The Copilot Tax: Why Your AI Strategy is Bleeding Cash

1
00:00:00,000 --> 00:00:03,100
Your CFO just asked the question nobody on your team can answer.

2
00:00:03,100 --> 00:00:05,000
What exactly are we paying for?

3
00:00:05,000 --> 00:00:06,900
Not a trick question, a real one.

4
00:00:06,900 --> 00:00:09,260
18 months into your co-pilot deployment,

5
00:00:09,260 --> 00:00:12,940
the promise was two to five hours saved per week per user.

6
00:00:12,940 --> 00:00:14,700
The invoice says something different,

7
00:00:14,700 --> 00:00:16,340
and somewhere between the pilot presentation

8
00:00:16,340 --> 00:00:18,940
and the quarterly budget review three hidden cost layers

9
00:00:18,940 --> 00:00:20,860
appeared that nobody warned you about.

10
00:00:20,860 --> 00:00:22,140
By the end of this episode,

11
00:00:22,140 --> 00:00:23,980
you'll have a framework to audit every dollar

12
00:00:23,980 --> 00:00:26,260
your organization is spending on AI,

13
00:00:26,260 --> 00:00:29,100
and a clear picture of where that money is actually going.

14
00:00:29,100 --> 00:00:31,100
The entry fee nobody reads carefully.

15
00:00:31,100 --> 00:00:32,700
Let's start with the number everyone knows.

16
00:00:32,700 --> 00:00:34,220
$30 per user per month.

17
00:00:34,220 --> 00:00:35,060
It sounds simple.

18
00:00:35,060 --> 00:00:36,980
It isn't that $30 is an add-on.

19
00:00:36,980 --> 00:00:39,460
It sits on top of whatever your organization already pays

20
00:00:39,460 --> 00:00:41,140
for Microsoft 365.

21
00:00:41,140 --> 00:00:42,980
And if you're a mid to large enterprise,

22
00:00:42,980 --> 00:00:45,860
that base is almost certainly E3 or E5.

23
00:00:45,860 --> 00:00:47,900
E3 runs roughly $36 per user per month.

24
00:00:47,900 --> 00:00:49,540
E5 runs around $57.

25
00:00:49,540 --> 00:00:52,100
So before a single co-pilot prompt gets typed,

26
00:00:52,100 --> 00:00:55,500
you're already at $66 to $87 per user per month

27
00:00:55,500 --> 00:00:56,980
just to reach the starting line.

28
00:00:56,980 --> 00:00:58,660
And the starting line keeps moving.

29
00:00:58,660 --> 00:01:00,700
As of July 1st, 2026,

30
00:01:00,700 --> 00:01:02,900
Microsoft has implemented list price increases

31
00:01:02,900 --> 00:01:06,340
of nine to 33% across most business and enterprise plans.

32
00:01:06,340 --> 00:01:08,900
Business basic went from $6 to $7.50.

33
00:01:08,900 --> 00:01:11,900
The frontline F1 license went from $2.25 to $3.

34
00:01:11,900 --> 00:01:15,060
That's a 33% increase for your lowest cost users.

35
00:01:15,060 --> 00:01:17,740
E3 and E5 are up roughly nine to 10%.

36
00:01:17,740 --> 00:01:19,820
And these aren't negotiable the way they used to be

37
00:01:19,820 --> 00:01:22,420
because Microsoft eliminated enterprise agreement volume

38
00:01:22,420 --> 00:01:24,860
discounts in November 2025.

39
00:01:24,860 --> 00:01:27,420
Every customer now pays level A list price

40
00:01:27,420 --> 00:01:29,660
regardless of how many seats they buy.

41
00:01:29,660 --> 00:01:31,300
The leverage that large enterprises used

42
00:01:31,300 --> 00:01:33,420
to have in renewal negotiations is gone.

43
00:01:33,420 --> 00:01:36,380
Now Microsoft has introduced the E7 Frontier Suite

44
00:01:36,380 --> 00:01:38,460
at $99 per user per month.

45
00:01:38,460 --> 00:01:42,100
It bundles E5 M365 co-pilot, the Entra Suite,

46
00:01:42,100 --> 00:01:45,340
and Agent 365 into a single SQ marketed

47
00:01:45,340 --> 00:01:49,100
as roughly 15% cheaper than buying each component separately.

48
00:01:49,100 --> 00:01:50,540
And that's probably accurate.

49
00:01:50,540 --> 00:01:52,220
But here's what the math doesn't show.

50
00:01:52,220 --> 00:01:54,340
You're now committing to a single vendor

51
00:01:54,340 --> 00:01:57,580
for identity, security, productivity, and AI

52
00:01:57,580 --> 00:01:58,900
in one multi-year contract.

53
00:01:58,900 --> 00:01:59,820
But that's not a discount.

54
00:01:59,820 --> 00:02:01,020
That's a dependency.

55
00:02:01,020 --> 00:02:04,380
Analyst watching Microsoft's commercial strategy

56
00:02:04,380 --> 00:02:07,980
since 2023 have summarized it in one phrase,

57
00:02:07,980 --> 00:02:09,940
charging more and giving less flexibility.

58
00:02:09,940 --> 00:02:12,740
The elimination of tiered volume pricing,

59
00:02:12,740 --> 00:02:14,860
the bundling of AI into high cost suites,

60
00:02:14,860 --> 00:02:17,660
the removal of standalone co-pilot-profile consumers.

61
00:02:17,660 --> 00:02:19,300
These aren't isolated decisions.

62
00:02:19,300 --> 00:02:21,180
They're a coherent commercial strategy

63
00:02:21,180 --> 00:02:25,420
to normalize AI spend as a fixed rising cost of doing business.

64
00:02:25,420 --> 00:02:28,020
For finance teams, the shock comes at renewal.

65
00:02:28,020 --> 00:02:31,540
Organizations are seeing 40 to 50% effective price increases

66
00:02:31,540 --> 00:02:33,500
when they account for both the list price hikes

67
00:02:33,500 --> 00:02:35,620
and the loss of historical discount structures.

68
00:02:35,620 --> 00:02:38,980
A company that was paying $45 per user per month two years ago

69
00:02:38,980 --> 00:02:41,420
can now be looking at $65 to $75

70
00:02:41,420 --> 00:02:43,820
before co-pilot even enters the conversation.

71
00:02:43,820 --> 00:02:44,940
Here's the critical point.

72
00:02:44,940 --> 00:02:46,500
None of this is the co-pilot tax.

73
00:02:46,500 --> 00:02:47,500
This is just admission.

74
00:02:47,500 --> 00:02:50,500
The base license, the E5 stack, the Frontier Suite commitment,

75
00:02:50,500 --> 00:02:52,460
this is what you pay to get into the building.

76
00:02:52,460 --> 00:02:54,500
The tax is what happens once you're inside

77
00:02:54,500 --> 00:02:56,940
when the billing model shifts from predictable seats

78
00:02:56,940 --> 00:02:58,460
to something far less visible.

79
00:02:58,460 --> 00:03:00,140
The seat cost is on your invoice.

80
00:03:00,140 --> 00:03:01,980
What comes next, isn't.

81
00:03:01,980 --> 00:03:04,140
Two billing systems running simultaneously,

82
00:03:04,140 --> 00:03:05,820
so you've accepted the entry fee.

83
00:03:05,820 --> 00:03:07,140
You know what the seats cost.

84
00:03:07,140 --> 00:03:09,700
And now you think you know what AI costs your organization.

85
00:03:09,700 --> 00:03:11,820
You don't because Microsoft doesn't run one billing system

86
00:03:11,820 --> 00:03:12,660
for AI.

87
00:03:12,660 --> 00:03:15,660
It runs two and they operate on completely different logic.

88
00:03:15,660 --> 00:03:17,420
The first is licensing based.

89
00:03:17,420 --> 00:03:19,740
This is the world most finance teams understand.

90
00:03:19,740 --> 00:03:21,940
You assign the seat, you pay a fixed monthly rate,

91
00:03:21,940 --> 00:03:24,380
you forecast it by multiplying headcount by price.

92
00:03:24,380 --> 00:03:27,180
M365 co-pilot falls here, so does GitHub co-pilot,

93
00:03:27,180 --> 00:03:30,460
so does Teams premium, predictable, structured, easy to model.

94
00:03:30,460 --> 00:03:33,180
The number on the invoice matches the number in the spreadsheet.

95
00:03:33,180 --> 00:03:35,020
The second system is consumption based.

96
00:03:35,020 --> 00:03:36,780
This is where Azure Open Eye tokens live,

97
00:03:36,780 --> 00:03:38,180
where co-pilot credits get billed,

98
00:03:38,180 --> 00:03:40,020
where security compute units get consumed.

99
00:03:40,020 --> 00:03:41,340
Nothing here is fixed.

100
00:03:41,340 --> 00:03:44,340
Everything scales with usage, with how many prompts get fired,

101
00:03:44,340 --> 00:03:45,980
how long the context windows are,

102
00:03:45,980 --> 00:03:48,700
how many agent loops complete before a task resolves.

103
00:03:48,700 --> 00:03:50,940
The billing is variable, the visibility is low,

104
00:03:50,940 --> 00:03:52,740
and the reconciliation is manual.

105
00:03:52,740 --> 00:03:54,420
Now here's where it gets complicated.

106
00:03:54,420 --> 00:03:56,100
Some of Microsoft's products don't fit

107
00:03:56,100 --> 00:03:57,660
cleanly into either system.

108
00:03:57,660 --> 00:04:00,100
Co-pilot studio for example hits both at once.

109
00:04:00,100 --> 00:04:02,460
You pay a per user license to access the platform.

110
00:04:02,460 --> 00:04:04,180
Then the agents you build on that platform,

111
00:04:04,180 --> 00:04:06,460
consume co-pilot credits every time they run.

112
00:04:06,460 --> 00:04:09,340
And those credits appear not in your M365 admin center,

113
00:04:09,340 --> 00:04:10,580
but on your Azure bill.

114
00:04:10,580 --> 00:04:13,740
Two separate invoices, two separate approval chains, one product.

115
00:04:13,740 --> 00:04:15,860
The same logic applies to security co-pilot,

116
00:04:15,860 --> 00:04:19,700
which Microsoft bundled into M365 e5 in January, 2026,

117
00:04:19,700 --> 00:04:22,180
with a monthly security compute unit allocation.

118
00:04:22,180 --> 00:04:24,780
Use more than your allocation and you spill into variable

119
00:04:24,780 --> 00:04:25,940
as your consumption.

120
00:04:25,940 --> 00:04:28,340
A security analyst running a complex threat investigation

121
00:04:28,340 --> 00:04:30,100
can exhaust a month's SCU budget

122
00:04:30,100 --> 00:04:32,260
in a single afternoon of intensive queries,

123
00:04:32,260 --> 00:04:33,780
and that overrate shows up somewhere

124
00:04:33,780 --> 00:04:35,940
other than where your IT team is looking.

125
00:04:35,940 --> 00:04:38,180
This fragmentation has a practical consequence

126
00:04:38,180 --> 00:04:40,220
that most organizations don't fully appreciate

127
00:04:40,220 --> 00:04:42,220
until they're sitting in a budget review trying

128
00:04:42,220 --> 00:04:43,420
to explain a spike.

129
00:04:43,420 --> 00:04:45,980
There is no unified AI consumption dashboard.

130
00:04:45,980 --> 00:04:49,460
Finance teams reconciling what AI actually cost last quarter

131
00:04:49,460 --> 00:04:52,100
have to pull reports from at least three separate places.

132
00:04:52,100 --> 00:04:55,220
The M365 admin center for seat level licensing,

133
00:04:55,220 --> 00:04:56,860
the Azure cost management console

134
00:04:56,860 --> 00:04:58,660
for token and compute consumption,

135
00:04:58,660 --> 00:05:02,340
and co-pilot studios analytics for agent level activity.

136
00:05:02,340 --> 00:05:04,180
These reports don't automatically align.

137
00:05:04,180 --> 00:05:06,860
The data formats differ, the attribution logic differs,

138
00:05:06,860 --> 00:05:08,420
and the billing cycles can differ.

139
00:05:08,420 --> 00:05:11,740
One user interaction can trigger all three simultaneously.

140
00:05:11,740 --> 00:05:14,340
A knowledge worker asks co-pilot to summarize a document

141
00:05:14,340 --> 00:05:15,700
and draft a response.

142
00:05:15,700 --> 00:05:18,700
That request touches the M365 co-pilot license.

143
00:05:18,700 --> 00:05:20,620
If the response pulls from a custom agent

144
00:05:20,620 --> 00:05:23,380
built in co-pilot studio, it consumes credits.

145
00:05:23,380 --> 00:05:25,460
If the underlying model call routes through a custom

146
00:05:25,460 --> 00:05:28,060
Azure open AI deployment, it burns tokens

147
00:05:28,060 --> 00:05:29,300
built directly to Azure.

148
00:05:29,300 --> 00:05:30,940
The user sees one interaction.

149
00:05:30,940 --> 00:05:33,900
The CFO sees three cost lines on three separate invoices.

150
00:05:33,900 --> 00:05:35,940
This isn't an accident of product design.

151
00:05:35,940 --> 00:05:38,620
It's a consequence of Microsoft building AI capabilities

152
00:05:38,620 --> 00:05:41,700
across an ecosystem that predates the AI era

153
00:05:41,700 --> 00:05:43,340
then layering new consumption models

154
00:05:43,340 --> 00:05:45,540
on top of legacy licensing structures.

155
00:05:45,540 --> 00:05:47,380
The architecture reflects how these products

156
00:05:47,380 --> 00:05:48,900
were acquired and integrated,

157
00:05:48,900 --> 00:05:50,940
not how a coherent billing system would be designed

158
00:05:50,940 --> 00:05:51,980
from scratch.

159
00:05:51,980 --> 00:05:53,860
The practical result is straightforward.

160
00:05:53,860 --> 00:05:56,660
Most organizations believe they know what AI costs them.

161
00:05:56,660 --> 00:05:58,740
They know the seat count, they know the per user rate,

162
00:05:58,740 --> 00:06:00,620
what they're missing is everything underneath.

163
00:06:00,620 --> 00:06:02,900
The variable consumption layer that runs in parallel,

164
00:06:02,900 --> 00:06:05,500
charges differently, and reports separately.

165
00:06:05,500 --> 00:06:07,580
That's the layer where the real money disappears.

166
00:06:07,580 --> 00:06:09,580
What a token actually costs you.

167
00:06:09,580 --> 00:06:11,020
To understand where the money goes,

168
00:06:11,020 --> 00:06:13,380
you first need to understand the unit of consumption.

169
00:06:13,380 --> 00:06:15,540
And for AI, that unit is a token.

170
00:06:15,540 --> 00:06:18,100
A token is roughly three to four characters of English text,

171
00:06:18,100 --> 00:06:20,260
not a word, not a sentence, a fragment.

172
00:06:20,260 --> 00:06:22,380
The word enterprise is about two tokens.

173
00:06:22,380 --> 00:06:25,940
A full paragraph is somewhere between 75 and 150 tokens

174
00:06:25,940 --> 00:06:27,860
depending on word length and punctuation.

175
00:06:27,860 --> 00:06:31,100
A standard page of business text, around 750 words,

176
00:06:31,100 --> 00:06:33,420
runs approximately 1,000 tokens.

177
00:06:33,420 --> 00:06:35,540
A dense technical document with longer words,

178
00:06:35,540 --> 00:06:37,900
code snippets, or specialized terminology

179
00:06:37,900 --> 00:06:40,540
can push 300 to 500 tokens per page.

180
00:06:40,540 --> 00:06:42,540
This matters because every co-pilot interaction

181
00:06:42,540 --> 00:06:45,580
has two token components, what goes in and what comes out.

182
00:06:45,580 --> 00:06:48,820
Input tokens carry the prompt, the context, any documents

183
00:06:48,820 --> 00:06:51,380
you've attached, any previous conversation history

184
00:06:51,380 --> 00:06:52,180
in the session.

185
00:06:52,180 --> 00:06:54,620
Output tokens are the response the model generates,

186
00:06:54,620 --> 00:06:56,100
and they're priced differently.

187
00:06:56,100 --> 00:06:58,500
On Azure OpenAI, where token pricing is actually visible

188
00:06:58,500 --> 00:07:02,900
to customers, GPT-5 runs 1.25 per million input tokens,

189
00:07:02,900 --> 00:07:04,980
and $10 per million output tokens.

190
00:07:04,980 --> 00:07:07,940
That's an 8x price differential between reading and writing.

191
00:07:07,940 --> 00:07:09,940
Output tokens aren't just more expensive.

192
00:07:09,940 --> 00:07:11,500
They're often where the waste concentrates

193
00:07:11,500 --> 00:07:14,220
because a vague prompt generates a long verbose response.

194
00:07:14,220 --> 00:07:15,260
More on that shortly.

195
00:07:15,260 --> 00:07:18,700
Now, M365 co-pilot doesn't expose this math to you.

196
00:07:18,700 --> 00:07:21,980
The $30 monthly fee abstracts all of it into a flat rate,

197
00:07:21,980 --> 00:07:25,380
and Microsoft manages the underlying token economics internally.

198
00:07:25,380 --> 00:07:26,540
But those economics are real,

199
00:07:26,540 --> 00:07:28,620
and they determine how Microsoft makes decisions

200
00:07:28,620 --> 00:07:30,300
about model selection, feature rollout,

201
00:07:30,300 --> 00:07:32,780
and the caps built into different licensed tiers.

202
00:07:32,780 --> 00:07:35,700
The abstraction benefits Microsoft's simplicity narrative.

203
00:07:35,700 --> 00:07:37,580
It doesn't necessarily benefit yours.

204
00:07:37,580 --> 00:07:39,940
Here's why the hidden economics matter for your organization

205
00:07:39,940 --> 00:07:41,540
specifically.

206
00:07:41,540 --> 00:07:44,500
A light knowledge worker, someone using co-pilot occasionally,

207
00:07:44,500 --> 00:07:47,780
maybe summarizing a few emails, running a quick document search,

208
00:07:47,780 --> 00:07:49,660
generating a short status update,

209
00:07:49,660 --> 00:07:53,700
consumes somewhere between 40,000 and 100,000 tokens per month.

210
00:07:53,700 --> 00:07:55,700
That's the low end of the usage curve.

211
00:07:55,700 --> 00:07:59,180
A heavy analytical user, a consultant building decks from research,

212
00:07:59,180 --> 00:08:01,740
a lawyer reviewing contracts, a financial analyst running

213
00:08:01,740 --> 00:08:04,100
iterative Excel queries with complex context,

214
00:08:04,100 --> 00:08:06,900
can reach 1.6 million tokens per month.

215
00:08:06,900 --> 00:08:10,780
That's a 16X to 40X usage gap between the lightest and heaviest users.

216
00:08:10,780 --> 00:08:14,660
And under flat per user licensing, they both pay exactly $30.

217
00:08:14,660 --> 00:08:17,140
What this creates is a structural cross subsidy.

218
00:08:17,140 --> 00:08:19,620
Your light users subsidize your heavy users.

219
00:08:19,620 --> 00:08:23,180
Microsoft prices the flat license around an assumed average distribution

220
00:08:23,180 --> 00:08:24,820
of usage across the organization.

221
00:08:24,820 --> 00:08:27,860
If your workforce skews heavily toward light users,

222
00:08:27,860 --> 00:08:31,140
which is true for most organizations deploying co-pilot broadly,

223
00:08:31,140 --> 00:08:33,420
you're paying frontier model prices for workloads

224
00:08:33,420 --> 00:08:35,700
that don't justify frontier model economics.

225
00:08:35,700 --> 00:08:38,620
You're covering the cost of summarizing one email thread per day

226
00:08:38,620 --> 00:08:41,380
with pricing designed for someone processing hundreds of pages

227
00:08:41,380 --> 00:08:42,460
of legal discovery.

228
00:08:42,460 --> 00:08:44,420
The inverse problem also exists.

229
00:08:44,420 --> 00:08:47,900
If your organization has a small concentration of extremely heavy users,

230
00:08:47,900 --> 00:08:49,620
developers running agente workflows,

231
00:08:49,620 --> 00:08:51,780
analysts processing large document sets,

232
00:08:51,780 --> 00:08:53,780
researchers doing intensive retrieval,

233
00:08:53,780 --> 00:08:55,900
those users may be generating token volumes

234
00:08:55,900 --> 00:09:00,060
that cost Microsoft far more to serve than the $30 license covers.

235
00:09:00,060 --> 00:09:03,020
Which is one reason Microsoft is moving consumption heavy products

236
00:09:03,020 --> 00:09:06,420
like GitHub co-pilot and co-pilot studio to what explicit meter billing.

237
00:09:06,420 --> 00:09:07,820
The flat license isn't permanent.

238
00:09:07,820 --> 00:09:11,260
It's a transitional model on the way to a consumption first architecture.

239
00:09:11,260 --> 00:09:14,420
Security co-pilot already uses security compute units.

240
00:09:14,420 --> 00:09:17,580
GitHub co-pilot switched to AI credits in June, 2026.

241
00:09:17,580 --> 00:09:18,580
The pattern is clear.

242
00:09:18,580 --> 00:09:21,660
As Microsoft's internal token costs become more variable

243
00:09:21,660 --> 00:09:23,300
with frontier model usage,

244
00:09:23,300 --> 00:09:27,020
they push that variability towards the customer through meter tiers.

245
00:09:27,020 --> 00:09:28,900
So even if you're currently paying a flat fee

246
00:09:28,900 --> 00:09:30,860
and token costs feel invisible,

247
00:09:30,860 --> 00:09:33,580
what you're seeing now is the pricing model before full transparency,

248
00:09:33,580 --> 00:09:34,420
not after it.

249
00:09:34,420 --> 00:09:35,660
The tokens are already there.

250
00:09:35,660 --> 00:09:38,020
The bill just hasn't fully arrived yet.

251
00:09:38,020 --> 00:09:39,660
The idle token problem.

252
00:09:39,660 --> 00:09:41,900
Knowing what tokens cost is one problem,

253
00:09:41,900 --> 00:09:44,220
knowing how many you're wasting is a different one entirely.

254
00:09:44,220 --> 00:09:47,460
The term idle token waste doesn't appear on any invoice.

255
00:09:47,460 --> 00:09:48,540
It's not a line item,

256
00:09:48,540 --> 00:09:50,420
Microsoft surfaces to admins,

257
00:09:50,420 --> 00:09:52,420
and there's no dashboard that calculates it for you.

258
00:09:52,420 --> 00:09:54,100
But it's real, it's measurable and aggregate,

259
00:09:54,100 --> 00:09:57,100
and it's almost certainly one of the largest controllable cost variables

260
00:09:57,100 --> 00:09:58,860
in your AI deployment.

261
00:09:58,860 --> 00:10:01,020
Idl token waste is tokens that get billed,

262
00:10:01,020 --> 00:10:02,740
but don't contribute meaningful value

263
00:10:02,740 --> 00:10:05,500
to any output the user actually sees or uses.

264
00:10:05,500 --> 00:10:06,900
Four categories drive most of it.

265
00:10:06,900 --> 00:10:09,660
The first is background completions that never reach the screen.

266
00:10:09,660 --> 00:10:11,180
The model generates a suggestion.

267
00:10:11,180 --> 00:10:13,620
The user switches tabs, closes the window,

268
00:10:13,620 --> 00:10:15,500
or types over it before it renders.

269
00:10:15,500 --> 00:10:16,940
The inference already ran.

270
00:10:16,940 --> 00:10:18,460
The tokens are already spent.

271
00:10:18,460 --> 00:10:20,020
The second is suggestions that display

272
00:10:20,020 --> 00:10:21,660
but get discarded immediately.

273
00:10:21,660 --> 00:10:23,860
The co-pilot auto-complete that appears,

274
00:10:23,860 --> 00:10:25,260
gets read in half a second,

275
00:10:25,260 --> 00:10:27,180
and gets rejected with a backspace.

276
00:10:27,180 --> 00:10:29,700
Generation cost is identical whether the suggestion takes

277
00:10:29,700 --> 00:10:31,380
10 seconds to read or zero.

278
00:10:31,380 --> 00:10:33,620
The third is unnecessary context overhead.

279
00:10:33,620 --> 00:10:35,380
Large persistent instruction blocks,

280
00:10:35,380 --> 00:10:37,780
system prompts injected on every single request,

281
00:10:37,780 --> 00:10:39,980
entire repository indexes that get scanned

282
00:10:39,980 --> 00:10:42,580
even when the task only touches two files.

283
00:10:42,580 --> 00:10:44,620
This context accumulates quietly in the background

284
00:10:44,620 --> 00:10:46,420
of every interaction and charges

285
00:10:46,420 --> 00:10:48,540
at the full input token rate every time.

286
00:10:48,540 --> 00:10:51,780
The fourth and the most expensive is a genetic chatter.

287
00:10:51,780 --> 00:10:53,580
Agent workflows that loop through planning,

288
00:10:53,580 --> 00:10:55,060
calling, evaluating, and replaning

289
00:10:55,060 --> 00:10:56,860
without converging on a better result.

290
00:10:56,860 --> 00:10:58,860
Each loop is a full token event.

291
00:10:58,860 --> 00:11:01,340
A poorly designed agent that runs six planning cycles

292
00:11:01,340 --> 00:11:03,580
when two with suffice doesn't just cost more.

293
00:11:03,580 --> 00:11:04,740
It costs three times more

294
00:11:04,740 --> 00:11:06,500
and it produces the same output.

295
00:11:06,500 --> 00:11:08,500
GitHub gave us a rare concrete look at this

296
00:11:08,500 --> 00:11:11,140
when it started publishing internal token consumption reports

297
00:11:11,140 --> 00:11:13,020
for co-pilot-powered workflows.

298
00:11:13,020 --> 00:11:17,220
A report from March 2nd, 2026 showed 104.5 million tokens

299
00:11:17,220 --> 00:11:18,620
consumed in a single day

300
00:11:18,620 --> 00:11:22,220
across 75 workflow runs covering 41 distinct workflows.

301
00:11:22,220 --> 00:11:25,020
One workflow alone, a CI failure doctor agent,

302
00:11:25,020 --> 00:11:28,660
consumed 28.1 million tokens across just 13 runs.

303
00:11:28,660 --> 00:11:31,940
That's 2.2 million tokens per run for a diagnostic task.

304
00:11:31,940 --> 00:11:34,380
What makes that number instructive isn't the scale.

305
00:11:34,380 --> 00:11:35,580
It's what happened next.

306
00:11:35,580 --> 00:11:38,740
That same reporting period showed usage 37% lower

307
00:11:38,740 --> 00:11:40,340
than the prior measurement window.

308
00:11:40,340 --> 00:11:45,620
And 56% lower than the all-time peak of 237.8 million tokens

309
00:11:45,620 --> 00:11:48,460
recorded on February 11th, 2026.

310
00:11:48,460 --> 00:11:50,660
The work being done hadn't meaningfully decreased.

311
00:11:50,660 --> 00:11:52,460
The workflows had simply been tuned.

312
00:11:52,460 --> 00:11:53,620
No teams were laid off.

313
00:11:53,620 --> 00:11:55,100
No features were removed.

314
00:11:55,100 --> 00:11:56,820
Engineers reviewed the agent chains,

315
00:11:56,820 --> 00:11:59,660
trimmed unnecessary loops, tightened context injection

316
00:11:59,660 --> 00:12:01,820
and improved routing logic, more than half

317
00:12:01,820 --> 00:12:03,300
the token consumption evaporated.

318
00:12:03,300 --> 00:12:05,420
The closest published analogy for enterprise co-pilot

319
00:12:05,420 --> 00:12:07,780
comes from anthropics developer tooling ecosystem.

320
00:12:07,780 --> 00:12:10,220
Claude code, a code assistant with usage patterns

321
00:12:10,220 --> 00:12:12,220
similar to co-pilot's agentic features,

322
00:12:12,220 --> 00:12:16,380
typically runs 150 to $250 per developer per month

323
00:12:16,380 --> 00:12:19,620
before any optimization work is applied, roughly $13

324
00:12:19,620 --> 00:12:20,980
per active day.

325
00:12:20,980 --> 00:12:24,180
After systematic optimization using documented practices,

326
00:12:24,180 --> 00:12:27,180
model routing, context management, session discipline,

327
00:12:27,180 --> 00:12:30,700
tool pruning, teams report 40 to 85% reductions

328
00:12:30,700 --> 00:12:32,340
in total token consumption.

329
00:12:32,340 --> 00:12:34,300
That isn't a marginal efficiency improvement.

330
00:12:34,300 --> 00:12:35,820
That's a fundamental restructuring

331
00:12:35,820 --> 00:12:37,740
of what the system actually needs to do

332
00:12:37,740 --> 00:12:39,780
versus what it was defaulting to do.

333
00:12:39,780 --> 00:12:42,380
For enterprise co-pilot, there are no formal published benchmarks

334
00:12:42,380 --> 00:12:44,060
yet on idle waste rates.

335
00:12:44,060 --> 00:12:46,180
The token-based billing model is too new

336
00:12:46,180 --> 00:12:48,580
and the telemetry most organizations have access to

337
00:12:48,580 --> 00:12:50,740
doesn't separate productive from wasted tokens

338
00:12:50,740 --> 00:12:52,300
at a granular level.

339
00:12:52,300 --> 00:12:54,220
But working from the GitHub workflow data,

340
00:12:54,220 --> 00:12:55,780
the Claude code optimization evidence

341
00:12:55,780 --> 00:12:57,340
and the cloud infrastructure waste patterns

342
00:12:57,340 --> 00:12:59,500
that FinOps teams have tracked for years,

343
00:12:59,500 --> 00:13:03,100
a reasonable working hypothesis puts unoptimized enterprise,

344
00:13:03,100 --> 00:13:06,100
idle token waste somewhere between 30 and 60%

345
00:13:06,100 --> 00:13:07,820
of total build tokens.

346
00:13:07,820 --> 00:13:10,020
That means somewhere between a third and more than half

347
00:13:10,020 --> 00:13:12,020
of what you're paying for AI computation right now

348
00:13:12,020 --> 00:13:14,140
isn't producing value anyone can point to.

349
00:13:14,140 --> 00:13:17,380
The billing infrastructure to make this visible is arriving.

350
00:13:17,380 --> 00:13:20,340
GitHub's switch to AI credits on June 1st, 2026

351
00:13:20,340 --> 00:13:23,500
means every unproductive token is now a direct cost line,

352
00:13:23,500 --> 00:13:24,540
not an abstraction.

353
00:13:24,540 --> 00:13:27,460
Other co-pilot surfaces are moving in the same direction.

354
00:13:27,460 --> 00:13:30,500
Idl waste is a structural problem with a structural fix.

355
00:13:30,500 --> 00:13:32,620
But there's a behavioral problem sitting on top of it

356
00:13:32,620 --> 00:13:34,780
that makes the structural problem worse.

357
00:13:34,780 --> 00:13:38,820
The lazy prompting tax, the idle token problem is structural.

358
00:13:38,820 --> 00:13:41,260
It lives in the architecture of how agents loop,

359
00:13:41,260 --> 00:13:44,580
how context gets injected, how background processes fire,

360
00:13:44,580 --> 00:13:46,260
you can fix it with engineering.

361
00:13:46,260 --> 00:13:49,060
What's harder to fix is what happens at the keyboard.

362
00:13:49,060 --> 00:13:51,540
Most users interact with co-pilot the way they learn

363
00:13:51,540 --> 00:13:54,700
to interact with Google, broad query, scan the results,

364
00:13:54,700 --> 00:13:56,780
refine, repeat, type something vague,

365
00:13:56,780 --> 00:13:59,540
see what comes back at a follow-up, narrow it down.

366
00:13:59,540 --> 00:14:01,220
This worked fine for search because the cost

367
00:14:01,220 --> 00:14:02,980
of a bad search query is zero.

368
00:14:02,980 --> 00:14:04,540
The cost of a bad prompt is not zero.

369
00:14:04,540 --> 00:14:06,620
Every clarification turn in a co-pilot conversation

370
00:14:06,620 --> 00:14:07,820
is a token event.

371
00:14:07,820 --> 00:14:09,300
The first message goes in as input.

372
00:14:09,300 --> 00:14:10,700
The response comes back as output.

373
00:14:10,700 --> 00:14:12,100
The follow-up message goes in carrying

374
00:14:12,100 --> 00:14:14,700
the full prior conversation as context.

375
00:14:14,700 --> 00:14:16,820
The next response comes back longer now

376
00:14:16,820 --> 00:14:19,340
because the model is synthesizing more history.

377
00:14:19,340 --> 00:14:20,980
A five turn conversation to arrive

378
00:14:20,980 --> 00:14:24,700
at one usable paragraph can consume 10 times the tokens

379
00:14:24,700 --> 00:14:27,220
that a single well-structured prompt would have required.

380
00:14:27,220 --> 00:14:29,660
Same end result, 10 times the compute.

381
00:14:29,660 --> 00:14:32,060
Under flat licensing, this was invisible.

382
00:14:32,060 --> 00:14:33,380
The meter wasn't running.

383
00:14:33,380 --> 00:14:35,540
Under consumption billing, every one of those turns

384
00:14:35,540 --> 00:14:36,420
has a price.

385
00:14:36,420 --> 00:14:38,860
The context window problem compounds this.

386
00:14:38,860 --> 00:14:41,260
When users aren't sure what information the model needs,

387
00:14:41,260 --> 00:14:43,500
the natural instinct is to give it everything,

388
00:14:43,500 --> 00:14:46,380
paste in the entire contract, attach the full email chain

389
00:14:46,380 --> 00:14:49,140
going back six months, drop the whole code base into the prompt

390
00:14:49,140 --> 00:14:51,860
and ask why one function is behaving unexpectedly.

391
00:14:51,860 --> 00:14:54,700
This feels thorough, it's actually expensive.

392
00:14:54,700 --> 00:14:57,420
Input tokens are the cheaper side of the pricing equation,

393
00:14:57,420 --> 00:15:00,300
but context window abuse drives them to scale in ways

394
00:15:00,300 --> 00:15:02,020
that quickly overwhelm any savings

395
00:15:02,020 --> 00:15:03,460
from using a cheaper model.

396
00:15:03,460 --> 00:15:05,740
And then the output arrives because the model received

397
00:15:05,740 --> 00:15:08,540
a large unstructured input, it generates a long,

398
00:15:08,540 --> 00:15:10,940
comprehensive response that addresses every dimension

399
00:15:10,940 --> 00:15:13,300
of what you gave it, including all the dimensions

400
00:15:13,300 --> 00:15:15,100
you didn't actually need.

401
00:15:15,100 --> 00:15:17,300
Output tokens cost three to 10 times more

402
00:15:17,300 --> 00:15:19,620
than input tokens depending on the model.

403
00:15:19,620 --> 00:15:21,620
The verbose response to the vague prompt

404
00:15:21,620 --> 00:15:24,340
is the single most expensive interaction pattern

405
00:15:24,340 --> 00:15:25,900
in any AI deployment.

406
00:15:25,900 --> 00:15:27,740
Reasoning models make this worse.

407
00:15:27,740 --> 00:15:30,620
When a model runs extended chain of thought processing,

408
00:15:30,620 --> 00:15:32,500
working through a problem step by step

409
00:15:32,500 --> 00:15:35,220
before producing a final answer, it generates

410
00:15:35,220 --> 00:15:36,980
what are called thinking tokens.

411
00:15:36,980 --> 00:15:39,460
These are internal reasoning steps that consume compute,

412
00:15:39,460 --> 00:15:41,420
but don't appear in the visible response.

413
00:15:41,420 --> 00:15:43,260
On some reasoning model configurations,

414
00:15:43,260 --> 00:15:46,420
effective cost per interaction can run three to nine times higher

415
00:15:46,420 --> 00:15:47,860
than the headline per token rate

416
00:15:47,860 --> 00:15:49,860
because of these hidden reasoning steps.

417
00:15:49,860 --> 00:15:51,780
A prompt that looks like it should cost a few cents

418
00:15:51,780 --> 00:15:54,660
can cost 30 to 90 cents when a reasoning model spins up

419
00:15:54,660 --> 00:15:56,940
a full deliberation chain for what was essentially

420
00:15:56,940 --> 00:15:57,900
a simple task.

421
00:15:57,900 --> 00:15:59,500
None of this is a user failure.

422
00:15:59,500 --> 00:16:01,740
Users weren't trained on prompt economics

423
00:16:01,740 --> 00:16:04,060
because under the legacy flat licensing model,

424
00:16:04,060 --> 00:16:06,020
prompt economics didn't exist for them.

425
00:16:06,020 --> 00:16:07,860
There was no feedback loop, no signal

426
00:16:07,860 --> 00:16:10,940
that a poorly structured prompt cost more than a well-structured one.

427
00:16:10,940 --> 00:16:12,820
The behavior developed in an environment

428
00:16:12,820 --> 00:16:15,860
where tokens were free and it persisted into an environment

429
00:16:15,860 --> 00:16:17,220
where they aren't.

430
00:16:17,220 --> 00:16:19,300
This is the core of the lazy prompting tax.

431
00:16:19,300 --> 00:16:21,100
It's not that users are careless.

432
00:16:21,100 --> 00:16:22,580
It's that the system never taught them

433
00:16:22,580 --> 00:16:24,220
that carelessness had a cost.

434
00:16:24,220 --> 00:16:27,060
Every vague prompt, every unnecessary attachment,

435
00:16:27,060 --> 00:16:29,100
every six turn clarification loop

436
00:16:29,100 --> 00:16:31,260
that could have been one well-scoped request.

437
00:16:31,260 --> 00:16:34,180
None of that registered as waste under the old model.

438
00:16:34,180 --> 00:16:35,980
So the behavior is a governance failure

439
00:16:35,980 --> 00:16:37,900
wearing a user behavior disguise.

440
00:16:37,900 --> 00:16:40,780
The organization deployed a consumption-sensitive tool

441
00:16:40,780 --> 00:16:43,580
without building consumption awareness into the deployment.

442
00:16:43,580 --> 00:16:46,620
No prompt guidelines, no context discipline policies,

443
00:16:46,620 --> 00:16:48,620
no training on when to use a reasoning model

444
00:16:48,620 --> 00:16:51,060
versus a standard one or when to reset a session

445
00:16:51,060 --> 00:16:53,020
that's accumulated irrelevant history.

446
00:16:53,020 --> 00:16:54,620
The prompting problem is fixable,

447
00:16:54,620 --> 00:16:57,420
but fixing it sits on a foundation that's already unstable

448
00:16:57,420 --> 00:16:59,500
because the outputs users are trying to verify

449
00:16:59,500 --> 00:17:01,220
aren't always correct.

450
00:17:01,220 --> 00:17:03,980
The verification tax, layer the hallucination problem

451
00:17:03,980 --> 00:17:05,540
on top of everything we've just covered

452
00:17:05,540 --> 00:17:07,980
and the cost picture gets significantly darker.

453
00:17:07,980 --> 00:17:10,420
Here's what the verification tax actually is.

454
00:17:10,420 --> 00:17:12,900
The extra human time required to check, correct,

455
00:17:12,900 --> 00:17:14,660
and document AI-generated work

456
00:17:14,660 --> 00:17:16,620
before it meets professional standards.

457
00:17:16,620 --> 00:17:18,180
Not the time to generate the output.

458
00:17:18,180 --> 00:17:20,020
The time spent after the output arrives

459
00:17:20,020 --> 00:17:22,420
by someone qualified enough to know when it's wrong

460
00:17:22,420 --> 00:17:24,420
and it is wrong with surprising frequency.

461
00:17:24,420 --> 00:17:27,100
Research benchmarking LLM accuracy

462
00:17:27,100 --> 00:17:29,020
on open-ended knowledge-heavy tasks

463
00:17:29,020 --> 00:17:31,700
puts hallucination rates between 50 and 82%

464
00:17:31,700 --> 00:17:34,300
depending on model, domain, and task structure,

465
00:17:34,300 --> 00:17:36,900
not edge cases, not unusual prompts.

466
00:17:36,900 --> 00:17:39,740
Routine professional tasks in areas like tax research,

467
00:17:39,740 --> 00:17:41,420
compliance review, legal analysis,

468
00:17:41,420 --> 00:17:43,140
and client-facing documentation.

469
00:17:43,140 --> 00:17:45,540
The model produces fluent, confident pros

470
00:17:45,540 --> 00:17:47,100
that reads like it was written by someone

471
00:17:47,100 --> 00:17:48,460
who knows what they're talking about.

472
00:17:48,460 --> 00:17:50,380
The problem is it sometimes wasn't

473
00:17:50,380 --> 00:17:53,060
the model pattern matched to what a correct answer looks like

474
00:17:53,060 --> 00:17:54,340
without actually being correct.

475
00:17:54,340 --> 00:17:55,820
This creates a verification burden

476
00:17:55,820 --> 00:17:58,700
that doesn't disappear because the AI got faster at drafting.

477
00:17:58,700 --> 00:18:01,260
A tax professional reviewing an AI-generated analysis

478
00:18:01,260 --> 00:18:04,380
of a multi-duration filing still has to read every line.

479
00:18:04,380 --> 00:18:06,900
A compliance officer reviewing an AI-drafted policy

480
00:18:06,900 --> 00:18:09,100
still has to verify every citation.

481
00:18:09,100 --> 00:18:11,540
A lawyer reviewing an AI-produced contract summary

482
00:18:11,540 --> 00:18:13,180
still has to check every clause reference

483
00:18:13,180 --> 00:18:14,580
against the source document.

484
00:18:14,580 --> 00:18:16,260
The AI may have cut the drafting time

485
00:18:16,260 --> 00:18:18,180
from two hours to 15 minutes.

486
00:18:18,180 --> 00:18:20,220
The review time didn't shrink proportionally

487
00:18:20,220 --> 00:18:22,300
because the reviewer has no way to know in advance

488
00:18:22,300 --> 00:18:24,780
which parts of the output are correct and which aren't.

489
00:18:24,780 --> 00:18:25,740
They have to check all of it.

490
00:18:25,740 --> 00:18:27,780
Practitioner analysis of verification costs

491
00:18:27,780 --> 00:18:30,700
in regulated workflows puts the overhead at up to 48%

492
00:18:30,700 --> 00:18:33,220
above baseline in judgment-heavy processes.

493
00:18:33,220 --> 00:18:36,740
Meaning, if the underlying task without AI takes 10 hours,

494
00:18:36,740 --> 00:18:38,620
a workflow that uses AI for drafting

495
00:18:38,620 --> 00:18:40,740
but requires professional verification

496
00:18:40,740 --> 00:18:43,220
can take up to 14.8 hours total.

497
00:18:43,220 --> 00:18:45,060
The AI saves time on generation.

498
00:18:45,060 --> 00:18:48,020
The verification absorbs a significant portion of those savings

499
00:18:48,020 --> 00:18:50,060
and in some cases, all of them.

500
00:18:50,060 --> 00:18:53,380
The math gets uncomfortable when you apply fully loaded labor costs

501
00:18:53,380 --> 00:18:54,820
to professional review time.

502
00:18:54,820 --> 00:18:57,740
At $80 to $200 per hour for a senior knowledge worker,

503
00:18:57,740 --> 00:19:00,100
the person qualified to catch what the AI got wrong.

504
00:19:00,100 --> 00:19:02,100
Two minutes of review per AI answer costs

505
00:19:02,100 --> 00:19:04,460
between two by six seven and six and 67.

506
00:19:04,460 --> 00:19:07,180
10 minutes cost between 13 and $32.

507
00:19:07,180 --> 00:19:09,820
Scale that across a day of AI assisted work

508
00:19:09,820 --> 00:19:13,420
in a regulated environment, across a team, across a quarter,

509
00:19:13,420 --> 00:19:16,820
and the free AI draft has a very real cost attached to it.

510
00:19:16,820 --> 00:19:18,420
It just doesn't appear on your Azure Bill.

511
00:19:18,420 --> 00:19:19,740
There's a compounding problem here

512
00:19:19,740 --> 00:19:21,660
that the token economics make worse.

513
00:19:21,660 --> 00:19:23,460
Per token pricing doesn't discriminate

514
00:19:23,460 --> 00:19:25,300
between right answers and wrong ones.

515
00:19:25,300 --> 00:19:28,180
A hallucinated response costs the same as a correct one.

516
00:19:28,180 --> 00:19:30,260
A long hallucinated reasoning chain

517
00:19:30,260 --> 00:19:32,420
where the model doubles down on an incorrect path

518
00:19:32,420 --> 00:19:33,780
and elaborates at length.

519
00:19:33,780 --> 00:19:35,780
Costs more than a short correct answer

520
00:19:35,780 --> 00:19:37,900
because it generates more output tokens.

521
00:19:37,900 --> 00:19:40,300
You are literally paying more in token terms

522
00:19:40,300 --> 00:19:41,900
for the wrong answer that then requires

523
00:19:41,900 --> 00:19:43,500
the most review time to untangle.

524
00:19:43,500 --> 00:19:46,020
What this means practically is that the use cases

525
00:19:46,020 --> 00:19:48,140
where the verification tax hits hardest

526
00:19:48,140 --> 00:19:50,580
are exactly the use cases that organizations

527
00:19:50,580 --> 00:19:53,860
tend to deploy AI first because they seem high value.

528
00:19:53,860 --> 00:19:57,940
Legal, compliance, tax, client facing deliverables,

529
00:19:57,940 --> 00:20:00,020
regulated communications.

530
00:20:00,020 --> 00:20:02,020
These are the domains where professional judgment

531
00:20:02,020 --> 00:20:05,260
carries the most weight where the cost of error is highest

532
00:20:05,260 --> 00:20:07,620
and where the reviewer's time is most expensive.

533
00:20:07,620 --> 00:20:10,020
They're also the domains where AI's hallucination rate

534
00:20:10,020 --> 00:20:11,740
tends to be most consequential

535
00:20:11,740 --> 00:20:14,700
because mistakes in these areas don't just require correction,

536
00:20:14,700 --> 00:20:17,540
they require documentation that the correction happened.

537
00:20:17,540 --> 00:20:19,180
The honest framing is this.

538
00:20:19,180 --> 00:20:21,100
In high stakes professional workflows,

539
00:20:21,100 --> 00:20:23,780
AI often shifts work rather than eliminates it.

540
00:20:23,780 --> 00:20:26,300
Time savings happen at the generation stage.

541
00:20:26,300 --> 00:20:27,900
They get partially or fully recaptured

542
00:20:27,900 --> 00:20:29,220
at the verification stage.

543
00:20:29,220 --> 00:20:30,300
The net benefit is real,

544
00:20:30,300 --> 00:20:32,940
but smaller than headline productivity numbers suggest.

545
00:20:32,940 --> 00:20:34,500
And it has a hard ceiling determined

546
00:20:34,500 --> 00:20:36,620
by how much of the task requires judgment

547
00:20:36,620 --> 00:20:39,220
that only a qualified human can apply.

548
00:20:39,220 --> 00:20:41,060
This isn't an argument against using AI

549
00:20:41,060 --> 00:20:42,420
and professional workflows.

550
00:20:42,420 --> 00:20:45,220
It's an argument for being precise about which workflows,

551
00:20:45,220 --> 00:20:46,900
which models, which risk thresholds

552
00:20:46,900 --> 00:20:49,860
and what verification process accompanies the deployment.

553
00:20:49,860 --> 00:20:51,420
The verification tax is worst

554
00:20:51,420 --> 00:20:53,860
where it was designed into the use case from the start

555
00:20:53,860 --> 00:20:56,660
and invisible until someone looks at the billing.

556
00:20:56,660 --> 00:20:58,460
Zombie seats and the adoption gap.

557
00:20:58,460 --> 00:21:01,260
The verification tax is worst in the wrong use cases,

558
00:21:01,260 --> 00:21:02,580
which raises the question,

559
00:21:02,580 --> 00:21:04,820
where is co-pilot actually being deployed?

560
00:21:04,820 --> 00:21:07,980
70% of Fortune 500 companies have officially adopted

561
00:21:07,980 --> 00:21:09,980
M365 co-pilot.

562
00:21:09,980 --> 00:21:11,540
That number gets cited constantly

563
00:21:11,540 --> 00:21:13,980
in Microsoft earnings calls and analyst briefings

564
00:21:13,980 --> 00:21:16,260
as evidence of enterprise AI momentum.

565
00:21:16,260 --> 00:21:17,940
What it actually describes is pilots,

566
00:21:17,940 --> 00:21:20,100
limited rollouts and licensing purchases

567
00:21:20,100 --> 00:21:22,580
that preceded any serious change management work.

568
00:21:22,580 --> 00:21:25,180
Adoption in the reporting sense means seats were bought.

569
00:21:25,180 --> 00:21:27,340
It doesn't mean those seats are producing anything.

570
00:21:27,340 --> 00:21:29,140
The conversion rate tells the real story.

571
00:21:29,140 --> 00:21:32,860
Across organizations with licensed M365 co-pilot access,

572
00:21:32,860 --> 00:21:36,020
only about 35.8% of employees actively use it.

573
00:21:36,020 --> 00:21:38,180
Roughly one in three people with a paid license

574
00:21:38,180 --> 00:21:40,700
opens co-pilot with any meaningful frequency.

575
00:21:40,700 --> 00:21:42,740
The other two thirds have a license assigned,

576
00:21:42,740 --> 00:21:45,340
a $30 monthly charge appearing on the invoice

577
00:21:45,340 --> 00:21:46,940
and an interaction history

578
00:21:46,940 --> 00:21:49,620
that might show two sessions in the last month.

579
00:21:49,620 --> 00:21:51,260
One to summarize an email,

580
00:21:51,260 --> 00:21:55,220
one out of curiosity after someone mentioned it in a meeting.

581
00:21:55,220 --> 00:21:57,740
Compare that to ChatGPT's voluntary conversion rate.

582
00:21:57,740 --> 00:22:01,180
Approximately 83.1% of users who have access to it

583
00:22:01,180 --> 00:22:02,620
actually use it regularly.

584
00:22:02,620 --> 00:22:05,020
That gap isn't explained by capability differences.

585
00:22:05,020 --> 00:22:07,460
Co-pilot has access to organizational data,

586
00:22:07,460 --> 00:22:08,820
deep office integration,

587
00:22:08,820 --> 00:22:10,980
and the same underlying model infrastructure.

588
00:22:10,980 --> 00:22:12,460
The gap is explained by fit.

589
00:22:12,460 --> 00:22:14,900
ChatGPT meets users where they already are,

590
00:22:14,900 --> 00:22:16,740
in a frictionless consumer interface,

591
00:22:16,740 --> 00:22:18,940
doing tasks they chose for themselves.

592
00:22:18,940 --> 00:22:21,180
Co-pilot arrives inside a workplace context

593
00:22:21,180 --> 00:22:23,540
where nobody redesigned the workflow around it.

594
00:22:23,540 --> 00:22:25,700
Nobody trained people on when to use it

595
00:22:25,700 --> 00:22:28,540
and the default answer to should I use co-pilot for this

596
00:22:28,540 --> 00:22:30,140
is still, I'm not sure.

597
00:22:30,140 --> 00:22:32,980
This is the zombie seat, a paid 30-up a month license

598
00:22:32,980 --> 00:22:35,420
assigned to a user who has it, knows it's there,

599
00:22:35,420 --> 00:22:37,820
and generates near zero measurable value from it.

600
00:22:37,820 --> 00:22:39,780
Not because they're resistant to AI

601
00:22:39,780 --> 00:22:42,380
because nobody connected the license to a specific task,

602
00:22:42,380 --> 00:22:44,260
a specific process, a specific outcome

603
00:22:44,260 --> 00:22:46,020
that made using it the obvious choice.

604
00:22:46,020 --> 00:22:47,620
The scale of this problem is concrete.

605
00:22:47,620 --> 00:22:50,940
At 5,000 licensed users with 35% active usage,

606
00:22:50,940 --> 00:22:54,140
you have roughly 3,750 people in the zombie category.

607
00:22:54,140 --> 00:22:58,700
That's $9,500 per month, $1.17 million per year,

608
00:22:58,700 --> 00:23:01,700
in license spend that is producing nothing you can point to.

609
00:23:01,700 --> 00:23:03,780
And that figure only counts the co-pilot add-on.

610
00:23:03,780 --> 00:23:06,220
It doesn't include the E3 or E5 base licenses

611
00:23:06,220 --> 00:23:07,700
those users are also sitting on,

612
00:23:07,700 --> 00:23:10,300
or the opportunity cost of having purchased AI access

613
00:23:10,300 --> 00:23:12,020
that isn't functioning as an asset.

614
00:23:12,020 --> 00:23:14,980
Organizations that have run formal license optimization cycles

615
00:23:14,980 --> 00:23:18,940
consistently find 10 to 20% of their total M365 costs

616
00:23:18,940 --> 00:23:21,140
are recoverable through identifying and reclaiming

617
00:23:21,140 --> 00:23:23,020
idle or redundant assignments.

618
00:23:23,020 --> 00:23:25,980
That initial pass usually surfaces the clearest waste.

619
00:23:25,980 --> 00:23:28,420
Departed employees who still hold licenses,

620
00:23:28,420 --> 00:23:30,660
users whose role changed and no longer justifies

621
00:23:30,660 --> 00:23:33,100
their current SKU add-ons assigned broadly

622
00:23:33,100 --> 00:23:35,500
for a project that concluded months ago.

623
00:23:35,500 --> 00:23:37,860
Organizations that move from periodic audits

624
00:23:37,860 --> 00:23:40,620
to continuous automated governance push that figure

625
00:23:40,620 --> 00:23:42,660
toward 20 to 30% over time.

626
00:23:42,660 --> 00:23:46,220
For a 5,000 seat environment with a fully loaded M365 stack

627
00:23:46,220 --> 00:23:48,580
averaging $65 per user per month,

628
00:23:48,580 --> 00:23:52,620
20% waste recovery represents $780,000 per year.

629
00:23:52,620 --> 00:23:55,020
That's not cost-cutting, that's a reallocation opportunity.

630
00:23:55,020 --> 00:23:57,580
Recovered budget that can fund the co-pilot expansion

631
00:23:57,580 --> 00:23:59,820
for roles where it actually generates return

632
00:23:59,820 --> 00:24:02,060
or the co-pilot studio agent development

633
00:24:02,060 --> 00:24:03,540
for the two or three workflows

634
00:24:03,540 --> 00:24:06,620
where AI delivers a genuinely measurable difference.

635
00:24:06,620 --> 00:24:08,020
This is the self-funding narrative

636
00:24:08,020 --> 00:24:10,380
that the most disciplined IT organizations are running

637
00:24:10,380 --> 00:24:11,620
in 2026.

638
00:24:11,620 --> 00:24:13,980
Don't ask the board for more AI budget.

639
00:24:13,980 --> 00:24:15,660
Audit the AI budget you already have,

640
00:24:15,660 --> 00:24:17,260
recover the waste and redirect it

641
00:24:17,260 --> 00:24:20,540
toward the 35% of users who will actually use what you give them.

642
00:24:20,540 --> 00:24:21,940
The money is already in the system.

643
00:24:21,940 --> 00:24:23,900
It's just sitting in the wrong seats.

644
00:24:23,900 --> 00:24:25,020
Zombie seats are a symptom.

645
00:24:25,020 --> 00:24:26,660
The root cause is a deployment model

646
00:24:26,660 --> 00:24:28,580
that prioritized speed over fit.

647
00:24:28,580 --> 00:24:31,060
The blanket rollout problem.

648
00:24:31,060 --> 00:24:33,220
The zombie seat problem didn't happen by accident.

649
00:24:33,220 --> 00:24:34,460
It was the predictable outcome

650
00:24:34,460 --> 00:24:36,300
of a specific procurement decision

651
00:24:36,300 --> 00:24:39,460
that most organizations made in 2023 and 2024

652
00:24:39,460 --> 00:24:40,900
and many are still making today.

653
00:24:40,900 --> 00:24:42,740
Call it the license first approach

654
00:24:42,740 --> 00:24:44,980
by the seats quickly, run a light training session,

655
00:24:44,980 --> 00:24:46,300
send an announcement email

656
00:24:46,300 --> 00:24:48,060
and let adoption happen organically.

657
00:24:48,060 --> 00:24:49,780
The reasoning felt sound at the time.

658
00:24:49,780 --> 00:24:51,820
Broad access creates broad familiarity.

659
00:24:51,820 --> 00:24:54,660
Familiarity creates usage, usage creates value,

660
00:24:54,660 --> 00:24:57,100
get the licenses in place and the behavior will follow.

661
00:24:57,100 --> 00:24:58,540
The data doesn't support it.

662
00:24:58,540 --> 00:25:02,180
The license first approach consistently produces ROI below 100%

663
00:25:02,180 --> 00:25:03,980
with payback timelines that stretch past

664
00:25:03,980 --> 00:25:05,820
any reasonable planning horizon

665
00:25:05,820 --> 00:25:07,180
and license utilization rates

666
00:25:07,180 --> 00:25:09,060
that mirror what we just described.

667
00:25:09,060 --> 00:25:11,340
Broad access didn't drive broad adoption.

668
00:25:11,340 --> 00:25:13,100
It drove broad license assignment,

669
00:25:13,100 --> 00:25:14,820
which is a different thing entirely.

670
00:25:14,820 --> 00:25:17,620
The failure isn't mysterious once you look at it structurally.

671
00:25:17,620 --> 00:25:19,460
Buying a license doesn't change a workflow.

672
00:25:19,460 --> 00:25:21,700
It doesn't redesign the process around the tool,

673
00:25:21,700 --> 00:25:24,340
identify the specific tasks where AI fits,

674
00:25:24,340 --> 00:25:26,060
build the muscle memory for using it

675
00:25:26,060 --> 00:25:28,940
or create accountability for whether it's being used effectively.

676
00:25:28,940 --> 00:25:30,060
A license is access.

677
00:25:30,060 --> 00:25:31,660
Access without a reason to use it

678
00:25:31,660 --> 00:25:33,980
produces the conversion rate we're looking at.

679
00:25:33,980 --> 00:25:35,700
One in three people actually engaging

680
00:25:35,700 --> 00:25:38,340
with a tool that costs the same whether it's touched or not.

681
00:25:38,340 --> 00:25:40,300
There's a second consequence of blanket rollout

682
00:25:40,300 --> 00:25:42,900
that gets less attention than the adoption gap,

683
00:25:42,900 --> 00:25:45,420
but it's equally damaging governance collapses.

684
00:25:45,420 --> 00:25:48,020
When every employee in the organization has access to an AI

685
00:25:48,020 --> 00:25:50,740
that can query organizational data across SharePoint teams

686
00:25:50,740 --> 00:25:53,900
and email and nobody has defined what that AI should

687
00:25:53,900 --> 00:25:56,500
and shouldn't touch, what gets logged, what gets reviewed,

688
00:25:56,500 --> 00:25:58,700
what constitutes appropriate use.

689
00:25:58,700 --> 00:26:00,740
You haven't deployed AI responsibly.

690
00:26:00,740 --> 00:26:04,060
You've deployed AI chaoticly and called it democratization.

691
00:26:04,060 --> 00:26:07,020
ROI data from 2026 deployments is consistent

692
00:26:07,020 --> 00:26:09,660
on the relationship between usage depth and return.

693
00:26:09,660 --> 00:26:12,300
Organizations where 70% or more of licensed users

694
00:26:12,300 --> 00:26:15,700
invoke co-pilot daily in at least one core workflow

695
00:26:15,700 --> 00:26:17,980
see three to four times the ROI of organizations

696
00:26:17,980 --> 00:26:20,260
stuck at 30% weekly active usage.

697
00:26:20,260 --> 00:26:21,540
The license cost is identical.

698
00:26:21,540 --> 00:26:23,180
The outcome is not what separates them

699
00:26:23,180 --> 00:26:24,420
isn't the quality of the AI.

700
00:26:24,420 --> 00:26:26,460
It's whether the deployment was built around a workflow

701
00:26:26,460 --> 00:26:28,540
transformation or just a software rollout.

702
00:26:28,540 --> 00:26:30,060
The correct model isn't complicated

703
00:26:30,060 --> 00:26:33,100
but it runs against how most procurement cycles operate.

704
00:26:33,100 --> 00:26:34,940
You start with a concentrated pilot,

705
00:26:34,940 --> 00:26:36,940
a defined set of high value roles,

706
00:26:36,940 --> 00:26:40,020
specific workflows with measurable before and after metrics,

707
00:26:40,020 --> 00:26:42,460
dedicated training and an eight to 12 week window

708
00:26:42,460 --> 00:26:44,300
to generate real data.

709
00:26:44,300 --> 00:26:46,860
Not a survey asking people how they feel about co-pilot.

710
00:26:46,860 --> 00:26:48,820
Actual telemetry, actual cycle times,

711
00:26:48,820 --> 00:26:51,020
actual throughput numbers, then you expand

712
00:26:51,020 --> 00:26:52,700
only where the data supports it,

713
00:26:52,700 --> 00:26:54,700
role by role, process by process,

714
00:26:54,700 --> 00:26:56,260
with governance controls in place

715
00:26:56,260 --> 00:26:59,500
before the next wave of licenses ships, not after.

716
00:26:59,500 --> 00:27:02,300
Most procurement cycles work on annual commitment timelines

717
00:27:02,300 --> 00:27:04,420
with volume based pricing pressure at renewal.

718
00:27:04,420 --> 00:27:06,620
The incentive structure pushes toward buying more seats

719
00:27:06,620 --> 00:27:09,380
than you can absorb because the price per seat improves

720
00:27:09,380 --> 00:27:12,900
with volume and the EA negotiation is a once a year window.

721
00:27:12,900 --> 00:27:15,620
This means organizations often acquire two or three years

722
00:27:15,620 --> 00:27:17,780
worth of AI licensing in a single transaction

723
00:27:17,780 --> 00:27:20,180
before the deployment maturity exists to justify it.

724
00:27:20,180 --> 00:27:22,940
That's how the zombie seat population scales so fast.

725
00:27:22,940 --> 00:27:24,860
And it's why the fix isn't a technology problem.

726
00:27:24,860 --> 00:27:26,500
It's a deployment model problem.

727
00:27:26,500 --> 00:27:27,780
To build the right model though,

728
00:27:27,780 --> 00:27:29,700
you need to understand what co-pilot is actually

729
00:27:29,700 --> 00:27:31,340
completing against when you're deciding

730
00:27:31,340 --> 00:27:33,940
where to spend that recovered seat budget.

731
00:27:33,940 --> 00:27:36,260
What traditional automation actually costs?

732
00:27:36,260 --> 00:27:37,780
Before you can make a rational decision

733
00:27:37,780 --> 00:27:41,020
about where AI spend belongs in your automation portfolio,

734
00:27:41,020 --> 00:27:42,340
you need an honest baseline

735
00:27:42,340 --> 00:27:44,260
for what the alternatives actually cost,

736
00:27:44,260 --> 00:27:47,380
not the marketing version, the unit economics version.

737
00:27:47,380 --> 00:27:49,020
Robotic process automation has been

738
00:27:49,020 --> 00:27:50,820
the enterprise automation workhorse

739
00:27:50,820 --> 00:27:52,300
for the better part of a decade.

740
00:27:52,300 --> 00:27:54,060
The cost structure is straightforward.

741
00:27:54,060 --> 00:27:57,540
Unattended bot licenses typically run $8,000 to $20,000

742
00:27:57,540 --> 00:28:00,540
per year depending on vendor, addition, and contract size.

743
00:28:00,540 --> 00:28:03,380
Add infrastructure, virtual machines, orchestrators,

744
00:28:03,380 --> 00:28:06,140
monitoring tools, and implementation costs,

745
00:28:06,140 --> 00:28:08,380
which can match or exceed the first year of licensing

746
00:28:08,380 --> 00:28:09,860
on complex deployments.

747
00:28:09,860 --> 00:28:11,380
Then factor in ongoing maintenance

748
00:28:11,380 --> 00:28:14,260
because RPA bots are brittle, when the UI changes,

749
00:28:14,260 --> 00:28:15,660
when the upstream system gets updated,

750
00:28:15,660 --> 00:28:18,380
when the form layout shifts, someone has to go fix the bot.

751
00:28:18,380 --> 00:28:20,380
That maintenance cost is real and recurring,

752
00:28:20,380 --> 00:28:23,860
but here's what makes RPA compelling for the right use cases.

753
00:28:23,860 --> 00:28:27,180
Once a bot is deployed against a stable, high volume process,

754
00:28:27,180 --> 00:28:30,580
the marginal cost per execution approaches zero.

755
00:28:30,580 --> 00:28:32,660
An unattended bot running invoice posting

756
00:28:32,660 --> 00:28:36,100
100,000 times a year, against a system that doesn't change,

757
00:28:36,100 --> 00:28:38,660
against a data format that stays consistent,

758
00:28:38,660 --> 00:28:41,780
that bot's effective cost per task can fall to $0,000,000,

759
00:28:41,780 --> 00:28:44,220
one-ballons to $0,000,000, five-ballons,

760
00:28:44,220 --> 00:28:46,020
depending on how you amortize the build

761
00:28:46,020 --> 00:28:48,180
and maintenance overhead across volume.

762
00:28:48,180 --> 00:28:50,100
At scale with process stability,

763
00:28:50,100 --> 00:28:52,500
RPA's cost per transaction is nearly free.

764
00:28:52,500 --> 00:28:54,460
Determinist scripts are even cheaper.

765
00:28:54,460 --> 00:28:56,460
A Python microservice that validates data

766
00:28:56,460 --> 00:28:59,180
against a business rule transforms a file format

767
00:28:59,180 --> 00:29:01,300
or roots a record based on a field value

768
00:29:01,300 --> 00:29:03,260
runs on commodity compute.

769
00:29:03,260 --> 00:29:06,380
The per-call cost is in the range of fractions of a millionth of a dollar,

770
00:29:06,380 --> 00:29:08,780
not fractions of a cent, fractions of a millionth of a dollar.

771
00:29:08,780 --> 00:29:11,140
For mechanical well-specified logic with no ambiguity,

772
00:29:11,140 --> 00:29:13,620
this is the floor and it's a very low floor.

773
00:29:13,620 --> 00:29:16,380
Now, 2026 changed part of this equation.

774
00:29:16,380 --> 00:29:21,580
LLM API costs dropped roughly 80% between early 2025 and early 2026,

775
00:29:21,580 --> 00:29:23,260
driven by model efficiency improvements,

776
00:29:23,260 --> 00:29:26,380
hardware scale out and intense competition across providers.

777
00:29:26,380 --> 00:29:29,220
That price collapse made AI agents genuinely competitive

778
00:29:29,220 --> 00:29:30,500
for unstructured work,

779
00:29:30,500 --> 00:29:33,740
the tasks that were never viable for RPA in the first place.

780
00:29:33,740 --> 00:29:35,660
Email classification, contract extraction,

781
00:29:35,660 --> 00:29:37,120
document summarization and big,

782
00:29:37,120 --> 00:29:38,500
US routing decisions,

783
00:29:38,500 --> 00:29:40,540
these are tasks where RPA has always struggled

784
00:29:40,540 --> 00:29:43,780
because the inputs don't conform to a predictable structure.

785
00:29:43,780 --> 00:29:47,140
For those tasks, an AI call at fractions of a cent is now often cheaper

786
00:29:47,140 --> 00:29:49,780
than building and maintaining the rules-based alternative,

787
00:29:49,780 --> 00:29:52,180
but for structured work, the comparison is still brutal.

788
00:29:52,180 --> 00:29:55,700
An AI model doesn't get cheaper with volume the way an RPA bot does.

789
00:29:55,700 --> 00:29:58,380
Every prompt costs tokens, every output costs tokens.

790
00:29:58,380 --> 00:30:02,020
Scale from 10,000 invoice validations to 100,000 invoice validations

791
00:30:02,020 --> 00:30:04,140
and your AI cost scales linearly.

792
00:30:04,140 --> 00:30:06,740
The RPA cost barely moves once the bot is running.

793
00:30:06,740 --> 00:30:09,620
For deterministic high-volume stable processes,

794
00:30:09,620 --> 00:30:12,660
AI's variable cost structure is a fundamental disadvantage

795
00:30:12,660 --> 00:30:15,460
that no model price reduction has yet overcome.

796
00:30:15,460 --> 00:30:18,020
This is where the strategic question sharpens.

797
00:30:18,020 --> 00:30:22,020
Organizations evaluating AI or RPA are asking the wrong question.

798
00:30:22,020 --> 00:30:23,300
The right question is,

799
00:30:23,300 --> 00:30:25,780
which layer of this workflow requires judgment

800
00:30:25,780 --> 00:30:27,340
and which requires execution?

801
00:30:28,300 --> 00:30:29,100
Judgment.

802
00:30:29,100 --> 00:30:32,620
Reading intent from an ambiguous email,

803
00:30:32,620 --> 00:30:35,100
extracting meaning from an unstructured document,

804
00:30:35,100 --> 00:30:38,380
deciding how to handle an exception that doesn't fit a predefined category,

805
00:30:38,380 --> 00:30:40,300
that's where LLMs earn their cost.

806
00:30:40,300 --> 00:30:43,260
They're flexible, they generalize across input variation,

807
00:30:43,260 --> 00:30:46,620
they handle the cases a rules engine would have to explicitly anticipate.

808
00:30:46,620 --> 00:30:49,980
For that layer, even today's AI prices are often justified.

809
00:30:49,980 --> 00:30:53,420
Execution, updating the CRM record, posting the journal entry,

810
00:30:53,420 --> 00:30:56,140
moving the file, triggering the downstream workflow,

811
00:30:56,140 --> 00:30:59,100
that's mechanical, it's deterministic, it follows rules,

812
00:30:59,100 --> 00:31:02,220
RPA handles it at near zero marginal cost once it's deployed,

813
00:31:02,220 --> 00:31:04,460
and a deterministic script handles it cheaper still.

814
00:31:04,460 --> 00:31:05,820
There is no judgment involved,

815
00:31:05,820 --> 00:31:08,460
which means there's no reason to pay judgment prices for it.

816
00:31:08,460 --> 00:31:10,940
Using a reasoning engine to perform a mechanical task

817
00:31:10,940 --> 00:31:13,820
is the most expensive mistake in modern enterprise automation.

818
00:31:13,820 --> 00:31:15,980
Not because the AI is bad at the task, it can do it,

819
00:31:15,980 --> 00:31:18,620
but it costs 500 times more than the alternative.

820
00:31:18,620 --> 00:31:20,460
For identical output quality,

821
00:31:20,460 --> 00:31:22,460
that cost gap becomes catastrophic

822
00:31:22,460 --> 00:31:25,580
when you move from individual tasks to automated workflows.

823
00:31:25,580 --> 00:31:27,340
And that's exactly where most organizations

824
00:31:27,340 --> 00:31:29,020
are pushing co-pilot right now.

825
00:31:29,020 --> 00:31:30,620
The agentec cost explosion,

826
00:31:30,620 --> 00:31:32,300
the unit economics shift dramatically

827
00:31:32,300 --> 00:31:35,820
when you move from co-pilot's chat interface to agentec workflows.

828
00:31:35,820 --> 00:31:38,300
Tasks where the AI doesn't just generate a response,

829
00:31:38,300 --> 00:31:42,140
but plans multiple steps, calls tools, evaluates outcomes,

830
00:31:42,140 --> 00:31:44,380
and iterates toward a solution.

831
00:31:44,380 --> 00:31:46,620
A standard conversational interaction with co-pilot

832
00:31:46,620 --> 00:31:48,540
consumes roughly 2,000 tokens.

833
00:31:48,540 --> 00:31:51,660
You ask a question, you get an answer, the conversation ends.

834
00:31:51,660 --> 00:31:54,780
A single agentec task, where the system researches a topic

835
00:31:54,780 --> 00:31:58,060
across multiple sources, drafts a comprehensive response,

836
00:31:58,060 --> 00:32:00,220
validates the output against requirements,

837
00:32:00,220 --> 00:32:02,700
identifies gaps, and refines the draft.

838
00:32:02,700 --> 00:32:06,300
Consumes between 200,000 and 1,000,000 tokens.

839
00:32:06,300 --> 00:32:07,660
That's not a 10x difference.

840
00:32:07,660 --> 00:32:10,380
That's up to a 500x multiplier on token consumption

841
00:32:10,380 --> 00:32:12,460
for what appears to be a single user request.

842
00:32:12,460 --> 00:32:14,380
This massive escalation was quantified

843
00:32:14,380 --> 00:32:17,340
in a 2025 Stanford Digital Economy Lab study

844
00:32:17,340 --> 00:32:19,500
in partnership with Microsoft Research.

845
00:32:19,500 --> 00:32:21,100
The finding was stark.

846
00:32:21,100 --> 00:32:23,340
Agentec workflows consume orders of magnitude

847
00:32:23,340 --> 00:32:26,700
more tokens than interactive chat because every planning cycle,

848
00:32:26,700 --> 00:32:29,180
every tool invocation, every evaluation loop,

849
00:32:29,180 --> 00:32:31,740
represents a full inference pass through the model.

850
00:32:31,740 --> 00:32:34,540
A three-step plan becomes three separate token events.

851
00:32:34,540 --> 00:32:37,100
A five-turn refinement cycle becomes five more.

852
00:32:37,100 --> 00:32:39,580
What users perceive as one task generates

853
00:32:39,580 --> 00:32:41,260
dozens of internal computations.

854
00:32:41,260 --> 00:32:42,620
Each one build independently.

855
00:32:42,620 --> 00:32:44,300
The cost math is unforgiving.

856
00:32:44,300 --> 00:32:46,300
At Claude Sonnet 4.6 pricing,

857
00:32:46,300 --> 00:32:47,580
three per million input tokens,

858
00:32:47,580 --> 00:32:49,260
15 per million output tokens,

859
00:32:49,260 --> 00:32:50,700
a mid-range agentec task,

860
00:32:50,700 --> 00:32:54,460
consuming 300,000 tokens costs approximately one or 50 salsons.

861
00:32:54,460 --> 00:32:57,740
A complex agentec task at one million tokens reaches five lies 40.

862
00:32:57,740 --> 00:32:59,260
These numbers are single instances,

863
00:32:59,260 --> 00:33:00,220
scale them across a team,

864
00:33:00,220 --> 00:33:01,580
across a day, across a quarter,

865
00:33:01,580 --> 00:33:04,540
and they accumulate into material line items on your Azure bill.

866
00:33:04,540 --> 00:33:07,020
But here's where it gets genuinely uncomfortable.

867
00:33:07,020 --> 00:33:08,860
Nvidia executives have publicly stated

868
00:33:08,860 --> 00:33:11,580
that AI token-based pricing for enterprise automation

869
00:33:11,580 --> 00:33:14,540
is in some cases already exceeding the annualized cost

870
00:33:14,540 --> 00:33:17,180
of employee salaries for comparable workloads.

871
00:33:17,180 --> 00:33:18,300
Let that sink in.

872
00:33:18,300 --> 00:33:20,940
A single sophisticated agentec workflow running daily,

873
00:33:20,940 --> 00:33:23,660
consuming a million tokens at frontier model prices

874
00:33:23,660 --> 00:33:26,140
can cost more per year than paying someone a salary

875
00:33:26,140 --> 00:33:27,660
to do the work manually.

876
00:33:27,660 --> 00:33:29,740
Without guardrails, without limits,

877
00:33:29,740 --> 00:33:32,540
without cost awareness built into how the system decides

878
00:33:32,540 --> 00:33:35,100
which workflows get to run at which model tiers.

879
00:33:35,100 --> 00:33:38,700
The cost explosion problem isn't unique to any single vendor

880
00:33:38,700 --> 00:33:40,220
or model, it's structural.

881
00:33:40,220 --> 00:33:44,060
Agentec systems work by decomposing complex goals into sub-tasks,

882
00:33:44,060 --> 00:33:46,940
then executing those sub-tasks with AI assistance.

883
00:33:46,940 --> 00:33:49,020
Every sub-task is a prompt response cycle.

884
00:33:49,020 --> 00:33:50,540
Every prompt carries context.

885
00:33:50,540 --> 00:33:52,460
Every response generates output.

886
00:33:52,460 --> 00:33:55,180
The system's intelligence, its ability to break down problems

887
00:33:55,180 --> 00:33:57,740
and iterate towards solutions is the same property

888
00:33:57,740 --> 00:33:59,340
that makes it expensive at scale.

889
00:33:59,340 --> 00:34:02,940
Organizations deploying agentec workflows in 2026

890
00:34:02,940 --> 00:34:04,620
are discovering this in real time.

891
00:34:04,620 --> 00:34:07,660
A knowledge management agent that crawls your document repository

892
00:34:07,660 --> 00:34:09,980
to answer questions can generate hundreds of thousands

893
00:34:09,980 --> 00:34:12,940
of tokens per user per week if nobody constrained its search scope.

894
00:34:12,940 --> 00:34:16,140
A customer service agent that roots inquiries

895
00:34:16,140 --> 00:34:19,260
to appropriate teams can spawn multiple parallel workflows

896
00:34:19,260 --> 00:34:23,180
investigating the same issue if nobody built in deduplication logic.

897
00:34:23,180 --> 00:34:25,740
A compliance agent that generates regulatory summaries

898
00:34:25,740 --> 00:34:28,860
across multiple jurisdictions can run unbounded searches

899
00:34:28,860 --> 00:34:30,860
across your entire legal repository

900
00:34:30,860 --> 00:34:33,260
unless someone explicitly capped the query depth.

901
00:34:33,260 --> 00:34:35,100
These aren't failures of model capability.

902
00:34:35,100 --> 00:34:36,540
They're failures of governance.

903
00:34:36,540 --> 00:34:38,940
An agentec system without spending controls

904
00:34:38,940 --> 00:34:41,580
is an agentec system that will eventually generate a bill

905
00:34:41,580 --> 00:34:42,700
nobody expected.

906
00:34:42,700 --> 00:34:45,740
The classic incident, now becoming common enough to have patterns,

907
00:34:45,740 --> 00:34:47,660
involves sub-agent fan-out.

908
00:34:47,660 --> 00:34:51,100
One agent spawns multiple child agents to parallelize work.

909
00:34:51,100 --> 00:34:54,540
Each child agent independently generates tokens consuming compute

910
00:34:54,540 --> 00:34:57,660
and without a circuit breaker, the system scales to absurd costs.

911
00:34:57,660 --> 00:35:01,580
GitHub users in early 2026 experienced incidents

912
00:35:01,580 --> 00:35:05,500
were a single agentec workflow spike generated $47,000

913
00:35:05,500 --> 00:35:07,580
in compute charges in the course of an afternoon

914
00:35:07,580 --> 00:35:09,820
with no human intervention or oversight.

915
00:35:09,820 --> 00:35:11,420
The agent worked exactly as designed.

916
00:35:11,420 --> 00:35:13,740
It just wasn't designed with cost as a constraint.

917
00:35:13,740 --> 00:35:16,140
Token costs must be managed like labor costs,

918
00:35:16,140 --> 00:35:18,220
with budgets, with utilization targets,

919
00:35:18,220 --> 00:35:20,780
with clear cost per business outcome metrics

920
00:35:20,780 --> 00:35:22,620
that force explicit trade-offs.

921
00:35:22,620 --> 00:35:25,100
Run this agent on the expensive reasoning model

922
00:35:25,100 --> 00:35:27,100
or run it on the cheaper general model.

923
00:35:27,100 --> 00:35:28,860
Root this workflow through a local model

924
00:35:28,860 --> 00:35:30,380
or call the Frontier API.

925
00:35:30,380 --> 00:35:31,740
These aren't technical questions.

926
00:35:31,740 --> 00:35:33,020
They're business questions,

927
00:35:33,020 --> 00:35:35,180
and they require the same governance apparatus

928
00:35:35,180 --> 00:35:37,740
you'd build for any significant operational cost.

929
00:35:37,740 --> 00:35:40,300
The enterprise without token governance infrastructure in place

930
00:35:40,300 --> 00:35:43,660
is the enterprise that discovers a $47,000 incident

931
00:35:43,660 --> 00:35:45,100
on its Azure bill,

932
00:35:45,100 --> 00:35:47,260
and realizes nobody owns the decision

933
00:35:47,260 --> 00:35:50,460
about which agents can run at what cost for what benefit.

934
00:35:50,460 --> 00:35:52,940
Now we have the full picture of what's bleeding cash.

935
00:35:52,940 --> 00:35:55,180
The question is where the money is actually well spent.

936
00:35:55,180 --> 00:35:57,500
Where Copilot actually wins.

937
00:35:57,500 --> 00:35:59,660
The argument so far has been relentlessly negative

938
00:35:59,660 --> 00:36:00,540
and for good reason.

939
00:36:00,540 --> 00:36:01,980
The structural flaws are real.

940
00:36:01,980 --> 00:36:03,420
The waste is quantifiable.

941
00:36:03,420 --> 00:36:05,180
The governance gaps are consequential,

942
00:36:05,180 --> 00:36:06,860
but Copilot ROI is not a myth.

943
00:36:06,860 --> 00:36:07,580
It's real.

944
00:36:07,580 --> 00:36:09,180
It's just highly concentrated,

945
00:36:09,180 --> 00:36:11,180
and most organizations are deploying it

946
00:36:11,180 --> 00:36:12,780
in a way that diffuses the return

947
00:36:12,780 --> 00:36:14,780
across too many seats to make it visible.

948
00:36:14,780 --> 00:36:16,940
The break-even threshold is simple to calculate.

949
00:36:16,940 --> 00:36:19,420
At a fully loaded labor cost of $50 per hour,

950
00:36:19,420 --> 00:36:22,220
a $30 per month license becomes economically rational

951
00:36:22,220 --> 00:36:25,420
if the user captures just 36 minutes of save time per week.

952
00:36:25,420 --> 00:36:27,260
36 minutes, not even an hour.

953
00:36:27,260 --> 00:36:28,860
For that 36-minute threshold,

954
00:36:28,860 --> 00:36:32,060
an organization has already crossed into positive ROI territory.

955
00:36:32,060 --> 00:36:34,780
The CFO doesn't need to believe in productivity magic.

956
00:36:34,780 --> 00:36:38,460
Just 36 minutes of actual time recapture per week per user.

957
00:36:38,460 --> 00:36:41,180
Real-world data from organizations that deployed Copilot

958
00:36:41,180 --> 00:36:42,940
with intentional adoption programming,

959
00:36:42,940 --> 00:36:44,220
not just seat assignment,

960
00:36:44,220 --> 00:36:46,380
but structured training, champion networks,

961
00:36:46,380 --> 00:36:48,060
role-specific prompting libraries,

962
00:36:48,060 --> 00:36:50,060
and measurable success metrics,

963
00:36:50,060 --> 00:36:53,420
shows that users consistently reach three to four hours per week

964
00:36:53,420 --> 00:36:54,540
saved at maturity.

965
00:36:54,540 --> 00:36:58,380
A forest-to-total economic impact study commissioned by Microsoft,

966
00:36:58,380 --> 00:37:00,220
working with client data from organizations

967
00:37:00,220 --> 00:37:03,500
that actively drove adoption and process redesign around Copilot,

968
00:37:03,500 --> 00:37:07,260
reported 116% ROI in the first year alone.

969
00:37:07,260 --> 00:37:08,620
That's a flaw, not a ceiling,

970
00:37:08,620 --> 00:37:10,300
but here's the critical constraint.

971
00:37:10,300 --> 00:37:12,860
ROI concentrates in specific work patterns,

972
00:37:12,860 --> 00:37:14,780
not all roles, not all tasks,

973
00:37:14,780 --> 00:37:16,700
specific roles, specific workflows where

974
00:37:16,700 --> 00:37:18,860
Copilot's strengths align with actual bottlenecks

975
00:37:18,860 --> 00:37:19,980
in how work gets done.

976
00:37:19,980 --> 00:37:21,820
The high-value profile is straightforward,

977
00:37:21,820 --> 00:37:24,140
document heavy roles, communication heavy roles,

978
00:37:24,140 --> 00:37:27,180
analysis heavy roles, knowledge workers who spend hours each week

979
00:37:27,180 --> 00:37:30,060
in outlook, teams, word, Excel, and PowerPoint,

980
00:37:30,060 --> 00:37:31,660
people who draft, revise,

981
00:37:31,660 --> 00:37:34,300
summarize, search, analyze, and synthesize

982
00:37:34,300 --> 00:37:35,580
as the core of their job,

983
00:37:35,580 --> 00:37:37,260
not as a marginal activity.

984
00:37:37,260 --> 00:37:39,260
The strongest use cases are concrete.

985
00:37:39,260 --> 00:37:41,660
Drafting proposals and RFP responses,

986
00:37:41,660 --> 00:37:43,420
Copilot can scaffold the structure,

987
00:37:43,420 --> 00:37:46,460
pull relevant prior language and generate first-pass content

988
00:37:46,460 --> 00:37:48,860
that a senior consultant refines into the final bid.

989
00:37:48,860 --> 00:37:52,060
Summarizing meeting transcripts and long documents,

990
00:37:52,060 --> 00:37:54,460
Copilot can extract decisions, action items,

991
00:37:54,460 --> 00:37:57,660
and key points from hours of recorded content in seconds.

992
00:37:57,660 --> 00:38:01,260
Cross-system search across email, teams, and SharePoint,

993
00:38:01,260 --> 00:38:04,060
retrieving relevant context from organizational memory

994
00:38:04,060 --> 00:38:05,740
without manual navigation.

995
00:38:05,740 --> 00:38:08,380
First-pass Excel analysis, generating formulas,

996
00:38:08,380 --> 00:38:10,940
suggesting pivot structures, identifying outliers

997
00:38:10,940 --> 00:38:13,500
before a financial analyst fine-tunes the model.

998
00:38:13,500 --> 00:38:15,420
These aren't speculative use cases,

999
00:38:15,420 --> 00:38:17,260
sales organizations, consulting firms,

1000
00:38:17,260 --> 00:38:19,580
legal departments, and product management teams

1001
00:38:19,580 --> 00:38:22,060
have reported consistent three to four hours per week

1002
00:38:22,060 --> 00:38:23,660
saved a deployment maturity.

1003
00:38:23,660 --> 00:38:26,460
At a $75,000 fully loaded annual salary

1004
00:38:26,460 --> 00:38:30,060
that's $5,000 since $25 in annual time value per user.

1005
00:38:30,060 --> 00:38:32,380
The license costs $360, the math is not close.

1006
00:38:32,380 --> 00:38:34,620
The ROI multiplier that distinguishes thriving

1007
00:38:34,620 --> 00:38:37,820
copilot deployments from failing ones is adoption depth.

1008
00:38:37,820 --> 00:38:40,460
Organizations where 70% or more of licensed users

1009
00:38:40,460 --> 00:38:43,500
invoke copilot daily in at least one core workflow

1010
00:38:43,500 --> 00:38:46,300
see 200 to 400% first year ROI.

1011
00:38:46,300 --> 00:38:49,340
Organizations stuck at 30% weekly active usage

1012
00:38:49,340 --> 00:38:50,780
see near zero return.

1013
00:38:50,780 --> 00:38:53,340
Same license costs, same model, same potential,

1014
00:38:53,340 --> 00:38:55,020
three to four times the actual outcome,

1015
00:38:55,020 --> 00:38:57,820
driven entirely by whether the organization engineered usage

1016
00:38:57,820 --> 00:39:00,860
into the workflow or just assigned seats and hoped.

1017
00:39:00,860 --> 00:39:02,620
This is where the tiered access framework

1018
00:39:02,620 --> 00:39:04,220
becomes not a cost-cutting exercise

1019
00:39:04,220 --> 00:39:05,660
but a concentration strategy.

1020
00:39:05,660 --> 00:39:07,340
You're not trying to minimize copilot spend

1021
00:39:07,340 --> 00:39:10,300
that you're trying to maximize copilot ROI per dollar spent

1022
00:39:10,300 --> 00:39:12,220
which means concentrating licenses

1023
00:39:12,220 --> 00:39:14,540
where the work pattern matches copilot's strengths

1024
00:39:14,540 --> 00:39:16,460
where process redesign is feasible

1025
00:39:16,460 --> 00:39:18,060
and where the labor costs saved

1026
00:39:18,060 --> 00:39:20,060
justifies the infrastructure cost.

1027
00:39:20,060 --> 00:39:22,860
And letting everyone else access the free or bundled tier

1028
00:39:22,860 --> 00:39:25,260
for the cases where copilot is a utility

1029
00:39:25,260 --> 00:39:26,700
not a critical capability.

1030
00:39:26,700 --> 00:39:28,700
The organization seeing the best outcomes

1031
00:39:28,700 --> 00:39:30,780
are not the ones that bought the most seats.

1032
00:39:30,780 --> 00:39:32,380
They're the ones that bought the right seats

1033
00:39:32,380 --> 00:39:34,540
in the right roles backed by structured enablement

1034
00:39:34,540 --> 00:39:35,980
governed by clear metrics

1035
00:39:35,980 --> 00:39:37,820
and resourced with change management.

1036
00:39:37,820 --> 00:39:40,460
That's how you move from a deployment that bleeds cash

1037
00:39:40,460 --> 00:39:42,540
to a deployment that generates return.

1038
00:39:42,540 --> 00:39:45,340
Knowing where it wins tells you how to build the framework

1039
00:39:45,340 --> 00:39:48,380
that captures that value instead of defusing it across the organization

1040
00:39:48,380 --> 00:39:49,820
like expensive overhead.

1041
00:39:49,820 --> 00:39:51,340
The tiered access model.

1042
00:39:51,340 --> 00:39:53,740
The structural answer to both the zombie seat problem

1043
00:39:53,740 --> 00:39:56,780
and the agente cost explosion isn't to deploy more broadly

1044
00:39:56,780 --> 00:39:58,060
or more conservatively.

1045
00:39:58,060 --> 00:39:59,660
It's to deploy differently

1046
00:39:59,660 --> 00:40:01,660
with explicit tiers that separate

1047
00:40:01,660 --> 00:40:04,140
who gets what capability under what conditions

1048
00:40:04,140 --> 00:40:05,580
and with what governance attached.

1049
00:40:05,580 --> 00:40:06,780
This isn't a new concept.

1050
00:40:06,780 --> 00:40:09,900
Financial institutions have used tiered access for years.

1051
00:40:09,900 --> 00:40:13,100
Basic customer service reps get red only access to accounts.

1052
00:40:13,100 --> 00:40:15,260
Supervisors get limited modification rights,

1053
00:40:15,260 --> 00:40:18,220
senior staff get unrestricted system access with logging.

1054
00:40:18,220 --> 00:40:19,980
You're applying the same principle to AI

1055
00:40:19,980 --> 00:40:21,820
but organized around value and risk

1056
00:40:21,820 --> 00:40:23,900
rather than organizational hierarchy.

1057
00:40:23,900 --> 00:40:25,580
The framework typically works like this.

1058
00:40:25,580 --> 00:40:28,300
tier three is broad access with minimal controls.

1059
00:40:28,300 --> 00:40:31,820
Think of this as the free copilot tier available to all staff.

1060
00:40:31,820 --> 00:40:33,500
These users get access to public

1061
00:40:33,500 --> 00:40:35,660
or well curated internal knowledge bases

1062
00:40:35,660 --> 00:40:38,540
but not direct access to sensitive corporate data.

1063
00:40:38,540 --> 00:40:40,540
They can use copilot for writing, translation,

1064
00:40:40,540 --> 00:40:43,580
summarization of non-sensitive material generic Q&A.

1065
00:40:43,580 --> 00:40:46,380
No agent creation, no custom integrations,

1066
00:40:46,380 --> 00:40:49,100
no ability to connect to line of business systems.

1067
00:40:49,100 --> 00:40:51,340
The model is constrained to prevent leakage

1068
00:40:51,340 --> 00:40:53,020
of confidential information

1069
00:40:53,020 --> 00:40:54,860
and the usage footprint is monitored

1070
00:40:54,860 --> 00:40:56,220
but not heavily governed.

1071
00:40:56,220 --> 00:40:58,460
This tier is cheap, either completely free

1072
00:40:58,460 --> 00:41:00,700
or bundled into existing licenses.

1073
00:41:00,700 --> 00:41:02,620
And the expectation is that a large portion

1074
00:41:02,620 --> 00:41:04,060
of the organization lands here.

1075
00:41:04,060 --> 00:41:05,900
tier two is for standard knowledge workers

1076
00:41:05,900 --> 00:41:08,940
whose role directly benefits from deeper AI capability.

1077
00:41:08,940 --> 00:41:11,100
They get full M365 copilot licensing

1078
00:41:11,100 --> 00:41:12,700
with access to departmental content,

1079
00:41:12,700 --> 00:41:14,620
their email, their teams channels,

1080
00:41:14,620 --> 00:41:16,300
their shared sharepoint sites.

1081
00:41:16,300 --> 00:41:17,900
They can ask copilot questions

1082
00:41:17,900 --> 00:41:19,340
about organizational policies

1083
00:41:19,340 --> 00:41:21,180
internal processes customer accounts

1084
00:41:21,180 --> 00:41:22,460
within their department.

1085
00:41:22,460 --> 00:41:23,980
Standard app integrations with word,

1086
00:41:23,980 --> 00:41:25,740
excel, PowerPoint, outlook, and teams.

1087
00:41:25,740 --> 00:41:27,660
They can run basic automations and workflows

1088
00:41:27,660 --> 00:41:29,180
but cannot create custom agents

1089
00:41:29,180 --> 00:41:30,620
or modify governance policies.

1090
00:41:30,620 --> 00:41:32,620
The cost is the 30-dap a month add-on

1091
00:41:32,620 --> 00:41:34,060
but it's concentrated on roles

1092
00:41:34,060 --> 00:41:37,180
where documented usage patterns show concrete ROI.

1093
00:41:37,180 --> 00:41:38,780
Sales teams, legal departments,

1094
00:41:38,780 --> 00:41:41,020
financial analysts, project managers.

1095
00:41:41,020 --> 00:41:42,540
Governance is active.

1096
00:41:42,540 --> 00:41:44,780
There are clear policies about what copilot can

1097
00:41:44,780 --> 00:41:45,980
and cannot access

1098
00:41:45,980 --> 00:41:48,620
with periodic reviews of usage patterns.

1099
00:41:48,620 --> 00:41:51,820
tier one is restricted access for high-value high-risk work.

1100
00:41:51,820 --> 00:41:53,340
This is where custom agents live.

1101
00:41:53,340 --> 00:41:55,900
This is where board-level or senior executive decisions

1102
00:41:55,900 --> 00:41:57,260
get AI assisted.

1103
00:41:57,260 --> 00:41:59,340
This is where sensitive data like M&A planning,

1104
00:41:59,340 --> 00:42:00,620
proprietary research,

1105
00:42:00,620 --> 00:42:03,500
or regulated compliance work flows through copilot.

1106
00:42:03,500 --> 00:42:05,180
Access is explicitly provisioned.

1107
00:42:05,180 --> 00:42:07,580
You don't get tier one access because of your job title.

1108
00:42:07,580 --> 00:42:10,780
You get it because a specific high-value workflow requires it

1109
00:42:10,780 --> 00:42:13,500
and the governance overhead is worth the business case.

1110
00:42:13,500 --> 00:42:16,380
These agents run under strict conditional access policies.

1111
00:42:16,380 --> 00:42:18,060
They have restricted sharepoint search scope.

1112
00:42:18,060 --> 00:42:19,020
They log everything.

1113
00:42:19,020 --> 00:42:20,780
Changes to their prompts or tool integrations

1114
00:42:20,780 --> 00:42:22,540
go through formal change management.

1115
00:42:22,540 --> 00:42:24,780
The compliance and legal review is heavyweight.

1116
00:42:24,780 --> 00:42:26,380
The cost per agent is higher

1117
00:42:26,380 --> 00:42:28,380
not because the license is more expensive

1118
00:42:28,380 --> 00:42:31,660
but because the governance infrastructure around it is more expensive.

1119
00:42:31,660 --> 00:42:33,660
The business case for tiering is straightforward.

1120
00:42:33,660 --> 00:42:36,060
Concentrate the $30 per month add-on spend

1121
00:42:36,060 --> 00:42:38,380
on the 35% of users who will actually generate

1122
00:42:38,380 --> 00:42:40,620
3-4 hours per week of time value.

1123
00:42:40,620 --> 00:42:43,660
Let the other 65% use free or bundled copilot

1124
00:42:43,660 --> 00:42:45,900
for the cases where they benefit incidentally.

1125
00:42:45,900 --> 00:42:48,300
The reallocation isn't a cost-cutting exercise.

1126
00:42:48,300 --> 00:42:50,380
It's a return optimization exercise.

1127
00:42:50,380 --> 00:42:52,300
You're moving budget from low utilization seats

1128
00:42:52,300 --> 00:42:53,820
to high impact workflows.

1129
00:42:53,820 --> 00:42:55,580
You're recovering the zombie seat spend

1130
00:42:55,580 --> 00:42:57,820
and redirecting it toward agents and capabilities

1131
00:42:57,820 --> 00:42:59,900
that drive measurable business outcomes.

1132
00:42:59,900 --> 00:43:02,140
Access tiers must be tied to governance controls

1133
00:43:02,140 --> 00:43:03,660
not just job titles.

1134
00:43:03,660 --> 00:43:06,460
Sales director doesn't automatically equal tier two.

1135
00:43:06,460 --> 00:43:07,500
Rather you ask,

1136
00:43:07,500 --> 00:43:10,540
does this role spend significant time in document heavy workflows

1137
00:43:10,540 --> 00:43:12,860
where AI can materially reduce cycle time?

1138
00:43:12,860 --> 00:43:14,380
Does the team have process discipline

1139
00:43:14,380 --> 00:43:16,220
around how AI outputs get used?

1140
00:43:16,220 --> 00:43:17,980
Are there clear success metrics?

1141
00:43:17,980 --> 00:43:20,060
If yes to all three, provision tier two.

1142
00:43:20,060 --> 00:43:21,820
If not, tier three with a clear path

1143
00:43:21,820 --> 00:43:23,820
to upgrade one's conditions are met.

1144
00:43:23,820 --> 00:43:25,580
This is a portfolio management decision

1145
00:43:25,580 --> 00:43:27,660
not an IT configuration decision.

1146
00:43:27,660 --> 00:43:30,780
You're not optimizing for how many people can we give AI to.

1147
00:43:30,780 --> 00:43:34,140
You're optimizing for how do we maximize business value per dollar

1148
00:43:34,140 --> 00:43:37,500
of AI spend while keeping governance risk under control.

1149
00:43:37,500 --> 00:43:40,140
The tiering framework is what enables that optimization

1150
00:43:40,140 --> 00:43:41,900
to actually function as a lived practice

1151
00:43:41,900 --> 00:43:44,140
instead of a theoretical ideal.

1152
00:43:44,140 --> 00:43:46,140
Token governance as a financial discipline.

1153
00:43:46,140 --> 00:43:48,140
Once you've built the tiered access framework,

1154
00:43:48,140 --> 00:43:49,580
you've solved the structural problem

1155
00:43:49,580 --> 00:43:51,260
of who gets what capability,

1156
00:43:51,260 --> 00:43:53,180
but you haven't solved the operational problem

1157
00:43:53,180 --> 00:43:55,900
of managing how much that capability costs once it's in use.

1158
00:43:55,900 --> 00:43:58,620
That requires a different kind of discipline entirely.

1159
00:43:58,620 --> 00:44:01,260
Think of token governance as FinOps applied to AI.

1160
00:44:01,260 --> 00:44:03,100
Cloud cost management emerged as a discipline

1161
00:44:03,100 --> 00:44:05,180
because cloud infrastructure is metered.

1162
00:44:05,180 --> 00:44:07,020
You get a bill for every compute hour,

1163
00:44:07,020 --> 00:44:09,820
every gigabyte stored, every terabyte transferred.

1164
00:44:09,820 --> 00:44:12,620
Organizations that didn't actively manage cloud spend

1165
00:44:12,620 --> 00:44:15,740
watched their bills grow 30 to 40% year over year,

1166
00:44:15,740 --> 00:44:18,460
even when actual workload demand stayed flat.

1167
00:44:18,460 --> 00:44:20,220
The solution wasn't to use less cloud.

1168
00:44:20,220 --> 00:44:21,900
It was to build operational practices

1169
00:44:21,900 --> 00:44:24,220
that made cloud cost visible, attributable,

1170
00:44:24,220 --> 00:44:25,900
and subject to budget discipline.

1171
00:44:25,900 --> 00:44:27,900
Token governance applies the same principle

1172
00:44:27,900 --> 00:44:29,260
to LLM consumption.

1173
00:44:29,260 --> 00:44:31,820
The first operational step is to segment your user base

1174
00:44:31,820 --> 00:44:33,180
into consumption profiles.

1175
00:44:33,180 --> 00:44:34,700
Not everyone is a heavy user.

1176
00:44:34,700 --> 00:44:36,860
Light users, those using co-pilot occasionally,

1177
00:44:36,860 --> 00:44:38,300
mostly for straightforward tasks,

1178
00:44:38,300 --> 00:44:40,940
like email summarization or quick document search,

1179
00:44:40,940 --> 00:44:44,220
consume between 40,000 and 100,000 tokens per month.

1180
00:44:44,220 --> 00:44:47,420
Heavy analytical users, consultants building research-backed

1181
00:44:47,420 --> 00:44:50,300
deliverables, engineers debugging complex systems

1182
00:44:50,300 --> 00:44:53,420
with extended context, analyst processing documents sets,

1183
00:44:53,420 --> 00:44:55,900
can reach 1.6 million tokens per month.

1184
00:44:55,900 --> 00:44:58,700
That's a 16 to 44 difference in consumption

1185
00:44:58,700 --> 00:45:00,220
from the same license type.

1186
00:45:00,220 --> 00:45:03,020
Understanding where your population actually falls on that curve

1187
00:45:03,020 --> 00:45:05,180
is the foundation for meaningful governance.

1188
00:45:05,180 --> 00:45:07,260
But consumption tracking alone is sterile

1189
00:45:07,260 --> 00:45:08,700
if you're just counting tokens.

1190
00:45:08,700 --> 00:45:10,780
What matters is cost per business outcome.

1191
00:45:10,780 --> 00:45:12,140
Don't track cost per token.

1192
00:45:12,140 --> 00:45:15,580
Track cost per merged pull request, cost per proposal drafted,

1193
00:45:15,580 --> 00:45:17,580
cost per support case resolved.

1194
00:45:17,580 --> 00:45:20,140
That reorientation accomplishes two things.

1195
00:45:20,140 --> 00:45:22,540
First, it ties spending to actual value,

1196
00:45:22,540 --> 00:45:25,260
making it clear when token consumption is growing faster

1197
00:45:25,260 --> 00:45:27,500
than the outcomes it's producing.

1198
00:45:27,500 --> 00:45:29,900
Second, it creates a natural conversation

1199
00:45:29,900 --> 00:45:32,140
about whether the outcome justifies the spend,

1200
00:45:32,140 --> 00:45:35,420
which is the core question any business discipline should answer.

1201
00:45:35,420 --> 00:45:38,220
Prompt governance is the most immediately actionable lever,

1202
00:45:38,220 --> 00:45:40,140
large persistent instruction blocks,

1203
00:45:40,140 --> 00:45:42,460
shared system prompts, global configuration files,

1204
00:45:42,460 --> 00:45:44,620
detailed context about your organization's policies

1205
00:45:44,620 --> 00:45:47,980
and processes, get injected into every single interaction.

1206
00:45:47,980 --> 00:45:50,220
If that shared context is 2,000 tokens,

1207
00:45:50,220 --> 00:45:52,940
and it lives in every prompt, then every co-pilot user

1208
00:45:52,940 --> 00:45:55,340
is paying a 2,000 token tax on all interactions,

1209
00:45:55,340 --> 00:45:58,540
whether that context is relevant to the specific query or not.

1210
00:45:58,540 --> 00:46:01,180
Trim that shared instruction set to essentials.

1211
00:46:01,180 --> 00:46:03,020
Segment context by role or department

1212
00:46:03,020 --> 00:46:05,740
so users only carry the instructions relevant to their work.

1213
00:46:05,740 --> 00:46:08,700
The Claude code case studies document 40% cost reductions

1214
00:46:08,700 --> 00:46:11,180
from simply moving large persistent context files

1215
00:46:11,180 --> 00:46:13,900
out of active prompts and into documentation or knowledge

1216
00:46:13,900 --> 00:46:15,980
bases that get consulted when needed

1217
00:46:15,980 --> 00:46:18,060
rather than injected always.

1218
00:46:18,060 --> 00:46:20,300
Session lifecycle control compounds this.

1219
00:46:20,300 --> 00:46:22,540
An IDE session left open overnight,

1220
00:46:22,540 --> 00:46:24,220
where a developer had a rich conversation

1221
00:46:24,220 --> 00:46:26,140
about a code base earlier in the day

1222
00:46:26,140 --> 00:46:28,140
carries the full conversation history

1223
00:46:28,140 --> 00:46:30,780
into every subsequent interaction the next morning.

1224
00:46:30,780 --> 00:46:33,020
When that developer switches to a completely different code base,

1225
00:46:33,020 --> 00:46:35,100
that old context is still being carried forward,

1226
00:46:35,100 --> 00:46:37,980
still being charged as input tokens on every new request.

1227
00:46:37,980 --> 00:46:40,060
Encourage users to reset or clear sessions

1228
00:46:40,060 --> 00:46:42,380
when switching tasks, make session reset

1229
00:46:42,380 --> 00:46:44,460
a deliberate action in workflow guidance.

1230
00:46:44,460 --> 00:46:47,340
The cognitive load on the model and the token bill

1231
00:46:47,340 --> 00:46:49,820
drops as soon as that irrelevant context is dropped.

1232
00:46:49,820 --> 00:46:52,220
Model rooting is where discipline token governance

1233
00:46:52,220 --> 00:46:54,540
becomes cost reduction theater.

1234
00:46:54,540 --> 00:46:56,380
Not every task requires a frontier model.

1235
00:46:56,380 --> 00:46:58,460
A routine code completion can run on a smaller,

1236
00:46:58,460 --> 00:46:59,660
faster, cheaper model.

1237
00:46:59,660 --> 00:47:01,580
Complex multi-step reasoning should escalate

1238
00:47:01,580 --> 00:47:03,020
to a more capable tier.

1239
00:47:03,020 --> 00:47:05,500
If 70% of your interactions are straightforward

1240
00:47:05,500 --> 00:47:08,220
and 30% are complex, root the straightforward work

1241
00:47:08,220 --> 00:47:10,060
to a cheaper tier automatically.

1242
00:47:10,060 --> 00:47:12,620
The documented case studies show this discipline alone.

1243
00:47:12,620 --> 00:47:15,260
Deliberate model selection based on task complexity,

1244
00:47:15,260 --> 00:47:18,140
rather than defaulting everything to your most capable model,

1245
00:47:18,140 --> 00:47:22,380
produces between 70 and 85% reductions in total token spend

1246
00:47:22,380 --> 00:47:25,660
while maintaining quality on the tasks that require it.

1247
00:47:25,660 --> 00:47:27,420
Finally, set spend limits and alerts,

1248
00:47:27,420 --> 00:47:30,300
treat token budgets the way you treat headcount budgets.

1249
00:47:30,300 --> 00:47:32,620
Department gets allocated X tokens per month.

1250
00:47:32,620 --> 00:47:35,020
Once they approach the limit, they get visibility

1251
00:47:35,020 --> 00:47:36,860
and a conversation with finance.

1252
00:47:36,860 --> 00:47:38,940
Individual workflows get spending thresholds.

1253
00:47:38,940 --> 00:47:40,700
Agentex systems get circuit breakers

1254
00:47:40,700 --> 00:47:42,700
that prevent runaway token consumption

1255
00:47:42,700 --> 00:47:45,500
without explicit limits, cost grows invisibly

1256
00:47:45,500 --> 00:47:47,580
until an invoice arrives that nobody expected.

1257
00:47:47,580 --> 00:47:50,540
Governance without measurement is policy theater.

1258
00:47:50,540 --> 00:47:53,740
Governance with measurement is operational reality.

1259
00:47:53,740 --> 00:47:55,500
The ROI-ordered framework,

1260
00:47:55,500 --> 00:47:58,540
measurement and governance work backward from a simple fact.

1261
00:47:58,540 --> 00:48:00,460
Most organizations deployed co-pilot

1262
00:48:00,460 --> 00:48:02,860
without capturing what before looked like.

1263
00:48:02,860 --> 00:48:06,300
You can't calculate ROI if you never measured the baseline.

1264
00:48:06,300 --> 00:48:07,740
You can't prove the time was saved

1265
00:48:07,740 --> 00:48:10,700
if you didn't document how long the task took before AI arrived.

1266
00:48:10,700 --> 00:48:13,420
This baseline problem compounds across every other dimension

1267
00:48:13,420 --> 00:48:14,620
of an audit framework.

1268
00:48:14,620 --> 00:48:18,620
A real co-pilot ROI audit operates across eight interconnected dimensions.

1269
00:48:18,620 --> 00:48:20,620
Each one answers a specific question

1270
00:48:20,620 --> 00:48:22,460
that feeds into the others.

1271
00:48:22,460 --> 00:48:24,700
Dimension 1 is strategic alignment.

1272
00:48:24,700 --> 00:48:27,900
Does every licensed user have a documented relevant use case?

1273
00:48:27,900 --> 00:48:30,300
Not maybe they'll use this eventually?

1274
00:48:30,300 --> 00:48:34,140
Documented, a concrete answer to where does this person use co-pilot?

1275
00:48:34,140 --> 00:48:35,420
If you can't articulate that,

1276
00:48:35,420 --> 00:48:38,620
the license is speculative spend, not strategic investment.

1277
00:48:38,620 --> 00:48:41,820
What percentage of your licenses align to high-impact processes

1278
00:48:41,820 --> 00:48:44,060
where documented ROI patterns exist?

1279
00:48:44,060 --> 00:48:47,100
Sales, legal consulting, financial analysis?

1280
00:48:47,100 --> 00:48:49,980
Versus roles where co-pilot adoption is still experimental?

1281
00:48:49,980 --> 00:48:53,740
Dimension 2 is baseline and benefit measurement.

1282
00:48:53,740 --> 00:48:55,500
This is where the actual time data lives.

1283
00:48:55,500 --> 00:48:58,140
How long did a specific task take before co-pilot?

1284
00:48:58,140 --> 00:48:59,180
How long after?

1285
00:48:59,180 --> 00:49:00,140
Take the time saved,

1286
00:49:00,140 --> 00:49:01,980
multiply by fully loaded labor cost,

1287
00:49:01,980 --> 00:49:03,340
multiply by adoption rate.

1288
00:49:03,340 --> 00:49:04,620
That's your monetized benefit,

1289
00:49:04,620 --> 00:49:07,500
compare it to license cost plus enablement overhead.

1290
00:49:07,500 --> 00:49:10,220
This is the one dimension you cannot eyeball or guess at.

1291
00:49:10,220 --> 00:49:12,300
The CFO will not accept survey responses

1292
00:49:12,300 --> 00:49:14,620
about how co-pilot feels faster.

1293
00:49:14,620 --> 00:49:17,660
They will accept before and after cycle times against defined tasks.

1294
00:49:17,660 --> 00:49:19,420
Dimension 3 is adoption behavior.

1295
00:49:19,420 --> 00:49:22,060
What fraction of licensed users invoke co-pilot daily

1296
00:49:22,060 --> 00:49:23,900
in at least one core workflow?

1297
00:49:23,900 --> 00:49:25,020
Weekly, monthly,

1298
00:49:25,020 --> 00:49:27,100
this is where usage telemetry becomes ground truth.

1299
00:49:27,100 --> 00:49:29,020
The admin center usage reports show this.

1300
00:49:29,020 --> 00:49:30,380
The graph APIs expose it.

1301
00:49:30,380 --> 00:49:31,820
You need to know the conversion rate,

1302
00:49:31,820 --> 00:49:32,700
the real number,

1303
00:49:32,700 --> 00:49:34,380
not the aspirational one.

1304
00:49:34,380 --> 00:49:37,420
Organizations with 70% or higher daily active usage

1305
00:49:37,420 --> 00:49:39,580
see fundamentally different ROI profiles

1306
00:49:39,580 --> 00:49:41,180
than organizations at 30%.

1307
00:49:41,180 --> 00:49:42,460
The metric is not controversial.

1308
00:49:42,460 --> 00:49:44,140
What organizations do with the metric

1309
00:49:44,140 --> 00:49:45,980
is where the real discipline lives.

1310
00:49:45,980 --> 00:49:48,700
Dimension 4 is technical readiness and performance.

1311
00:49:48,700 --> 00:49:51,660
Are all users on supported M365 update channels?

1312
00:49:51,660 --> 00:49:53,340
Do they have the right base licenses?

1313
00:49:53,340 --> 00:49:55,340
Is there latency or reliability friction

1314
00:49:55,340 --> 00:49:56,940
that suppresses adoption?

1315
00:49:56,940 --> 00:49:58,860
Technical friction doesn't feel like a cost

1316
00:49:58,860 --> 00:50:01,100
until you realize it's preventing the time savings

1317
00:50:01,100 --> 00:50:02,380
you're supposed to be capturing.

1318
00:50:02,380 --> 00:50:04,380
Dimension 5 is data governance quality.

1319
00:50:04,380 --> 00:50:06,220
Can co-pilot deliver useful answers

1320
00:50:06,220 --> 00:50:08,540
if your SharePoint is a maze of overshared folders

1321
00:50:08,540 --> 00:50:09,820
and unlabeled documents?

1322
00:50:09,820 --> 00:50:12,220
A cleaner data estate increases ROI

1323
00:50:12,220 --> 00:50:13,740
by improving answer quality

1324
00:50:13,740 --> 00:50:16,140
while simultaneously reducing compliance risk.

1325
00:50:16,140 --> 00:50:17,340
This isn't theoretical.

1326
00:50:17,340 --> 00:50:18,780
Organizations that skip this step

1327
00:50:18,780 --> 00:50:20,540
discover it through a security incident,

1328
00:50:20,540 --> 00:50:21,900
not a success metric.

1329
00:50:21,900 --> 00:50:24,540
Dimension 6 is cost and licensing optimization.

1330
00:50:24,540 --> 00:50:27,180
Identify which licenses are actually generating value

1331
00:50:27,180 --> 00:50:28,540
and which are occupying seats.

1332
00:50:28,540 --> 00:50:29,740
Find users whose role changed

1333
00:50:29,740 --> 00:50:31,820
and no longer justifies their current skew.

1334
00:50:31,820 --> 00:50:34,860
Reclaim licenses from consistently low usage accounts.

1335
00:50:34,860 --> 00:50:36,540
Reallocate to high value roles.

1336
00:50:36,540 --> 00:50:37,660
This is not cost-cutting.

1337
00:50:37,660 --> 00:50:41,420
This is capital redeployment from low return to high return uses.

1338
00:50:41,420 --> 00:50:43,900
Dimension 7 is change management and enablement.

1339
00:50:43,900 --> 00:50:46,940
Did you run formal training beyond an email announcement?

1340
00:50:46,940 --> 00:50:48,300
Do you have champions in each department

1341
00:50:48,300 --> 00:50:49,820
who can model usage and troubleshoot?

1342
00:50:49,820 --> 00:50:51,580
Is there an ongoing enablement program

1343
00:50:51,580 --> 00:50:53,260
or just an initial onboarding?

1344
00:50:53,260 --> 00:50:55,980
Change management has a direct correlation with adoption depth

1345
00:50:55,980 --> 00:50:58,940
and adoption depth drives ROI by three to four times.

1346
00:50:58,940 --> 00:51:01,500
Dimension 8 is continuous improvement and governance.

1347
00:51:01,500 --> 00:51:03,580
How often are you reviewing ROI metrics?

1348
00:51:03,580 --> 00:51:05,740
Are you adjusting which tiers users belong to

1349
00:51:05,740 --> 00:51:07,340
based on actual usage patterns?

1350
00:51:07,340 --> 00:51:08,620
They have you built feedback loops

1351
00:51:08,620 --> 00:51:12,300
where usage data informs policy and policy informs usage allocation.

1352
00:51:12,300 --> 00:51:14,300
The quarterly license audit is not optional.

1353
00:51:14,300 --> 00:51:16,540
It's the mechanism that converts recovered spend

1354
00:51:16,540 --> 00:51:17,820
into funded expansion.

1355
00:51:17,820 --> 00:51:19,980
Without it, waste accumulates invisibly

1356
00:51:19,980 --> 00:51:23,100
and the next budget cycle perpetuates the same inefficiency.

1357
00:51:23,100 --> 00:51:24,540
The audit tells you what's broken.

1358
00:51:24,540 --> 00:51:26,380
The next step is fixing the deployment model.

1359
00:51:26,380 --> 00:51:28,860
The self-funding AI strategy.

1360
00:51:28,860 --> 00:51:30,460
The audit identifies the waste.

1361
00:51:30,460 --> 00:51:33,580
The tiered framework concentrates where you actually want to spend.

1362
00:51:33,580 --> 00:51:36,780
But where does the budget come from to fund the expansion?

1363
00:51:36,780 --> 00:51:38,860
This is where the self-funding narrative shifts

1364
00:51:38,860 --> 00:51:41,180
from cost consciousness to capital strategy.

1365
00:51:41,180 --> 00:51:42,860
The mechanism is straightforward.

1366
00:51:42,860 --> 00:51:45,900
Independent advisors and SAM specialists consistently report

1367
00:51:45,900 --> 00:51:48,140
that organizations uncover between 10

1368
00:51:48,140 --> 00:51:50,460
and 20% in unnecessary license costs

1369
00:51:50,460 --> 00:51:52,860
during their first formal optimization cycle.

1370
00:51:52,860 --> 00:51:54,140
Not through aggressive cuts,

1371
00:51:54,140 --> 00:51:56,060
through reclaiming seats that are actively assigned

1372
00:51:56,060 --> 00:51:58,460
but generating zero measurable value?

1373
00:51:58,460 --> 00:52:01,500
Users who left the organization but still hold licenses.

1374
00:52:01,500 --> 00:52:04,460
Roles that changed and no longer justify E5 level access.

1375
00:52:04,460 --> 00:52:06,300
Adon's purchased for short term projects

1376
00:52:06,300 --> 00:52:09,500
that concluded months ago, continuous automated governance.

1377
00:52:09,500 --> 00:52:11,580
Where systems continuously audit usage

1378
00:52:11,580 --> 00:52:13,260
and reallocate based on behavior

1379
00:52:13,260 --> 00:52:15,420
rather than waiting for annual cycles,

1380
00:52:15,420 --> 00:52:18,620
pushes that recovery range to 20% to 30% over time.

1381
00:52:18,620 --> 00:52:20,780
For a 5,000 user organization

1382
00:52:20,780 --> 00:52:22,780
where 30% of seats are genuinely idle

1383
00:52:22,780 --> 00:52:25,020
at the current per seat cost of $30 per month,

1384
00:52:25,020 --> 00:52:28,540
that idle population is carrying $540,000 per year

1385
00:52:28,540 --> 00:52:31,020
in unproductive co-pilot spend alone.

1386
00:52:31,020 --> 00:52:32,940
At the base E3 or E5 licenses,

1387
00:52:32,940 --> 00:52:34,300
those users are also holding

1388
00:52:34,300 --> 00:52:36,860
and the total recoverable cost climbs significantly higher.

1389
00:52:36,860 --> 00:52:38,540
That's not a marginal adjustment.

1390
00:52:38,540 --> 00:52:41,500
That's substantial capital sitting in the wrong allocation.

1391
00:52:41,500 --> 00:52:42,860
The self-funding narrative works

1392
00:52:42,860 --> 00:52:45,180
because that recovered budget doesn't disappear.

1393
00:52:45,180 --> 00:52:47,980
It funds exactly the capabilities that generate return.

1394
00:52:47,980 --> 00:52:50,540
Targeted co-pilot expansion to the high ROI roles

1395
00:52:50,540 --> 00:52:52,860
you identified in your strategic alignment audit.

1396
00:52:52,860 --> 00:52:54,860
The sales teams and consulting practices

1397
00:52:54,860 --> 00:52:58,300
where documented usage patterns show three to four hours per week saved.

1398
00:52:58,300 --> 00:53:01,100
Co-pilot's studio agent development

1399
00:53:01,100 --> 00:53:03,180
for the two or three specific workflows

1400
00:53:03,180 --> 00:53:05,100
where you measured concrete business outcomes.

1401
00:53:05,100 --> 00:53:06,540
The compliance research agent,

1402
00:53:06,540 --> 00:53:07,900
the CRM routing agent,

1403
00:53:07,900 --> 00:53:09,580
the proposal scaffold agent,

1404
00:53:09,580 --> 00:53:12,220
Azure Open AI Capacity for Custom Applications,

1405
00:53:12,220 --> 00:53:14,220
where token economics favor direct billing

1406
00:53:14,220 --> 00:53:15,820
over per user licenses.

1407
00:53:15,820 --> 00:53:17,260
The specialized domain co-pilot

1408
00:53:17,260 --> 00:53:18,460
that needs fine tuning,

1409
00:53:18,460 --> 00:53:21,420
the vertical specific agent built on proprietary data.

1410
00:53:21,420 --> 00:53:23,020
The reallocation works at scale

1411
00:53:23,020 --> 00:53:26,060
because concentration beats coverage in terms of ROI.

1412
00:53:26,060 --> 00:53:27,820
We've already established that organizations

1413
00:53:27,820 --> 00:53:30,700
hitting 70% or higher weekly active usage

1414
00:53:30,700 --> 00:53:32,620
see three to four times the return

1415
00:53:32,620 --> 00:53:35,900
on the same co-pilot license compared to those stuck at 30%.

1416
00:53:35,900 --> 00:53:37,500
That's not because the model improved.

1417
00:53:37,500 --> 00:53:39,180
That's because the deployment changed.

1418
00:53:39,180 --> 00:53:41,180
The same $30 monthly investment produces

1419
00:53:41,180 --> 00:53:43,020
a four to one difference in outcome

1420
00:53:43,020 --> 00:53:45,580
based on whether you're spreading it thin across a population

1421
00:53:45,580 --> 00:53:47,260
or concentrating it in roles

1422
00:53:47,260 --> 00:53:50,300
where the work pattern makes AI genuinely useful.

1423
00:53:50,300 --> 00:53:53,260
The phased expansion model operationalizes this principle.

1424
00:53:53,260 --> 00:53:55,020
You don't recover zombie seat licenses

1425
00:53:55,020 --> 00:53:57,740
and immediately redeploy them to the next 1,000 people.

1426
00:53:57,740 --> 00:53:59,500
You pilot with a concentrated group

1427
00:53:59,500 --> 00:54:02,540
100 to 500 users in genuinely high value roles.

1428
00:54:02,540 --> 00:54:03,980
You measure rigorously.

1429
00:54:03,980 --> 00:54:06,940
Time data, adoption metrics, cost per outcome.

1430
00:54:06,940 --> 00:54:08,460
Only when the data supports it,

1431
00:54:08,460 --> 00:54:10,380
do you expand to the next cohort.

1432
00:54:10,380 --> 00:54:12,540
Then repeat, each phase is a funding gate.

1433
00:54:12,540 --> 00:54:13,500
You don't get more budget

1434
00:54:13,500 --> 00:54:15,660
unless the prior phase demonstrated return.

1435
00:54:15,660 --> 00:54:18,460
This approach inverts the risk profile of AI spending.

1436
00:54:18,460 --> 00:54:19,820
The blanket rollout model,

1437
00:54:19,820 --> 00:54:22,060
by many seats, hope adoption follows,

1438
00:54:22,060 --> 00:54:23,900
puts all the risk up front.

1439
00:54:23,900 --> 00:54:25,820
You've committed the budget and the infrastructure

1440
00:54:25,820 --> 00:54:27,980
before you know if it will generate return.

1441
00:54:27,980 --> 00:54:29,660
The self-funding model defers the risk.

1442
00:54:29,660 --> 00:54:31,420
You prove return with a constrained group.

1443
00:54:31,420 --> 00:54:34,060
You recover waste from the broader organization.

1444
00:54:34,060 --> 00:54:36,940
You reinvest that recovered capital into expansion

1445
00:54:36,940 --> 00:54:38,460
but only in the directions

1446
00:54:38,460 --> 00:54:40,620
where data already validates the outcome.

1447
00:54:40,620 --> 00:54:43,740
This is not a cost-cutting exercise disguised as strategy.

1448
00:54:43,740 --> 00:54:44,540
It's the opposite.

1449
00:54:44,540 --> 00:54:47,580
You're preserving or expanding your AI capability footprint

1450
00:54:47,580 --> 00:54:50,700
while dramatically improving the efficiency of every dollar deployed.

1451
00:54:50,700 --> 00:54:52,940
The total AI budget doesn't necessarily shrink.

1452
00:54:52,940 --> 00:54:55,020
The return per dollar dramatically improves.

1453
00:54:55,020 --> 00:54:56,540
And the funding mechanism,

1454
00:54:56,540 --> 00:54:59,340
reallocation from low return to high return uses,

1455
00:54:59,340 --> 00:55:00,540
is internally sustainable,

1456
00:55:00,540 --> 00:55:03,500
which means you're not asking the board for more budget

1457
00:55:03,500 --> 00:55:05,180
to fund better AI strategy.

1458
00:55:05,180 --> 00:55:07,660
You're funding it by not wasting what you already have.

1459
00:55:07,660 --> 00:55:10,220
Before you can execute this reallocation effectively, though,

1460
00:55:10,220 --> 00:55:12,700
you need to address one structural prerequisite

1461
00:55:12,700 --> 00:55:15,020
that most organizations haven't actually solved.

1462
00:55:15,020 --> 00:55:17,500
The data governance prerequisite

1463
00:55:17,500 --> 00:55:19,260
Before you can execute any of this,

1464
00:55:19,260 --> 00:55:22,620
the tiered framework, the token governance, the ROI audit,

1465
00:55:22,620 --> 00:55:23,820
there's a foundational layer

1466
00:55:23,820 --> 00:55:26,300
that almost every organization underestimates.

1467
00:55:26,300 --> 00:55:28,060
Data quality and access control.

1468
00:55:28,060 --> 00:55:29,420
Co-pilot's value is tightly coupled

1469
00:55:29,420 --> 00:55:31,740
to how clean your data estate actually is,

1470
00:55:31,740 --> 00:55:32,780
not theoretically.

1471
00:55:32,780 --> 00:55:35,340
Operationally, when you grant co-pilot access

1472
00:55:35,340 --> 00:55:37,340
to your SharePoint teams and email archive,

1473
00:55:37,340 --> 00:55:39,180
you're not just enabling productivity.

1474
00:55:39,180 --> 00:55:40,460
You're exposing every permission,

1475
00:55:40,460 --> 00:55:41,900
misconfiguration, every document

1476
00:55:41,900 --> 00:55:43,340
that shouldn't be broadly accessible,

1477
00:55:43,340 --> 00:55:45,660
every folder structure that nobody's touched in three years

1478
00:55:45,660 --> 00:55:47,820
because nobody remembers why it was created.

1479
00:55:47,820 --> 00:55:50,140
Co-pilot doesn't create data access problems.

1480
00:55:50,140 --> 00:55:53,260
It surfaces existing ones at unprecedented speed and scale.

1481
00:55:53,260 --> 00:55:55,340
A human searching SharePoint manually might dig

1482
00:55:55,340 --> 00:55:57,740
through five or six documents before giving up.

1483
00:55:57,740 --> 00:55:59,740
Co-pilot can ingest 50 documents

1484
00:55:59,740 --> 00:56:02,540
in the time it takes a human to open the search interface.

1485
00:56:02,540 --> 00:56:04,940
If half those documents were never intended to be accessed

1486
00:56:04,940 --> 00:56:06,540
by the person running the query,

1487
00:56:06,540 --> 00:56:08,220
if they contain M&A preparation,

1488
00:56:08,220 --> 00:56:09,500
unreleased product roadmaps,

1489
00:56:09,500 --> 00:56:11,340
or confidential client information,

1490
00:56:11,340 --> 00:56:13,740
Co-pilot surfaces all of it with zero friction.

1491
00:56:14,620 --> 00:56:17,020
The access control failed long before Co-pilot arrived.

1492
00:56:17,020 --> 00:56:18,700
Co-pilot just made the failure visible.

1493
00:56:18,700 --> 00:56:20,860
The pre-deployment requirements are concrete.

1494
00:56:20,860 --> 00:56:24,140
Run a permissions audit across SharePoint, OneDrive, and Teams.

1495
00:56:24,140 --> 00:56:26,460
Find folders shared with anyone with the link.

1496
00:56:26,460 --> 00:56:28,620
Find Teams channels where guest access is active

1497
00:56:28,620 --> 00:56:31,260
for people long since departed from the organization.

1498
00:56:31,260 --> 00:56:33,180
Identify documents marked sensitive

1499
00:56:33,180 --> 00:56:34,620
but shared broadly anyway.

1500
00:56:34,620 --> 00:56:36,060
Remediate the oversharing.

1501
00:56:36,060 --> 00:56:38,940
Implement sensitivity labels that actually cover active content

1502
00:56:38,940 --> 00:56:41,340
and not the aspirational labels you defined two years ago,

1503
00:56:41,340 --> 00:56:43,340
but the labels that reflect how your organization

1504
00:56:43,340 --> 00:56:45,260
actually classifies information today.

1505
00:56:45,260 --> 00:56:46,940
Setup restricted search parameters,

1506
00:56:46,940 --> 00:56:49,420
so Co-pilot doesn't return results from data.

1507
00:56:49,420 --> 00:56:51,900
The requesting user doesn't have legitimate access to.

1508
00:56:51,900 --> 00:56:53,180
This work is not glamorous.

1509
00:56:53,180 --> 00:56:55,180
It doesn't show up in a productivity demo.

1510
00:56:55,180 --> 00:56:57,100
It's also not optional if you want Co-pilot

1511
00:56:57,100 --> 00:56:58,700
to function as a capability,

1512
00:56:58,700 --> 00:57:00,220
rather than a liability.

1513
00:57:00,220 --> 00:57:01,820
Here's the practical consequence.

1514
00:57:01,820 --> 00:57:04,380
Organizations that skip the data governance prerequisite

1515
00:57:04,380 --> 00:57:07,500
and deploy Co-pilot into a dirty data environment experience

1516
00:57:07,500 --> 00:57:09,340
two simultaneous problems.

1517
00:57:09,340 --> 00:57:11,180
First, Co-pilot's quality degrades

1518
00:57:11,180 --> 00:57:13,100
because it's pulling from a messy knowledge base

1519
00:57:13,100 --> 00:57:16,060
full of outdated, redundant, and conflicting information.

1520
00:57:16,060 --> 00:57:19,020
Users ask questions and get answers that are technically accurate

1521
00:57:19,020 --> 00:57:21,420
but strategically obsolete or contradictory

1522
00:57:21,420 --> 00:57:24,780
depending on which document Co-pilot happened to rank higher.

1523
00:57:24,780 --> 00:57:27,420
Second, compliance and security exposure spike

1524
00:57:27,420 --> 00:57:29,900
because Co-pilot is surfacing sensitive information

1525
00:57:29,900 --> 00:57:32,700
that should never have been broadly accessible in the first place.

1526
00:57:32,700 --> 00:57:35,260
The security incident is how most organizations discover

1527
00:57:35,260 --> 00:57:36,300
they skipped this step,

1528
00:57:36,300 --> 00:57:37,900
not through an architecture review,

1529
00:57:37,900 --> 00:57:39,420
through a breach or an audit finding.

1530
00:57:39,420 --> 00:57:41,660
Someone accesses information through Co-pilot

1531
00:57:41,660 --> 00:57:43,580
that they shouldn't have had permission to see

1532
00:57:43,580 --> 00:57:46,300
and suddenly governance becomes urgent in a very different way.

1533
00:57:46,300 --> 00:57:49,820
A cleaner data estate increases ROI by improving answer quality

1534
00:57:49,820 --> 00:57:51,980
while simultaneously reducing compliance risk.

1535
00:57:51,980 --> 00:57:53,500
It's not a trade-off where you improve one

1536
00:57:53,500 --> 00:57:55,580
at the expense of the other, they move together.

1537
00:57:55,580 --> 00:57:57,500
Better classification and access control means

1538
00:57:57,500 --> 00:57:59,900
Co-pilot returns more relevant, more recent,

1539
00:57:59,900 --> 00:58:01,420
more trustworthy information.

1540
00:58:01,420 --> 00:58:03,900
Better access control means Co-pilot's search scope

1541
00:58:03,900 --> 00:58:06,140
is constrained by legitimate permission boundaries

1542
00:58:06,140 --> 00:58:07,980
rather than exploding across everything

1543
00:58:07,980 --> 00:58:09,740
the organization ever documented.

1544
00:58:09,740 --> 00:58:11,900
Data readiness is not an IT prerequisite,

1545
00:58:11,900 --> 00:58:13,100
it's a business prerequisite.

1546
00:58:13,100 --> 00:58:16,060
Your CFO needs to understand that before Co-pilot can work,

1547
00:58:16,060 --> 00:58:17,900
your SharePoint environment needs to function

1548
00:58:17,900 --> 00:58:19,500
like an actual information architecture

1549
00:58:19,500 --> 00:58:22,460
rather than a filing cabinet where everything eventually gets dumped.

1550
00:58:22,460 --> 00:58:23,980
Your legal team needs to confirm

1551
00:58:23,980 --> 00:58:26,220
that the content Co-pilot will have access to

1552
00:58:26,220 --> 00:58:27,580
is classified appropriately.

1553
00:58:27,580 --> 00:58:29,740
Your business leaders need to own the decision

1554
00:58:29,740 --> 00:58:32,220
about what data Co-pilot can surface to whom.

1555
00:58:32,220 --> 00:58:33,980
That's the foundation before anything else

1556
00:58:33,980 --> 00:58:35,900
in the framework actually lands cleanly.

1557
00:58:35,900 --> 00:58:37,740
The hybrid automation architecture,

1558
00:58:37,740 --> 00:58:39,980
the architecture question is no longer binary,

1559
00:58:39,980 --> 00:58:42,380
you're not choosing between deploy Co-pilot everywhere

1560
00:58:42,380 --> 00:58:44,540
and stick with RPA.

1561
00:58:44,540 --> 00:58:47,180
The organization seeing the best outcomes in 2026

1562
00:58:47,180 --> 00:58:49,100
are building something different entirely.

1563
00:58:49,100 --> 00:58:51,020
A layered stack where each component handles

1564
00:58:51,020 --> 00:58:53,740
the work is cheapest and most effective at performing.

1565
00:58:53,740 --> 00:58:55,500
Think of it as four distinct layers,

1566
00:58:55,500 --> 00:58:58,060
each with its own cost profile and responsibility.

1567
00:58:58,060 --> 00:58:59,980
The interpretation layer sits at the top.

1568
00:58:59,980 --> 00:59:01,740
This is where language understanding lives.

1569
00:59:01,740 --> 00:59:04,060
You have incoming email with unstructured text

1570
00:59:04,060 --> 00:59:06,700
or a document in PDF format with inconsistent layouts

1571
00:59:06,700 --> 00:59:09,980
or a customer support chat thread mixing technical details

1572
00:59:09,980 --> 00:59:11,420
with emotional context.

1573
00:59:11,420 --> 00:59:13,740
These inputs don't fit predefined templates.

1574
00:59:13,740 --> 00:59:14,700
They require judgment.

1575
00:59:14,700 --> 00:59:16,860
They require someone or something to read intent

1576
00:59:16,860 --> 00:59:17,740
and extract meaning.

1577
00:59:17,740 --> 00:59:19,340
This is where LLMs operate.

1578
00:59:19,340 --> 00:59:20,380
Classification.

1579
00:59:20,380 --> 00:59:22,780
Is this email a complaint or a question

1580
00:59:22,780 --> 00:59:24,780
or a request for information?

1581
00:59:24,780 --> 00:59:25,660
Extraction.

1582
00:59:25,660 --> 00:59:28,940
What are the key entities and what does the customer actually need?

1583
00:59:28,940 --> 00:59:29,980
Summarization.

1584
00:59:29,980 --> 00:59:32,780
What's the one paragraph version of this 50 page contract?

1585
00:59:32,780 --> 00:59:35,180
The interpretation layer is pure AI work.

1586
00:59:35,180 --> 00:59:37,260
The cost scales with token consumption

1587
00:59:37,260 --> 00:59:38,860
but the value it produces,

1588
00:59:38,860 --> 00:59:42,540
converting messy unstructured input into clean, structured signal

1589
00:59:42,540 --> 00:59:45,260
is what makes the entire downstream workflow possible.

1590
00:59:45,260 --> 00:59:47,420
The execution layer does the mechanical work.

1591
00:59:47,420 --> 00:59:49,420
Once you've extracted meaning from that email,

1592
00:59:49,420 --> 00:59:51,500
someone needs to update the CRM record,

1593
00:59:51,500 --> 00:59:53,420
create a ticket in the helpdesk system,

1594
00:59:53,420 --> 00:59:55,660
move a file from inbox to archive folder,

1595
00:59:55,660 --> 00:59:57,580
post a transaction to the general ledger,

1596
00:59:57,580 --> 00:59:59,580
log an action in the project management system.

1597
00:59:59,580 --> 01:00:00,940
These are deterministic,

1598
01:00:00,940 --> 01:00:02,780
rules-based predictable operations.

1599
01:00:02,780 --> 01:00:03,980
They don't require judgment.

1600
01:00:03,980 --> 01:00:06,220
They require precision and consistency.

1601
01:00:06,220 --> 01:00:08,860
This is where RPA or deterministic scripts operate.

1602
01:00:08,860 --> 01:00:11,180
The cost here approaches zero once deployed

1603
01:00:11,180 --> 01:00:12,780
because you're not reasoning about anything.

1604
01:00:12,780 --> 01:00:14,700
You're just moving data and triggering workflows.

1605
01:00:14,700 --> 01:00:17,500
The marginal cost per execution is fractions of a cent

1606
01:00:17,500 --> 01:00:18,940
or in the case of a Python script,

1607
01:00:18,940 --> 01:00:20,540
fractions of a millionth of a dollar.

1608
01:00:20,540 --> 01:00:23,740
The supervisory layer handles what neither AI nor RPA

1609
01:00:23,740 --> 01:00:24,780
can safely own.

1610
01:00:24,780 --> 01:00:25,740
Exceptions.

1611
01:00:25,740 --> 01:00:26,700
Approvals.

1612
01:00:26,700 --> 01:00:28,140
High stakes decisions.

1613
01:00:28,140 --> 01:00:30,780
The email that came in wasn't a standard request.

1614
01:00:30,780 --> 01:00:33,740
It's a special circumstance requiring human judgment.

1615
01:00:33,740 --> 01:00:36,380
The transaction looks like a duplicate, but might not be.

1616
01:00:36,380 --> 01:00:39,820
The contract contains a clause that deviates from standard language

1617
01:00:39,820 --> 01:00:41,980
and needs a lawyer's opinion before it's committed.

1618
01:00:41,980 --> 01:00:43,260
Humans handle these cases.

1619
01:00:43,260 --> 01:00:45,340
The cost is whatever the fully loaded labor rate is

1620
01:00:45,340 --> 01:00:47,100
for the person doing the review.

1621
01:00:47,100 --> 01:00:49,180
It's expensive compared to AI or RPA,

1622
01:00:49,180 --> 01:00:51,900
but it's necessary because the reputational cost

1623
01:00:51,900 --> 01:00:54,460
and legal liability of getting these cases wrong

1624
01:00:54,460 --> 01:00:57,660
far exceeds the labor cost of handling them thoughtfully.

1625
01:00:57,660 --> 01:00:59,260
The control layer is the infrastructure

1626
01:00:59,260 --> 01:01:01,180
that makes the whole stack defensible.

1627
01:01:01,180 --> 01:01:03,660
Governance, audit trails, compliance logging,

1628
01:01:03,660 --> 01:01:06,060
observability into what's happening at every stage.

1629
01:01:06,060 --> 01:01:07,500
Which AI model ran?

1630
01:01:07,500 --> 01:01:09,500
What were the input tokens and output tokens?

1631
01:01:09,500 --> 01:01:10,940
Which RPA bot executed?

1632
01:01:10,940 --> 01:01:13,500
Did a human approve this transaction before it was posted?

1633
01:01:13,500 --> 01:01:15,900
What happened when the system encountered an exception?

1634
01:01:15,900 --> 01:01:17,420
This layer doesn't execute work.

1635
01:01:17,420 --> 01:01:19,180
It certifies that work happened correctly.

1636
01:01:19,180 --> 01:01:20,300
The decisions are traceable.

1637
01:01:20,300 --> 01:01:22,060
That mistakes can be identified and corrected.

1638
01:01:22,060 --> 01:01:24,540
It's not cheap, but it's how you prove the regulators,

1639
01:01:24,540 --> 01:01:26,460
auditors and your own executive team

1640
01:01:26,460 --> 01:01:28,300
that automation is happening under control.

1641
01:01:28,300 --> 01:01:30,700
The practical mechanics are worth walking through.

1642
01:01:30,700 --> 01:01:31,980
An email arrives.

1643
01:01:31,980 --> 01:01:33,420
The interpretation layer.

1644
01:01:33,420 --> 01:01:37,580
An LLM running a classifier task reads it, determines intent

1645
01:01:37,580 --> 01:01:39,420
and extracts structured data.

1646
01:01:39,420 --> 01:01:41,020
Cost fractions of a cent.

1647
01:01:41,020 --> 01:01:43,340
Output, a decision about what to do next

1648
01:01:43,340 --> 01:01:45,420
and the key information needed to do it.

1649
01:01:45,420 --> 01:01:48,540
The execution layer, RPA, or a deterministic script

1650
01:01:48,540 --> 01:01:50,860
takes that structured data and executes the workflow.

1651
01:01:50,860 --> 01:01:53,180
It updates the CRM, logs the action,

1652
01:01:53,180 --> 01:01:55,660
roots the case to the appropriate queue or person,

1653
01:01:55,660 --> 01:01:57,820
cost near zero marginal expense.

1654
01:01:57,820 --> 01:02:00,140
The supervisory layer reviews exceptions.

1655
01:02:00,140 --> 01:02:02,780
The 5 or 10% of cases where the interpretation

1656
01:02:02,780 --> 01:02:05,340
wasn't confident or the situation requires judgment.

1657
01:02:05,340 --> 01:02:07,180
The human person makes the final call.

1658
01:02:07,180 --> 01:02:09,820
The control layer logs everything, creates an audit trail,

1659
01:02:09,820 --> 01:02:10,940
flags anomalies.

1660
01:02:10,940 --> 01:02:13,420
The cost advantage of this hybrid model is compound.

1661
01:02:13,420 --> 01:02:16,700
You pay AI prices, which are variable and token dependent,

1662
01:02:16,700 --> 01:02:18,460
only for the interpretation work,

1663
01:02:18,460 --> 01:02:20,700
where human labor would be even more expensive.

1664
01:02:20,700 --> 01:02:23,740
You pay near zero execution prices for the deterministic work,

1665
01:02:23,740 --> 01:02:26,780
where RPA or scripts are vastly cheaper than asking an AI

1666
01:02:26,780 --> 01:02:28,140
to do mechanical tasks.

1667
01:02:28,140 --> 01:02:30,540
You pay human labor rates, which are expensive,

1668
01:02:30,540 --> 01:02:34,060
but only for the exceptions and decisions that actually require human judgment.

1669
01:02:34,060 --> 01:02:37,100
You're not forcing expensive reasoning engines to do mechanical work

1670
01:02:37,100 --> 01:02:39,580
and you're not using cheap deterministic scripts for tasks

1671
01:02:39,580 --> 01:02:41,740
that require contextual understanding.

1672
01:02:41,740 --> 01:02:43,660
The result is an automation stack,

1673
01:02:43,660 --> 01:02:46,940
where cost aligns to value and risk aligns to control,

1674
01:02:46,940 --> 01:02:48,460
not hype, not hope.

1675
01:02:48,460 --> 01:02:49,980
Structural economics.

1676
01:02:49,980 --> 01:02:51,900
Building your co-pilot business case.

1677
01:02:51,900 --> 01:02:54,380
A defensible business case in 2026

1678
01:02:54,380 --> 01:02:59,020
looks materially different from the AI business cases of 2024 and 2025.

1679
01:02:59,020 --> 01:03:00,460
The difference isn't academic.

1680
01:03:00,460 --> 01:03:02,220
It's the difference between getting budget approved

1681
01:03:02,220 --> 01:03:04,380
and getting budget revoked after the first renewal.

1682
01:03:04,380 --> 01:03:07,980
The five components that make a case defensible are baseline metrics,

1683
01:03:07,980 --> 01:03:09,980
conservative time savings assumptions,

1684
01:03:09,980 --> 01:03:13,580
pilot design, measurement methodology, and scale criteria.

1685
01:03:13,580 --> 01:03:17,020
You need all five, and they need to work together as an integrated model,

1686
01:03:17,020 --> 01:03:19,420
not as separate justifications that you stack together

1687
01:03:19,420 --> 01:03:21,820
and hope nobody asks hard questions about.

1688
01:03:21,820 --> 01:03:24,540
Baseline metrics means you've captured how long specific tasks

1689
01:03:24,540 --> 01:03:26,380
take before co-pilot arrives,

1690
01:03:26,380 --> 01:03:28,220
not estimates, actual data.

1691
01:03:28,220 --> 01:03:30,380
If proposal drafting is your target use case,

1692
01:03:30,380 --> 01:03:32,620
you've timed how long a senior consultant spends

1693
01:03:32,620 --> 01:03:35,180
on a typical RFP from intake to delivery.

1694
01:03:35,180 --> 01:03:37,260
If meeting summarization is the focus,

1695
01:03:37,260 --> 01:03:40,620
you've measured how long it takes someone to review a recorded meeting

1696
01:03:40,620 --> 01:03:42,220
and produce actionable notes.

1697
01:03:42,220 --> 01:03:43,660
If email triage is the driver,

1698
01:03:43,660 --> 01:03:47,100
you've documented cycle time from message arrival to initial response.

1699
01:03:47,100 --> 01:03:48,780
This baseline is your control group.

1700
01:03:48,780 --> 01:03:50,620
Without it, you can't measure the delta.

1701
01:03:50,620 --> 01:03:52,780
Without the delta, you can't calculate ROI.

1702
01:03:52,780 --> 01:03:55,900
The CFO will accept nothing less than documented time data.

1703
01:03:55,900 --> 01:03:59,420
Conservative modeling is where you protect yourself against the optimism trap.

1704
01:03:59,420 --> 01:04:03,580
The Forrestor study reported three to four hours per week saved at maturity.

1705
01:04:03,580 --> 01:04:04,700
That's the upper bound.

1706
01:04:04,700 --> 01:04:07,580
Your business case should assume 30 to 60 minutes per week

1707
01:04:07,580 --> 01:04:09,260
saved as your primary model.

1708
01:04:09,260 --> 01:04:12,860
Run sensitivity analysis up to three to five hours per week for mature users,

1709
01:04:12,860 --> 01:04:14,780
but lead with conservative assumptions.

1710
01:04:14,780 --> 01:04:17,900
If your case only works if you hit the optimistic scenario,

1711
01:04:17,900 --> 01:04:21,100
you've built a case that will fail the moment adoption under performs

1712
01:04:21,100 --> 01:04:24,140
or task complexity turns out to be higher than expected.

1713
01:04:24,140 --> 01:04:28,060
Build the case so it works at conservative assumptions and outperforms from there.

1714
01:04:28,060 --> 01:04:30,300
Pilot design is the operational linchpin.

1715
01:04:30,300 --> 01:04:34,140
You're not deploying co-pilot to the entire organization and measuring what happens.

1716
01:04:34,140 --> 01:04:37,100
You're running a controlled experiment with 100 to 500 users

1717
01:04:37,100 --> 01:04:39,500
in specific roles for eight to 12 weeks.

1718
01:04:39,500 --> 01:04:40,940
The roles are chosen explicitly,

1719
01:04:40,940 --> 01:04:43,660
sales, legal consulting, product management, customer service,

1720
01:04:43,660 --> 01:04:46,140
because they're document heavy and communication heavy

1721
01:04:46,140 --> 01:04:49,340
and you've already hypothesized that co-pilot fits their workflow.

1722
01:04:49,340 --> 01:04:51,740
The users in those roles get dedicated training,

1723
01:04:51,740 --> 01:04:53,580
not a generic email announcement.

1724
01:04:53,580 --> 01:04:55,500
You create role-specific prompting libraries,

1725
01:04:55,500 --> 01:04:57,100
so people aren't starting from scratch.

1726
01:04:57,100 --> 01:04:59,740
You establish clear success metrics per process.

1727
01:04:59,740 --> 01:05:01,340
Not did people like it?

1728
01:05:01,340 --> 01:05:04,860
Metrics like did proposal cycle time drop 20%

1729
01:05:04,860 --> 01:05:08,540
or did email response time improve 15 minutes per interaction?

1730
01:05:08,540 --> 01:05:12,540
Or did document revision cycles shrink from three rounds to two?

1731
01:05:12,540 --> 01:05:15,420
Measurement methodology is the operational discipline

1732
01:05:15,420 --> 01:05:17,500
that captures those metrics rigorously.

1733
01:05:17,500 --> 01:05:20,700
You're tracking cycle time, throughput, error rates, and adoption depth.

1734
01:05:20,700 --> 01:05:23,580
Adoption depth means daily active usage in core workflows,

1735
01:05:23,580 --> 01:05:25,020
not license assignment.

1736
01:05:25,020 --> 01:05:27,340
You're capturing this data before the pilot starts,

1737
01:05:27,340 --> 01:05:29,020
continuously during the pilot,

1738
01:05:29,020 --> 01:05:30,380
and at a defined endpoint,

1739
01:05:30,380 --> 01:05:33,180
so you can quantify the before and after gap.

1740
01:05:33,180 --> 01:05:36,060
You're not conducting surveys asking people how productive they feel.

1741
01:05:36,060 --> 01:05:38,860
The CFO will dismiss productivity feelings as anecdotes.

1742
01:05:38,860 --> 01:05:41,820
You're showing measurable change in how work actually flows.

1743
01:05:41,820 --> 01:05:44,540
Scale criteria is where you define the graduation gate.

1744
01:05:44,540 --> 01:05:47,500
You've run your pilot and gathered 12 weeks of data.

1745
01:05:47,500 --> 01:05:52,060
You've calculated ROI based on measured time savings multiplied by fully loaded labor cost

1746
01:05:52,060 --> 01:05:55,180
and compared to license costs plus enablement overhead.

1747
01:05:55,180 --> 01:05:59,260
You've set a threshold, maybe 150% ROI minimum,

1748
01:05:59,260 --> 01:06:01,100
maybe six month payback requirement,

1749
01:06:01,100 --> 01:06:04,060
maybe 60% adoption depth in the pilot cohort.

1750
01:06:04,060 --> 01:06:05,180
Whatever your threshold is,

1751
01:06:05,180 --> 01:06:08,220
it needs to be explicit and documented before the pilot starts.

1752
01:06:08,220 --> 01:06:10,700
Once the pilot ends, you measure against that threshold.

1753
01:06:10,700 --> 01:06:12,220
If you hit it, you expand.

1754
01:06:12,220 --> 01:06:16,140
If you don't, you diagnose why and iterate before committing to a wider rollout.

1755
01:06:16,140 --> 01:06:20,780
The critical insight that most organizations miss is that ROI tracks adoption,

1756
01:06:20,780 --> 01:06:21,740
not licenses.

1757
01:06:21,740 --> 01:06:23,500
You can buy all the seats you want.

1758
01:06:23,500 --> 01:06:25,420
ROI doesn't follow adoption follows,

1759
01:06:25,420 --> 01:06:27,420
change management and training are not overhead.

1760
01:06:27,420 --> 01:06:31,180
They are the line items that determine whether the license actually generates return.

1761
01:06:31,180 --> 01:06:33,260
Include them in your cost model from the beginning,

1762
01:06:33,260 --> 01:06:34,540
not as an afterthought.

1763
01:06:34,540 --> 01:06:36,860
A business case built this way is defensible with the CFO

1764
01:06:36,860 --> 01:06:40,460
because it's built on the same logic finance uses to evaluate any investment.

1765
01:06:40,460 --> 01:06:43,260
Documented baseline, conservative assumptions,

1766
01:06:43,260 --> 01:06:44,460
measurable outcomes,

1767
01:06:44,460 --> 01:06:45,740
gated expansion.

1768
01:06:45,740 --> 01:06:47,420
What the next 12 months require.

1769
01:06:47,420 --> 01:06:51,020
The structural shifts happening in 2026 aren't temporary, they're directional.

1770
01:06:51,020 --> 01:06:55,740
Microsoft's consumption first strategy is accelerating across the entire stack.

1771
01:06:55,740 --> 01:07:00,220
The hybrid model, mixing per-user licenses with meter capacity consumption,

1772
01:07:00,220 --> 01:07:03,020
is hardening into the standard rather than the exception.

1773
01:07:03,020 --> 01:07:06,860
Security co-pilot, which launched bundled into M365E5

1774
01:07:06,860 --> 01:07:09,180
with monthly security compute unit allocations,

1775
01:07:09,180 --> 01:07:13,980
is essentially a technical preview of how all co-pilot consumption will eventually flow.

1776
01:07:13,980 --> 01:07:17,260
You're looking at a future where co-pilot licensing works like cloud infrastructure

1777
01:07:17,260 --> 01:07:18,700
did five years ago.

1778
01:07:18,700 --> 01:07:22,060
Predictable base costs plus variable consumption on top

1779
01:07:22,060 --> 01:07:24,060
with no unified dashboard to track it all.

1780
01:07:24,060 --> 01:07:27,660
Organizations that build token governance infrastructure now,

1781
01:07:27,660 --> 01:07:31,340
the observability systems, the budget controls, the cost per outcome metrics,

1782
01:07:31,340 --> 01:07:34,220
will have a structural cost advantage as consumption billing expands.

1783
01:07:34,220 --> 01:07:35,580
You're not betting on a future model,

1784
01:07:35,580 --> 01:07:37,580
you're preparing for the one that's already here.

1785
01:07:37,580 --> 01:07:39,820
The one still operating on hope and assumption

1786
01:07:39,820 --> 01:07:42,220
will discover that they need governance infrastructure,

1787
01:07:42,220 --> 01:07:45,100
the moment an unexpected invoice arrives and nobody can explain it.

1788
01:07:45,100 --> 01:07:48,460
Three immediate actions compress the urgency.

1789
01:07:48,460 --> 01:07:51,820
First, conduct a full license and usage audit across M365

1790
01:07:51,820 --> 01:07:53,900
and any GitHub co-pilot deployments.

1791
01:07:53,900 --> 01:07:56,540
Not aspirational metrics about what you think people use.

1792
01:07:56,540 --> 01:08:00,060
Actual usage data from the admin center and usage reporting APIs.

1793
01:08:00,060 --> 01:08:02,460
Identify which licenses are generating measurable value

1794
01:08:02,460 --> 01:08:03,820
and which are seats in name only.

1795
01:08:03,820 --> 01:08:06,860
This isn't theoretical, this is capital redeployment preparation.

1796
01:08:06,860 --> 01:08:10,220
Second, define roll-based license templates.

1797
01:08:10,780 --> 01:08:16,220
Explicit mappings that say sales roll gets tier two with x capabilities.

1798
01:08:16,220 --> 01:08:19,340
Frontline staff gets tier three with y constraints.

1799
01:08:19,340 --> 01:08:22,380
Engineering leadership gets tier one with z governance.

1800
01:08:22,380 --> 01:08:23,900
These templates aren't permanent.

1801
01:08:23,900 --> 01:08:25,980
They are the framework that governs initial provisioning

1802
01:08:25,980 --> 01:08:29,260
and guides upgrade decisions as actual usage data accumulates.

1803
01:08:29,260 --> 01:08:33,740
Third, integrate co-pilot usage telemetry into your quarterly business reviews.

1804
01:08:33,740 --> 01:08:37,340
Make token consumption and cost per outcome visible to the business owners

1805
01:08:37,340 --> 01:08:38,620
who control the budget.

1806
01:08:38,620 --> 01:08:39,980
Finance sees the line item.

1807
01:08:39,980 --> 01:08:41,660
Engineering sees the adoption curve.

1808
01:08:41,660 --> 01:08:44,220
Sales sees time impact on deal cycles.

1809
01:08:44,220 --> 01:08:46,060
Visibility precedes discipline.

1810
01:08:46,060 --> 01:08:50,780
The 12-month horizon is the bridge between where you are now and where you need to be.

1811
01:08:50,780 --> 01:08:53,020
12 months from today, the conversation shouldn't be.

1812
01:08:53,020 --> 01:08:54,540
We have co-pilot licenses.

1813
01:08:54,540 --> 01:08:58,300
But we have a governed AI portfolio with measurable ROI-per-tier,

1814
01:08:58,300 --> 01:09:00,300
recovering 30% of idle spend,

1815
01:09:00,300 --> 01:09:03,180
and reinvesting that capital into high-impact agent development.

1816
01:09:03,180 --> 01:09:04,860
That's not aspirational.

1817
01:09:04,860 --> 01:09:08,300
That's operational reality for the organization's moving at scale right now.

1818
01:09:08,300 --> 01:09:11,740
The competitive gap is widening, but not in the direction most people assume.

1819
01:09:11,740 --> 01:09:14,860
It's not widening between organizations that have AI and those that don't.

1820
01:09:14,860 --> 01:09:16,140
Everyone has AI now.

1821
01:09:16,140 --> 01:09:18,780
The gap is widening between organizations that govern it,

1822
01:09:18,780 --> 01:09:21,420
with explicit tiers, clear metrics, token budgets,

1823
01:09:21,420 --> 01:09:22,860
and cost per outcome discipline,

1824
01:09:22,860 --> 01:09:24,540
and organizations that don't.

1825
01:09:24,540 --> 01:09:27,900
The governed ones see three to four times the return on the same license spend.

1826
01:09:27,900 --> 01:09:30,540
They're recapturing 30% of idle investment.

1827
01:09:30,540 --> 01:09:34,300
They are deploying agents confidently because they understand the cost profile

1828
01:09:34,300 --> 01:09:36,940
and have the controls in place to prevent surprises.

1829
01:09:36,940 --> 01:09:40,060
The ungoverned ones are treating co-pilot like an infinite utility,

1830
01:09:40,060 --> 01:09:41,900
wondering why their invoices are climbing,

1831
01:09:41,900 --> 01:09:44,700
and unable to articulate what they're getting in return.

1832
01:09:44,700 --> 01:09:46,220
The co-pilot tax is not inevitable.

1833
01:09:46,220 --> 01:09:47,180
It's a choice.

1834
01:09:47,180 --> 01:09:49,500
Organizations that deploy without a framework

1835
01:09:49,500 --> 01:09:52,220
that assume broad access will naturally drive adoption,

1836
01:09:52,220 --> 01:09:54,540
that don't measure baselines before deployment,

1837
01:09:54,540 --> 01:09:57,100
that don't tier access by ROI or risk,

1838
01:09:57,100 --> 01:09:59,180
that lack token governance infrastructure,

1839
01:09:59,180 --> 01:10:01,500
those organizations are choosing to bleed cash.

1840
01:10:01,500 --> 01:10:04,860
The organizations that deploy with the framework we've outlined in this episode

1841
01:10:04,860 --> 01:10:06,060
are choosing differently.

1842
01:10:06,060 --> 01:10:08,220
They're extracting return, they're managing risk,

1843
01:10:08,220 --> 01:10:11,820
they're making capital allocation decisions based on evidence rather than hope.

1844
01:10:11,820 --> 01:10:15,020
The work starts now, not after you've built the perfect solution.

1845
01:10:15,020 --> 01:10:18,940
Now, the best time to establish governance was before you deployed co-pilot.

1846
01:10:18,940 --> 01:10:20,780
The second best time is today.

1847
01:10:20,780 --> 01:10:23,980
The co-pilot tax isn't just a line item on your Microsoft invoice,

1848
01:10:23,980 --> 01:10:27,660
it's the gap between the potential you bought and the value you're actually capturing.

1849
01:10:27,660 --> 01:10:31,500
It's the 48% verification burden, it's the 60% idle token waste.

1850
01:10:31,500 --> 01:10:35,900
It's the zombie seed sitting in roles that don't need reasoning engines to do mechanical work,

1851
01:10:35,900 --> 01:10:38,780
but as we've seen, this isn't an inevitable cost of doing business,

1852
01:10:38,780 --> 01:10:39,820
it's a governance choice.

1853
01:10:39,820 --> 01:10:42,380
If you built the tiers, ordered the baselines,

1854
01:10:42,380 --> 01:10:44,620
and treat tokens like a financial discipline,

1855
01:10:44,620 --> 01:10:47,180
you don't just stop the bleeding, you fund the future.

1856
01:10:47,180 --> 01:10:50,220
If this episode changed how you think about your AI strategy,

1857
01:10:50,220 --> 01:10:53,180
or if it gave you the data you need for your next budget review,

1858
01:10:53,180 --> 01:10:55,260
follow me, Mirko Peters, on LinkedIn.

1859
01:10:55,260 --> 01:10:57,740
I'm always looking for the next structural flaw to deconstruct,

1860
01:10:57,740 --> 01:11:00,780
and your feedback helps me find the topics that matter most right now,

1861
01:11:00,780 --> 01:11:04,540
and if you want more deep dives into the reality of the M365 ecosystem,

1862
01:11:04,540 --> 01:11:06,540
subscribe to the M365FM podcast.

1863
01:11:06,540 --> 01:11:07,900
Leave a review while you're there.

1864
01:11:07,900 --> 01:11:11,500
It helps more leaders find this framework before their next renewal.

1865
01:11:11,500 --> 01:11:16,300
Share this with your team, especially if you're looking at your 2026 budget right now.