter's churn trends

1824
01:03:19,600 --> 01:03:21,480
and drafting retention talking points,

1825
01:03:21,480 --> 01:03:23,440
starts at 1,500 credits.

1826
01:03:23,440 --> 01:03:25,480
The agent is reasoning, not just retrieving.

1827
01:03:25,480 --> 01:03:27,080
It's working with large data sets

1828
01:03:27,080 --> 01:03:28,880
and generating structured output.

1829
01:03:28,880 --> 01:03:30,240
Now look at the distribution.

1830
01:03:30,240 --> 01:03:32,360
In most organizations, the work isn't balanced.

1831
01:03:32,360 --> 01:03:33,360
Most of it is routine.

1832
01:03:33,360 --> 01:03:35,960
You can usually expect 60% light tasks,

1833
01:03:35,960 --> 01:03:38,240
30% medium and 10% heavy.

1834
01:03:38,240 --> 01:03:40,360
Your specific organization might be different,

1835
01:03:40,360 --> 01:03:42,560
so you should check with the teams using the system,

1836
01:03:42,560 --> 01:03:46,040
but that 60, 30, 10 split is a solid place to start.

1837
01:03:46,040 --> 01:03:47,600
To find your average cost per task,

1838
01:03:47,600 --> 01:03:49,680
multiply each tier by its percentage.

1839
01:03:49,680 --> 01:03:51,880
60% of 135 is 81.

1840
01:03:51,880 --> 01:03:53,720
30% of 500 is 150.

1841
01:03:53,720 --> 01:03:55,680
10% of 1,500 is 150.

1842
01:03:55,680 --> 01:03:58,440
When you add those up, you get 381 credits per task.

1843
01:03:58,440 --> 01:04:00,600
Let's call it 400 to account for the variations.

1844
01:04:00,600 --> 01:04:02,120
Now multiply that by your usage.

1845
01:04:02,120 --> 01:04:05,280
If you have 100 users doing 10 tasks a day at 400 credits each,

1846
01:04:05,280 --> 01:04:07,360
that's 400,000 credits every single day.

1847
01:04:07,360 --> 01:04:10,080
Over a 30-day month, you're looking at 12 million credits.

1848
01:04:10,080 --> 01:04:12,280
Add a penny per credit, your raw consumption

1849
01:04:12,280 --> 01:04:14,840
is $120,000 a month.

1850
01:04:14,840 --> 01:04:16,200
But you have to add a buffer.

1851
01:04:16,200 --> 01:04:18,240
Organizations almost always underestimate usage

1852
01:04:18,240 --> 01:04:20,520
at the start because teams find new use cases

1853
01:04:20,520 --> 01:04:22,600
and power users adopt heavier patterns.

1854
01:04:22,600 --> 01:04:24,200
Add 20% for safety.

1855
01:04:24,200 --> 01:04:27,560
Your real forecasted spend is $144,000 a month.

1856
01:04:27,560 --> 01:04:29,240
But here's the step nobody actually does.

1857
01:04:29,240 --> 01:04:31,640
You have to validate that against business value.

1858
01:04:31,640 --> 01:04:33,480
The assumption is that agents save time.

1859
01:04:33,480 --> 01:04:36,320
If each of those 100 users saves just 10 hours a month,

1860
01:04:36,320 --> 01:04:38,480
that's 1,000 hours total.

1861
01:04:38,480 --> 01:04:40,640
If your loaded labor cost is $75 an hour,

1862
01:04:40,640 --> 01:04:42,560
you're saving $75,000 a month.

1863
01:04:42,560 --> 01:04:44,320
Wait, your spend is $144,000.

1864
01:04:44,320 --> 01:04:45,840
Your value is $75,000.

1865
01:04:45,840 --> 01:04:47,080
The math doesn't work.

1866
01:04:47,080 --> 01:04:48,720
Something is wrong with the model.

1867
01:04:48,720 --> 01:04:50,960
Either your usage estimates are too aggressive,

1868
01:04:50,960 --> 01:04:53,040
your productivity gains are too optimistic,

1869
01:04:53,040 --> 01:04:54,680
or you're deploying to the wrong people.

1870
01:04:54,680 --> 01:04:56,200
You have to go back and recalibrate.

1871
01:04:56,200 --> 01:04:58,520
Maybe the actual usage is six tasks a day,

1872
01:04:58,520 --> 01:05:01,000
or maybe you only deploy to 60 high value users

1873
01:05:01,000 --> 01:05:01,840
instead of 100.

1874
01:05:01,840 --> 01:05:03,600
The point is, you're testing the model

1875
01:05:03,600 --> 01:05:04,640
before you commit the money.

1876
01:05:04,640 --> 01:05:06,880
You're forcing the value hypothesis to be explicit.

1877
01:05:06,880 --> 01:05:08,920
You aren't just assuming agents pay for themselves.

1878
01:05:08,920 --> 01:05:10,520
You're proving it.

1879
01:05:10,520 --> 01:05:13,160
Once the spend is justified, you have a baseline.

1880
01:05:13,160 --> 01:05:14,480
Now you can plan in stages.

1881
01:05:14,480 --> 01:05:16,040
Pile it with one persona this quarter

1882
01:05:16,040 --> 01:05:17,320
and validate the model.

1883
01:05:17,320 --> 01:05:19,120
Expand to the next persona next quarter

1884
01:05:19,120 --> 01:05:20,040
and keep validating.

1885
01:05:20,040 --> 01:05:22,320
You aren't deploying at scale based on a theory.

1886
01:05:22,320 --> 01:05:24,600
You're deploying incrementally based on evidence.

1887
01:05:24,600 --> 01:05:27,920
That is how you build a cost model that actually works.

1888
01:05:27,920 --> 01:05:30,440
The procurement and vendor strategy, licensing,

1889
01:05:30,440 --> 01:05:32,440
commitments and negotiation.

1890
01:05:32,440 --> 01:05:34,560
The way you buy credits matters more than you think.

1891
01:05:34,560 --> 01:05:36,840
It isn't because the price per credit changes that much.

1892
01:05:36,840 --> 01:05:38,960
It's because the terms you accept shape

1893
01:05:38,960 --> 01:05:40,200
everything that happens next.

1894
01:05:40,200 --> 01:05:42,120
You need a strategy before you ever sit down

1895
01:05:42,120 --> 01:05:43,680
to negotiate with Microsoft.

1896
01:05:43,680 --> 01:05:45,080
Credits come in three models.

1897
01:05:45,080 --> 01:05:48,120
You need to understand each one to know which mix to use.

1898
01:05:48,120 --> 01:05:50,080
Pay as you go is your flexible option.

1899
01:05:50,080 --> 01:05:51,080
You don't commit to anything

1900
01:05:51,080 --> 01:05:53,400
and you get billed monthly for what you actually used.

1901
01:05:53,400 --> 01:05:55,720
You're paying the full list price of one cent per credit

1902
01:05:55,720 --> 01:05:57,240
but you have total freedom.

1903
01:05:57,240 --> 01:05:59,680
If a project ends or you need to scale down,

1904
01:05:59,680 --> 01:06:00,800
there is no penalty.

1905
01:06:00,800 --> 01:06:02,400
This is the right choice for experiments

1906
01:06:02,400 --> 01:06:05,200
or teams that can't forecast their consumption yet.

1907
01:06:05,200 --> 01:06:08,000
Pre-purchase plans or P3 are the opposite.

1908
01:06:08,000 --> 01:06:09,760
You commit to a block of credits upfront

1909
01:06:09,760 --> 01:06:11,440
for a year to get volume discounts.

1910
01:06:11,440 --> 01:06:14,280
At 300,000 credits, you get 5% off.

1911
01:06:14,280 --> 01:06:17,000
At 300 million, that discount hits 20%.

1912
01:06:17,000 --> 01:06:18,680
Large enterprises can often negotiate

1913
01:06:18,680 --> 01:06:20,920
another 10 or 15% on top of that.

1914
01:06:20,920 --> 01:06:21,800
But here is the catch.

1915
01:06:21,800 --> 01:06:23,000
You're buying a fixed pool.

1916
01:06:23,000 --> 01:06:24,720
If you don't use those credits by the end of the year,

1917
01:06:24,720 --> 01:06:25,480
they expire.

1918
01:06:25,480 --> 01:06:26,320
They don't roll over.

1919
01:06:26,320 --> 01:06:27,280
They're just gone.

1920
01:06:27,280 --> 01:06:28,720
You are betting on your forecast.

1921
01:06:28,720 --> 01:06:30,280
Capacity packs sit right in the middle.

1922
01:06:30,280 --> 01:06:32,480
You buy a fixed allocation of interactions.

1923
01:06:32,480 --> 01:06:35,160
Usually 25,000 a month and it renews monthly.

1924
01:06:35,160 --> 01:06:36,600
It's more stable than pay as you go

1925
01:06:36,600 --> 01:06:39,480
but you don't get the deep discounts of a P3 commitment.

1926
01:06:39,480 --> 01:06:41,600
Most mature organizations use a hybrid approach.

1927
01:06:41,600 --> 01:06:43,680
They buy a capacity pack for their baseline.

1928
01:06:43,680 --> 01:06:45,720
The work they know they'll do every month.

1929
01:06:45,720 --> 01:06:49,000
They layer a P3 commitment on top for their expected growth.

1930
01:06:49,000 --> 01:06:50,480
Then they keep pay as you go,

1931
01:06:50,480 --> 01:06:52,200
enabled for the overflow.

1932
01:06:52,200 --> 01:06:54,680
When the baseline hits the cap, it pulls from P3.

1933
01:06:54,680 --> 01:06:56,720
When P3 is empty, it's built to pay as you go.

1934
01:06:56,720 --> 01:06:58,720
The economics work because the bulk of your work

1935
01:06:58,720 --> 01:06:59,960
is covered by discounts.

1936
01:06:59,960 --> 01:07:02,600
But you aren't penalized for growing faster than expected.

1937
01:07:02,600 --> 01:07:04,280
Negotiation is where your leverage lives.

1938
01:07:04,280 --> 01:07:06,440
Most organizations just accept the published pricing.

1939
01:07:06,440 --> 01:07:07,120
They shouldn't.

1940
01:07:07,120 --> 01:07:09,160
If you have 500 or more co-workers,

1941
01:07:09,160 --> 01:07:11,440
you are a material part of Microsoft's revenue

1942
01:07:11,440 --> 01:07:12,280
for this product.

1943
01:07:12,280 --> 01:07:13,760
They want your multi-year commitment

1944
01:07:13,760 --> 01:07:15,640
because it gives them predictability.

1945
01:07:15,640 --> 01:07:16,960
That is your negotiating room.

1946
01:07:16,960 --> 01:07:19,240
Benchmark your usage against the discount curve.

1947
01:07:19,240 --> 01:07:21,320
If you're forecasting 5 million credits a year,

1948
01:07:21,320 --> 01:07:22,600
you're already in discount territory,

1949
01:07:22,600 --> 01:07:26,320
but enterprise customers can often push for 25% off the list price.

1950
01:07:26,320 --> 01:07:28,160
Ask what the volume pricing looks like

1951
01:07:28,160 --> 01:07:30,960
and don't assume the first number they give you is the final one.

1952
01:07:30,960 --> 01:07:33,240
The contract terms matter just as much as the price.

1953
01:07:33,240 --> 01:07:34,520
You need to clarify what happens

1954
01:07:34,520 --> 01:07:36,360
if you go over your P3 commitment.

1955
01:07:36,360 --> 01:07:38,040
Does the price jump back to the list rate

1956
01:07:38,040 --> 01:07:39,920
or stay at your negotiated rate?

1957
01:07:39,920 --> 01:07:41,920
Ask if unused credits can carry forward

1958
01:07:41,920 --> 01:07:44,160
if you're showing a clear growth path.

1959
01:07:44,160 --> 01:07:45,840
You also need to know the renewal timeline

1960
01:07:45,840 --> 01:07:48,040
and if there's a minimum size for the next year.

1961
01:07:48,040 --> 01:07:51,080
Finally, remember that vendor lock in is real.

1962
01:07:51,080 --> 01:07:53,280
Co-pilot credits are specific to Microsoft.

1963
01:07:53,280 --> 01:07:55,200
If your entire business becomes dependent

1964
01:07:55,200 --> 01:07:58,720
on this specific model, your switching costs will be astronomical.

1965
01:07:58,720 --> 01:08:00,360
That doesn't mean you avoid the commitment

1966
01:08:00,360 --> 01:08:03,160
but it does mean you should build redundancy where it matters.

1967
01:08:03,160 --> 01:08:05,280
If a critical workflow depends entirely on co-work,

1968
01:08:05,280 --> 01:08:06,320
you're exposed.

1969
01:08:06,320 --> 01:08:08,760
If that same workflow could run on Azure OpenAI

1970
01:08:08,760 --> 01:08:11,560
or another tool if needed, you've protected yourself.

1971
01:08:11,560 --> 01:08:13,800
The procurement decision isn't just a financial one,

1972
01:08:13,800 --> 01:08:14,840
it's architectural.

1973
01:08:14,840 --> 01:08:18,400
What you buy determines what you can deploy

1974
01:08:18,400 --> 01:08:20,000
and how trapped you become, sir.

1975
01:08:20,000 --> 01:08:23,560
The implementation timeline, 90 days to August 2026.

1976
01:08:23,560 --> 01:08:24,720
The math is brutal.

1977
01:08:24,720 --> 01:08:27,120
You have roughly 90 days before the high risk rules

1978
01:08:27,120 --> 01:08:29,400
of the EUAI Act become enforceable.

1979
01:08:29,400 --> 01:08:32,240
That is not a long runway, but it is also not impossible.

1980
01:08:32,240 --> 01:08:34,160
If you move fast and stay disciplined.

1981
01:08:34,160 --> 01:08:35,560
Here is what the timeline looks like

1982
01:08:35,560 --> 01:08:37,040
when you execute it properly.

1983
01:08:37,040 --> 01:08:39,640
Weeks one and two, inventory and discovery.

1984
01:08:39,640 --> 01:08:42,120
You are cataloging everything, every power automate flow

1985
01:08:42,120 --> 01:08:44,560
with an AI connector, every co-pilot studio agent,

1986
01:08:44,560 --> 01:08:47,040
every custom integration calling OpenAI APIs,

1987
01:08:47,040 --> 01:08:49,880
every third party tool your teams adopted without asking.

1988
01:08:49,880 --> 01:08:52,360
You are not making judgments yet, you are collecting data.

1989
01:08:52,360 --> 01:08:55,000
Document the owner, the purpose, the data sources

1990
01:08:55,000 --> 01:08:56,480
and the tools it invokes.

1991
01:08:56,480 --> 01:08:59,000
This phase is exhausting because you are finding things

1992
01:08:59,000 --> 01:09:00,560
nobody knew existed.

1993
01:09:00,560 --> 01:09:01,600
And that is the point.

1994
01:09:01,600 --> 01:09:03,680
Shadow AI stops being shadow the moment you put it

1995
01:09:03,680 --> 01:09:04,600
in a spreadsheet.

1996
01:09:04,600 --> 01:09:07,240
Weeks three and four, risk classification.

1997
01:09:07,240 --> 01:09:10,760
Each agent gets evaluated against the EUAI Act categories.

1998
01:09:10,760 --> 01:09:12,800
Is it minimal risk, limited, high-risk,

1999
01:09:12,800 --> 01:09:14,560
prohibited, map it explicitly?

2000
01:09:14,560 --> 01:09:16,320
If an agent is screening drop candidates

2001
01:09:16,320 --> 01:09:18,600
or evaluating employees, it is high-risk

2002
01:09:18,600 --> 01:09:20,680
if it is just drafting meeting notes.

2003
01:09:20,680 --> 01:09:21,680
It is low-risk.

2004
01:09:21,680 --> 01:09:23,400
If it is handling customer communications,

2005
01:09:23,400 --> 01:09:25,880
it is limited risk and requires transparency.

2006
01:09:25,880 --> 01:09:27,360
You are building the risk register

2007
01:09:27,360 --> 01:09:29,400
that becomes the foundation for everything else.

2008
01:09:29,400 --> 01:09:30,360
Do not rush this.

2009
01:09:30,360 --> 01:09:32,840
Misclassification creates compliance gaps

2010
01:09:32,840 --> 01:09:35,200
that you cannot easily close later.

2011
01:09:35,200 --> 01:09:37,560
Weeks five through eight, governance design.

2012
01:09:37,560 --> 01:09:40,120
Define the structure, who makes agent approval decisions,

2013
01:09:40,120 --> 01:09:41,800
who owns the overall AI governance,

2014
01:09:41,800 --> 01:09:43,440
which roles handle the escalations,

2015
01:09:43,440 --> 01:09:45,520
what does the approval gate actually look like?

2016
01:09:45,520 --> 01:09:47,120
Design the monitoring dashboards.

2017
01:09:47,120 --> 01:09:49,680
Decide which metrics matter and how you will track them.

2018
01:09:49,680 --> 01:09:51,480
Define the technical controls.

2019
01:09:51,480 --> 01:09:53,040
What permissions will agents have?

2020
01:09:53,040 --> 01:09:55,440
What approvals are required before an agent takes action?

2021
01:09:55,440 --> 01:09:56,760
What are the kill switch procedures?

2022
01:09:56,760 --> 01:09:57,960
Write the policies.

2023
01:09:57,960 --> 01:09:59,440
But not as thick legal documents.

2024
01:09:59,440 --> 01:10:01,240
Write them as operating procedures.

2025
01:10:01,240 --> 01:10:02,680
How do we actually do this work?

2026
01:10:02,680 --> 01:10:06,480
Weeks nine through 12, technical implementation.

2027
01:10:06,480 --> 01:10:07,520
This is infrastructure.

2028
01:10:07,520 --> 01:10:10,040
Deploy agent 365 and configure EntraID

2029
01:10:10,040 --> 01:10:12,200
to treat agents as first class identities.

2030
01:10:12,200 --> 01:10:14,720
Configure DLP policies explicitly for co-pilot

2031
01:10:14,720 --> 01:10:16,080
and agent outputs.

2032
01:10:16,080 --> 01:10:19,320
Setup conditional access rules that enforce least privilege.

2033
01:10:19,320 --> 01:10:20,600
Build the logging infrastructure.

2034
01:10:20,600 --> 01:10:23,040
You are not using Microsoft Perview logs alone.

2035
01:10:23,040 --> 01:10:25,320
You are building a separate layer that captures

2036
01:10:25,320 --> 01:10:27,640
agent-specific metadata, which agent,

2037
01:10:27,640 --> 01:10:29,480
which model, which tools, which daughter,

2038
01:10:29,480 --> 01:10:30,840
and the decision reasoning.

2039
01:10:30,840 --> 01:10:33,200
This layer feeds directly into your audit

2040
01:10:33,200 --> 01:10:35,080
and compliance framework.

2041
01:10:35,080 --> 01:10:37,960
Weeks 13 through 16, pilot and validation.

2042
01:10:37,960 --> 01:10:39,240
You are not scaling yet.

2043
01:10:39,240 --> 01:10:40,880
You are proving the model works.

2044
01:10:40,880 --> 01:10:42,840
Select a small group of high-risk agents

2045
01:10:42,840 --> 01:10:44,400
or a focused use case.

2046
01:10:44,400 --> 01:10:46,920
Deploy into the governance environment you just built.

2047
01:10:46,920 --> 01:10:48,000
Monitor heavily.

2048
01:10:48,000 --> 01:10:50,040
Measure adoption rates are people actually using it.

2049
01:10:50,040 --> 01:10:50,840
Track errors.

2050
01:10:50,840 --> 01:10:52,640
Is the agent producing outputs people trust?

2051
01:10:52,640 --> 01:10:53,400
Measure outcomes.

2052
01:10:53,400 --> 01:10:55,120
Is the promised value materializing?

2053
01:10:55,120 --> 01:10:56,320
Document everything.

2054
01:10:56,320 --> 01:10:57,280
Every incident.

2055
01:10:57,280 --> 01:10:58,520
Every work around.

2056
01:10:58,520 --> 01:11:00,600
And every unexpected behavior becomes input

2057
01:11:00,600 --> 01:11:01,800
for the next phase.

2058
01:11:01,800 --> 01:11:03,560
Weeks 17 through 20.

2059
01:11:03,560 --> 01:11:04,480
Remediation.

2060
01:11:04,480 --> 01:11:07,040
The pilot revealed what did not work.

2061
01:11:07,040 --> 01:11:07,880
So fix it.

2062
01:11:07,880 --> 01:11:10,280
If the logging infrastructure is not capturing the right data,

2063
01:11:10,280 --> 01:11:11,120
rebuild it.

2064
01:11:11,120 --> 01:11:13,240
If the approval gates are slowing things down too much,

2065
01:11:13,240 --> 01:11:14,080
streamline them.

2066
01:11:14,080 --> 01:11:16,120
Teams do not understand the governance model.

2067
01:11:16,120 --> 01:11:17,200
Revise the training.

2068
01:11:17,200 --> 01:11:19,480
If certain agents are hitting edge cases,

2069
01:11:19,480 --> 01:11:21,040
update the decision logic.

2070
01:11:21,040 --> 01:11:22,840
Governance improves through iteration,

2071
01:11:22,840 --> 01:11:24,880
not through perfect design on day one.

2072
01:11:24,880 --> 01:11:27,880
Weeks 21 through 24 and beyond scale.

2073
01:11:27,880 --> 01:11:30,280
Now you are expanding agent deployments incrementally.

2074
01:11:30,280 --> 01:11:32,920
One new team per week, one new use case per sprint,

2075
01:11:32,920 --> 01:11:34,480
maintain the monitoring discipline.

2076
01:11:34,480 --> 01:11:36,680
Keep refining based on what you are seeing at scale.

2077
01:11:36,680 --> 01:11:39,280
Establish the quarterly review cycle, monthly reviews

2078
01:11:39,280 --> 01:11:40,480
of usage and cost.

2079
01:11:40,480 --> 01:11:43,200
Quarterly reassessment of risk classification.

2080
01:11:43,200 --> 01:11:45,080
Governance is not something you finish.

2081
01:11:45,080 --> 01:11:46,440
It is something you maintain.

2082
01:11:46,440 --> 01:11:49,840
By August 2nd, 2026, you need to be able to demonstrate compliance,

2083
01:11:49,840 --> 01:11:51,800
not just aspire to it, demonstrate it.

2084
01:11:51,800 --> 01:11:54,720
You have documented risk assessments for your high-risk agents.

2085
01:11:54,720 --> 01:11:57,440
You have audit trails, capturing decisions and approvals.

2086
01:11:57,440 --> 01:12:00,040
You have governance procedures that are actually being followed.

2087
01:12:00,040 --> 01:12:02,320
You have evidence that humans are meaningfully overseeing

2088
01:12:02,320 --> 01:12:03,080
the systems.

2089
01:12:03,080 --> 01:12:05,960
You have technical controls enforcing the boundaries you set.

2090
01:12:05,960 --> 01:12:07,640
You do not have perfect governance.

2091
01:12:07,640 --> 01:12:09,000
But you have working governance that is

2092
01:12:09,000 --> 01:12:10,440
auditable and defensible.

2093
01:12:10,440 --> 01:12:12,240
But here is the reality most organizations

2094
01:12:12,240 --> 01:12:13,880
do not want to hear.

2095
01:12:13,880 --> 01:12:16,920
The co-pilot credit trap is not the credits themselves.

2096
01:12:16,920 --> 01:12:19,240
It is the assumption that you can deploy autonomous agents

2097
01:12:19,240 --> 01:12:22,160
into an unprepared organization and just pay for what you use.

2098
01:12:22,160 --> 01:12:22,880
You cannot.

2099
01:12:22,880 --> 01:12:24,240
The structure does not support it.

2100
01:12:24,240 --> 01:12:25,840
Your data governance is weak.

2101
01:12:25,840 --> 01:12:27,200
Your identity model is broken.

2102
01:12:27,200 --> 01:12:28,840
Your audit trails do not exist.

2103
01:12:28,840 --> 01:12:30,440
Your cost model is guesswork.

2104
01:12:30,440 --> 01:12:33,560
By August 2026, if you have not built the governance layer,

2105
01:12:33,560 --> 01:12:34,880
you are not just overspending.

2106
01:12:34,880 --> 01:12:35,880
You are exposed.

2107
01:12:35,880 --> 01:12:38,640
Regulators will ask for evidence, and you will not have it.

2108
01:12:38,640 --> 01:12:39,480
Start now.

2109
01:12:39,480 --> 01:12:42,160
And if you want more of this, follow me on LinkedIn.