The Death of the UI: Why CUA is the End of SaaS as We Know It

1
00:00:00,000 --> 00:00:02,160
For two decades, we built enterprise software

2
00:00:02,160 --> 00:00:03,600
around a single assumption.

3
00:00:03,600 --> 00:00:07,080
Humans need graphical interfaces to navigate complexity.

4
00:00:07,080 --> 00:00:10,520
Buttons, forms, dashboards, navigation trees.

5
00:00:10,520 --> 00:00:12,840
That entire apparatus existed because the human brain

6
00:00:12,840 --> 00:00:14,400
cannot reason about data at scale

7
00:00:14,400 --> 00:00:15,800
without visual representation.

8
00:00:15,800 --> 00:00:16,680
It was necessary.

9
00:00:16,680 --> 00:00:18,040
It was real.

10
00:00:18,040 --> 00:00:19,800
That assumption is collapsing.

11
00:00:19,800 --> 00:00:22,240
Computer, using agents don't need dashboards.

12
00:00:22,240 --> 00:00:24,360
They don't need navigation menus or search filters

13
00:00:24,360 --> 00:00:25,600
or workflow builders.

14
00:00:25,600 --> 00:00:29,280
They need something different, logic, context, and stable APIs.

15
00:00:29,280 --> 00:00:31,040
They need to understand what they're looking at,

16
00:00:31,040 --> 00:00:33,120
decide what to do, and execute.

17
00:00:33,120 --> 00:00:34,560
No clicks required.

18
00:00:34,560 --> 00:00:37,080
And right now, your entire SAS valuation model

19
00:00:37,080 --> 00:00:38,520
is compressing because of it.

20
00:00:38,520 --> 00:00:41,080
The seat-based pricing model, where you multiply headcount

21
00:00:41,080 --> 00:00:43,360
by a monthly fee to calculate recurring revenue,

22
00:00:43,360 --> 00:00:45,960
is structurally incompatible with agent workflows.

23
00:00:45,960 --> 00:00:48,240
When one agent replaces 50 human seats,

24
00:00:48,240 --> 00:00:49,840
the vendor's revenue collapses.

25
00:00:49,840 --> 00:00:52,440
When that happens across the entire software industry,

26
00:00:52,440 --> 00:00:55,080
$300 billion in equity value evaporates,

27
00:00:55,080 --> 00:00:56,640
not because software is broken.

28
00:00:56,640 --> 00:00:59,120
But because the operating model underneath has shifted,

29
00:00:59,120 --> 00:01:00,680
this isn't about software disappearing.

30
00:01:00,680 --> 00:01:02,920
This is about the assumption that broke the moment agents

31
00:01:02,920 --> 00:01:04,600
became economically viable.

32
00:01:04,600 --> 00:01:06,040
By the end of this conversation,

33
00:01:06,040 --> 00:01:08,240
you'll understand why your current software investments

34
00:01:08,240 --> 00:01:10,520
are becoming bottlenecks instead of assets.

35
00:01:10,520 --> 00:01:12,320
You'll see where the real mode lives now,

36
00:01:12,320 --> 00:01:14,240
and you'll understand exactly what has to change

37
00:01:14,240 --> 00:01:16,360
if you want to compete in the next decade.

38
00:01:16,360 --> 00:01:18,320
The UI was always a workaround.

39
00:01:18,320 --> 00:01:21,640
The graphical interface solved a real problem in the 1980s.

40
00:01:21,640 --> 00:01:23,640
Before that, you interfaced with computers

41
00:01:23,640 --> 00:01:25,360
through command lines and punch cards.

42
00:01:25,360 --> 00:01:27,720
You had to translate your intent into machine syntax

43
00:01:27,720 --> 00:01:30,960
and memorize commands, which meant the cognitive load was immense

44
00:01:30,960 --> 00:01:32,200
when the GUI arrived.

45
00:01:32,200 --> 00:01:33,960
The desktop, the window, the mouse pointer.

46
00:01:33,960 --> 00:01:36,160
It was revolutionary because it compressed the gap

47
00:01:36,160 --> 00:01:37,360
between what you wanted to do

48
00:01:37,360 --> 00:01:39,360
and what you had to tell the computer to do.

49
00:01:39,360 --> 00:01:40,640
But it was always a workaround.

50
00:01:40,640 --> 00:01:42,760
It was a workaround for the fact that humans and machines

51
00:01:42,760 --> 00:01:44,600
reason about information differently.

52
00:01:44,600 --> 00:01:45,960
Humans think visually.

53
00:01:45,960 --> 00:01:48,080
We see a button and know it's clickable.

54
00:01:48,080 --> 00:01:49,560
We see a form and nowhere to type.

55
00:01:49,560 --> 00:01:51,480
We scan a dashboard and extract the pattern.

56
00:01:51,480 --> 00:01:52,600
Machines don't think that way.

57
00:01:52,600 --> 00:01:54,200
They process symbols and logic.

58
00:01:54,200 --> 00:01:56,040
Every SAS product built since then

59
00:01:56,040 --> 00:01:58,440
has been constructed on the same architecture.

60
00:01:58,440 --> 00:02:01,280
Database plus business logic plus a UI layer

61
00:02:01,280 --> 00:02:03,440
designed specifically for human interaction.

62
00:02:03,440 --> 00:02:05,880
Salesforce didn't invent a new paradigm.

63
00:02:05,880 --> 00:02:09,120
It took the three-tier architecture and made it cloud native.

64
00:02:09,120 --> 00:02:09,840
Same pattern.

65
00:02:09,840 --> 00:02:11,480
Humans click, machines respond.

66
00:02:11,480 --> 00:02:12,760
Humans interpret the response.

67
00:02:12,760 --> 00:02:14,120
That cycle repeats.

68
00:02:14,120 --> 00:02:17,360
The problem is that cycle has become the constraint,

69
00:02:17,360 --> 00:02:18,520
not the solution.

70
00:02:18,520 --> 00:02:20,960
When a human uses software, they're translating intent

71
00:02:20,960 --> 00:02:22,000
into clicks.

72
00:02:22,000 --> 00:02:24,120
You type in the search box, click the filter button,

73
00:02:24,120 --> 00:02:26,920
and wait for the page to load so you can scan the results.

74
00:02:26,920 --> 00:02:28,960
If you don't find what you need, you try again.

75
00:02:28,960 --> 00:02:31,560
That translation loop intent to interface, interface

76
00:02:31,560 --> 00:02:34,360
to action, action to result, result to interpretation

77
00:02:34,360 --> 00:02:35,720
is where latency lives.

78
00:02:35,720 --> 00:02:37,000
It's where errors accumulate.

79
00:02:37,000 --> 00:02:38,360
It's where cost hides.

80
00:02:38,360 --> 00:02:39,680
An agent doesn't operate that way.

81
00:02:39,680 --> 00:02:41,160
An agent reads data directly.

82
00:02:41,160 --> 00:02:43,840
It understands the structure without visual representation.

83
00:02:43,840 --> 00:02:45,640
It operates on the logic, not the interface.

84
00:02:45,640 --> 00:02:46,600
It doesn't wait.

85
00:02:46,600 --> 00:02:48,240
It doesn't misinterpret what it's looking at

86
00:02:48,240 --> 00:02:49,880
because the layout confused it.

87
00:02:49,880 --> 00:02:51,240
There's no translation loop.

88
00:02:51,240 --> 00:02:52,360
The dashboard was scaffolding.

89
00:02:52,360 --> 00:02:53,320
So is the form.

90
00:02:53,320 --> 00:02:55,480
The workflow builder, the drag and drop interface,

91
00:02:55,480 --> 00:02:57,880
and the admin panel were all tools designed

92
00:02:57,880 --> 00:03:00,680
to make software usable for humans who couldn't code.

93
00:03:00,680 --> 00:03:02,880
They worked because they mapped the underlying logic

94
00:03:02,880 --> 00:03:05,720
into visual metaphors that the human brain could grasp.

95
00:03:05,720 --> 00:03:07,120
But that scaffolding only exists

96
00:03:07,120 --> 00:03:09,320
because of a constraint, the human user.

97
00:03:09,320 --> 00:03:10,880
The moment you remove that constraint,

98
00:03:10,880 --> 00:03:13,320
the moment you have a system that can reason about data

99
00:03:13,320 --> 00:03:14,920
without visual representation.

100
00:03:14,920 --> 00:03:16,360
The scaffolding becomes dead weight.

101
00:03:16,360 --> 00:03:17,600
It slows things down.

102
00:03:17,600 --> 00:03:19,640
It adds latency, where there should be none.

103
00:03:19,640 --> 00:03:22,040
It introduces opacity, where there should be clarity.

104
00:03:22,040 --> 00:03:24,360
This is why the entire software industry is being forced

105
00:03:24,360 --> 00:03:26,600
to rebuild, not because software is broken,

106
00:03:26,600 --> 00:03:27,760
but because for the first time,

107
00:03:27,760 --> 00:03:30,640
there's a viable alternative to the human-driven UI model.

108
00:03:30,640 --> 00:03:33,280
And that alternative doesn't need any of the visual apparatus

109
00:03:33,280 --> 00:03:35,320
we've spent decades perfecting.

110
00:03:35,320 --> 00:03:37,640
Understanding why UI's exist is the first step

111
00:03:37,640 --> 00:03:40,880
to understanding why they're about to become irrelevant.

112
00:03:40,880 --> 00:03:43,200
The seat-based model is mathematically broken.

113
00:03:43,200 --> 00:03:45,080
The math has been simple for 20 years.

114
00:03:45,080 --> 00:03:46,720
You count the people using your software.

115
00:03:46,720 --> 00:03:48,560
You multiply that by a monthly fee.

116
00:03:48,560 --> 00:03:49,720
You add it all up for the year.

117
00:03:49,720 --> 00:03:51,640
That is your annual recurring revenue.

118
00:03:51,640 --> 00:03:53,360
For two decades, we build business models

119
00:03:53,360 --> 00:03:55,360
around predictable growth in that one number.

120
00:03:55,360 --> 00:03:56,560
You hire more sales reps.

121
00:03:56,560 --> 00:03:59,280
You build features that force people to add more users.

122
00:03:59,280 --> 00:04:02,080
The more headcount your customer has, the more they pay you.

123
00:04:02,080 --> 00:04:02,600
It is clean.

124
00:04:02,600 --> 00:04:03,600
It is predictable.

125
00:04:03,600 --> 00:04:05,000
Wall Street understands it.

126
00:04:05,000 --> 00:04:07,360
That formula worked because of one stable assumption.

127
00:04:07,360 --> 00:04:10,280
We assumed headcount growth was tied to business expansion.

128
00:04:10,280 --> 00:04:13,240
When a company hires new employees, they buy more licenses.

129
00:04:13,240 --> 00:04:15,920
When they open a department, they add more seats.

130
00:04:15,920 --> 00:04:18,400
The variable you price against, the number of humans,

131
00:04:18,400 --> 00:04:20,600
moves in the same direction as the organization.

132
00:04:20,600 --> 00:04:21,600
You can forecast it.

133
00:04:21,600 --> 00:04:22,600
You can model it.

134
00:04:22,600 --> 00:04:24,880
You can build a repeatable sales machine around it.

135
00:04:24,880 --> 00:04:26,040
Then agents show up.

136
00:04:26,040 --> 00:04:28,000
One agent handling invoice reconciliation

137
00:04:28,000 --> 00:04:30,680
can process what used to take a team of five people.

138
00:04:30,680 --> 00:04:31,760
That is not a projection.

139
00:04:31,760 --> 00:04:34,360
That is what we are seeing in pilot programs right now.

140
00:04:34,360 --> 00:04:35,280
One system.

141
00:04:35,280 --> 00:04:36,680
Five seats of value destroyed.

142
00:04:36,680 --> 00:04:38,320
The company that bought your software

143
00:04:38,320 --> 00:04:40,800
no longer needs five people for that function.

144
00:04:40,800 --> 00:04:42,840
They need one person to oversee the agent

145
00:04:42,840 --> 00:04:44,720
and one person to handle exceptions.

146
00:04:44,720 --> 00:04:46,480
Now calculate what happens to your revenue.

147
00:04:46,480 --> 00:04:47,920
The customer still needs your software,

148
00:04:47,920 --> 00:04:49,760
but they need four fewer licenses.

149
00:04:49,760 --> 00:04:52,400
That is an 80% reduction in what they pay you.

150
00:04:52,400 --> 00:04:54,400
If you have 100 customers and this pattern

151
00:04:54,400 --> 00:04:57,560
hits even half of them, your revenue collapses by 40%.

152
00:04:57,560 --> 00:05:01,080
Your retention metrics crater, your expansion revenue evaporates,

153
00:05:01,080 --> 00:05:02,520
your growth curve inverts.

154
00:05:02,520 --> 00:05:04,960
This is why traditional SaaS companies are trapped.

155
00:05:04,960 --> 00:05:06,320
If they do not adopt agents

156
00:05:06,320 --> 00:05:07,960
and integrate them into their platform,

157
00:05:07,960 --> 00:05:10,200
they lose market share to competitors who do.

158
00:05:10,200 --> 00:05:12,200
A vendor that builds agente capabilities

159
00:05:12,200 --> 00:05:14,000
into their product becomes more valuable

160
00:05:14,000 --> 00:05:15,160
than one that does not.

161
00:05:15,160 --> 00:05:16,200
Customers migrate.

162
00:05:16,200 --> 00:05:18,160
Market consolidation happens fast.

163
00:05:18,160 --> 00:05:20,520
The laggards get acquired or they simply disappear.

164
00:05:20,520 --> 00:05:22,000
But if they do adopt agents,

165
00:05:22,000 --> 00:05:24,080
they cannibalize their own revenue model.

166
00:05:24,080 --> 00:05:26,000
You cannot price a system the same way

167
00:05:26,000 --> 00:05:28,520
when the unit of value has shifted from the number of humans

168
00:05:28,520 --> 00:05:30,280
to the amount of work completed.

169
00:05:30,280 --> 00:05:31,600
Those are not the same thing.

170
00:05:31,600 --> 00:05:33,960
A company with 50 employees might process

171
00:05:33,960 --> 00:05:35,560
10,000 invoices a month.

172
00:05:35,560 --> 00:05:38,360
With agents, they might still process 10,000 invoices,

173
00:05:38,360 --> 00:05:39,840
but with 30 employees.

174
00:05:39,840 --> 00:05:42,360
Same output, 30% of the license revenue is gone.

175
00:05:42,360 --> 00:05:44,000
The market understands this dynamic

176
00:05:44,000 --> 00:05:46,280
and it is responding with brutal efficiency.

177
00:05:46,280 --> 00:05:48,840
SaaS multiples have compressed from 12 times revenue

178
00:05:48,840 --> 00:05:51,080
down to four times revenue across the board.

179
00:05:51,080 --> 00:05:53,480
This is not because individual products are performing worse

180
00:05:53,480 --> 00:05:55,480
or because customer satisfaction has declined.

181
00:05:55,480 --> 00:05:57,840
It is happening because investors are reprising

182
00:05:57,840 --> 00:06:00,760
the entire category based on the structural incompatibility

183
00:06:00,760 --> 00:06:03,600
between seat-based pricing and agente workflows.

184
00:06:03,600 --> 00:06:04,800
If you are a SaaS vendor

185
00:06:04,800 --> 00:06:07,080
and your revenue model is still based on headcount,

186
00:06:07,080 --> 00:06:09,400
your valuation is deteriorating in real time.

187
00:06:09,400 --> 00:06:11,480
Mid-market SaaS is getting hit hardest.

188
00:06:11,480 --> 00:06:13,080
Enterprise vendors have enough margin

189
00:06:13,080 --> 00:06:14,760
to absorb transition costs.

190
00:06:14,760 --> 00:06:16,640
They can experiment with usage-based pricing.

191
00:06:16,640 --> 00:06:18,360
They can layer on consumption models.

192
00:06:18,360 --> 00:06:20,760
They have enough cash flow to survive the reprising.

193
00:06:20,760 --> 00:06:23,520
But mid-market companies selling into mid-size organizations

194
00:06:23,520 --> 00:06:26,440
at 10,000 to $100,000 annual contract values

195
00:06:26,440 --> 00:06:27,560
do not have that cushion.

196
00:06:27,560 --> 00:06:29,600
They need growth to be clean and predictable.

197
00:06:29,600 --> 00:06:31,240
They cannot afford the margin compression

198
00:06:31,240 --> 00:06:33,080
that comes with agente integration.

199
00:06:33,080 --> 00:06:36,000
The vendors moving fastest are scrambling toward new models.

200
00:06:36,000 --> 00:06:37,840
They are looking at consumption-based pricing,

201
00:06:37,840 --> 00:06:39,720
outcome-based pricing and hybrid models

202
00:06:39,720 --> 00:06:41,200
that blend seats with usage.

203
00:06:41,200 --> 00:06:43,920
But these require completely different operational infrastructure.

204
00:06:43,920 --> 00:06:45,280
You need metering systems.

205
00:06:45,280 --> 00:06:47,360
You need billing that can handle variable charges.

206
00:06:47,360 --> 00:06:49,880
You need finance teams that can forecast revenue

207
00:06:49,880 --> 00:06:52,560
when it is no longer tied to a simple headcount multiple.

208
00:06:52,560 --> 00:06:54,440
Most of these systems do not exist yet.

209
00:06:54,440 --> 00:06:56,320
The reprising is not a temporary correction.

210
00:06:56,320 --> 00:06:57,760
It is a structural shift.

211
00:06:57,760 --> 00:07:00,160
The market has decided that the old model is obsolete.

212
00:07:00,160 --> 00:07:02,080
What is being priced in right now is the assumption

213
00:07:02,080 --> 00:07:04,120
that SaaS vendors will successfully transition

214
00:07:04,120 --> 00:07:06,440
to agente architectures and new pricing models.

215
00:07:06,440 --> 00:07:08,840
But that transition is harder than anyone anticipated

216
00:07:08,840 --> 00:07:11,280
and the cost of failure is extinction.

217
00:07:11,280 --> 00:07:12,600
This is why the economic pressure

218
00:07:12,600 --> 00:07:15,200
is forcing a fundamental shift in how software gets built

219
00:07:15,200 --> 00:07:16,000
in price.

220
00:07:16,000 --> 00:07:17,600
That shift starts with understanding

221
00:07:17,600 --> 00:07:19,920
what agents actually require.

222
00:07:19,920 --> 00:07:21,480
What agents actually require?

223
00:07:21,480 --> 00:07:23,560
So what does an agent actually need to do its job?

224
00:07:23,560 --> 00:07:24,600
It is not what you think.

225
00:07:24,600 --> 00:07:27,160
Your instinct is probably to say it needs a good UI

226
00:07:27,160 --> 00:07:28,560
or a clever interface.

227
00:07:28,560 --> 00:07:29,320
That is wrong.

228
00:07:29,320 --> 00:07:30,960
An agent does not need a dashboard.

229
00:07:30,960 --> 00:07:33,000
It does not care what color the buttons are.

230
00:07:33,000 --> 00:07:34,280
It needs something completely different.

231
00:07:34,280 --> 00:07:36,120
Stable, predictable APIs.

232
00:07:36,120 --> 00:07:37,320
It needs explicit logic.

233
00:07:37,320 --> 00:07:39,440
It needs clear boundaries on what it is allowed to do

234
00:07:39,440 --> 00:07:41,160
and what it is forbidden from doing.

235
00:07:41,160 --> 00:07:42,960
A dashboard is a visualization layer.

236
00:07:42,960 --> 00:07:46,240
It exists because humans need to see information to process it.

237
00:07:46,240 --> 00:07:49,320
An agent processes information without visualization.

238
00:07:49,320 --> 00:07:51,240
It can read data structures directly.

239
00:07:51,240 --> 00:07:53,600
It understands what a JSON object represents

240
00:07:53,600 --> 00:07:56,120
without needing a table or a chart or a graph.

241
00:07:56,120 --> 00:07:57,800
Give it a field called invoice amount

242
00:07:57,800 --> 00:07:59,600
and it knows it is dealing with a number.

243
00:07:59,600 --> 00:08:01,360
Give it a field called approval status

244
00:08:01,360 --> 00:08:03,920
and it understands the possible states without a legend.

245
00:08:03,920 --> 00:08:07,040
This distinction matters because it changes what usability means.

246
00:08:07,040 --> 00:08:09,480
For a human usability means the interface is intuitive.

247
00:08:09,480 --> 00:08:11,040
The controls are discoverable.

248
00:08:11,040 --> 00:08:12,840
The information architecture makes sense.

249
00:08:12,840 --> 00:08:15,320
For an agent usability means the API is stable.

250
00:08:15,320 --> 00:08:17,600
The scheme is consistent and the business logic

251
00:08:17,600 --> 00:08:19,720
is exposed as calbable operations.

252
00:08:19,720 --> 00:08:21,840
When we talk about computer using agents,

253
00:08:21,840 --> 00:08:24,320
we are talking about systems that can read a screen,

254
00:08:24,320 --> 00:08:26,720
interpret the visual layout and execute actions

255
00:08:26,720 --> 00:08:27,720
the way a human would.

256
00:08:27,720 --> 00:08:30,160
That is, CUA, it is powerful because it works

257
00:08:30,160 --> 00:08:32,280
with legacy systems that do not have APIs.

258
00:08:32,280 --> 00:08:33,800
But it is also a workaround.

259
00:08:33,800 --> 00:08:38,640
CUA succeeds on simple UI tasks, about 67% to 85% of the time.

260
00:08:38,640 --> 00:08:40,200
On complex multi-step workflows,

261
00:08:40,200 --> 00:08:42,080
it drops to 9 to 19%.

262
00:08:42,080 --> 00:08:44,480
That is not because the agent is not intelligent enough.

263
00:08:44,480 --> 00:08:46,800
It is because the agent is trying to infer intent

264
00:08:46,800 --> 00:08:49,960
from visual design instead of operating on explicit logic.

265
00:08:49,960 --> 00:08:52,760
CUA is friction masquerading as a solution.

266
00:08:52,760 --> 00:08:55,040
The real requirement is API first architecture.

267
00:08:55,040 --> 00:08:57,040
Every piece of business logic needs to be exposed

268
00:08:57,040 --> 00:08:58,040
as a calbable service.

269
00:08:58,040 --> 00:08:59,640
It cannot be hidden behind the UI.

270
00:08:59,640 --> 00:09:01,800
It cannot be locked in a workflow builder.

271
00:09:01,800 --> 00:09:04,160
It cannot be embedded in a proprietary data format.

272
00:09:04,160 --> 00:09:07,040
It must be exposed as an API that an agent can invoke,

273
00:09:07,040 --> 00:09:09,760
understand and chain together with other operations.

274
00:09:09,760 --> 00:09:11,200
But here is where it gets harder.

275
00:09:11,200 --> 00:09:13,520
An agent does not just need access to operations.

276
00:09:13,520 --> 00:09:14,440
It needs context.

277
00:09:14,440 --> 00:09:16,960
It needs to understand not just the data, but the rules,

278
00:09:16,960 --> 00:09:19,480
the constraints, the organizational knowledge

279
00:09:19,480 --> 00:09:21,640
that determines what actions are valid.

280
00:09:21,640 --> 00:09:24,160
When a financial agent is processing an expense report,

281
00:09:24,160 --> 00:09:27,080
it cannot just check if the amount is under $10,000.

282
00:09:27,080 --> 00:09:30,080
It needs to know that expenses over $5,000

283
00:09:30,080 --> 00:09:32,800
in the EMEA region require a different approval chain

284
00:09:32,800 --> 00:09:34,560
than expenses under $5,000.

285
00:09:34,560 --> 00:09:36,600
It needs to know that travel expenses get flagged

286
00:09:36,600 --> 00:09:38,440
differently than software subscriptions.

287
00:09:38,440 --> 00:09:40,280
It needs to know that the Requestors Department

288
00:09:40,280 --> 00:09:43,040
has a specific budget that is already 60% consumed.

289
00:09:43,040 --> 00:09:43,800
That is context.

290
00:09:43,800 --> 00:09:45,040
That is institutional logic.

291
00:09:45,040 --> 00:09:47,040
That is the difference between a generic agent

292
00:09:47,040 --> 00:09:49,120
and one that actually understands your business.

293
00:09:49,120 --> 00:09:50,680
An agent also needs governance.

294
00:09:50,680 --> 00:09:53,320
It needs clear boundaries on what a canon cannot do.

295
00:09:53,320 --> 00:09:55,440
It needs an audit trail for every decision.

296
00:09:55,440 --> 00:09:57,880
If an agent approves an expense, you need to know

297
00:09:57,880 --> 00:09:59,520
which agent made that decision.

298
00:09:59,520 --> 00:10:02,320
What data it was operating on, what policy it applied,

299
00:10:02,320 --> 00:10:02,920
and when.

300
00:10:02,920 --> 00:10:04,360
This is not for data analytics.

301
00:10:04,360 --> 00:10:06,320
It is for accountability.

302
00:10:06,320 --> 00:10:08,600
Because when something goes wrong and it will,

303
00:10:08,600 --> 00:10:09,960
you need to understand what happened

304
00:10:09,960 --> 00:10:11,120
and why the agent did it.

305
00:10:11,120 --> 00:10:13,480
Most critically, an agent needs to be treated

306
00:10:13,480 --> 00:10:15,440
as a first class identity in your system.

307
00:10:15,440 --> 00:10:16,360
It is not a script.

308
00:10:16,360 --> 00:10:17,720
It is not a background process.

309
00:10:17,720 --> 00:10:19,400
It is an entity with its own identity,

310
00:10:19,400 --> 00:10:21,400
its own permissions, and its own life cycle.

311
00:10:21,400 --> 00:10:23,280
It needs to be provisioned like a user.

312
00:10:23,280 --> 00:10:25,000
It needs to have its access managed.

313
00:10:25,000 --> 00:10:27,520
It needs to be deprovisioned when it is no longer needed.

314
00:10:27,520 --> 00:10:29,880
It needs audit logs that show what it did and when.

315
00:10:29,880 --> 00:10:31,680
It needs approval workflows that confirm

316
00:10:31,680 --> 00:10:34,040
it is authorized to perform sensitive operations.

317
00:10:34,040 --> 00:10:36,080
This is why Foundry Agent Service exists.

318
00:10:36,080 --> 00:10:37,280
It is not just a runtime.

319
00:10:37,280 --> 00:10:39,000
It is an orchestration layer that enforces

320
00:10:39,000 --> 00:10:41,160
these requirements at the platform level.

321
00:10:41,160 --> 00:10:45,680
Context management, governance, identity, observability.

322
00:10:45,680 --> 00:10:47,280
These are not nice to have features.

323
00:10:47,280 --> 00:10:48,960
They are foundational requirements.

324
00:10:48,960 --> 00:10:51,000
Once you understand what agents actually need,

325
00:10:51,000 --> 00:10:53,320
you start seeing how the current enterprise software stack

326
00:10:53,320 --> 00:10:56,880
is fundamentally misaligned with that reality.

327
00:10:56,880 --> 00:10:58,680
The API first imperative.

328
00:10:58,680 --> 00:11:00,960
Here is the uncomfortable truth about enterprise architecture

329
00:11:00,960 --> 00:11:01,800
today.

330
00:11:01,800 --> 00:11:05,080
82% of organizations claim they have an API first strategy.

331
00:11:05,080 --> 00:11:07,120
They say it in their CTO presentations.

332
00:11:07,120 --> 00:11:09,400
They put it in their modernization roadmaps.

333
00:11:09,400 --> 00:11:11,600
And they have likely spent millions on consultants

334
00:11:11,600 --> 00:11:13,720
who told them this approach is non-negotiable.

335
00:11:13,720 --> 00:11:16,040
But only 25% actually operate that way.

336
00:11:16,040 --> 00:11:17,760
The gap between what companies claim

337
00:11:17,760 --> 00:11:20,360
and what they actually do is where the real bottleneck lives.

338
00:11:20,360 --> 00:11:22,000
The reason is historical.

339
00:11:22,000 --> 00:11:24,840
We built software around human interaction for so long

340
00:11:24,840 --> 00:11:26,320
that the practice became embedded

341
00:11:26,320 --> 00:11:27,880
in how we think about architecture.

342
00:11:27,880 --> 00:11:30,320
When you design a system with a UI first mindset,

343
00:11:30,320 --> 00:11:32,480
the interface becomes the primary contract.

344
00:11:32,480 --> 00:11:33,840
The database lives at the bottom.

345
00:11:33,840 --> 00:11:35,600
The business logic lives in the middle.

346
00:11:35,600 --> 00:11:37,720
And then you eventually realize you need an API.

347
00:11:37,720 --> 00:11:40,120
So you build one, you expose whatever the UI needs,

348
00:11:40,120 --> 00:11:42,680
and make sure the API can support the same operations,

349
00:11:42,680 --> 00:11:43,680
the dashboard supports.

350
00:11:43,680 --> 00:11:44,520
You call it done.

351
00:11:44,520 --> 00:11:45,840
You think you are API first.

352
00:11:45,840 --> 00:11:48,960
But in reality, you are UI first with an API bolted on.

353
00:11:48,960 --> 00:11:50,800
The difference matters because it changes everything

354
00:11:50,800 --> 00:11:52,680
about how an agent works with your system.

355
00:11:52,680 --> 00:11:55,440
When the UI is primary and the API is secondary,

356
00:11:55,440 --> 00:11:58,240
the API is shaped by what humans find convenient rather

357
00:11:58,240 --> 00:11:59,200
than what machines need.

358
00:11:59,200 --> 00:12:01,920
You get thick endpoints that combine multiple operations

359
00:12:01,920 --> 00:12:04,440
and you get response formats that include visual hints

360
00:12:04,440 --> 00:12:06,920
and metadata that make sense in a table,

361
00:12:06,920 --> 00:12:08,840
but add noise in a service call.

362
00:12:08,840 --> 00:12:10,240
You get rate limits and pagination

363
00:12:10,240 --> 00:12:13,040
designed for human browsing instead of machine automation.

364
00:12:13,040 --> 00:12:16,480
But when you build API first, the UI becomes the afterthought.

365
00:12:16,480 --> 00:12:18,000
It is just another consumer.

366
00:12:18,000 --> 00:12:20,040
The API is designed around discreet,

367
00:12:20,040 --> 00:12:22,920
calibable operations where the business logic is exposed

368
00:12:22,920 --> 00:12:24,320
as atomic services.

369
00:12:24,320 --> 00:12:25,080
A tool is a tool.

370
00:12:25,080 --> 00:12:26,520
An operation is an operation.

371
00:12:26,520 --> 00:12:28,160
The UI consumes those operations

372
00:12:28,160 --> 00:12:30,080
and arranges them visually for humans,

373
00:12:30,080 --> 00:12:32,960
while another system, an agent or a workflow engine,

374
00:12:32,960 --> 00:12:36,680
consumes the same operations and chains them together logically.

375
00:12:36,680 --> 00:12:38,720
For agents to work at scale, this is the requirement.

376
00:12:38,720 --> 00:12:40,840
Every business capability needs to be atomized

377
00:12:40,840 --> 00:12:42,800
into discreet, calibable services.

378
00:12:42,800 --> 00:12:44,920
They cannot be hidden behind a dashboard

379
00:12:44,920 --> 00:12:46,640
or embedded in a workflow builder.

380
00:12:46,640 --> 00:12:48,200
They must be exposed as operations

381
00:12:48,200 --> 00:12:51,080
that an agent can invoke, understand, and chain.

382
00:12:51,080 --> 00:12:52,680
This is not just a technical change.

383
00:12:52,680 --> 00:12:54,320
It is an organizational change

384
00:12:54,320 --> 00:12:56,960
and it is harder than most enterprises realize.

385
00:12:56,960 --> 00:12:58,600
Your CRM is no longer just a system

386
00:12:58,600 --> 00:13:00,640
for humans to manage customer relationships.

387
00:13:00,640 --> 00:13:03,520
It is a service that agents can query for customer data,

388
00:13:03,520 --> 00:13:05,160
update with interaction history,

389
00:13:05,160 --> 00:13:07,160
and orchestrate across your entire workflow.

390
00:13:07,160 --> 00:13:09,920
Your ERP is not just a place where financial data lives,

391
00:13:09,920 --> 00:13:12,960
but an API that agents can call to make decisions,

392
00:13:12,960 --> 00:13:15,080
validate transactions, and execute them

393
00:13:15,080 --> 00:13:16,800
with full audit trails.

394
00:13:16,800 --> 00:13:18,840
The enterprises moving fastest are not the ones

395
00:13:18,840 --> 00:13:20,440
with the most advanced AI.

396
00:13:20,440 --> 00:13:23,600
They are the ones that already invested in deep API architecture.

397
00:13:23,600 --> 00:13:26,040
They did not do it because they were planning for agents,

398
00:13:26,040 --> 00:13:29,040
but because they needed APIs for mobile apps or integrations

399
00:13:29,040 --> 00:13:30,320
or external partners.

400
00:13:30,320 --> 00:13:32,680
Now those APIs exist and they can plug agents

401
00:13:32,680 --> 00:13:36,160
into existing services without rebuilding the entire stack,

402
00:13:36,160 --> 00:13:38,840
companies that are still UI first are facing a choice.

403
00:13:38,840 --> 00:13:40,320
Modernize your architecture now

404
00:13:40,320 --> 00:13:42,120
or watch your software become a constraint

405
00:13:42,120 --> 00:13:43,480
on what agents can do.

406
00:13:43,480 --> 00:13:45,280
The cost of waiting is accelerating.

407
00:13:45,280 --> 00:13:47,520
Every month without API first architecture

408
00:13:47,520 --> 00:13:49,560
is a month your competitors are gaining density

409
00:13:49,560 --> 00:13:51,000
in their agent deployments.

410
00:13:51,000 --> 00:13:52,600
They can build new agents faster

411
00:13:52,600 --> 00:13:55,120
because the underlying services are already exposed

412
00:13:55,120 --> 00:13:56,520
and they can iterate more quickly

413
00:13:56,520 --> 00:13:59,520
because they are not fighting the friction of UI first design.

414
00:13:59,520 --> 00:14:01,600
They can scale automation to more use cases

415
00:14:01,600 --> 00:14:03,240
because the architecture supports it.

416
00:14:03,240 --> 00:14:05,920
This is why API first has become a competitive necessity

417
00:14:05,920 --> 00:14:07,280
rather than a preference.

418
00:14:07,280 --> 00:14:08,520
It is not about being trendy.

419
00:14:08,520 --> 00:14:10,160
It is because agents require it

420
00:14:10,160 --> 00:14:11,840
and agents are becoming the operating model

421
00:14:11,840 --> 00:14:13,200
of enterprise software.

422
00:14:13,200 --> 00:14:15,800
The next question is, once you have built the architecture

423
00:14:15,800 --> 00:14:17,680
who actually owns the advantage,

424
00:14:17,680 --> 00:14:19,520
the answer is simpler than you would think,

425
00:14:19,520 --> 00:14:21,080
but it requires a new way of thinking

426
00:14:21,080 --> 00:14:23,640
about what makes a company defensible.

427
00:14:23,640 --> 00:14:25,800
Workflow Capital is your real mode.

428
00:14:25,800 --> 00:14:28,760
There is a concept that has become the defining strategic insight

429
00:14:28,760 --> 00:14:31,640
for enterprises thinking about what actually matters anymore.

430
00:14:31,640 --> 00:14:33,760
I've only introduced it and once you understand it,

431
00:14:33,760 --> 00:14:36,080
you cannot unsee how everything in enterprise software

432
00:14:36,080 --> 00:14:38,520
is being reorganized around this single idea.

433
00:14:38,520 --> 00:14:40,240
It is called Workflow Capital.

434
00:14:40,240 --> 00:14:42,000
Workflow Capital is not software.

435
00:14:42,000 --> 00:14:43,200
It is not a feature set

436
00:14:43,200 --> 00:14:45,000
and it is not something you can buy.

437
00:14:45,000 --> 00:14:46,640
It is the accumulated operational logic

438
00:14:46,640 --> 00:14:48,200
that makes your company work.

439
00:14:48,200 --> 00:14:49,400
It is the unwritten rules

440
00:14:49,400 --> 00:14:51,320
and the edge cases everyone knows about

441
00:14:51,320 --> 00:14:52,800
but nobody documented.

442
00:14:52,800 --> 00:14:54,280
It includes the decision criteria

443
00:14:54,280 --> 00:14:56,360
that separates a good approval from a bad one,

444
00:14:56,360 --> 00:14:58,040
the approval chains and the sequences

445
00:14:58,040 --> 00:15:00,480
of who talks to whom and in what order.

446
00:15:00,480 --> 00:15:01,880
It is the informal workarounds

447
00:15:01,880 --> 00:15:03,840
that actually make the formal process function

448
00:15:03,840 --> 00:15:06,360
and the judgment calls that experienced employees make

449
00:15:06,360 --> 00:15:07,880
without even thinking about it.

450
00:15:07,880 --> 00:15:10,320
It is how you use the software, not the software itself.

451
00:15:10,320 --> 00:15:12,440
Here is where it gets concrete, take two banks.

452
00:15:12,440 --> 00:15:14,400
Both run the same CRM with the same vendor,

453
00:15:14,400 --> 00:15:16,360
the same features and the same interface,

454
00:15:16,360 --> 00:15:18,080
but they have completely different workflows

455
00:15:18,080 --> 00:15:20,200
because they have different risk tolerances,

456
00:15:20,200 --> 00:15:21,440
different customer segments

457
00:15:21,440 --> 00:15:23,680
and different regulatory constraints.

458
00:15:23,680 --> 00:15:25,680
One bank might approve a commercial credit line

459
00:15:25,680 --> 00:15:29,080
at the branch level if the amount is under $100,000

460
00:15:29,080 --> 00:15:31,320
while the other requires regional approval

461
00:15:31,320 --> 00:15:33,720
for anything over $50,000.

462
00:15:33,720 --> 00:15:35,640
One bank flags international transactions

463
00:15:35,640 --> 00:15:37,560
for enhanced due diligence immediately

464
00:15:37,560 --> 00:15:39,560
but the other has a more granular risk model

465
00:15:39,560 --> 00:15:41,080
that factors in the customer history

466
00:15:41,080 --> 00:15:42,960
and counterparty reputation.

467
00:15:42,960 --> 00:15:44,680
One bank routes small business lending

468
00:15:44,680 --> 00:15:47,640
through a different approval chain than commercial lending

469
00:15:47,640 --> 00:15:50,120
while the other uses the same process for everything.

470
00:15:50,120 --> 00:15:52,000
Both banks are running the same software

471
00:15:52,000 --> 00:15:54,360
but their workflow capital is completely different

472
00:15:54,360 --> 00:15:57,080
and that difference is what actually drives their competitive advantage.

473
00:15:57,080 --> 00:15:59,240
It is not the CRM, it is the way they use it.

474
00:15:59,240 --> 00:16:01,240
In a WorldWare agent's execute workflows,

475
00:16:01,240 --> 00:16:02,480
that shifts everything.

476
00:16:02,480 --> 00:16:05,080
An agent that understands your specific risk model,

477
00:16:05,080 --> 00:16:07,400
your approval hierarchy, your exception handling

478
00:16:07,400 --> 00:16:08,880
and your decision heuristics

479
00:16:08,880 --> 00:16:11,360
is worth infinitely more than a generic agent

480
00:16:11,360 --> 00:16:13,160
because the agent is not just executing,

481
00:16:13,160 --> 00:16:14,520
it is embodying your logic,

482
00:16:14,520 --> 00:16:17,320
it is making decisions the way your organization makes decisions,

483
00:16:17,320 --> 00:16:19,320
handling exceptions the way you handle them

484
00:16:19,320 --> 00:16:21,720
and escalating the right things to the right people.

485
00:16:21,720 --> 00:16:24,680
It is following your playbook instead of some vendor template

486
00:16:24,680 --> 00:16:26,040
but here is where it gets dangerous.

487
00:16:26,040 --> 00:16:29,520
If you let vendors train on your workflow data to make their agents better,

488
00:16:29,520 --> 00:16:32,480
you are teaching them how to replicate your competitive advantage,

489
00:16:32,480 --> 00:16:34,000
you are handing them the playbook.

490
00:16:34,000 --> 00:16:35,560
The vendor sees your approval patterns,

491
00:16:35,560 --> 00:16:37,880
your risk thresholds and your decision criteria.

492
00:16:37,880 --> 00:16:39,320
They use that to train their models,

493
00:16:39,320 --> 00:16:40,800
they make their agents smarter

494
00:16:40,800 --> 00:16:43,320
and then they sell that intelligence to your competitors.

495
00:16:43,320 --> 00:16:45,320
You have just commoditized your secret source.

496
00:16:45,320 --> 00:16:46,640
This is happening right now.

497
00:16:46,640 --> 00:16:50,200
Vendors are asking customers for access to their real work artifacts like email,

498
00:16:50,200 --> 00:16:52,400
documents, meeting transcripts and transaction logs.

499
00:16:52,400 --> 00:16:56,200
They want this data to train agents that work better in real office environments.

500
00:16:56,200 --> 00:16:59,320
What they are actually doing is learning how your organization works

501
00:16:59,320 --> 00:17:02,120
so they can productize that knowledge and sell it to everyone else.

502
00:17:02,120 --> 00:17:05,920
This is why organizations need to be deliberate about what workflow data they share

503
00:17:05,920 --> 00:17:06,960
with external vendors.

504
00:17:06,960 --> 00:17:11,280
The enterprises that will win in the agentic era are not the ones buying generic SaaS tools.

505
00:17:11,280 --> 00:17:15,480
They are the ones encoding their workflow capital into private or controlled agents.

506
00:17:15,480 --> 00:17:17,840
These agents reflect how they actually do business.

507
00:17:17,840 --> 00:17:21,440
They embody institutional knowledge and competitors cannot replicate them

508
00:17:21,440 --> 00:17:24,520
because the logic is proprietary, the context is internal

509
00:17:24,520 --> 00:17:26,760
and the training data never left the company.

510
00:17:26,760 --> 00:17:28,320
This requires a fundamental shift.

511
00:17:28,320 --> 00:17:32,200
We have to move from buying software to building agents that embody our logic.

512
00:17:32,200 --> 00:17:37,240
We have to move from configuring the vendor workflow to coding our workflow into our agents.

513
00:17:37,240 --> 00:17:39,840
We have to stop treating workflow as configuration

514
00:17:39,840 --> 00:17:42,040
and start treating it as intellectual property

515
00:17:42,040 --> 00:17:44,720
that needs to be protected and compounded over time.

516
00:17:44,720 --> 00:17:46,400
It is a different operating model entirely.

517
00:17:46,400 --> 00:17:49,040
Most software companies today optimize for feature parity.

518
00:17:49,040 --> 00:17:52,840
They ask what their competitors offer and they build that plus a little more.

519
00:17:52,840 --> 00:17:54,640
It is an arms race of functionality.

520
00:17:54,640 --> 00:17:57,680
But in an agentic world, that race becomes irrelevant.

521
00:17:57,680 --> 00:18:00,320
The features are fast to copy, but the workflows are not.

522
00:18:00,320 --> 00:18:05,320
An agent trained on generic best practices performs worse than an agent trained on your specific logic.

523
00:18:05,320 --> 00:18:08,760
An agent that embodies your risk model, your customer segmentation

524
00:18:08,760 --> 00:18:11,000
and your operational heuristics is defensible.

525
00:18:11,000 --> 00:18:11,840
That is durable.

526
00:18:11,840 --> 00:18:12,760
That is your mode.

527
00:18:12,760 --> 00:18:15,640
The organizations that understand this are already moving.

528
00:18:15,640 --> 00:18:19,360
They are inventorying their workflow capital and asking what makes them unique.

529
00:18:19,360 --> 00:18:22,880
They are looking for the operational logic they have that competitors do not,

530
00:18:22,880 --> 00:18:27,200
where their decisions are different and what their experience people know that they could encode.

531
00:18:27,200 --> 00:18:29,600
They are treating that as their most valuable asset.

532
00:18:29,600 --> 00:18:32,720
Protecting and compounding workflow capital requires a governance model

533
00:18:32,720 --> 00:18:34,600
that most enterprises do not have yet.

534
00:18:34,600 --> 00:18:36,480
The non-human identity problem.

535
00:18:36,480 --> 00:18:40,040
For decades, enterprise identity management was a solved problem.

536
00:18:40,040 --> 00:18:45,160
You have employees, you provision them when they joined, you manage their permissions as they move around.

537
00:18:45,160 --> 00:18:50,040
You deprovision them when they leave platforms like Entry, D and Octa were built on this one assumption.

538
00:18:50,040 --> 00:18:52,760
Everything flows from the fact that you are managing humans.

539
00:18:52,760 --> 00:18:54,560
Humans have names and email addresses.

540
00:18:54,560 --> 00:18:56,800
They belong to departments and report to managers.

541
00:18:56,800 --> 00:18:57,800
They have tenure.

542
00:18:57,800 --> 00:19:01,200
Because humans are predictable, you can write policies around their behavior.

543
00:19:01,200 --> 00:19:04,840
But agents break this model completely, an agent isn't a user.

544
00:19:04,840 --> 00:19:07,520
It doesn't have an email address or a performance review.

545
00:19:07,520 --> 00:19:12,200
But it's also not a service account. In the old model, a service account was the answer for non-human access.

546
00:19:12,200 --> 00:19:13,960
You'd create one account and give it a password.

547
00:19:13,960 --> 00:19:16,000
It lived forever. It had broad permissions.

548
00:19:16,000 --> 00:19:18,080
So it could do everything it might ever need.

549
00:19:18,080 --> 00:19:20,880
Multiple applications or scripts would share that same account.

550
00:19:20,880 --> 00:19:24,240
It was convenient and simple, but at scale, it's a security nightmare.

551
00:19:24,240 --> 00:19:27,520
The problem is that service accounts are too rigid and too privileged.

552
00:19:27,520 --> 00:19:32,600
A service account created in 2015 might still have access to systems that don't even exist anymore.

553
00:19:32,600 --> 00:19:35,440
It might hold permissions for operations it hasn't performed in years.

554
00:19:35,440 --> 00:19:38,640
If that account is compromised, an attacker has access to everything.

555
00:19:38,640 --> 00:19:41,520
And they have all the time in the world to figure out what that is.

556
00:19:41,520 --> 00:19:46,480
There is no audit trail of which application used it or what specific operation happened.

557
00:19:46,480 --> 00:19:49,560
It's just a blob of standing privilege sitting in your infrastructure.

558
00:19:49,560 --> 00:19:50,560
An agent is different.

559
00:19:50,560 --> 00:19:54,480
An agent needs to perform a specific operation for a specific period of time.

560
00:19:54,480 --> 00:19:56,000
It doesn't need broad permissions.

561
00:19:56,000 --> 00:19:59,760
It needs minimal permissions just enough to do the job and only while it's doing it.

562
00:19:59,760 --> 00:20:01,360
This is the non-human identity problem.

563
00:20:01,360 --> 00:20:05,120
How do you govern an entity that acts on its own and makes its own decisions?

564
00:20:05,120 --> 00:20:08,960
The traditional answer of giving it a service account and hoping for the best doesn't work.

565
00:20:08,960 --> 00:20:12,640
You end up with hundreds of accounts with no clear owner and no audit trail.

566
00:20:12,640 --> 00:20:14,800
Microsoft's answer is "entraagent id".

567
00:20:14,800 --> 00:20:17,520
Each agent gets a unique identity just like a human user.

568
00:20:17,520 --> 00:20:21,600
That identity can be provisioned and governed using the same frameworks you already use.

569
00:20:21,600 --> 00:20:25,040
You can assign it to groups or apply conditional access policies.

570
00:20:25,040 --> 00:20:29,280
You can monitor it for suspicious activity and revoke access the moment the agent is done.

571
00:20:29,280 --> 00:20:30,720
But here's the critical difference.

572
00:20:30,720 --> 00:20:33,680
Agents should operate on the principle of zero standing privilege.

573
00:20:33,680 --> 00:20:38,720
This means an agent only has access to the specific resources it needs for the task it's doing right now.

574
00:20:38,720 --> 00:20:40,240
Not everything it might ever need.

575
00:20:40,240 --> 00:20:41,840
Not everything that would be convenient.

576
00:20:41,840 --> 00:20:43,520
Just what it's doing in this moment.

577
00:20:43,520 --> 00:20:46,480
After the task is complete, the access should be revoked.

578
00:20:46,480 --> 00:20:49,120
This requires just-in-time access and continuous monitoring.

579
00:20:49,120 --> 00:20:53,040
You need the ability to say that an agent can read customer data from a specific database,

580
00:20:53,040 --> 00:20:55,600
but only if a specific business process invokes it.

581
00:20:55,600 --> 00:20:57,440
And only for the next 30 minutes.

582
00:20:57,440 --> 00:21:00,800
Most enterprises don't have this level of granularity in their systems.

583
00:21:00,800 --> 00:21:04,080
They were built to assign permissions to a role that stays active all day.

584
00:21:04,080 --> 00:21:07,920
You need granularity measured in minutes and tied to specific data contexts.

585
00:21:07,920 --> 00:21:09,840
This is not how traditional platforms work.

586
00:21:09,840 --> 00:21:12,400
It's not how your current infrastructure was designed.

587
00:21:12,400 --> 00:21:13,680
They're going to have to build it.

588
00:21:13,680 --> 00:21:15,600
And that's where agent 365 comes in.

589
00:21:15,600 --> 00:21:18,480
Where agent 365 is the control plane.

590
00:21:18,480 --> 00:21:22,160
Agent 365 reached general availability in May of 2026.

591
00:21:22,160 --> 00:21:25,600
It's the first enterprise product designed to solve one problem.

592
00:21:25,600 --> 00:21:28,560
How do you govern autonomous agents at an organizational scale?

593
00:21:28,560 --> 00:21:30,080
It's not a tool for building agents.

594
00:21:30,080 --> 00:21:31,360
It's a tool for governing them.

595
00:21:31,360 --> 00:21:32,160
Think of it this way.

596
00:21:32,160 --> 00:21:33,760
Agents are emerging everywhere right now.

597
00:21:33,760 --> 00:21:36,000
You have copilot studio agents, foundry agents,

598
00:21:36,000 --> 00:21:38,480
and custom ones built by your engineering teams.

599
00:21:38,480 --> 00:21:40,400
Some are authorized, some are experiments,

600
00:21:40,400 --> 00:21:42,720
and some were built six months ago and forgotten.

601
00:21:42,720 --> 00:21:45,760
Without visibility into what's running, you have a governance crisis.

602
00:21:45,760 --> 00:21:49,360
In agent 365, every agent in your tenant gets registered.

603
00:21:49,360 --> 00:21:50,800
They are classified and monitored.

604
00:21:50,800 --> 00:21:52,080
You can see who owns them,

605
00:21:52,080 --> 00:21:54,640
what systems they touch, and how often they run.

606
00:21:54,640 --> 00:21:56,480
This matters because without that visibility,

607
00:21:56,480 --> 00:21:59,120
you end up with shadow agents across your infrastructure.

608
00:21:59,120 --> 00:22:00,880
An engineer might spin up a custom agent

609
00:22:00,880 --> 00:22:02,960
to automate their workflow without telling anyone.

610
00:22:02,960 --> 00:22:06,880
Suddenly, an unmanaged entity is querying customer data in your CRM.

611
00:22:06,880 --> 00:22:09,680
You have no idea it exists because nobody reported it.

612
00:22:09,680 --> 00:22:12,000
Agent 365 forces that into the light.

613
00:22:12,000 --> 00:22:13,760
Visibility is just the foundation though.

614
00:22:13,760 --> 00:22:15,440
The real power is in the governance layer

615
00:22:15,440 --> 00:22:17,360
because each agent has an intra-ID,

616
00:22:17,360 --> 00:22:20,880
you can apply conditional access policies like you would for a human.

617
00:22:20,880 --> 00:22:24,400
But these policies are measured against what the agent is actually doing.

618
00:22:24,400 --> 00:22:25,920
You can set a capability boundary,

619
00:22:25,920 --> 00:22:28,640
stating an agent can read customer data but not write it.

620
00:22:28,640 --> 00:22:30,000
You can set a temporal boundary,

621
00:22:30,000 --> 00:22:33,680
meaning the agent only has access from 9am to 5pm.

622
00:22:33,680 --> 00:22:36,160
You can even set a data classification boundary.

623
00:22:36,160 --> 00:22:38,960
Per view, sensitivity labels are enforced at the platform level,

624
00:22:38,960 --> 00:22:40,240
not in the agent code.

625
00:22:40,240 --> 00:22:41,680
If the agent tries to touch something,

626
00:22:41,680 --> 00:22:42,480
it shouldn't.

627
00:22:42,480 --> 00:22:45,120
The platform blocks it before the request even executes.

628
00:22:45,120 --> 00:22:46,640
You can also set a control boundary.

629
00:22:46,640 --> 00:22:48,640
This means an agent might require human approval

630
00:22:48,640 --> 00:22:50,560
before it can execute a financial transaction.

631
00:22:50,560 --> 00:22:53,120
It can draft the transaction and validate the data,

632
00:22:53,120 --> 00:22:55,360
but then it stops and waits for a person to review it.

633
00:22:55,360 --> 00:22:58,160
These policies are enforced as hard stops at the platform level.

634
00:22:58,160 --> 00:23:00,000
They aren't just guardrails or recommendations.

635
00:23:00,000 --> 00:23:02,960
The agent cannot do the thing because the platform won't let it.

636
00:23:02,960 --> 00:23:04,800
That's the difference between hoping agents behave

637
00:23:04,800 --> 00:23:06,480
and actually ensuring they do.

638
00:23:06,480 --> 00:23:09,760
Agent 365 integrates with the rest of your security stack.

639
00:23:09,760 --> 00:23:11,360
Per view handles the data governance

640
00:23:11,360 --> 00:23:13,360
while defender watches for threat signals.

641
00:23:13,360 --> 00:23:15,360
If an agent starts accessing resources,

642
00:23:15,360 --> 00:23:17,520
it doesn't normally touch, you get an alert.

643
00:23:17,520 --> 00:23:19,040
Then there's the observability piece.

644
00:23:19,040 --> 00:23:22,080
You can audit every decision, every tool call, and every interaction.

645
00:23:22,080 --> 00:23:25,600
You can replay a conversation to understand exactly what the agent did and why.

646
00:23:25,600 --> 00:23:26,800
When something goes wrong,

647
00:23:26,800 --> 00:23:29,760
you can see the data it was using and the policy it applied.

648
00:23:29,760 --> 00:23:33,440
This helps you figure out if there's a genuine problem or just a false positive.

649
00:23:33,440 --> 00:23:36,000
The licensing model is per user, not per agent.

650
00:23:36,000 --> 00:23:38,720
You aren't paying $100 for every agent you create.

651
00:23:38,720 --> 00:23:41,520
Instead, you pay for the governance capability itself.

652
00:23:41,520 --> 00:23:44,480
Everyone in your organization gets access to manage agents

653
00:23:44,480 --> 00:23:46,080
under a single license pool.

654
00:23:46,080 --> 00:23:49,280
This aligns the incentives because you aren't penalized for having hundreds of agents.

655
00:23:49,280 --> 00:23:52,080
You're simply paying for the ability to govern them effectively.

656
00:23:52,080 --> 00:23:55,360
The more agents you have, the more critical this infrastructure becomes.

657
00:23:55,360 --> 00:23:56,960
The pricing reflects that reality.

658
00:23:56,960 --> 00:24:00,880
This is how you move from shadow agents everywhere to governed agents everywhere.

659
00:24:00,880 --> 00:24:03,120
The platform becomes the final word on what's allowed.

660
00:24:03,120 --> 00:24:05,600
Foundry agent service as the runtime.

661
00:24:05,600 --> 00:24:07,520
Agent 365 handles the governance.

662
00:24:07,520 --> 00:24:10,240
It defines the rules and keeps your agents within the lines.

663
00:24:10,240 --> 00:24:13,920
But governance without a place to run is just a set of rules with no game.

664
00:24:13,920 --> 00:24:16,160
You need a home for these agents to actually do the work.

665
00:24:16,160 --> 00:24:18,000
That home is the Foundry agent service.

666
00:24:18,000 --> 00:24:20,240
Foundry is the managed platform from Microsoft,

667
00:24:20,240 --> 00:24:22,560
built for running agents at a massive scale.

668
00:24:22,560 --> 00:24:23,920
It is not co-pilot studio.

669
00:24:23,920 --> 00:24:27,120
That's a tool for building chatbots that answer questions in teams.

670
00:24:27,120 --> 00:24:28,720
It is also not Azure Functions.

671
00:24:28,720 --> 00:24:32,080
That's just a place to run snippets of code when something triggers them.

672
00:24:32,080 --> 00:24:34,960
Foundry is different because it was built for one specific purpose.

673
00:24:34,960 --> 00:24:37,760
It runs agents that need to manage complex workflows.

674
00:24:37,760 --> 00:24:39,760
Remember what happened in previous steps

675
00:24:39,760 --> 00:24:41,680
and talked to multiple systems at once.

676
00:24:41,680 --> 00:24:43,360
The way it works is actually very simple.

677
00:24:43,360 --> 00:24:44,960
You start by defining your agent.

678
00:24:44,960 --> 00:24:47,040
You write the prompt that gives it instructions.

679
00:24:47,040 --> 00:24:49,600
You give it tools so it can interact with the world.

680
00:24:49,600 --> 00:24:51,040
You connect it to knowledge sources.

681
00:24:51,040 --> 00:24:52,240
So it has the right context.

682
00:24:52,240 --> 00:24:54,000
Once that's done, you have an agent.

683
00:24:54,000 --> 00:24:56,320
Foundry takes care of the rest of the infrastructure.

684
00:24:56,320 --> 00:24:58,720
The tools are the most important part of the setup.

685
00:24:58,720 --> 00:25:01,520
A tool is how your agent reaches out and touches another system.

686
00:25:01,520 --> 00:25:03,440
It might be an API call to your CRM,

687
00:25:03,440 --> 00:25:04,720
a quick database query,

688
00:25:04,720 --> 00:25:06,560
or even a command to move a file.

689
00:25:06,560 --> 00:25:09,280
You define what these tools are and what they can do.

690
00:25:09,280 --> 00:25:11,680
The agent learns what is available in its toolbox.

691
00:25:11,680 --> 00:25:12,960
When it needs to get a job done,

692
00:25:12,960 --> 00:25:14,960
it picks the right tool and uses it.

693
00:25:14,960 --> 00:25:16,880
Foundry manages the actual connection,

694
00:25:16,880 --> 00:25:18,800
handles the errors if something breaks,

695
00:25:18,800 --> 00:25:20,800
and keeps a log of every single move.

696
00:25:20,800 --> 00:25:23,200
One of the biggest shifts here is how Foundry handles

697
00:25:23,200 --> 00:25:24,960
multiple agents working together.

698
00:25:24,960 --> 00:25:28,240
You aren't stuck with one giant agent trying to do everything.

699
00:25:28,240 --> 00:25:29,920
Instead, you can build an orchestrator.

700
00:25:29,920 --> 00:25:31,600
This is a lead agent that takes a request

701
00:25:31,600 --> 00:25:33,200
and figures out where it needs to go.

702
00:25:33,200 --> 00:25:35,440
If the task needs a deep dive into data,

703
00:25:35,440 --> 00:25:37,120
it sends it to the retrieval agent.

704
00:25:37,120 --> 00:25:38,560
If it needs to check a legal rule,

705
00:25:38,560 --> 00:25:40,320
it hands it off to the policy agent.

706
00:25:40,320 --> 00:25:43,040
If it needs to actually move money or update a record,

707
00:25:43,040 --> 00:25:44,640
it calls the action agent.

708
00:25:44,640 --> 00:25:46,240
Foundry manages the handoffs

709
00:25:46,240 --> 00:25:48,240
and makes sure the context stays consistent

710
00:25:48,240 --> 00:25:49,760
as the task moves between them.

711
00:25:49,760 --> 00:25:52,160
This is how you actually scale an AI strategy.

712
00:25:52,160 --> 00:25:53,920
You don't try to make one agent smarter.

713
00:25:53,920 --> 00:25:56,960
You build a team of specialists that are great at one narrow thing

714
00:25:56,960 --> 00:25:59,360
and you let an orchestrator coordinate the workflow.

715
00:25:59,360 --> 00:26:02,720
The pricing model also shows how much the value of software has changed.

716
00:26:02,720 --> 00:26:04,320
You aren't paying for seats anymore.

717
00:26:04,320 --> 00:26:06,480
You are paying for what the agent actually does.

718
00:26:06,480 --> 00:26:08,400
Foundry charges for compute by the hour

719
00:26:08,400 --> 00:26:10,640
based on the resources your agent containers use.

720
00:26:10,640 --> 00:26:12,880
You pay for tokens based on the model you choose.

721
00:26:12,880 --> 00:26:15,120
If you use GPT-4, you pay those rates.

722
00:26:15,120 --> 00:26:17,920
If you use a smaller, cheaper model, your bill goes down.

723
00:26:17,920 --> 00:26:19,360
Even the tools are meted.

724
00:26:19,360 --> 00:26:22,000
If your agent calls an API a thousand times,

725
00:26:22,000 --> 00:26:23,600
you pay for a thousand calls.

726
00:26:23,600 --> 00:26:24,960
This is a true consumption model.

727
00:26:24,960 --> 00:26:27,840
It aligns what the vendor wants with what the customer needs.

728
00:26:27,840 --> 00:26:29,520
In the old world of seat-based pricing,

729
00:26:29,520 --> 00:26:31,200
efficiency didn't really matter.

730
00:26:31,200 --> 00:26:32,560
Now, if your agent is wasteful,

731
00:26:32,560 --> 00:26:35,040
if it uses an expensive model for a tiny task

732
00:26:35,040 --> 00:26:37,040
or makes the same API call over and over,

733
00:26:37,040 --> 00:26:38,560
you see that cost immediately,

734
00:26:38,560 --> 00:26:40,560
this creates a real pressure to optimize.

735
00:26:40,560 --> 00:26:41,920
The agent has to be smart.

736
00:26:41,920 --> 00:26:44,560
It needs to cache data and pick the right model for the right job.

737
00:26:44,560 --> 00:26:45,920
These aren't just technical goals.

738
00:26:45,920 --> 00:26:47,280
They are financial ones.

739
00:26:47,280 --> 00:26:49,680
That alignment changes how everyone behaves.

740
00:26:49,680 --> 00:26:51,680
Vendors want things to be efficient

741
00:26:51,680 --> 00:26:53,600
because it saves them support costs.

742
00:26:53,600 --> 00:26:56,400
Customers want efficiency because it saves them hard cache.

743
00:26:56,400 --> 00:26:58,000
Everyone is finally on the same team.

744
00:26:58,000 --> 00:27:00,240
This is the future of the operating model.

745
00:27:00,240 --> 00:27:02,320
It isn't software that humans have to click through.

746
00:27:02,320 --> 00:27:04,400
It is infrastructure that agents live in.

747
00:27:04,400 --> 00:27:05,920
It is managed, it is metered,

748
00:27:05,920 --> 00:27:07,760
and it is priced based on the outcome.

749
00:27:07,760 --> 00:27:09,520
The runtime is where the work happens,

750
00:27:09,520 --> 00:27:11,200
but agents still have to talk to systems

751
00:27:11,200 --> 00:27:12,640
that were never meant for them.

752
00:27:12,640 --> 00:27:14,080
They have to deal with old software

753
00:27:14,080 --> 00:27:16,240
and apps that don't have a modern interface.

754
00:27:16,240 --> 00:27:17,920
That is where the bridge comes in.

755
00:27:17,920 --> 00:27:20,080
Computer using agents as the bridge.

756
00:27:20,080 --> 00:27:23,440
The reality is that not every system in your company has an API.

757
00:27:23,440 --> 00:27:26,640
This is the part of the job that people hate to talk about in meetings.

758
00:27:26,640 --> 00:27:28,240
Your old mainframe doesn't have an API

759
00:27:28,240 --> 00:27:30,160
because it was built in the 1970s.

760
00:27:30,160 --> 00:27:31,920
Your custom desktop app doesn't have one

761
00:27:31,920 --> 00:27:33,760
because nobody ever got around to building it.

762
00:27:33,760 --> 00:27:36,320
Even your modern SaaS tools might have limited APIs

763
00:27:36,320 --> 00:27:39,120
that don't let you touch the specific workflow you actually need.

764
00:27:39,120 --> 00:27:40,960
You can see the data, but you can't pull the lever.

765
00:27:40,960 --> 00:27:41,840
You're stuck.

766
00:27:41,840 --> 00:27:44,320
You cannot wait for every system to have a perfect API

767
00:27:44,320 --> 00:27:45,920
before you start using agents.

768
00:27:45,920 --> 00:27:47,920
If you wait for that, you will be waiting forever.

769
00:27:47,920 --> 00:27:50,080
Some of these systems are never going to change.

770
00:27:50,080 --> 00:27:53,280
They are in maintenance mode and it's too expensive to rebuild them

771
00:27:53,280 --> 00:27:54,960
just so a machine can talk to them.

772
00:27:54,960 --> 00:27:56,160
You need a different way in.

773
00:27:56,160 --> 00:27:57,360
You need a way to work with systems

774
00:27:57,360 --> 00:27:59,360
that only have a screen meant for humans.

775
00:27:59,360 --> 00:28:02,400
This is why we have computer using agents or CUAs.

776
00:28:02,400 --> 00:28:05,520
A CUAs is an AI model that looks at a screen exactly like you do.

777
00:28:05,520 --> 00:28:06,400
It sees the buttons.

778
00:28:06,400 --> 00:28:08,080
It recognizes the text boxes.

779
00:28:08,080 --> 00:28:10,000
It understands how a table is laid out.

780
00:28:10,000 --> 00:28:12,160
It knows that if a message pops up in the corner,

781
00:28:12,160 --> 00:28:14,000
it's probably a confirmation of the button

782
00:28:14,000 --> 00:28:16,080
it just clicked. It can then take action.

783
00:28:16,080 --> 00:28:17,680
It clicks, it types,

784
00:28:17,680 --> 00:28:19,200
and it navigates through a process

785
00:28:19,200 --> 00:28:21,280
by chaining those visual steps together.

786
00:28:21,280 --> 00:28:22,880
It isn't the fastest way to work

787
00:28:22,880 --> 00:28:24,320
and it isn't the most elegant.

788
00:28:24,320 --> 00:28:26,000
But it works when nothing else does.

789
00:28:26,000 --> 00:28:29,200
This tech actually started as a way to help people with vision loss.

790
00:28:29,200 --> 00:28:32,320
The goal was to have an AI describe what was happening on a screen.

791
00:28:32,320 --> 00:28:35,520
It turns out that if an AI is smart enough to describe a screen,

792
00:28:35,520 --> 00:28:36,960
it is smart enough to use it.

793
00:28:36,960 --> 00:28:38,960
That mix of computer vision and understanding

794
00:28:38,960 --> 00:28:40,880
is what makes this automation possible.

795
00:28:40,880 --> 00:28:41,680
But there is a catch.

796
00:28:41,680 --> 00:28:44,080
Recent research shows that CUA is a bridge,

797
00:28:44,080 --> 00:28:45,600
not the final destination.

798
00:28:45,600 --> 00:28:47,200
When you give a CUA a simple task,

799
00:28:47,200 --> 00:28:49,360
like filling out a single form or reading a table,

800
00:28:49,360 --> 00:28:50,320
it does pretty well.

801
00:28:50,320 --> 00:28:53,440
The success rate is usually between 67 and 85%.

802
00:28:53,440 --> 00:28:54,640
That is good enough to be useful.

803
00:28:54,640 --> 00:28:56,320
But when the task gets complex,

804
00:28:56,320 --> 00:28:58,800
when the agent has to remember something from three screens ago

805
00:28:58,800 --> 00:29:00,560
or handle an unexpected pop-up,

806
00:29:00,560 --> 00:29:03,600
the success rate drops to between 9 and 19%.

807
00:29:03,600 --> 00:29:06,960
That is nowhere near good enough for a critical business process.

808
00:29:06,960 --> 00:29:09,120
That gap tells us exactly what CUA is.

809
00:29:09,120 --> 00:29:10,240
It's a pattern matcher.

810
00:29:10,240 --> 00:29:11,760
It works when things stay the same.

811
00:29:11,760 --> 00:29:13,280
It breaks when things get complicated.

812
00:29:13,280 --> 00:29:16,160
The long term goal is still to have APIs for everything.

813
00:29:16,160 --> 00:29:19,920
But for right now, CUA stops you from being blocked by your oldest software.

814
00:29:19,920 --> 00:29:24,000
This is why Microsoft includes CUA as a tool in their responses API.

815
00:29:24,000 --> 00:29:27,600
Your agents in Foundry can call on a CUA whenever they hit a wall.

816
00:29:27,600 --> 00:29:30,160
You don't have to build a special CUA agent as...

817
00:29:30,160 --> 00:29:31,600
It's just another tool in the box.

818
00:29:31,600 --> 00:29:33,440
When the agent can use an API, it does.

819
00:29:33,440 --> 00:29:35,760
When it hits a legacy system with no other way in,

820
00:29:35,760 --> 00:29:37,120
it switches to CUA.

821
00:29:37,120 --> 00:29:39,040
Foundry tracks both parts in the same lock.

822
00:29:39,040 --> 00:29:42,800
You don't have to fix your entire technical debt before you start using AI.

823
00:29:42,800 --> 00:29:45,040
You can start with the systems that are ready right now.

824
00:29:45,040 --> 00:29:46,480
Use Foundry to coordinate them.

825
00:29:46,480 --> 00:29:47,920
Use CUA to fill in the hole.

826
00:29:47,920 --> 00:29:51,680
You can build one agent that talks to a modern database through an API

827
00:29:51,680 --> 00:29:55,280
and then jumps into an old desktop app using CUA to finish the job.

828
00:29:55,280 --> 00:29:57,280
As you eventually modernize those old systems,

829
00:29:57,280 --> 00:29:59,760
you just swap the CUA tool for a direct API.

830
00:29:59,760 --> 00:30:01,120
The workflow stays the same.

831
00:30:01,120 --> 00:30:02,800
You just made it faster and more reliable.

832
00:30:02,800 --> 00:30:04,800
This is how you move forward one step at a time

833
00:30:04,800 --> 00:30:06,720
instead of waiting for a perfect world.

834
00:30:06,720 --> 00:30:10,560
CUA is important because it kills the excuse that you can't use agents yet.

835
00:30:10,560 --> 00:30:12,240
You can, you just have to build the bridge.

836
00:30:12,240 --> 00:30:14,800
The path is right there and you can start on it today.

837
00:30:14,800 --> 00:30:17,040
But as these agents start moving through your systems,

838
00:30:17,040 --> 00:30:18,320
you have to know what they are doing.

839
00:30:18,320 --> 00:30:19,760
You need to see where they are going

840
00:30:19,760 --> 00:30:21,600
and make sure they aren't making mistakes.

841
00:30:21,600 --> 00:30:24,160
That brings us to the problem of observability.

842
00:30:24,160 --> 00:30:26,000
The safety and observability layer.

843
00:30:26,000 --> 00:30:28,080
When an agent starts acting on its own,

844
00:30:28,080 --> 00:30:31,200
it creates a problem that traditional software never had to deal with.

845
00:30:31,200 --> 00:30:33,360
You have to see what it is doing while it is happening

846
00:30:33,360 --> 00:30:35,840
and you have to prove exactly what it did after the fact.

847
00:30:35,840 --> 00:30:37,120
This is not just monitoring.

848
00:30:37,120 --> 00:30:38,000
Monitoring is basic.

849
00:30:38,000 --> 00:30:39,600
It tells you if the power is on,

850
00:30:39,600 --> 00:30:42,160
if the system is slow or if the code crashed.

851
00:30:42,160 --> 00:30:45,120
But observability for an agent goes much deeper than that.

852
00:30:45,120 --> 00:30:47,360
You need to see every single decision the agent made

853
00:30:47,360 --> 00:30:49,280
and the specific data that led to that choice.

854
00:30:49,280 --> 00:30:51,840
You need to know why it picked one tool over another.

855
00:30:51,840 --> 00:30:53,680
When things go wrong and they will,

856
00:30:53,680 --> 00:30:56,240
you have to understand the logic that led to the mistake,

857
00:30:56,240 --> 00:30:58,560
not just the fact that an error popped up on a dashboard.

858
00:30:58,560 --> 00:31:02,240
Foundry builds this tracing directly into the system.

859
00:31:02,240 --> 00:31:04,720
Every time a tool is called the system logs it.

860
00:31:04,720 --> 00:31:07,360
Every time the agent makes a choice, the record is saved.

861
00:31:07,360 --> 00:31:09,200
You do not have to build your own logging system

862
00:31:09,200 --> 00:31:11,200
because the runtime handles it for you.

863
00:31:11,200 --> 00:31:13,600
You can open up a full trace of how the agent thinks

864
00:31:13,600 --> 00:31:16,000
and see the exact order of every action it took.

865
00:31:16,000 --> 00:31:18,080
You see what data came back from a database

866
00:31:18,080 --> 00:31:20,640
and how the agent used that information to reach a conclusion.

867
00:31:20,640 --> 00:31:22,720
This creates a clear chain of cause and effect

868
00:31:22,720 --> 00:31:24,800
from the first question to the final result.

869
00:31:24,800 --> 00:31:26,800
This is the only way to handle compliance.

870
00:31:26,800 --> 00:31:29,200
If an agent approves a loan or moves money around,

871
00:31:29,200 --> 00:31:31,520
a regulator is going to ask how that happened.

872
00:31:31,520 --> 00:31:32,800
They want a trail of evidence,

873
00:31:32,800 --> 00:31:34,560
they want to see which rules were followed

874
00:31:34,560 --> 00:31:36,000
and which data points were checked.

875
00:31:36,000 --> 00:31:38,640
You cannot just tell them that the AI made a choice.

876
00:31:38,640 --> 00:31:39,600
You have to explain it.

877
00:31:39,600 --> 00:31:42,560
Foundry gives you that explanation in a format you can actually use.

878
00:31:42,560 --> 00:31:44,400
But seeing what happened is only the start.

879
00:31:44,400 --> 00:31:45,920
The real shift is in the guardrails.

880
00:31:45,920 --> 00:31:48,560
An agent should never move money without a human saying yes.

881
00:31:48,560 --> 00:31:50,000
It is not a technical limitation.

882
00:31:50,000 --> 00:31:52,240
It is because a mistake costs real dollars.

883
00:31:52,240 --> 00:31:55,360
An agent should never delete a file without a confirmation

884
00:31:55,360 --> 00:31:57,360
because you cannot undo that action.

885
00:31:57,360 --> 00:31:58,720
It should never touch data.

886
00:31:58,720 --> 00:32:00,080
It isn't allowed to see.

887
00:32:00,080 --> 00:32:01,760
Because that is how secrets get out.

888
00:32:01,760 --> 00:32:03,520
You cannot put these rules in the agent code.

889
00:32:03,520 --> 00:32:06,480
If you just put a line in a prompt saying do not delete data,

890
00:32:06,480 --> 00:32:08,960
the agent might find a creative way to ignore you.

891
00:32:08,960 --> 00:32:11,120
It might follow the words but break the intent.

892
00:32:11,120 --> 00:32:13,360
You need the platform to step in and stop it.

893
00:32:13,360 --> 00:32:16,320
Foundry connects directly to your data rules through purview.

894
00:32:16,320 --> 00:32:17,840
If an agent tries to grab a file,

895
00:32:17,840 --> 00:32:20,800
mark the secret, the platform kills the request before it even starts.

896
00:32:20,800 --> 00:32:22,480
The agent never even gets a look at the data.

897
00:32:22,480 --> 00:32:24,320
The boundary is enforced at the gate.

898
00:32:24,320 --> 00:32:26,800
Now you have a system where you can see what the agent tried to do

899
00:32:26,800 --> 00:32:28,640
and exactly how the platform blocked it.

900
00:32:28,640 --> 00:32:30,240
Then you have defender watching for patterns.

901
00:32:30,240 --> 00:32:33,360
If an agent suddenly starts asking for resources it never used before,

902
00:32:33,360 --> 00:32:34,560
you get an alert.

903
00:32:34,560 --> 00:32:36,480
If it starts working 10 times faster than usual

904
00:32:36,480 --> 00:32:38,720
or logs in at 3am from a weird location,

905
00:32:38,720 --> 00:32:40,000
the system flags it.

906
00:32:40,000 --> 00:32:42,240
These shifts might mean the agent is compromised

907
00:32:42,240 --> 00:32:43,920
or they might just mean the model is drifting.

908
00:32:43,920 --> 00:32:46,640
Either way you see it happening in real time so you can step in.

909
00:32:46,640 --> 00:32:49,280
You also have to test for safety before you go live.

910
00:32:49,280 --> 00:32:51,600
The assert framework lets you run scenarios

911
00:32:51,600 --> 00:32:54,240
to see how the agent acts in a safe environment.

912
00:32:54,240 --> 00:32:56,000
You can throw weird requests at it

913
00:32:56,000 --> 00:32:57,760
or try to trick it with bad data.

914
00:32:57,760 --> 00:32:59,680
This matters because agents can surprise you.

915
00:32:59,680 --> 00:33:01,440
A model might look at all data and decide

916
00:33:01,440 --> 00:33:03,200
that certain customers are always a risk.

917
00:33:03,200 --> 00:33:04,960
Even when that logic is outdated,

918
00:33:04,960 --> 00:33:06,800
it might find a loophole in your instructions

919
00:33:06,800 --> 00:33:08,080
that you never saw coming.

920
00:33:08,080 --> 00:33:10,960
Testing catches those gaps before they turn into a headline.

921
00:33:10,960 --> 00:33:13,920
These layers of safety inside are what make agents ready for work.

922
00:33:13,920 --> 00:33:16,560
Without them, you are just making the same old mistakes.

923
00:33:16,560 --> 00:33:19,280
Only you are doing it faster and at a much higher scale.

924
00:33:19,280 --> 00:33:21,520
The organizational shift.

925
00:33:21,520 --> 00:33:23,360
Running agents at scale takes a set of skills

926
00:33:23,360 --> 00:33:25,760
that most software teams just do not have yet.

927
00:33:25,760 --> 00:33:27,520
This is where most companies get stuck.

928
00:33:27,520 --> 00:33:30,400
You need people who treat prompt engineering like a science.

929
00:33:30,400 --> 00:33:32,080
Not a trick they found on the internet.

930
00:33:32,080 --> 00:33:34,080
They need to know how to structure instructions

931
00:33:34,080 --> 00:33:36,400
so the behavior is the same every single time.

932
00:33:36,400 --> 00:33:38,080
They have to know how to fix a prompt

933
00:33:38,080 --> 00:33:39,760
when the output is slightly off.

934
00:33:39,760 --> 00:33:41,760
They have to tell the difference between an agent

935
00:33:41,760 --> 00:33:43,760
that actually understands a task

936
00:33:43,760 --> 00:33:46,240
and one that is just guessing based on a pattern.

937
00:33:46,240 --> 00:33:48,800
You also need people who understand data at a very deep level.

938
00:33:48,800 --> 00:33:51,360
It is no longer enough to just say who can see a folder.

939
00:33:51,360 --> 00:33:54,480
You have to know what specific context an agent needs to finish a job.

940
00:33:54,480 --> 00:33:55,680
You have to label data

941
00:33:55,680 --> 00:33:57,760
so the agent knows where the boundaries are.

942
00:33:57,760 --> 00:34:00,400
If you don't, the agent might accidentally leak a secret

943
00:34:00,400 --> 00:34:02,720
while it is explaining its reasoning to a user.

944
00:34:02,720 --> 00:34:04,000
Then there is the business logic.

945
00:34:04,000 --> 00:34:05,920
You need people who can take your company rules

946
00:34:05,920 --> 00:34:08,240
and turn them into something a machine can follow.

947
00:34:08,240 --> 00:34:10,400
This isn't about dragging boxes in a workflow tool.

948
00:34:10,400 --> 00:34:12,000
It is about taking your heuristics

949
00:34:12,000 --> 00:34:13,120
and your exception rules

950
00:34:13,120 --> 00:34:15,440
and turning them into prompts and policies.

951
00:34:15,440 --> 00:34:17,520
Most companies do not have these people on staff.

952
00:34:17,520 --> 00:34:19,440
They are going to have to train them or find them

953
00:34:19,440 --> 00:34:21,040
and this is not a small project.

954
00:34:21,040 --> 00:34:23,120
This is a change to the DNA of the company.

955
00:34:23,120 --> 00:34:25,920
It takes time that most leaders think they don't have.

956
00:34:25,920 --> 00:34:27,680
The reason this is so hard is that

957
00:34:27,680 --> 00:34:30,080
the old way of building software is dead.

958
00:34:30,080 --> 00:34:31,920
You cannot just write a list of requirements,

959
00:34:31,920 --> 00:34:33,840
give them to a coder and wait for the app.

960
00:34:33,840 --> 00:34:36,080
In the old world, the developer makes the choices

961
00:34:36,080 --> 00:34:37,360
and writes the code.

962
00:34:37,360 --> 00:34:38,880
If it is wrong, you fix the code.

963
00:34:38,880 --> 00:34:40,480
Agents do not work like that.

964
00:34:40,480 --> 00:34:42,240
Agents are trained and guided.

965
00:34:42,240 --> 00:34:44,160
You write a prompt, you give it some tools

966
00:34:44,160 --> 00:34:46,000
and you test it with real info.

967
00:34:46,000 --> 00:34:47,520
If the model does something weird,

968
00:34:47,520 --> 00:34:49,520
you don't rewrite a thousand lines of code.

969
00:34:49,520 --> 00:34:51,360
You refine the prompt and try again.

970
00:34:51,360 --> 00:34:52,800
You are iterating on behavior.

971
00:34:52,800 --> 00:34:54,240
You are watching how the model thinks

972
00:34:54,240 --> 00:34:56,640
and making tweaks based on what you see.

973
00:34:56,640 --> 00:34:58,720
This feels more like data science than engineering.

974
00:34:58,720 --> 00:35:00,320
You are looking for patterns in the failures

975
00:35:00,320 --> 00:35:02,080
and tuning the system based on results.

976
00:35:02,080 --> 00:35:03,760
But it isn't pure data science either.

977
00:35:03,760 --> 00:35:05,680
You aren't building a model from scratch.

978
00:35:05,680 --> 00:35:07,600
You are taking a massive foundation model

979
00:35:07,600 --> 00:35:10,000
and steering it with guardrails and fine tuning.

980
00:35:10,000 --> 00:35:12,640
It is a new kind of job that doesn't even have a real name yet.

981
00:35:12,640 --> 00:35:14,560
The way companies are organizing to handle this

982
00:35:14,560 --> 00:35:15,920
is through a center of excellence.

983
00:35:15,920 --> 00:35:17,360
One central team owns the platform.

984
00:35:17,360 --> 00:35:19,680
They manage Foundry and Agent 365.

985
00:35:19,680 --> 00:35:22,640
They set the big rules for what is allowed and what is banned.

986
00:35:22,640 --> 00:35:25,520
They provide the templates so everyone isn't starting from zero.

987
00:35:25,520 --> 00:35:28,320
But the actual business units own the specific agents.

988
00:35:28,320 --> 00:35:31,360
The people in sales or HR are the ones who define the workflows.

989
00:35:31,360 --> 00:35:33,600
They are the ones who understand their own processes

990
00:35:33,600 --> 00:35:35,600
well enough to teach them to an agent.

991
00:35:35,600 --> 00:35:38,160
They are the ones who keep the quality high as the business changes.

992
00:35:38,160 --> 00:35:40,240
This creates a balance between control and speed.

993
00:35:40,240 --> 00:35:42,960
You aren't trying to build one giant agent

994
00:35:42,960 --> 00:35:44,560
that does everything for everyone.

995
00:35:44,560 --> 00:35:46,880
You are building a platform that lets you run hundreds

996
00:35:46,880 --> 00:35:49,200
of small specialized agents.

997
00:35:49,200 --> 00:35:51,360
Each one is built for one specific job

998
00:35:51,360 --> 00:35:52,960
in one specific department.

999
00:35:52,960 --> 00:35:55,760
The central team just makes sure all of them are safe and easy to watch.

1000
00:35:55,760 --> 00:35:58,320
This is a massive change in how big companies work.

1001
00:35:58,320 --> 00:36:00,960
Usually the IT department controls every single thing.

1002
00:36:00,960 --> 00:36:03,360
They make the choices and everyone else waits in line.

1003
00:36:03,360 --> 00:36:06,160
If you want to change, you put in a ticket and hope for the best.

1004
00:36:06,160 --> 00:36:08,160
In the agent model, ownership is spread out.

1005
00:36:08,160 --> 00:36:10,000
The business teams are the ones actually building.

1006
00:36:10,000 --> 00:36:11,760
They define what the agent does and they decide

1007
00:36:11,760 --> 00:36:13,520
if it is working well enough to use.

1008
00:36:13,520 --> 00:36:16,800
The central IT team sets the fences, but they don't drive the car.

1009
00:36:16,800 --> 00:36:19,760
This takes a level of trust that a lot of companies just haven't built yet.

1010
00:36:19,760 --> 00:36:22,960
It requires very clear rules so people know what they can and cannot do.

1011
00:36:22,960 --> 00:36:25,360
It requires tools that are easy enough for a business user

1012
00:36:25,360 --> 00:36:26,960
to use without breaking a policy.

1013
00:36:26,960 --> 00:36:29,600
The way you organize your team determines what you can build,

1014
00:36:29,600 --> 00:36:32,560
but it also determines exactly how much risk you are taking on.

1015
00:36:32,560 --> 00:36:34,800
Measuring agent value and ROI.

1016
00:36:34,800 --> 00:36:38,800
Calculating ROI for traditional SaaS is almost insultingly simple.

1017
00:36:38,800 --> 00:36:39,760
You count your users.

1018
00:36:39,760 --> 00:36:41,280
You multiply by the license fee.

1019
00:36:41,280 --> 00:36:44,320
You compare that to productivity gains and divide by the cost.

1020
00:36:44,320 --> 00:36:47,840
If you save 500 hours a year and each hour costs $40 in labor,

1021
00:36:47,840 --> 00:36:49,920
you've generated $20,000 in benefit.

1022
00:36:49,920 --> 00:36:53,040
If the software costs $5,000, you get a 4-to-1 return.

1023
00:36:53,040 --> 00:36:55,760
It's a rough calculation, but it works for most tools.

1024
00:36:55,760 --> 00:36:58,480
With agents, that calculation explodes in complexity.

1025
00:36:58,480 --> 00:37:01,040
The problem is the fundamental unit of measurement.

1026
00:37:01,040 --> 00:37:01,760
It's wrong.

1027
00:37:01,760 --> 00:37:03,440
An agent doesn't have a cost per user.

1028
00:37:03,440 --> 00:37:04,960
It has a cost per execution.

1029
00:37:04,960 --> 00:37:08,800
That shift matters because the volume of work changes the entire equation.

1030
00:37:08,800 --> 00:37:11,600
You aren't measuring productivity gains per person anymore.

1031
00:37:11,600 --> 00:37:13,760
You're measuring productivity gains per task.

1032
00:37:13,760 --> 00:37:15,680
You aren't dividing benefit by headcount.

1033
00:37:15,680 --> 00:37:17,200
You're dividing it by workload.

1034
00:37:17,200 --> 00:37:20,240
To get a real answer, you have to ask different questions.

1035
00:37:20,240 --> 00:37:22,320
How many tasks did the agent actually finish?

1036
00:37:22,320 --> 00:37:25,280
How much time did it save compared to a human doing it manually?

1037
00:37:25,280 --> 00:37:26,880
What was the quality of that output?

1038
00:37:26,880 --> 00:37:30,080
What was the actual cost in compute, tokens, and infrastructure?

1039
00:37:30,080 --> 00:37:32,240
These metrics have to be tracked every single day.

1040
00:37:32,240 --> 00:37:34,400
You can't just check them at deployment and walk away.

1041
00:37:34,400 --> 00:37:37,280
An agent that works perfectly on day one can drift over six months

1042
00:37:37,280 --> 00:37:39,440
as models update or business rules evolve.

1043
00:37:39,440 --> 00:37:41,200
If you aren't watching, you won't see the drift

1044
00:37:41,200 --> 00:37:43,040
until it becomes a massive problem.

1045
00:37:43,040 --> 00:37:45,760
The most mature organizations are now using a concept called

1046
00:37:45,760 --> 00:37:47,760
"Agent work units" to standardize this.

1047
00:37:47,760 --> 00:37:50,640
An "Agent work unit" is just a discrete measurable piece of work.

1048
00:37:50,640 --> 00:37:52,720
For customer service, it's a resolved ticket.

1049
00:37:52,720 --> 00:37:55,040
For finance, it's a processed expense report.

1050
00:37:55,040 --> 00:37:56,960
For data teams, it's a query answered.

1051
00:37:56,960 --> 00:37:58,960
You define what a unit looks like in your world,

1052
00:37:58,960 --> 00:38:01,040
then you measure the cost per unit

1053
00:38:01,040 --> 00:38:02,640
and compare it to the manual cost.

1054
00:38:02,640 --> 00:38:04,960
If a human resolving a ticket costs $30,

1055
00:38:04,960 --> 00:38:07,600
and the agent does it for five, you have a 6-to-1 return.

1056
00:38:07,600 --> 00:38:10,720
But if the manual cost is 20, and the agent costs 18,

1057
00:38:10,720 --> 00:38:12,880
the return is only 1.1 to 1.

1058
00:38:12,880 --> 00:38:15,360
In that case, you probably shouldn't deploy it at all.

1059
00:38:15,360 --> 00:38:17,280
But there's a subtlety here that most people miss.

1060
00:38:17,280 --> 00:38:18,640
Not all units are equal.

1061
00:38:18,640 --> 00:38:21,280
An agent might resolve a ticket, but the quality

1062
00:38:21,280 --> 00:38:23,120
might be lower than a human's work.

1063
00:38:23,120 --> 00:38:25,200
Maybe the customer isn't fully satisfied

1064
00:38:25,200 --> 00:38:27,680
or the first contact resolution rate drops.

1065
00:38:27,680 --> 00:38:30,800
You have to adjust your benefit calculation to account for that gap.

1066
00:38:30,800 --> 00:38:33,520
If the agent hits 90% quality compared to a person,

1067
00:38:33,520 --> 00:38:34,880
you don't count it as a full unit.

1068
00:38:34,880 --> 00:38:36,800
You count it as 0.9 units.

1069
00:38:36,800 --> 00:38:38,480
Suddenly, the economics look different.

1070
00:38:38,480 --> 00:38:39,600
It might still be worth it,

1071
00:38:39,600 --> 00:38:42,960
but now you're making a decision based on reality instead of a guess.

1072
00:38:42,960 --> 00:38:46,000
Some teams also track human efforts saved as a companion metric.

1073
00:38:46,000 --> 00:38:48,480
They ask how much time a person would have spent on this task

1074
00:38:48,480 --> 00:38:50,080
before the agent existed.

1075
00:38:50,080 --> 00:38:52,880
Then they measure how much time is actually being saved now.

1076
00:38:52,880 --> 00:38:55,040
This is more subjective because you're estimating a world

1077
00:38:55,040 --> 00:38:56,080
that doesn't exist anymore,

1078
00:38:56,080 --> 00:38:58,560
but it captures the gap between theoretical savings

1079
00:38:58,560 --> 00:38:59,920
and actual results.

1080
00:38:59,920 --> 00:39:02,800
If your team is spending 40 hours a week on a task,

1081
00:39:02,800 --> 00:39:04,880
and the agent brings that down to 20,

1082
00:39:04,880 --> 00:39:06,880
you've saved 20 hours of labor.

1083
00:39:06,880 --> 00:39:07,920
That is your benefit.

1084
00:39:07,920 --> 00:39:10,720
Your cost is simply what you pay foundry to run the agent.

1085
00:39:10,720 --> 00:39:13,520
The critical insight is that you have to measure continuously.

1086
00:39:13,520 --> 00:39:16,400
An agent that is 80% as capable as a human,

1087
00:39:16,400 --> 00:39:18,960
but costs 10% as much is a great investment.

1088
00:39:18,960 --> 00:39:21,760
You should deploy that and use the savings to build something else,

1089
00:39:21,760 --> 00:39:24,640
but an agent that is 60% as capable

1090
00:39:24,640 --> 00:39:27,520
and costs 50% as much is a bad investment.

1091
00:39:27,520 --> 00:39:28,960
Do you either kill that project

1092
00:39:28,960 --> 00:39:30,960
or only use it in non-critical spots

1093
00:39:30,960 --> 00:39:32,800
where good enough is actually fine?

1094
00:39:32,800 --> 00:39:35,680
You also have to define what good looks like for different roles.

1095
00:39:35,680 --> 00:39:37,760
A document agent needs high quality

1096
00:39:37,760 --> 00:39:39,440
because people rely on the text.

1097
00:39:39,440 --> 00:39:41,120
A rooting agent can be lower quality

1098
00:39:41,120 --> 00:39:43,200
because a human catches the mistakes later.

1099
00:39:43,200 --> 00:39:45,840
A financial agent needs near-perfect accuracy.

1100
00:39:45,840 --> 00:39:47,200
An infobot can be rougher

1101
00:39:47,200 --> 00:39:50,240
because the user evaluates the answer on the fly.

1102
00:39:50,240 --> 00:39:52,720
Your metrics framework has to account for these differences.

1103
00:39:52,720 --> 00:39:54,720
You aren't measuring performance in a vacuum.

1104
00:39:54,720 --> 00:39:57,520
You're measuring it against the specific context of the job

1105
00:39:57,520 --> 00:40:00,000
and what happens downstream when the agent fails.

1106
00:40:00,000 --> 00:40:03,280
That context is what determines if the performance is acceptable.

1107
00:40:03,280 --> 00:40:04,800
This is where most companies struggle.

1108
00:40:04,800 --> 00:40:06,800
They don't have the infrastructure to measure every day.

1109
00:40:06,800 --> 00:40:08,720
They don't have clear definitions of success.

1110
00:40:08,720 --> 00:40:11,520
They don't have a process to pivot when conditions change.

1111
00:40:11,520 --> 00:40:13,760
Building that measurement capability is hard work.

1112
00:40:13,760 --> 00:40:15,200
But it's the only way to know

1113
00:40:15,200 --> 00:40:17,360
if your agent investment is returning value

1114
00:40:17,360 --> 00:40:20,640
or if you're just automating costs without seeing the benefit.

1115
00:40:20,640 --> 00:40:22,080
Ownership and accountability.

1116
00:40:22,080 --> 00:40:23,360
The moment an agent fails,

1117
00:40:23,360 --> 00:40:25,360
you find out what ownership actually means.

1118
00:40:25,360 --> 00:40:28,240
In the old software world, ownership was easy to map out.

1119
00:40:28,240 --> 00:40:29,920
The product owner owned the product,

1120
00:40:29,920 --> 00:40:31,440
the engineer's owned the code,

1121
00:40:31,440 --> 00:40:32,960
the data team owned the database.

1122
00:40:32,960 --> 00:40:34,240
Everyone had a defined territory.

1123
00:40:34,240 --> 00:40:36,480
When a bug appeared, you knew exactly who to call.

1124
00:40:36,480 --> 00:40:39,520
You could trace a feature failure back through the teams to find the root cause

1125
00:40:39,520 --> 00:40:41,520
and the person responsible for the fix.

1126
00:40:41,520 --> 00:40:43,440
With agents, ownership becomes ambiguous

1127
00:40:43,440 --> 00:40:44,800
and that ambiguity is dangerous.

1128
00:40:44,800 --> 00:40:46,160
Who actually owns an agent?

1129
00:40:46,160 --> 00:40:47,760
Is it the business unit that asked for it?

1130
00:40:47,760 --> 00:40:50,000
Is it the platform team that built the infrastructure?

1131
00:40:50,000 --> 00:40:52,480
Is it the data team that provided the training sets?

1132
00:40:52,480 --> 00:40:54,320
Is it the specialist who tuned the prompts?

1133
00:40:54,320 --> 00:40:56,880
All of these people played a part in how the agent behaves

1134
00:40:56,880 --> 00:40:59,120
but none of them feels like they owned it completely.

1135
00:40:59,120 --> 00:41:02,560
When something breaks, nobody has the clear responsibility to fix it.

1136
00:41:02,560 --> 00:41:04,320
When the quality starts to slip over time,

1137
00:41:04,320 --> 00:41:07,120
nobody notices because nobody was tasked with monitoring it.

1138
00:41:07,120 --> 00:41:10,560
This confusion creates friction that slows down your response to every problem.

1139
00:41:10,560 --> 00:41:14,480
The enterprises moving the fastest are setting up explicit ownership models.

1140
00:41:14,480 --> 00:41:18,160
A common pattern is emerging where the business unit owns the agent and the outcome.

1141
00:41:18,160 --> 00:41:21,200
They are the ones who requested the tool and defined the workflow.

1142
00:41:21,200 --> 00:41:23,200
Since they are the ones who benefit when it works

1143
00:41:23,200 --> 00:41:25,760
and suffer when it fails, they own the results.

1144
00:41:25,760 --> 00:41:29,040
Meanwhile, the platform team owns the infrastructure and the safety rules.

1145
00:41:29,040 --> 00:41:32,160
They maintain foundry, they enforce agent 365 policies,

1146
00:41:32,160 --> 00:41:34,480
they make sure the agent stays within the guardrails.

1147
00:41:34,480 --> 00:41:36,160
They aren't deciding what the agent does.

1148
00:41:36,160 --> 00:41:38,400
They are deciding the constraints it operates within.

1149
00:41:38,400 --> 00:41:41,760
Then you have the data team who owns the training data and its quality.

1150
00:41:41,760 --> 00:41:45,120
If the agent starts failing because the data is stale or biased,

1151
00:41:45,120 --> 00:41:47,520
that is a data problem, not an agent problem.

1152
00:41:47,520 --> 00:41:49,680
This creates clear lines of accountability.

1153
00:41:49,680 --> 00:41:52,000
The business unit knows they are responsible.

1154
00:41:52,000 --> 00:41:53,760
The platform team knows their boundaries.

1155
00:41:53,760 --> 00:41:55,200
The data team knows their domain,

1156
00:41:55,200 --> 00:41:57,200
but this clarity requires a level of trust

1157
00:41:57,200 --> 00:41:58,960
that doesn't always exist between teams.

1158
00:41:58,960 --> 00:42:02,960
The business unit has to trust that the platform team won't block a valuable agent

1159
00:42:02,960 --> 00:42:04,800
just because it doesn't fit a perfect template.

1160
00:42:04,800 --> 00:42:07,600
Real work is messy and sometimes you need an exception.

1161
00:42:07,600 --> 00:42:11,840
On the flip side, the platform team has to trust that the business won't deploy a risky agent

1162
00:42:11,840 --> 00:42:12,800
just to move faster.

1163
00:42:12,800 --> 00:42:15,200
You build this trust through transparent policies.

1164
00:42:15,200 --> 00:42:16,720
When the platform team says no,

1165
00:42:16,720 --> 00:42:19,680
they explain it through actual risk instead of arbitrary rules.

1166
00:42:19,680 --> 00:42:22,560
When the business pushes back, the platform team actually listens.

1167
00:42:22,560 --> 00:42:25,760
Some organizations are now creating a former role called the agent owner.

1168
00:42:25,760 --> 00:42:27,200
It's similar to a data owner.

1169
00:42:27,200 --> 00:42:30,960
This person isn't necessarily the one who wrote the prompt or tuned the parameters.

1170
00:42:30,960 --> 00:42:35,280
Instead, the agent owner is responsible for making sure the agent is working correctly in production.

1171
00:42:35,280 --> 00:42:39,280
They ensure it's delivering the value at promised and staying compliant with company policy.

1172
00:42:39,280 --> 00:42:42,160
For a small agent, this might be a part-time task.

1173
00:42:42,160 --> 00:42:44,720
For a critical agent processing thousands of transactions,

1174
00:42:44,720 --> 00:42:45,760
it's a full-time job.

1175
00:42:45,760 --> 00:42:49,840
The important thing is that a specific person is accountable for the agent's behavior.

1176
00:42:49,840 --> 00:42:53,920
Clear ownership sets expectations and gives you a point of escalation when things go wrong.

1177
00:42:53,920 --> 00:42:56,640
But ownership without the right tools is just a title.

1178
00:42:56,640 --> 00:43:00,400
You need the operational framework to make that ownership actually mean something.

1179
00:43:00,400 --> 00:43:02,160
The operational model.

1180
00:43:02,160 --> 00:43:05,840
Running agents at scale requires a level of discipline most companies haven't built yet,

1181
00:43:05,840 --> 00:43:09,200
because the reality is, building an agent is easy.

1182
00:43:09,200 --> 00:43:11,280
Running it in production is where the work starts,

1183
00:43:11,280 --> 00:43:13,200
and that is a completely different skill set.

1184
00:43:13,200 --> 00:43:15,200
You need a way to deploy these agents consistently.

1185
00:43:15,200 --> 00:43:18,400
Not a manual process where a single laptop is the source of truth.

1186
00:43:18,400 --> 00:43:21,840
Not a mess of ad hoc deployments where every team follows their own rules.

1187
00:43:21,840 --> 00:43:26,400
You need a repeatable, auditable process that puts an agent into production reliably.

1188
00:43:26,400 --> 00:43:28,880
Every single time, and once it's there, you have to monitor it.

1189
00:43:28,880 --> 00:43:30,400
But not just system uptime.

1190
00:43:30,400 --> 00:43:32,640
You need to know if the agent is actually doing its job.

1191
00:43:32,640 --> 00:43:34,400
Is it handling the volume you expected?

1192
00:43:34,400 --> 00:43:36,560
Is the quality of the answers starting to slip?

1193
00:43:36,560 --> 00:43:37,920
Are the error rates trending up?

1194
00:43:37,920 --> 00:43:38,880
You navigate.

1195
00:43:38,880 --> 00:43:39,680
You monitor.

1196
00:43:39,680 --> 00:43:40,560
You adjust.

1197
00:43:40,560 --> 00:43:43,440
Because business doesn't stop just because you want to make an improvement.

1198
00:43:43,440 --> 00:43:46,640
You have to be able to update the logic, test the changes,

1199
00:43:46,640 --> 00:43:50,320
and roll them out without breaking the workflows that people are counting on.

1200
00:43:50,320 --> 00:43:53,600
And if that new prompt you deployed yesterday starts making weird decisions,

1201
00:43:53,600 --> 00:43:56,000
you need a way to get back to the old version fast.

1202
00:43:56,000 --> 00:43:58,720
But before you even get to production, you need evaluation.

1203
00:43:58,720 --> 00:44:01,600
Not a gut check where you look at a few outputs and feel good about them.

1204
00:44:01,600 --> 00:44:04,960
You need a structured systematic way to test against real criteria.

1205
00:44:04,960 --> 00:44:07,280
When you add new training data or refine a tool,

1206
00:44:07,280 --> 00:44:09,040
you have to know if it's actually better.

1207
00:44:09,040 --> 00:44:10,960
Or if it's worse in ways you haven't noticed yet.

1208
00:44:10,960 --> 00:44:11,600
You can't guess.

1209
00:44:11,600 --> 00:44:12,560
You have to test.

1210
00:44:12,560 --> 00:44:15,040
Foundry handles the runtime and the tool orchestration,

1211
00:44:15,040 --> 00:44:17,440
but the DevOps side of this is still mostly on you.

1212
00:44:17,440 --> 00:44:19,280
The platform doesn't have built-in AB testing.

1213
00:44:19,280 --> 00:44:22,640
If you want to deploy two versions to different users and see which ones wins,

1214
00:44:22,640 --> 00:44:25,360
you have to build that yourself or plug in a third-party tool.

1215
00:44:25,360 --> 00:44:27,360
It's not a click button feature.

1216
00:44:27,360 --> 00:44:30,400
And there is another layer here because agents aren't static software.

1217
00:44:30,400 --> 00:44:31,920
They are continuously learning.

1218
00:44:31,920 --> 00:44:34,400
If you're fine tuning an agent on real-world data,

1219
00:44:34,400 --> 00:44:36,160
you are constantly updating the model.

1220
00:44:36,160 --> 00:44:38,080
Every week you're looking at where it failed,

1221
00:44:38,080 --> 00:44:41,520
adding those cases to the training set and retraining the system.

1222
00:44:41,520 --> 00:44:43,440
The agent gets better incrementally.

1223
00:44:43,440 --> 00:44:44,960
But now you have to manage the versions.

1224
00:44:44,960 --> 00:44:48,320
You have to track exactly which weights and parameters are running right now.

1225
00:44:48,320 --> 00:44:52,160
Some teams use MLOps platforms like MLflow to handle this metadata.

1226
00:44:52,160 --> 00:44:54,960
Others build custom solutions because their requirements are specific

1227
00:44:54,960 --> 00:44:56,080
to their context.

1228
00:44:56,080 --> 00:44:59,760
Then there is the part that isn't glamorous, but it's critical, human in the loop.

1229
00:44:59,760 --> 00:45:02,000
Not every agent should be fully autonomous.

1230
00:45:02,000 --> 00:45:06,000
Some should draft an action and then wait for a human to hit approve.

1231
00:45:06,000 --> 00:45:08,240
The agent does 80% of the heavy lifting,

1232
00:45:08,240 --> 00:45:10,400
gathering data and making a recommendation,

1233
00:45:10,400 --> 00:45:12,240
and the human makes the final call.

1234
00:45:12,240 --> 00:45:15,680
But managing those approvals at scale is a massive operational task.

1235
00:45:15,680 --> 00:45:19,360
If you have 100 agents generating 10,000 requests a day,

1236
00:45:19,360 --> 00:45:21,520
you can't let those get lost in an inbox.

1237
00:45:21,520 --> 00:45:23,680
You need a system to root them, track them,

1238
00:45:23,680 --> 00:45:25,360
and escalate them when they stall.

1239
00:45:25,360 --> 00:45:27,600
Tools like Power Automate or Logic Apps

1240
00:45:27,600 --> 00:45:30,080
can sit on top of your infrastructure to handle that flow.

1241
00:45:30,080 --> 00:45:32,320
This is where most enterprises are going to struggle,

1242
00:45:32,320 --> 00:45:34,160
because the model is fundamentally different.

1243
00:45:34,160 --> 00:45:37,360
In the old model, you shipped software once a quarter and then maintained it.

1244
00:45:37,360 --> 00:45:38,560
It was static.

1245
00:45:38,560 --> 00:45:41,600
With agents, deployment is more like running a live service.

1246
00:45:41,600 --> 00:45:43,280
You are constantly watching for drift.

1247
00:45:43,280 --> 00:45:45,200
You are evaluating improvements every day.

1248
00:45:45,200 --> 00:45:46,880
This requires a different mindset.

1249
00:45:46,880 --> 00:45:51,280
The companies that are ready for this are the ones already living in a cloud-native world.

1250
00:45:51,280 --> 00:45:54,640
If you are used to shipping code multiple times a day and watching metrics obsessively,

1251
00:45:54,640 --> 00:45:55,840
agents aren't a big jump.

1252
00:45:55,840 --> 00:45:58,880
But if you are used to long-testing cycles and quarterly releases,

1253
00:45:58,880 --> 00:46:00,560
you are going to feel the friction.

1254
00:46:00,560 --> 00:46:03,680
Operational readiness starts before the first agent ever goes live.

1255
00:46:03,680 --> 00:46:06,080
The Incident Response Model

1256
00:46:06,080 --> 00:46:09,120
When an agent fails in production, it happens at machine speed.

1257
00:46:09,120 --> 00:46:11,520
That is the reality that changes how you respond.

1258
00:46:11,520 --> 00:46:14,320
Imagine an agent designed to archive old customer records.

1259
00:46:14,320 --> 00:46:15,440
The logic is simple.

1260
00:46:15,440 --> 00:46:18,320
Find anything older than 90 days and move it to storage.

1261
00:46:18,320 --> 00:46:20,320
But instead, the agent starts deleting them.

1262
00:46:20,320 --> 00:46:22,640
By the time you see the spike in database activity,

1263
00:46:22,640 --> 00:46:24,240
thousands of records are gone forever.

1264
00:46:24,240 --> 00:46:25,440
That isn't just a bug.

1265
00:46:25,440 --> 00:46:28,080
It's a catastrophe for compliance and customer trust.

1266
00:46:28,080 --> 00:46:33,280
Or imagine an expense agent that suddenly starts approving every single requested sees.

1267
00:46:33,280 --> 00:46:36,240
Within hours, millions of dollars are leaving your accounts.

1268
00:46:36,240 --> 00:46:37,760
The financial impact is immediate.

1269
00:46:37,760 --> 00:46:39,600
These aren't what-if stories.

1270
00:46:39,600 --> 00:46:42,240
This is what happens when you scale without an Incident Response Plan.

1271
00:46:42,240 --> 00:46:43,520
The first rule is simple.

1272
00:46:43,520 --> 00:46:44,480
You need a kill switch.

1273
00:46:44,480 --> 00:46:46,800
You have to be able to kill an agent in seconds.

1274
00:46:46,800 --> 00:46:48,240
Not after an hour of meetings.

1275
00:46:48,240 --> 00:46:50,080
Not after waiting for a manager to sign off.

1276
00:46:50,080 --> 00:46:53,040
You need one action that stops the execution immediately.

1277
00:46:53,040 --> 00:46:55,920
And you need clear policies on what triggers that switch.

1278
00:46:55,920 --> 00:46:58,400
If a financial agent hits a 5% error rate,

1279
00:46:58,400 --> 00:47:00,880
it should probably shut itself down automatically.

1280
00:47:00,880 --> 00:47:03,280
For a low stakes agent, you might give it more room.

1281
00:47:03,280 --> 00:47:05,680
But you have to define those rules before the emergency happens.

1282
00:47:05,680 --> 00:47:08,160
You also need monitoring that catches the slow bleed.

1283
00:47:08,160 --> 00:47:10,080
A catastrophic crash is easy to see.

1284
00:47:10,080 --> 00:47:12,080
A gradual degradation is much more dangerous.

1285
00:47:12,080 --> 00:47:14,640
You need dashboards that show you what's happening in real time.

1286
00:47:14,640 --> 00:47:16,720
Volume, error rates, latency.

1287
00:47:17,440 --> 00:47:20,400
If the patents deviate even a little bit, someone needs to know.

1288
00:47:20,400 --> 00:47:23,440
And those alerts shouldn't go to a general Slack channel where they get ignored.

1289
00:47:23,440 --> 00:47:26,240
They need to reach the people who have the authority to act.

1290
00:47:26,240 --> 00:47:29,200
In a real Incident, the first 30 seconds are everything.

1291
00:47:29,200 --> 00:47:31,600
If the error rate jumps from 1% to 15,

1292
00:47:31,600 --> 00:47:33,600
the escalation path has to be automatic.

1293
00:47:33,600 --> 00:47:34,320
Who gets the call?

1294
00:47:34,320 --> 00:47:35,280
What can they do?

1295
00:47:35,280 --> 00:47:37,440
That authority has to be baked into the structure.

1296
00:47:37,440 --> 00:47:40,720
Then, after the fire is out, you have to figure out why it happened.

1297
00:47:40,720 --> 00:47:41,920
Was it bad data?

1298
00:47:41,920 --> 00:47:43,840
A model update that changed the behavior?

1299
00:47:43,840 --> 00:47:46,000
Did a downstream API change without warning?

1300
00:47:46,000 --> 00:47:49,440
You have to be able to reproduce the failure so you can fix the causal chain.

1301
00:47:49,440 --> 00:47:51,200
Maybe you add validation logic.

1302
00:47:51,200 --> 00:47:52,560
Maybe you tighten the guardrails.

1303
00:47:52,560 --> 00:47:55,440
But every failure has to be encoded back into the agent's logic.

1304
00:47:55,440 --> 00:47:56,720
So it never happens again.

1305
00:47:56,720 --> 00:47:58,720
Some mature teams even use chaos engineering.

1306
00:47:58,720 --> 00:48:01,600
They deliberately inject bad data or adversarial inputs

1307
00:48:01,600 --> 00:48:03,360
just to see how the agent breaks.

1308
00:48:03,360 --> 00:48:05,600
They want to find the weaknesses before production does.

1309
00:48:05,600 --> 00:48:08,720
This is the divide between the people taking agents seriously

1310
00:48:08,720 --> 00:48:10,480
and the people who are just experimenting.

1311
00:48:10,480 --> 00:48:11,840
The serious teams have playbooks.

1312
00:48:11,840 --> 00:48:12,880
They have the monitoring.

1313
00:48:12,880 --> 00:48:14,480
They have the decision authority.

1314
00:48:14,480 --> 00:48:17,520
They treat an agent failure like any other critical production incident

1315
00:48:17,520 --> 00:48:19,200
because they know the damage is just as real.

1316
00:48:19,200 --> 00:48:21,280
The experimenters are just hoping it works

1317
00:48:21,280 --> 00:48:23,680
and hope is not an operational model.

1318
00:48:23,680 --> 00:48:25,280
The continuous improvement cycle.

1319
00:48:25,280 --> 00:48:26,880
Agents are not static artifacts.

1320
00:48:26,880 --> 00:48:29,120
You don't just build them once and let them run forever.

1321
00:48:29,120 --> 00:48:30,320
They are living systems.

1322
00:48:30,320 --> 00:48:33,040
And they require active refinement to stay useful.

1323
00:48:33,040 --> 00:48:34,880
An agent that performs perfectly on day one

1324
00:48:34,880 --> 00:48:37,120
might be a total disaster six months later.

1325
00:48:37,120 --> 00:48:39,440
Business conditions shift, data become stale,

1326
00:48:39,440 --> 00:48:41,520
the rules governing your operations change.

1327
00:48:41,520 --> 00:48:43,280
The moment you stop improving an agent,

1328
00:48:43,280 --> 00:48:45,040
it starts becoming a liability.

1329
00:48:45,040 --> 00:48:46,720
Because while your business moves forward,

1330
00:48:46,720 --> 00:48:48,480
the agent stays exactly where it was

1331
00:48:48,480 --> 00:48:50,560
and that's how your competitive advantage erodes.

1332
00:48:50,560 --> 00:48:53,360
This is why the continuous improvement cycle is non-negotiable.

1333
00:48:53,360 --> 00:48:55,120
It starts with evaluation.

1334
00:48:55,120 --> 00:48:57,440
Not a feeling that the agent is working

1335
00:48:57,440 --> 00:49:00,400
but structured evaluation against explicit criteria.

1336
00:49:00,400 --> 00:49:01,760
For a customer service agent,

1337
00:49:01,760 --> 00:49:03,760
that might mean resolving 90% of tickets

1338
00:49:03,760 --> 00:49:07,120
on the first contact while maintaining 95% satisfaction.

1339
00:49:07,120 --> 00:49:10,160
For a financial agent, you might demand 99% accuracy

1340
00:49:10,160 --> 00:49:13,200
on expense categorization and zero false approvals.

1341
00:49:13,200 --> 00:49:14,960
You have to define what good looks like

1342
00:49:14,960 --> 00:49:16,400
before you start measuring.

1343
00:49:16,400 --> 00:49:18,560
And you don't just measure once a deployment.

1344
00:49:18,560 --> 00:49:19,680
You do it continuously.

1345
00:49:19,680 --> 00:49:21,440
Weekly evaluations are the standard

1346
00:49:21,440 --> 00:49:24,080
but some organizations are already doing this daily.

1347
00:49:24,080 --> 00:49:26,160
Foundry has built in evaluation capabilities

1348
00:49:26,160 --> 00:49:27,680
but they are just the foundation.

1349
00:49:27,680 --> 00:49:29,040
You define test cases.

1350
00:49:29,040 --> 00:49:30,400
You run the agent through them.

1351
00:49:30,400 --> 00:49:32,720
You check if the output matches the result you expected.

1352
00:49:32,720 --> 00:49:35,600
But this only works if you actually understand what to test for.

1353
00:49:35,600 --> 00:49:38,000
Defining good test cases is much harder than it sounds

1354
00:49:38,000 --> 00:49:39,920
because you have to anticipate the edge cases.

1355
00:49:39,920 --> 00:49:42,320
You have to understand how the system fails.

1356
00:49:42,320 --> 00:49:45,440
And you have to think like an adversary trying to break your own agent.

1357
00:49:45,440 --> 00:49:48,560
If you only test the happy path where everything goes smoothly,

1358
00:49:48,560 --> 00:49:50,400
you will never catch the problems that emerge

1359
00:49:50,400 --> 00:49:51,840
in the messiness of the real world.

1360
00:49:51,840 --> 00:49:54,480
That's why many organizations supplement platform tools

1361
00:49:54,480 --> 00:49:55,760
with human evaluation.

1362
00:49:55,760 --> 00:49:58,960
You have real people review a sample of what the agent produces.

1363
00:49:58,960 --> 00:50:01,120
Not every single output because that's too expensive

1364
00:50:01,120 --> 00:50:02,320
but a representative sample.

1365
00:50:02,320 --> 00:50:04,960
Maybe you review 50 approvals a week for an expense agent

1366
00:50:04,960 --> 00:50:07,120
or 100 tickets for customer service.

1367
00:50:07,120 --> 00:50:10,400
Humans read the responses and provide the nuance that machines miss.

1368
00:50:10,400 --> 00:50:13,920
They can see when an agent made a judgment call that is technically correct

1369
00:50:13,920 --> 00:50:16,640
but actually violates the spirit of the company policy.

1370
00:50:16,640 --> 00:50:18,160
Yes, human feedback is expensive.

1371
00:50:18,160 --> 00:50:21,280
You're paying people to spend their time grading the agent's homework.

1372
00:50:21,280 --> 00:50:23,040
But it gives you a high quality signal

1373
00:50:23,040 --> 00:50:26,080
that you just can't get from clean automated test cases.

1374
00:50:26,080 --> 00:50:29,040
Once you see where the agent is failing, you can actually fix it.

1375
00:50:29,040 --> 00:50:31,120
The fix depends on what the analysis shows.

1376
00:50:31,120 --> 00:50:32,560
Sometimes it's just a prompt change

1377
00:50:32,560 --> 00:50:35,680
because the agent misunderstood an instruction that was too ambiguous.

1378
00:50:35,680 --> 00:50:37,040
You clarify the language.

1379
00:50:37,040 --> 00:50:38,480
You provide better examples.

1380
00:50:38,480 --> 00:50:40,480
And the agent learns the distinction.

1381
00:50:40,480 --> 00:50:42,240
Other times it's a tool change.

1382
00:50:42,240 --> 00:50:45,200
The agent might be failing because it doesn't have the right information

1383
00:50:45,200 --> 00:50:47,440
so you give it a new tool to provide that context

1384
00:50:47,440 --> 00:50:49,040
or maybe it's the training data.

1385
00:50:49,040 --> 00:50:52,320
You fine tune the agent on specific examples of yes and no

1386
00:50:52,320 --> 00:50:54,560
until it understands the decision boundary.

1387
00:50:54,560 --> 00:50:56,560
Sometimes you just need a policy guardrail.

1388
00:50:56,560 --> 00:50:58,640
A piece of logic that catches a specific mistake

1389
00:50:58,640 --> 00:51:00,080
before the agent can commit to it.

1390
00:51:00,080 --> 00:51:02,960
But before you deploy any of these improvements, you have to test them.

1391
00:51:02,960 --> 00:51:04,880
This is where A/B testing is vital.

1392
00:51:04,880 --> 00:51:08,000
You run the improved agent in parallel with the one currently in production.

1393
00:51:08,000 --> 00:51:09,680
You send the same requests to both

1394
00:51:09,680 --> 00:51:12,320
and measure if the new version actually performs better.

1395
00:51:12,320 --> 00:51:14,080
If it does, you validated the change.

1396
00:51:14,080 --> 00:51:14,960
So you deploy it.

1397
00:51:14,960 --> 00:51:17,600
If it doesn't, you just learned that your intuition was wrong.

1398
00:51:17,600 --> 00:51:18,320
And that's fine.

1399
00:51:18,320 --> 00:51:21,920
It just means you iterate further instead of pushing a change that doesn't work.

1400
00:51:21,920 --> 00:51:25,200
The continuous improvement cycle is where agents become genuinely valuable.

1401
00:51:25,200 --> 00:51:26,320
Think about it this way.

1402
00:51:26,320 --> 00:51:28,800
An agent that starts at 70% of human capability

1403
00:51:28,800 --> 00:51:32,560
but improves by 5% every month will eventually outperform a person.

1404
00:51:32,560 --> 00:51:35,520
After 12 months, that 5% monthly improvement compounds

1405
00:51:35,520 --> 00:51:37,360
into a 60% cumulative gain.

1406
00:51:37,360 --> 00:51:40,080
Now the agent is substantially better than the human baseline.

1407
00:51:40,080 --> 00:51:44,880
But an agent that stays at 70% and never improves is just a liability.

1408
00:51:44,880 --> 00:51:46,640
It's permanently mediocre.

1409
00:51:46,640 --> 00:51:50,000
It costs money to run, but it's never good enough to fully trust.

1410
00:51:50,000 --> 00:51:52,400
The difference between an agent that transforms your business

1411
00:51:52,400 --> 00:51:54,720
and one that becomes dead weight is simple.

1412
00:51:54,720 --> 00:51:57,520
It's whether or not you build the systems to keep it getting better.

1413
00:51:57,520 --> 00:51:59,680
The investment case.

1414
00:51:59,680 --> 00:52:02,560
Building agent capabilities requires an upfront investment

1415
00:52:02,560 --> 00:52:04,880
that most enterprises aren't ready to quantify.

1416
00:52:04,880 --> 00:52:06,800
You have to build the platform infrastructure.

1417
00:52:06,800 --> 00:52:07,920
Foundry isn't free.

1418
00:52:07,920 --> 00:52:09,680
And neither is agent 365.

1419
00:52:09,680 --> 00:52:10,800
The governance systems.

1420
00:52:10,800 --> 00:52:14,000
The monitoring and the integration points between PerView and Entra

1421
00:52:14,000 --> 00:52:15,440
don't just appear for free.

1422
00:52:15,440 --> 00:52:17,440
You aren't just licensing software.

1423
00:52:17,440 --> 00:52:20,480
You are building operational systems that didn't exist before.

1424
00:52:20,480 --> 00:52:22,720
And you need to hire or train people with skills

1425
00:52:22,720 --> 00:52:25,120
that most organizations simply don't have yet.

1426
00:52:25,120 --> 00:52:28,560
Prompt engineers, data governance experts,

1427
00:52:28,560 --> 00:52:30,080
business process designers.

1428
00:52:30,080 --> 00:52:33,680
These people have to encode institutional logic into agent instructions

1429
00:52:33,680 --> 00:52:36,640
and those roles don't exist in a traditional IT department.

1430
00:52:36,640 --> 00:52:38,880
You're either building new capability areas

1431
00:52:38,880 --> 00:52:40,560
or recruiting from the outside.

1432
00:52:40,560 --> 00:52:41,920
And both are expensive.

1433
00:52:41,920 --> 00:52:45,760
You also have to modernize your software architecture to be API first.

1434
00:52:45,760 --> 00:52:49,200
Legacy systems usually don't have the APIs that agents need to function

1435
00:52:49,200 --> 00:52:50,480
so you have to build them.

1436
00:52:50,480 --> 00:52:51,600
That is engineering work.

1437
00:52:51,600 --> 00:52:53,200
Often thousands of hours of it.

1438
00:52:53,200 --> 00:52:55,680
Then you have to establish the processes for development,

1439
00:52:55,680 --> 00:52:57,840
testing and safety that don't exist yet.

1440
00:52:57,840 --> 00:52:59,280
These are processes not products.

1441
00:52:59,280 --> 00:53:01,840
You have to invent them based on your specific context.

1442
00:53:01,840 --> 00:53:02,880
This is not cheap.

1443
00:53:02,880 --> 00:53:04,800
Organizations that are serious about this

1444
00:53:04,800 --> 00:53:06,880
are investing millions in infrastructure

1445
00:53:06,880 --> 00:53:08,480
and people before they ever see a return.

1446
00:53:08,480 --> 00:53:10,320
But the payoff when it finally hits

1447
00:53:10,320 --> 00:53:12,480
is big enough to justify every cent.

1448
00:53:12,480 --> 00:53:17,040
An agent that replaces 30 human seeds can save a company millions of dollars every single year.

1449
00:53:17,040 --> 00:53:19,040
Those seeds have a fully loaded cost.

1450
00:53:19,040 --> 00:53:22,080
Salary, benefits, office space, training.

1451
00:53:22,080 --> 00:53:27,600
A typical knowledge worker costs the company between 120 and 150,000 dollars a year.

1452
00:53:27,600 --> 00:53:31,840
30 of those seeds represent a cost of 3.6 to 4.5 million dollars.

1453
00:53:31,840 --> 00:53:38,000
An agent that does the work of those 30 people running 24/7 probably costs less than 500,000 to operate.

1454
00:53:38,000 --> 00:53:40,480
The payoff in labor cost alone is 7 to 1.

1455
00:53:40,480 --> 00:53:43,840
But an agent that enables a new capability doesn't just cut costs.

1456
00:53:43,840 --> 00:53:45,360
It creates revenue.

1457
00:53:45,360 --> 00:53:49,120
A customer service agent that handles interactions humans couldn't do profitably

1458
00:53:49,120 --> 00:53:50,720
allows you to expand your market.

1459
00:53:50,720 --> 00:53:52,240
You can serve smaller customers.

1460
00:53:52,240 --> 00:53:53,600
You can handle higher volumes.

1461
00:53:53,600 --> 00:53:55,040
You can open new sales channels.

1462
00:53:55,040 --> 00:53:56,720
That is growth, not just efficiency.

1463
00:53:56,720 --> 00:53:59,600
And when an agent reduces errors, the impact compounds.

1464
00:53:59,600 --> 00:54:04,240
Fewer mistakes mean lower costs for rework while higher quality leads to better customer retention

1465
00:54:04,240 --> 00:54:05,440
and premium pricing.

1466
00:54:05,440 --> 00:54:08,640
The real question is how do you decide which agents to build first?

1467
00:54:08,640 --> 00:54:11,840
This is where most organizations fail because they try to do too much.

1468
00:54:11,840 --> 00:54:16,240
They want the perfect sophisticated agent that handles everything so they spend a year planning.

1469
00:54:16,240 --> 00:54:18,800
By the time they are ready to build, the window has closed.

1470
00:54:18,800 --> 00:54:20,160
The answer is actually simpler.

1471
00:54:20,160 --> 00:54:22,880
Start with high volume, well-defined workflows.

1472
00:54:22,880 --> 00:54:27,600
If a process has repetitive steps and clear decision criteria, it's a perfect candidate.

1473
00:54:27,600 --> 00:54:28,960
Expense processing.

1474
00:54:28,960 --> 00:54:30,640
Invoice categorization.

1475
00:54:30,640 --> 00:54:32,320
Document classification.

1476
00:54:32,320 --> 00:54:34,640
These are workflows where the logic is explicit,

1477
00:54:34,640 --> 00:54:36,960
where a person could sit down and write out the rules.

1478
00:54:36,960 --> 00:54:39,760
If a human can write the rules, an agent can learn them.

1479
00:54:39,760 --> 00:54:42,960
Workflows that rely on gut feeling or complex judgment

1480
00:54:42,960 --> 00:54:45,200
are much harder to automate at the start.

1481
00:54:45,200 --> 00:54:49,120
Sales negotiations or complex disputes involve nuance that is hard to codify.

1482
00:54:49,120 --> 00:54:52,000
You'll get to those eventually, but they shouldn't be your first project.

1483
00:54:52,000 --> 00:54:53,360
You want the easy wins.

1484
00:54:53,360 --> 00:54:57,120
Build an agent that handles 80% of your basic inquiries.

1485
00:54:57,120 --> 00:55:00,160
Build an agent that processes 80% of your expense reports.

1486
00:55:00,160 --> 00:55:03,200
These are the projects that show a clear return on investment immediately.

1487
00:55:03,200 --> 00:55:05,440
They build the momentum you need to keep going.

1488
00:55:05,440 --> 00:55:08,080
Once you've proven the model works in your specific context,

1489
00:55:08,080 --> 00:55:11,200
you can tackle the harder problems with the whole organization behind you.

1490
00:55:11,200 --> 00:55:13,680
The investment case is also about your strategic position.

1491
00:55:13,680 --> 00:55:16,880
If your competitors are building agents and you aren't, you are going to fall behind.

1492
00:55:16,880 --> 00:55:20,160
They are automating their labor costs while your stay fixed.

1493
00:55:20,160 --> 00:55:23,840
They are building capabilities that you won't be able to match if you start three years late.

1494
00:55:23,840 --> 00:55:26,480
But if you are the one building agents now, you pull ahead.

1495
00:55:26,480 --> 00:55:30,800
The fastest moving enterprises are treating agent development as a strategic imperative.

1496
00:55:30,800 --> 00:55:34,320
They are investing heavily because they know this is how you build a competitive advantage

1497
00:55:34,320 --> 00:55:35,200
in the next decade.

1498
00:55:35,200 --> 00:55:38,000
The investment pays off, but only if you actually execute.

1499
00:55:38,000 --> 00:55:39,520
The strategic choices.

1500
00:55:39,520 --> 00:55:42,960
Building agent capabilities requires a series of directional decisions

1501
00:55:42,960 --> 00:55:45,280
that ripple through your entire organization.

1502
00:55:45,280 --> 00:55:47,360
None of these choices lock you in forever,

1503
00:55:47,360 --> 00:55:50,880
but they do shape a trajectory that becomes very difficult to reverse later on.

1504
00:55:50,880 --> 00:55:54,720
Making these decisions deliberately is always better than making them by default.

1505
00:55:54,720 --> 00:55:56,720
The first fork in the road is your sourcing model.

1506
00:55:56,720 --> 00:55:58,800
Do you build agents in-house using Foundry?

1507
00:55:58,800 --> 00:56:00,560
Buy pre-built agents from vendors?

1508
00:56:00,560 --> 00:56:01,920
Or blend both approaches?

1509
00:56:01,920 --> 00:56:04,320
Building entirely in-house gives you maximum control

1510
00:56:04,320 --> 00:56:06,480
because you own the prompts and the training process.

1511
00:56:06,480 --> 00:56:09,120
You understand exactly how every decision is made,

1512
00:56:09,120 --> 00:56:11,680
and you can adapt the agent to your specific context

1513
00:56:11,680 --> 00:56:13,920
without negotiating with a third party.

1514
00:56:13,920 --> 00:56:16,720
But the downside is the massive investment in speed.

1515
00:56:16,720 --> 00:56:20,640
You are starting from scratch and need a team that actually knows how to build these systems.

1516
00:56:20,640 --> 00:56:22,400
Because building agents takes months,

1517
00:56:22,400 --> 00:56:26,160
the business has often moved and the market has shifted by the time you actually finish.

1518
00:56:26,160 --> 00:56:28,560
Buying pre-built agents from vendors is much faster.

1519
00:56:28,560 --> 00:56:30,480
Someone else has already solved the problem,

1520
00:56:30,480 --> 00:56:33,040
so you just configure it for your context and deploy.

1521
00:56:33,040 --> 00:56:34,880
You aren't building from first principles,

1522
00:56:34,880 --> 00:56:36,240
but the downside is the fit.

1523
00:56:36,240 --> 00:56:38,640
A vendor's agent is built for a generic use case,

1524
00:56:38,640 --> 00:56:40,160
and your business probably isn't generic.

1525
00:56:40,160 --> 00:56:41,920
Your workflows have specific quirks,

1526
00:56:41,920 --> 00:56:43,600
and your decision-making has nuances

1527
00:56:43,600 --> 00:56:46,000
that a pre-built agent simply won't capture.

1528
00:56:46,000 --> 00:56:49,200
You end up trying to configure something that was built for someone else.

1529
00:56:49,200 --> 00:56:52,720
And you're always constrained by what the vendor decided to optimize for.

1530
00:56:52,720 --> 00:56:55,200
Most enterprises are settling on a hybrid model.

1531
00:56:55,200 --> 00:56:58,880
They use Foundry as the platform and build custom agents for their specific workflows.

1532
00:56:58,880 --> 00:57:01,360
You aren't inventing the underlying infrastructure,

1533
00:57:01,360 --> 00:57:04,080
and you aren't arguing with a vendor about customization options.

1534
00:57:04,080 --> 00:57:05,680
You're using a platform designed for this

1535
00:57:05,680 --> 00:57:07,920
and building what makes sense in your own context.

1536
00:57:07,920 --> 00:57:11,200
This requires less upfront work than pure in-house builds

1537
00:57:11,200 --> 00:57:13,680
and produces better results than buying off the shelf.

1538
00:57:13,680 --> 00:57:16,400
But it does require a commitment to learning the platform

1539
00:57:16,400 --> 00:57:18,000
and maintaining your own agents.

1540
00:57:18,000 --> 00:57:19,920
The second choice is your governance structure.

1541
00:57:19,920 --> 00:57:22,480
Do you centralize agent development in a single team,

1542
00:57:22,480 --> 00:57:25,840
distributed across business units, or find the middle ground?

1543
00:57:25,840 --> 00:57:27,760
Centralized teams make governance much easier

1544
00:57:27,760 --> 00:57:29,280
because one team sets the standards

1545
00:57:29,280 --> 00:57:31,280
and reviews every single deployment.

1546
00:57:31,280 --> 00:57:33,440
Policy is consistent and quality is predictable,

1547
00:57:33,440 --> 00:57:34,640
but the problem is speed.

1548
00:57:34,640 --> 00:57:36,560
Every request goes through a central queue.

1549
00:57:36,560 --> 00:57:39,440
And every business unit has to wait for that team's capacity.

1550
00:57:39,440 --> 00:57:40,960
The organization can't move fast

1551
00:57:40,960 --> 00:57:42,480
because you've created a bottleneck.

1552
00:57:42,480 --> 00:57:44,720
A business unit, needing a specific agent,

1553
00:57:44,720 --> 00:57:46,720
might wait months while the central team

1554
00:57:46,720 --> 00:57:48,400
works on a different priority.

1555
00:57:48,400 --> 00:57:51,120
Decentralized development is fast because every business unit

1556
00:57:51,120 --> 00:57:52,320
builds its own agents.

1557
00:57:52,320 --> 00:57:53,840
There are no bottlenecks and no waiting.

1558
00:57:53,840 --> 00:57:56,880
Each team deploys what makes sense for their specific workflow.

1559
00:57:56,880 --> 00:57:58,480
But this is where governance breaks down.

1560
00:57:58,480 --> 00:58:00,320
Different teams use different approaches.

1561
00:58:00,320 --> 00:58:02,480
And some teams give their agents too many privileges.

1562
00:58:02,480 --> 00:58:04,480
Some teams don't monitor their agents properly.

1563
00:58:04,480 --> 00:58:05,920
So quality varies dramatically.

1564
00:58:05,920 --> 00:58:07,360
You end up with some excellent agents

1565
00:58:07,360 --> 00:58:09,840
and some that are total operational disasters.

1566
00:58:09,840 --> 00:58:11,440
The middle ground is a federated model.

1567
00:58:11,440 --> 00:58:13,040
A central team owns the platform

1568
00:58:13,040 --> 00:58:14,240
and sets the guardrails,

1569
00:58:14,240 --> 00:58:16,320
providing templates and tools for everyone else.

1570
00:58:16,320 --> 00:58:18,080
Business units own their specific agents

1571
00:58:18,080 --> 00:58:19,200
within those boundaries.

1572
00:58:19,200 --> 00:58:21,440
The central team defines what you can and cannot do.

1573
00:58:21,440 --> 00:58:24,160
And they provide a process for requesting exceptions.

1574
00:58:24,160 --> 00:58:26,160
Business units build fast because they aren't waiting

1575
00:58:26,160 --> 00:58:28,240
for central approvals on routine builds.

1576
00:58:28,240 --> 00:58:30,720
And quality stays consistent because the boundaries

1577
00:58:30,720 --> 00:58:32,560
enforce the standards.

1578
00:58:32,560 --> 00:58:33,440
Third choice.

1579
00:58:33,440 --> 00:58:35,920
Do you use shared foundation models from vendors

1580
00:58:35,920 --> 00:58:38,400
or deploy models in your own infrastructure?

1581
00:58:38,400 --> 00:58:40,160
Shared models are operationally simpler

1582
00:58:40,160 --> 00:58:41,680
because Microsoft or another provider

1583
00:58:41,680 --> 00:58:42,960
manages the infrastructure.

1584
00:58:42,960 --> 00:58:44,400
You call the APIs and you don't have

1585
00:58:44,400 --> 00:58:46,320
to run data centers or maintain hardware.

1586
00:58:46,320 --> 00:58:47,840
The concern here is your data.

1587
00:58:47,840 --> 00:58:49,840
You are showing the vendor your workflows.

1588
00:58:49,840 --> 00:58:51,680
And you're indirectly training their models

1589
00:58:51,680 --> 00:58:53,920
on your data through your usage patterns.

1590
00:58:53,920 --> 00:58:55,440
The vendor understands your business better

1591
00:58:55,440 --> 00:58:57,040
with every single interaction.

1592
00:58:57,040 --> 00:58:59,200
For some organizations, that's acceptable.

1593
00:58:59,200 --> 00:59:01,280
For others, it's a deal breaker.

1594
00:59:01,280 --> 00:59:04,320
Private model deployment means you own the entire infrastructure.

1595
00:59:04,320 --> 00:59:06,000
You deploy models in your own environment

1596
00:59:06,000 --> 00:59:07,760
so no data ever leaves your boundary.

1597
00:59:07,760 --> 00:59:09,280
You maintain complete control.

1598
00:59:09,280 --> 00:59:10,640
But the cost is much higher.

1599
00:59:10,640 --> 00:59:13,840
You're managing GPU clusters and handling updates yourself.

1600
00:59:13,840 --> 00:59:15,600
You are responsible for reliability.

1601
00:59:15,600 --> 00:59:17,120
And you need infrastructure expertise

1602
00:59:17,120 --> 00:59:18,560
that you might not currently have.

1603
00:59:18,560 --> 00:59:21,360
That trade-off is usually worth it for sensitive workflows

1604
00:59:21,360 --> 00:59:23,760
where your logic is your competitive advantage.

1605
00:59:23,760 --> 00:59:25,840
Many organizations start with shared models

1606
00:59:25,840 --> 00:59:27,280
for standard workflows

1607
00:59:27,280 --> 00:59:30,320
and move to private deployment for their most important tasks.

1608
00:59:30,320 --> 00:59:32,000
This gives you control where it matters

1609
00:59:32,000 --> 00:59:34,000
and simplicity where it doesn't.

1610
00:59:34,000 --> 00:59:34,800
Fourth choice.

1611
00:59:34,800 --> 00:59:35,920
Focus or breath.

1612
00:59:35,920 --> 00:59:38,640
Do you narrow down on two or three high-value use cases

1613
00:59:38,640 --> 00:59:41,280
or try to build agents across your entire operation?

1614
00:59:41,280 --> 00:59:42,800
Focus leads to deeper success.

1615
00:59:42,800 --> 00:59:44,240
You really understand one workflow

1616
00:59:44,240 --> 00:59:46,080
and build an agent that is genuinely excellent.

1617
00:59:46,080 --> 00:59:48,960
You prove the model works and build organizational credibility.

1618
00:59:48,960 --> 00:59:50,400
The risk is limited impact

1619
00:59:50,400 --> 00:59:52,560
because you're only optimizing one workflow

1620
00:59:52,560 --> 00:59:54,480
instead of transforming the whole company.

1621
00:59:54,480 --> 00:59:56,560
Breath means trying to automate everything at once.

1622
00:59:56,560 --> 00:59:58,880
You spread your resources across too many initiatives

1623
00:59:58,880 --> 01:00:01,360
so you get some progress everywhere but excellence nowhere.

1624
01:00:01,360 --> 01:00:03,760
Nothing ever finishes and nothing shows clear value.

1625
01:00:03,760 --> 01:00:06,160
The pattern that actually works is focused breath.

1626
01:00:06,160 --> 01:00:07,920
You pick one or two areas where agents

1627
01:00:07,920 --> 01:00:10,000
will have the maximum impact.

1628
01:00:10,000 --> 01:00:11,040
Build them excellently

1629
01:00:11,040 --> 01:00:14,000
and then expand to adjacent areas once you've proven success.

1630
01:00:14,000 --> 01:00:16,720
These four choices compound into your overall strategy.

1631
01:00:16,720 --> 01:00:19,520
They determine your speed, your control, and your risk.

1632
01:00:19,520 --> 01:00:21,600
They are worth thinking through deliberately

1633
01:00:21,600 --> 01:00:24,480
rather than letting the defaults choose for you.

1634
01:00:24,480 --> 01:00:25,680
The execution roadmap.

1635
01:00:25,680 --> 01:00:27,520
The transition from decision to action

1636
01:00:27,520 --> 01:00:29,760
is where most strategic initiatives collapse.

1637
01:00:29,760 --> 01:00:31,680
You've accepted that agents are necessary

1638
01:00:31,680 --> 01:00:33,440
and you understand the architecture.

1639
01:00:33,440 --> 01:00:34,800
You've even allocated the budget.

1640
01:00:34,800 --> 01:00:36,320
Now you need a sequence of execution

1641
01:00:36,320 --> 01:00:37,600
that produces visible winds

1642
01:00:37,600 --> 01:00:39,680
while building toward a sustainable transformation.

1643
01:00:39,680 --> 01:00:41,440
The organization is moving the fastest.

1644
01:00:41,440 --> 01:00:42,880
Follow a three phase roadmap.

1645
01:00:42,880 --> 01:00:44,560
Though the timeline always varies

1646
01:00:44,560 --> 01:00:46,240
based on your starting position

1647
01:00:46,240 --> 01:00:48,720
and how much change your organization can handle.

1648
01:00:48,720 --> 01:00:50,000
Phase one is the foundation.

1649
01:00:50,000 --> 01:00:51,920
This usually runs three to six months.

1650
01:00:51,920 --> 01:00:53,360
You start with an honest assessment

1651
01:00:53,360 --> 01:00:54,800
of where you actually are today.

1652
01:00:54,800 --> 01:00:55,760
What APIs exist?

1653
01:00:55,760 --> 01:00:57,600
How mature is your data governance?

1654
01:00:57,600 --> 01:00:59,760
Where are the biggest operational bottlenecks?

1655
01:00:59,760 --> 01:01:01,440
You aren't trying to fix everything at once.

1656
01:01:01,440 --> 01:01:03,120
You're looking for the specific place

1657
01:01:03,120 --> 01:01:05,040
where an agent delivers immediate value

1658
01:01:05,040 --> 01:01:07,120
with a reasonable amount of complexity.

1659
01:01:07,120 --> 01:01:08,640
Parallel to that assessment.

1660
01:01:08,640 --> 01:01:10,080
You have to build the platform.

1661
01:01:10,080 --> 01:01:11,440
This means implementing foundry

1662
01:01:11,440 --> 01:01:13,200
and configuring agent 365

1663
01:01:13,200 --> 01:01:15,440
to provide visibility into what your agents are doing.

1664
01:01:15,440 --> 01:01:17,200
You need to establish governance policies

1665
01:01:17,200 --> 01:01:18,400
before you actually need them.

1666
01:01:18,400 --> 01:01:19,600
It's tempting to skip this

1667
01:01:19,600 --> 01:01:21,040
and worry about governance

1668
01:01:21,040 --> 01:01:22,640
after your first agent is built.

1669
01:01:22,640 --> 01:01:23,680
But don't do that.

1670
01:01:23,680 --> 01:01:25,360
An ungoverned agent in production

1671
01:01:25,360 --> 01:01:26,960
is an operational liability.

1672
01:01:26,960 --> 01:01:28,480
Get the infrastructure in place first

1673
01:01:28,480 --> 01:01:30,000
then build the agents within it.

1674
01:01:30,000 --> 01:01:31,360
The pilot is where you prove

1675
01:01:31,360 --> 01:01:34,080
the model actually works in your specific context.

1676
01:01:34,080 --> 01:01:35,360
This isn't a proof of concept

1677
01:01:35,360 --> 01:01:36,640
that everyone forgets about.

1678
01:01:36,640 --> 01:01:37,520
It's a live pilot

1679
01:01:37,520 --> 01:01:39,040
where the agent executes real work

1680
01:01:39,040 --> 01:01:40,480
and produces real results.

1681
01:01:40,480 --> 01:01:42,080
Pick something high value but contained.

1682
01:01:42,080 --> 01:01:43,440
Like a specific document type,

1683
01:01:43,440 --> 01:01:44,640
the agent can categorize

1684
01:01:44,640 --> 01:01:46,880
or a subset of customer inquiries.

1685
01:01:46,880 --> 01:01:49,360
If a workflow currently consumes 20 hours a week,

1686
01:01:49,360 --> 01:01:51,920
that's a perfect candidate for an agent to accelerate,

1687
01:01:51,920 --> 01:01:53,520
deploy the agent to limited traffic

1688
01:01:53,520 --> 01:01:55,360
and let it run for about a month or two.

1689
01:01:55,360 --> 01:01:56,720
Measure everything systematically.

1690
01:01:56,720 --> 01:01:57,680
Is it faster?

1691
01:01:57,680 --> 01:01:58,560
What is the quality?

1692
01:01:58,560 --> 01:02:00,640
This pilot is your truth telling mechanism

1693
01:02:00,640 --> 01:02:02,000
and it surfaces the problems

1694
01:02:02,000 --> 01:02:04,320
before you've committed massive resources.

1695
01:02:04,320 --> 01:02:05,600
Phase two is scale.

1696
01:02:05,600 --> 01:02:07,280
This runs six to 12 months.

1697
01:02:07,280 --> 01:02:08,560
You've proven the model works

1698
01:02:08,560 --> 01:02:10,000
and now you're building momentum.

1699
01:02:10,000 --> 01:02:12,080
You're creating the operational muscle memory

1700
01:02:12,080 --> 01:02:14,080
that makes building agents feel routine

1701
01:02:14,080 --> 01:02:15,040
instead of heroic.

1702
01:02:15,040 --> 01:02:17,200
Establish templates and best practices

1703
01:02:17,200 --> 01:02:20,080
so building the next agent is faster and more consistent.

1704
01:02:20,080 --> 01:02:21,600
Don't reinvent the prompt structure

1705
01:02:21,600 --> 01:02:24,080
or rethink the governance model every single time.

1706
01:02:24,080 --> 01:02:25,440
Use what worked and adapted.

1707
01:02:25,440 --> 01:02:27,120
You should be compounding your learning,

1708
01:02:27,120 --> 01:02:28,480
not starting over from scratch.

1709
01:02:28,480 --> 01:02:30,000
Scale the governance infrastructure

1710
01:02:30,000 --> 01:02:31,520
to handle the increased volume.

1711
01:02:31,520 --> 01:02:33,040
You had one agent in the pilot

1712
01:02:33,040 --> 01:02:34,320
but soon you'll have 50.

1713
01:02:34,320 --> 01:02:35,600
You're monitoring dashboard

1714
01:02:35,600 --> 01:02:37,200
and your evaluation processes

1715
01:02:37,200 --> 01:02:38,640
need to handle that throughput.

1716
01:02:38,640 --> 01:02:40,320
You're not hiring one person per agent.

1717
01:02:40,320 --> 01:02:41,360
You're building systems

1718
01:02:41,360 --> 01:02:44,000
that let one person govern many agents at once.

1719
01:02:44,000 --> 01:02:46,320
Train your business units to build their own agents.

1720
01:02:46,320 --> 01:02:48,880
You aren't trying to make every employee in engineer.

1721
01:02:48,880 --> 01:02:50,560
But the operational leaders who understand

1722
01:02:50,560 --> 01:02:51,840
where the friction lives need to know

1723
01:02:51,840 --> 01:02:53,440
how to collaborate on agent design.

1724
01:02:53,440 --> 01:02:55,360
They don't need to write the prompts themselves.

1725
01:02:55,360 --> 01:02:57,200
They just need to translate their workflows

1726
01:02:57,200 --> 01:02:58,480
into agent requirements.

1727
01:02:58,480 --> 01:02:59,760
That is a learnable skill.

1728
01:02:59,760 --> 01:03:01,120
Phase three is transform.

1729
01:03:01,120 --> 01:03:03,200
This is the phase that never really ends.

1730
01:03:03,200 --> 01:03:04,480
Agents stop being a project

1731
01:03:04,480 --> 01:03:06,880
and become how the organization actually operates.

1732
01:03:06,880 --> 01:03:09,920
You're building agents for increasingly complex workflows

1733
01:03:09,920 --> 01:03:11,760
because you've finally learned how.

1734
01:03:11,760 --> 01:03:13,600
You aren't waiting for central capacity

1735
01:03:13,600 --> 01:03:15,760
or debating whether agents are worthwhile

1736
01:03:15,760 --> 01:03:17,120
that conversation is over.

1737
01:03:17,120 --> 01:03:19,760
Now you're just debating which workflow to automate next.

1738
01:03:19,760 --> 01:03:21,600
Start thinking about how agents enable

1739
01:03:21,600 --> 01:03:23,600
entirely new business capabilities.

1740
01:03:23,600 --> 01:03:26,240
It's not just about faster invoice processing anymore.

1741
01:03:26,240 --> 01:03:29,040
Can agents enable you to serve smaller customers profitably?

1742
01:03:29,040 --> 01:03:30,640
Can they expand your support capacity

1743
01:03:30,640 --> 01:03:31,840
for emerging markets?

1744
01:03:31,840 --> 01:03:33,520
The productivity gains you've made

1745
01:03:33,520 --> 01:03:35,760
become the foundation for your future growth.

1746
01:03:35,760 --> 01:03:37,120
Throughout all three phases.

1747
01:03:37,120 --> 01:03:39,440
You are also modernizing your underlying architecture.

1748
01:03:39,440 --> 01:03:41,280
You're building APIs where they don't exist

1749
01:03:41,280 --> 01:03:42,720
and improving your data governance.

1750
01:03:42,720 --> 01:03:43,920
This isn't a side project.

1751
01:03:43,920 --> 01:03:46,800
It's the structural change that makes the agent platform actually work.

1752
01:03:46,800 --> 01:03:49,440
You cannot run sophisticated agents on infrastructure

1753
01:03:49,440 --> 01:03:50,880
that was built for human users.

1754
01:03:50,880 --> 01:03:52,800
You have to evolve the infrastructure.

1755
01:03:52,800 --> 01:03:55,360
The execution roadmap is where your strategy

1756
01:03:55,360 --> 01:03:57,520
finally becomes a reality.

1757
01:03:57,520 --> 01:03:58,560
He culture shift.

1758
01:03:58,560 --> 01:04:00,640
The biggest barrier to agents isn't the tech.

1759
01:04:00,640 --> 01:04:01,440
It's the culture.

1760
01:04:03,040 --> 01:04:06,080
For 30 years, we operated on one fixed assumption.

1761
01:04:06,080 --> 01:04:07,040
Humans do the work.

1762
01:04:07,040 --> 01:04:08,480
Software helps them do it better.

1763
01:04:08,480 --> 01:04:10,000
We built everything around that idea.

1764
01:04:10,000 --> 01:04:10,800
We hired for it.

1765
01:04:10,800 --> 01:04:11,600
We trained for it.

1766
01:04:11,600 --> 01:04:13,840
We built entire career paths and identities

1767
01:04:13,840 --> 01:04:15,280
around human expertise.

1768
01:04:15,280 --> 01:04:17,040
Now that assumption is inverting.

1769
01:04:17,040 --> 01:04:18,160
Agents do the work.

1770
01:04:18,160 --> 01:04:19,760
Humans oversee and guide them.

1771
01:04:19,760 --> 01:04:22,080
For most people, this doesn't feel like an improvement.

1772
01:04:22,080 --> 01:04:23,440
It feels like a threat.

1773
01:04:23,440 --> 01:04:26,480
The people most at risk aren't the ones doing high-level strategy.

1774
01:04:26,480 --> 01:04:28,480
They're the ones executing the routine tasks.

1775
01:04:28,480 --> 01:04:29,680
Invoice processors.

1776
01:04:29,680 --> 01:04:31,440
See agents handling the billing.

1777
01:04:31,440 --> 01:04:33,840
Customer service reps see agents taking the calls.

1778
01:04:33,840 --> 01:04:37,040
Data analysts see agents answering the questions they used to answer.

1779
01:04:37,040 --> 01:04:38,800
These fears aren't a sign of paranoia.

1780
01:04:38,800 --> 01:04:40,480
They're a rational response to a massive shift.

1781
01:04:40,480 --> 01:04:41,920
Some of these roles will disappear.

1782
01:04:41,920 --> 01:04:43,440
Others will be completely reshaped.

1783
01:04:43,440 --> 01:04:45,120
The people in those seats know it.

1784
01:04:45,120 --> 01:04:48,080
And they're watching the technology very closely for that exact reason.

1785
01:04:48,080 --> 01:04:49,920
This is where leadership actually matters.

1786
01:04:49,920 --> 01:04:51,520
Not with inspirational speeches,

1787
01:04:51,520 --> 01:04:52,960
but with honest specific talk.

1788
01:04:52,960 --> 01:04:55,040
You have to say, here is what's changing.

1789
01:04:55,040 --> 01:04:56,320
Here is how we're planning for it.

1790
01:04:56,320 --> 01:04:58,720
Here is what we expect from you.

1791
01:04:58,720 --> 01:04:59,920
And most importantly,

1792
01:04:59,920 --> 01:05:02,320
here is how we're investing in you so you can adapt.

1793
01:05:02,320 --> 01:05:05,600
The companies moving the fastest are transparent about displacement.

1794
01:05:05,600 --> 01:05:06,880
They don't sugar-coded.

1795
01:05:06,880 --> 01:05:08,800
They admit that certain work patterns are going away.

1796
01:05:08,800 --> 01:05:12,000
But then, they immediately show what's being built instead.

1797
01:05:12,000 --> 01:05:13,520
New roles are appearing right now.

1798
01:05:13,520 --> 01:05:14,480
Agent trainers.

1799
01:05:14,480 --> 01:05:16,000
People who ordered agent performance.

1800
01:05:16,000 --> 01:05:18,880
People who refine instructions based on what the agent is learning.

1801
01:05:18,880 --> 01:05:21,920
People who handle the complex workflows that agents can't touch yet.

1802
01:05:21,920 --> 01:05:23,440
These jobs didn't exist two years ago.

1803
01:05:23,440 --> 01:05:24,320
They exist now.

1804
01:05:24,320 --> 01:05:25,040
They pay well.

1805
01:05:25,040 --> 01:05:26,480
And while they require new skills,

1806
01:05:26,480 --> 01:05:28,240
those skills aren't impossible to learn.

1807
01:05:28,240 --> 01:05:32,160
The investment in re-skilling is the line between being serious and just experimenting.

1808
01:05:32,160 --> 01:05:34,400
Serious organizations put training in the budget.

1809
01:05:34,400 --> 01:05:37,680
They set up mentorships between the old guard and the new practitioners.

1810
01:05:37,680 --> 01:05:39,760
They provide real career counseling.

1811
01:05:39,760 --> 01:05:43,200
They help people see that while the market for old skills is shrinking,

1812
01:05:43,200 --> 01:05:45,040
the market for new skills is exploding.

1813
01:05:45,040 --> 01:05:46,480
They fund the certifications.

1814
01:05:46,480 --> 01:05:48,720
They create a path where an invoice processor

1815
01:05:48,720 --> 01:05:51,280
becomes the person who validates the invoice agents.

1816
01:05:51,280 --> 01:05:52,240
That's not a demotion.

1817
01:05:52,240 --> 01:05:53,280
It pays just as well.

1818
01:05:53,280 --> 01:05:54,640
It has more growth potential.

1819
01:05:54,640 --> 01:05:56,160
It just requires a different mindset.

1820
01:05:56,160 --> 01:05:58,800
This shift also changes how we look at failure.

1821
01:05:58,800 --> 01:06:00,080
In traditional software.

1822
01:06:00,080 --> 01:06:01,040
Failure is bad.

1823
01:06:01,040 --> 01:06:02,800
A bug in production is a problem.

1824
01:06:02,800 --> 01:06:03,840
Someone messed up.

1825
01:06:03,840 --> 01:06:05,680
There's a meeting to find out who's to blame.

1826
01:06:05,680 --> 01:06:07,760
With agents, failures are just experiments.

1827
01:06:07,760 --> 01:06:09,680
If the agent tries something and it doesn't work,

1828
01:06:09,680 --> 01:06:10,560
that's just information.

1829
01:06:10,560 --> 01:06:14,080
You learn something about how it reads instructions or what data it needs.

1830
01:06:14,080 --> 01:06:16,880
You take that lesson and you put it into the next version.

1831
01:06:16,880 --> 01:06:18,960
If the agent makes a mistake, nobody saw coming.

1832
01:06:18,960 --> 01:06:20,480
That's valuable data on an edge case.

1833
01:06:20,480 --> 01:06:23,040
You build a guardrail to stop it from happening again.

1834
01:06:23,040 --> 01:06:25,600
Failure becomes feedback, not shame.

1835
01:06:25,600 --> 01:06:27,760
This is a massive change for most companies.

1836
01:06:27,760 --> 01:06:30,400
It requires leaders who know the difference between being sloppy

1837
01:06:30,400 --> 01:06:31,520
and being a learner.

1838
01:06:31,520 --> 01:06:34,480
Deploying an agent without any oversight is being sloppy.

1839
01:06:34,480 --> 01:06:36,560
But an agent hitting an unexpected wall

1840
01:06:36,560 --> 01:06:38,240
while operating inside its limits.

1841
01:06:38,240 --> 01:06:39,040
That's learning.

1842
01:06:39,040 --> 01:06:40,400
The distinction is everything.

1843
01:06:40,400 --> 01:06:41,840
The enterprises winning right now

1844
01:06:41,840 --> 01:06:43,840
are the ones that already have this learning culture.

1845
01:06:43,840 --> 01:06:45,040
They've done the agile work.

1846
01:06:45,040 --> 01:06:46,800
They've embraced continuous deployment.

1847
01:06:46,800 --> 01:06:49,440
They built organizations where testing is expected.

1848
01:06:49,440 --> 01:06:51,280
We're small mistakes teach big lessons.

1849
01:06:51,280 --> 01:06:54,640
And where failing fast is a reality, not a poster on the wall.

1850
01:06:54,640 --> 01:06:56,720
For them agents are just the next step.

1851
01:06:56,720 --> 01:06:59,120
But for companies still stuck in command and control

1852
01:06:59,120 --> 01:07:02,400
where mistakes are punished, agents are going to feel dangerous at every level.

1853
01:07:02,400 --> 01:07:05,520
The future operating model, we're moving from a world

1854
01:07:05,520 --> 01:07:07,040
where humans navigate software

1855
01:07:07,040 --> 01:07:08,800
to a world where agents navigate software.

1856
01:07:08,800 --> 01:07:10,000
The impact is structural.

1857
01:07:10,000 --> 01:07:11,520
It isn't just a surface change.

1858
01:07:11,520 --> 01:07:14,240
The user interface is becoming a legacy artifact.

1859
01:07:14,240 --> 01:07:16,400
The seat-based pricing model is becoming a relic.

1860
01:07:16,400 --> 01:07:18,640
Work flows through agents instead of through dashboards.

1861
01:07:18,640 --> 01:07:20,720
Humans stay at the top, guiding the process

1862
01:07:20,720 --> 01:07:23,120
and making the calls where nuance actually matters.

1863
01:07:23,120 --> 01:07:25,840
Software is being built for machines to consume first.

1864
01:07:25,840 --> 01:07:28,320
Organizations are being structured around who owns the agent

1865
01:07:28,320 --> 01:07:29,600
and who governs the output.

1866
01:07:29,600 --> 01:07:32,720
Your competitive advantage won't come from your software features.

1867
01:07:32,720 --> 01:07:34,560
It will come from your workflow capital,

1868
01:07:34,560 --> 01:07:36,720
the institutional logic that your agents carry.

1869
01:07:36,720 --> 01:07:38,400
The companies that win in this model

1870
01:07:38,400 --> 01:07:39,920
are the ones moving the fastest.

1871
01:07:39,920 --> 01:07:41,360
It's not because they're smarter.

1872
01:07:41,360 --> 01:07:45,440
It's because they realize this is an organizational change disguised as a tech change.

1873
01:07:45,440 --> 01:07:46,480
They invest in their people.

1874
01:07:46,480 --> 01:07:48,240
They build the right architecture.

1875
01:07:48,240 --> 01:07:49,440
They set up the governance.

1876
01:07:49,440 --> 01:07:51,360
They accept that this is a decade of work,

1877
01:07:51,360 --> 01:07:52,640
not a two-year project.

1878
01:07:52,640 --> 01:07:55,200
By 2030, the organizations that haven't made this move

1879
01:07:55,200 --> 01:07:56,560
will be at a massive disadvantage.

1880
01:07:56,560 --> 01:07:58,160
Their software will become a constraint.

1881
01:07:58,160 --> 01:07:59,920
Their costs will make them uncompetitive.

1882
01:07:59,920 --> 01:08:02,320
Their ability to innovate will be choked by infrastructure

1883
01:08:02,320 --> 01:08:04,400
built for a generation that's already over.

1884
01:08:04,400 --> 01:08:05,680
But you can start today.

1885
01:08:05,680 --> 01:08:09,440
One agent, one workflow, learn, iterate, build.

1886
01:08:09,440 --> 01:08:11,920
The question isn't whether you should move toward agents.

1887
01:08:11,920 --> 01:08:15,280
The question is how much time you're willing to lose before you do.