1
00:00:00,000 --> 00:00:03,251
Every launch planning cycle
ends with the same question.

2
00:00:03,680 --> 00:00:04,680
Are we ready.

3
00:00:05,317 --> 00:00:11,134
It sounds like one question, but it is
really eleven answers from eleven places,

4
00:00:11,645 --> 00:00:16,382
each one coming from a different owner,
and all of them reporting on their

5
00:00:16,428 --> 00:00:17,428
own schedule.

6
00:00:17,891 --> 00:00:22,779
What is still broken, what has not
been tested, who has not been trained,

7
00:00:23,464 --> 00:00:26,730
whether the data moved,
whether the vendor confirmed.

8
00:00:27,330 --> 00:00:30,070
Collecting all of that is the easy half.

9
00:00:30,755 --> 00:00:34,726
The hard half is knowing what to
look into, because you are the only

10
00:00:34,784 --> 00:00:37,396
person who sees how the pieces connect.

11
00:00:37,988 --> 00:00:41,100
Three teams can each report
yellow on their own piece,

12
00:00:41,425 --> 00:00:45,465
and on its own each one is survivable,
but stacked on the same weekend

13
00:00:45,500 --> 00:00:49,111
they add up to a red, and none
of those three can see that

14
00:00:49,145 --> 00:00:50,145
from where they sit.

15
00:00:50,933 --> 00:00:55,844
Or the data team is yellow on
reconciliation and the vendor is yellow on

16
00:00:55,891 --> 00:00:59,235
the file format, and those
two are the same problem,

17
00:00:59,664 --> 00:01:02,504
which makes the interface behind them red.

18
00:01:03,104 --> 00:01:06,285
And the colors only ever
lie in one direction.

19
00:01:07,121 --> 00:01:10,058
A red means a conversation
with your sponsor

20
00:01:10,894 --> 00:01:13,181
and somebody's name on a delay.

21
00:01:13,797 --> 00:01:18,278
A yellow costs nothing, and nobody
wants to be the person who stopped the

22
00:01:18,348 --> 00:01:19,348
launch.

23
00:01:19,915 --> 00:01:23,363
Status reporting has been studied for
years, and the finding is always the

24
00:01:23,421 --> 00:01:24,421
same.

25
00:01:24,664 --> 00:01:27,880
Optimistic bias is far more
common than pessimistic,

26
00:01:28,565 --> 00:01:32,315
and the color softens by
a grade on its way up.

27
00:01:33,255 --> 00:01:35,495
So some of those yellows are reds.

28
00:01:36,006 --> 00:01:39,977
And part of your job is investigating
until you know which ones,

29
00:01:40,348 --> 00:01:42,927
and then being the one
who says it out loud.

30
00:01:43,526 --> 00:01:48,367
Which is why the launches that went wrong
were almost never cases where nobody

31
00:01:48,414 --> 00:01:49,414
knew.

32
00:01:49,749 --> 00:01:50,749
Somebody knew.

33
00:01:51,456 --> 00:01:55,612
It was in a test report, or
a risk log, or two people in

34
00:01:55,658 --> 00:01:58,433
a meeting who did not want
to be the one to say it.

35
00:01:59,374 --> 00:02:03,072
The warning existed and it
never reached the decision.

36
00:02:03,672 --> 00:02:06,586
So be clear about what
you are building here.

37
00:02:07,271 --> 00:02:11,962
It does not make the decision, and it
does not replace your judgment about

38
00:02:11,996 --> 00:02:13,041
what to look into.

39
00:02:13,842 --> 00:02:17,871
It gathers the pieces, it
tells you how old each one is,

40
00:02:18,672 --> 00:02:21,238
it points you at what to open first,

41
00:02:22,039 --> 00:02:24,456
and it keeps a record of what it read.

42
00:02:25,056 --> 00:02:27,599
Most of us already do the
first version of this.

43
00:02:28,376 --> 00:02:33,264
The night before, you open a
chatbot, you paste in your notes,

44
00:02:33,845 --> 00:02:36,329
and you ask it what your risks are.

45
00:02:37,246 --> 00:02:41,391
And it gives you a real answer, better
than what you would have written

46
00:02:41,670 --> 00:02:44,224
at eleven at night on your own.

47
00:02:44,909 --> 00:02:48,659
It sorts the problems, it
finds the two that matter,

48
00:02:48,938 --> 00:02:51,294
it writes the paragraph you were dreading.

49
00:02:52,096 --> 00:02:55,857
But the whole answer rests on
the context you put in by hand.

50
00:02:56,449 --> 00:02:59,932
Anything you left out is
missing from the answer too,

51
00:03:00,304 --> 00:03:02,150
and it will never tell you that.

52
00:03:02,750 --> 00:03:05,269
So stop retyping that prompt.

53
00:03:06,210 --> 00:03:10,169
You are going to write some version
of it before every launch for the

54
00:03:10,250 --> 00:03:11,365
rest of your career.

55
00:03:11,875 --> 00:03:15,498
Save it once, with the questions
you always mean to ask,

56
00:03:16,009 --> 00:03:19,898
including the ones you skip
when the week has been long.

57
00:03:20,583 --> 00:03:24,739
And now your judgment rests on
the same evidence every time.

58
00:03:25,541 --> 00:03:30,544
You are asking the same questions of the
same sources against the same standard,

59
00:03:31,346 --> 00:03:35,881
instead of whatever happened to
make it into your notes that night.

60
00:03:36,481 --> 00:03:38,048
Then you give it your standard.

61
00:03:38,849 --> 00:03:40,486
Not a general idea of ready,

62
00:03:41,288 --> 00:03:44,132
the one your organization actually uses.

63
00:03:44,933 --> 00:03:48,683
No critical defects open,
training above ninety percent,

64
00:03:49,484 --> 00:03:52,839
the data reconciled and signed
off by the business owner

65
00:03:53,640 --> 00:03:58,458
rather than by you, and a rollback plan
somebody has actually rehearsed rather

66
00:03:58,505 --> 00:04:00,989
than one that only exists as a document.

67
00:04:01,582 --> 00:04:05,622
Those are your thresholds, and they
get set when you plan the launch,

68
00:04:05,993 --> 00:04:06,993
not at the end.

69
00:04:07,410 --> 00:04:09,697
Set them early and they are a standard.

70
00:04:10,289 --> 00:04:14,178
Set them in the last week and they
are whatever the schedule needs them

71
00:04:14,248 --> 00:04:15,248
to be.

72
00:04:15,455 --> 00:04:19,972
Be careful with pass rates while you
are at it, because they hide things.

73
00:04:20,564 --> 00:04:26,102
Ninety-five percent of tests passed
is not a pass if five of the failures

74
00:04:26,160 --> 00:04:27,160
are critical.

75
00:04:27,533 --> 00:04:30,087
Some of these you can write in a minute.

76
00:04:30,679 --> 00:04:34,522
Others have been in your head for
years and you have never had to

77
00:04:34,615 --> 00:04:36,008
explain them to anybody.

78
00:04:36,693 --> 00:04:40,931
You just know, and you will sit
there a while looking for the words.

79
00:04:41,442 --> 00:04:46,632
Now, a written rule can be changed,
and sooner or later somebody will want

80
00:04:46,667 --> 00:04:47,667
to change one.

81
00:04:47,874 --> 00:04:51,670
Usually late, and usually
because the date is the date.

82
00:04:51,949 --> 00:04:53,006
That is allowed.

83
00:04:53,516 --> 00:04:57,800
What you want is for it to be
a decision rather than a drift.

84
00:04:58,485 --> 00:05:01,411
So the rules have an owner,
you keep the old version,

85
00:05:01,783 --> 00:05:04,894
and when the bar moves you
know who moved it and when,

86
00:05:05,324 --> 00:05:07,738
before the meeting rather than during it.

87
00:05:08,331 --> 00:05:09,921
And write down what yellow means.

88
00:05:10,606 --> 00:05:13,416
In a formal gate, yellow is not a feeling.

89
00:05:14,008 --> 00:05:17,967
It means the rule has not been met,
and there is a mitigation plan,

90
00:05:18,478 --> 00:05:21,113
and that plan has an
owner, and it has a date.

91
00:05:21,705 --> 00:05:24,828
If any of those three is
missing, it was never yellow.

92
00:05:25,258 --> 00:05:26,258
It is red.

93
00:05:26,837 --> 00:05:30,877
Put that in the rules and every
yellow that arrives without a plan,

94
00:05:31,307 --> 00:05:35,138
an owner and a date comes
back to you printed in red,

95
00:05:35,509 --> 00:05:37,672
and you did not have to
be the one to argue it.

96
00:05:38,272 --> 00:05:41,233
You are still carrying all
of this in by hand, though.

97
00:05:41,918 --> 00:05:46,493
So the next step is that you
stop pasting, and you point it at

98
00:05:46,538 --> 00:05:49,928
the sources and let it read them itself.

99
00:05:50,522 --> 00:05:52,204
It goes into the defect list

100
00:05:53,145 --> 00:05:58,787
and the test results, the training
report, the cutover checklist and the risk

101
00:05:58,857 --> 00:06:01,772
log, and it reads what
the teams actually wrote,

102
00:06:02,142 --> 00:06:05,358
before anybody rounds it
off for the steering deck.

103
00:06:05,869 --> 00:06:09,619
And then ask it for the right
thing, which is not a verdict.

104
00:06:10,130 --> 00:06:13,799
Ask it for three places to start this
morning, ordered by how much they

105
00:06:13,834 --> 00:06:17,665
would change the answer, with the
reason each one is on the list sitting

106
00:06:17,712 --> 00:06:18,712
right next to it.

107
00:06:19,534 --> 00:06:23,111
Something earns a place there because
a green has nothing behind it,

108
00:06:23,702 --> 00:06:28,695
or a yellow turned up with no
plan, or nobody has opened a source

109
00:06:28,822 --> 00:06:34,430
since August, or the same risk is sitting
in two different reports under two

110
00:06:34,488 --> 00:06:35,488
different names.

111
00:06:35,777 --> 00:06:37,727
You still decide what to look into.

112
00:06:38,157 --> 00:06:42,174
The list is only there so it
includes the ones you would not have

113
00:06:42,208 --> 00:06:43,208
thought of.

114
00:06:43,366 --> 00:06:47,615
Then you put it on a schedule, so
that through launch week it runs

115
00:06:47,708 --> 00:06:52,433
every morning at seven and the page
is waiting for you before the first

116
00:06:52,503 --> 00:06:53,503
meeting.

117
00:06:53,722 --> 00:06:57,356
That is the point where it
stops being something you open.

118
00:06:58,157 --> 00:07:03,057
In Copilot the simplest version
of that is a scheduled prompt.

119
00:07:03,568 --> 00:07:07,747
You get ten of them, and the
answer is sitting in your chat list

120
00:07:08,119 --> 00:07:09,119
when you sit down.

121
00:07:09,616 --> 00:07:13,471
If you want a real event to
set it off instead of a clock,

122
00:07:13,900 --> 00:07:15,711
that is Copilot Studio.

123
00:07:16,222 --> 00:07:19,067
It can start when somebody
closes a Planner task,

124
00:07:19,496 --> 00:07:24,268
or when one of your tables changes,
or when a document lands in the

125
00:07:24,314 --> 00:07:27,066
SharePoint folder your team actually uses.

126
00:07:27,658 --> 00:07:31,236
And every one of those triggers
is billed as a message.

127
00:07:31,836 --> 00:07:36,828
Everything after that is design, and
it is the same design you would do

128
00:07:36,910 --> 00:07:39,812
for any process you
own, in four questions.

129
00:07:40,393 --> 00:07:42,807
First, what is it allowed to touch.

130
00:07:43,609 --> 00:07:48,218
Start with one that only reads,
because reading can be undone.

131
00:07:49,019 --> 00:07:53,767
Nothing gets sent, nobody gets told,
and if it is wrong you close the

132
00:07:53,814 --> 00:07:54,726
tab.

133
00:07:54,766 --> 00:07:59,131
Then keep reading and writing apart,
so the thing that gathers your evidence

134
00:07:59,503 --> 00:08:02,416
is not also the thing that
can move a date in the plan.

135
00:08:03,218 --> 00:08:04,413
And it never gets delete.

136
00:08:05,006 --> 00:08:08,779
Name the sources yourself,
all eleven, one at a time.

137
00:08:09,371 --> 00:08:13,469
A workflow left to wander will find
things you did not ask about and

138
00:08:13,562 --> 00:08:18,042
hand them to you with exactly the
same confidence as the things you did.

139
00:08:18,642 --> 00:08:22,938
If you work in Microsoft three sixty
five, that is what the agent builder

140
00:08:23,031 --> 00:08:24,493
inside Copilot is for.

141
00:08:25,086 --> 00:08:28,580
You give it instructions, and
you give it knowledge sources,

142
00:08:28,731 --> 00:08:33,607
which is their phrase for the particular
SharePoint sites and files you point

143
00:08:33,642 --> 00:08:39,238
it at, so it answers from those
instead of from whatever it can find.

144
00:08:40,179 --> 00:08:42,651
When it needs something that
lives outside Microsoft,

145
00:08:43,336 --> 00:08:47,621
that is either a Copilot connector,
which your administrator has to set up,

146
00:08:48,050 --> 00:08:49,710
or it is an M C P server.

147
00:08:49,943 --> 00:08:53,426
And M C P is worth a minute,
because you are going to keep

148
00:08:53,484 --> 00:08:54,484
hearing it.

149
00:08:54,726 --> 00:08:59,521
It is an open standard for letting
an AI tool reach another system.

150
00:09:00,032 --> 00:09:03,457
It was published a couple of years
ago and it sits with the Linux

151
00:09:03,515 --> 00:09:08,031
Foundation now, and the line
people use is that it is a U S

152
00:09:08,089 --> 00:09:10,376
B C port for AI applications.

153
00:09:10,968 --> 00:09:16,344
One kind of plug, many systems, instead
of somebody building custom glue for

154
00:09:16,402 --> 00:09:17,402
every one.

155
00:09:17,540 --> 00:09:21,359
The pieces are called servers
and tools, and Jira has one,

156
00:09:21,638 --> 00:09:25,037
ServiceNow has one, and your
own warehouse can have one.

157
00:09:25,637 --> 00:09:29,027
What matters for you is
simpler than the acronym.

158
00:09:29,805 --> 00:09:34,960
An M C P server is a door into
a system, so the questions

159
00:09:35,320 --> 00:09:37,758
are the ones you would ask about any door.

160
00:09:38,536 --> 00:09:41,323
You want to know who owns
it, and what is behind it,

161
00:09:41,740 --> 00:09:44,736
and whether it can only read or
whether it can write as well.

162
00:09:45,421 --> 00:09:49,577
And Microsoft made that
concrete in a way worth copying.

163
00:09:50,378 --> 00:09:54,744
In Copilot Studio an M C P
server shows up as a connector,

164
00:09:55,254 --> 00:09:59,643
so it falls under the same data policy
your administrator already applies to

165
00:09:59,701 --> 00:10:00,701
connectors.

166
00:10:00,792 --> 00:10:04,577
Block the connector, and the tools
behind it are blocked with it.

167
00:10:05,680 --> 00:10:08,815
There is a published list
of the ways this goes wrong,

168
00:10:09,186 --> 00:10:11,543
and two of them are worth knowing by name.

169
00:10:12,135 --> 00:10:17,626
The first is tool poisoning, where a
server describes its own tools in a

170
00:10:17,696 --> 00:10:22,491
way that steers the model into
doing something you never asked for.

171
00:10:23,176 --> 00:10:28,540
The second is the shadow server, the
useful one somebody wired up that nobody

172
00:10:28,609 --> 00:10:30,351
owns and nobody logs.

173
00:10:30,932 --> 00:10:34,833
So if your workflow leans on
an M C P server, it needs an

174
00:10:34,914 --> 00:10:38,635
owner and a log, like
everything else in your program.

175
00:10:39,235 --> 00:10:44,169
Then draw a line through the middle
of it, because most of this workflow

176
00:10:44,680 --> 00:10:47,060
is not judgment at all.

177
00:10:47,745 --> 00:10:51,193
Counting how many critical
defects are open is arithmetic.

178
00:10:51,994 --> 00:10:57,614
Comparing eighty-seven percent against
your ninety percent line is arithmetic.

179
00:10:58,415 --> 00:11:03,419
Checking whether a yellow arrived
with a plan, an owner and a date is

180
00:11:03,477 --> 00:11:04,477
arithmetic.

181
00:11:04,800 --> 00:11:09,107
Checking whether a source was touched
in the last seven days is arithmetic.

182
00:11:10,210 --> 00:11:13,183
None of that should be a
language model's opinion,

183
00:11:13,693 --> 00:11:16,631
because an opinion can come
back different on Tuesday.

184
00:11:17,002 --> 00:11:20,532
Let ordinary code do the
counting and the comparing,

185
00:11:20,683 --> 00:11:23,782
and keep the model for the part
that actually needs reading.

186
00:11:24,061 --> 00:11:26,278
What the test lead wrote in the comments.

187
00:11:26,708 --> 00:11:30,504
Whether two owners are describing
the same problem in different words.

188
00:11:31,097 --> 00:11:32,745
The paragraph at the end.

189
00:11:33,256 --> 00:11:37,447
Build it that way and the only place
anything can wobble is the place

190
00:11:37,482 --> 00:11:38,482
you chose to put it.

191
00:11:39,065 --> 00:11:43,837
And decide whose account it runs under,
because that decides whether everybody

192
00:11:43,872 --> 00:11:44,872
gets the same answer.

193
00:11:45,300 --> 00:11:50,060
Run it as you, and it sees what
you can see, and your colleague

194
00:11:50,269 --> 00:11:53,601
opens the same workflow and
gets a quieter version of it.

195
00:11:54,286 --> 00:11:56,991
Give it its own identity
with its own named access,

196
00:11:57,362 --> 00:12:02,842
and every run reads the same eleven
sources no matter who pressed the button.

197
00:12:03,783 --> 00:12:06,047
For anything other people
are going to depend on,

198
00:12:06,476 --> 00:12:10,714
you want the second one, and it comes
with the obligations any account has.

199
00:12:11,225 --> 00:12:15,474
An owner, the least access that does
the job, and somebody who turns it

200
00:12:15,509 --> 00:12:16,670
off when the work is done.

201
00:12:17,355 --> 00:12:21,464
Microsoft has caught up with this,
and an agent can now hold its own

202
00:12:21,499 --> 00:12:25,737
identity the way a service account
does, and it shows up in the Microsoft

203
00:12:25,969 --> 00:12:29,173
three sixty five admin center
with an owner beside it.

204
00:12:29,858 --> 00:12:32,424
This also answers a question
that comes up constantly.

205
00:12:33,365 --> 00:12:36,186
Somebody writes a prompt
that works beautifully,

206
00:12:36,465 --> 00:12:40,342
shares it with the team, and it behaves
differently for everybody who runs

207
00:12:40,400 --> 00:12:41,243
it.

208
00:12:41,283 --> 00:12:42,444
That is not the prompt.

209
00:12:43,129 --> 00:12:47,575
Copilot grounds its answer in what you
personally have permission to open,

210
00:12:48,167 --> 00:12:51,500
and it will never hand back something
you could not have found yourself.

211
00:12:52,301 --> 00:12:57,084
So the prompt is identical and
the material underneath it is not.

212
00:12:57,769 --> 00:13:00,858
Which means a prompt was never
the right thing to share.

213
00:13:01,960 --> 00:13:07,139
Publish an agent with named knowledge
sources, so every answer is built on

214
00:13:07,185 --> 00:13:10,354
the same approved content,
and hand over the sources,

215
00:13:10,587 --> 00:13:15,652
the rules, and one example with
the right answer written beside it.

216
00:13:16,252 --> 00:13:19,259
Second, how will you
know when it is wrong.

217
00:13:20,060 --> 00:13:23,706
It goes wrong in three ways
and all three are quiet.

218
00:13:24,077 --> 00:13:25,122
It goes stale.

219
00:13:25,923 --> 00:13:30,788
A source nobody has updated in six
weeks, training numbers from August.

220
00:13:31,299 --> 00:13:34,573
It did not lie to you, it
read exactly what was there.

221
00:13:35,165 --> 00:13:39,356
So freshness becomes a rule, and
the age of every source goes at the

222
00:13:39,402 --> 00:13:40,402
top of the page.

223
00:13:40,621 --> 00:13:41,632
It softens.

224
00:13:42,142 --> 00:13:46,879
A dependency that has not moved in
three weeks comes back to you as

225
00:13:46,961 --> 00:13:48,470
monitoring ongoing.

226
00:13:49,062 --> 00:13:50,455
Every word of that is true.

227
00:13:51,140 --> 00:13:54,693
It is just a few degrees gentler
than the thing it describes,

228
00:13:54,972 --> 00:13:59,244
because the material it was built from
was written by people being careful

229
00:13:59,314 --> 00:14:00,261
too.

230
00:14:00,301 --> 00:14:04,898
So every morning you take one line
and open the source behind it,

231
00:14:05,328 --> 00:14:08,997
and that means the source
itself, not the page again.

232
00:14:09,597 --> 00:14:10,597
And it guesses.

233
00:14:10,793 --> 00:14:13,962
The vendor has not confirmed,
the results are half in,

234
00:14:14,287 --> 00:14:15,704
nobody actually knows yet.

235
00:14:16,075 --> 00:14:18,757
If you have not told it what
to do there, it will write a

236
00:14:18,850 --> 00:14:20,197
confident sentence anyway.

237
00:14:20,568 --> 00:14:21,741
So decide that up front.

238
00:14:22,171 --> 00:14:25,572
When it cannot tell, it says
so, and it names who to ask.

239
00:14:26,083 --> 00:14:30,031
I do not know, check with the migration
lead, is an answer you designed.

240
00:14:30,460 --> 00:14:33,107
There is one more test
and it costs you nothing.

241
00:14:33,339 --> 00:14:35,034
Ask it the same question twice.

242
00:14:35,719 --> 00:14:38,680
The counting should come
back identical, every time,

243
00:14:39,005 --> 00:14:42,883
and if it does not then something
is wrong in the wiring and you

244
00:14:42,952 --> 00:14:43,952
go and find it.

245
00:14:44,531 --> 00:14:49,094
The wording will not come back identical,
because that part is not deterministic.

246
00:14:49,605 --> 00:14:53,668
What you are watching for is
whether the conclusion moves.

247
00:14:54,470 --> 00:14:57,569
If two runs disagree about
anything you would act on,

248
00:14:58,162 --> 00:14:59,682
you have not pinned down enough yet.

249
00:15:00,275 --> 00:15:05,301
And the reason you sample instead of
trusting your gut is that the research

250
00:15:05,348 --> 00:15:10,805
on automation says we over-trust these
outputs, and that when the output explains

251
00:15:10,851 --> 00:15:14,485
its reasoning we trust it more, not less.

252
00:15:14,996 --> 00:15:18,943
The dangerous morning is not the one
where the page is obviously wrong.

253
00:15:19,373 --> 00:15:22,526
It is the one where it is
wrong and it argues well.

254
00:15:23,125 --> 00:15:25,853
Third, where does a person stand.

255
00:15:26,538 --> 00:15:29,069
It drafts to you, and only to you.

256
00:15:29,755 --> 00:15:33,713
It does not post the readiness summary
to the committee and it does not

257
00:15:33,795 --> 00:15:37,870
email the vendor, because the
moment it talks to other people,

258
00:15:38,670 --> 00:15:42,548
that is not a tool talking,
that is your program talking.

259
00:15:43,350 --> 00:15:46,449
The decision to launch
has a person's name on it,

260
00:15:47,250 --> 00:15:50,478
and in regulated work
that name is written down.

261
00:15:51,071 --> 00:15:54,286
The official who accepts the
risk cannot hand that off,

262
00:15:54,658 --> 00:15:59,000
and the most automation is allowed
to do is tell them and then hold

263
00:15:59,034 --> 00:16:00,928
the change until they answer.

264
00:16:01,728 --> 00:16:05,675
The workflow moves the work, but
it cannot carry the consequence.

265
00:16:06,779 --> 00:16:10,587
In the Microsoft stack that
gate is an approval action,

266
00:16:11,178 --> 00:16:15,520
and it finds you in Teams or in
Outlook, which is where you already

267
00:16:15,591 --> 00:16:17,808
are at seven in the morning.

268
00:16:18,179 --> 00:16:23,055
And decide now what you do the first
morning it is wrong and somebody

269
00:16:23,207 --> 00:16:26,144
has already read it, because
that morning is coming.

270
00:16:26,828 --> 00:16:30,938
You correct it in a way that
leaves the first version visible,

271
00:16:31,309 --> 00:16:36,361
because a page that quietly changes
overnight is worse than one that was wrong

272
00:16:36,731 --> 00:16:37,846
and says so.

273
00:16:38,446 --> 00:16:42,986
Fourth, what does it leave behind,
and what is it costing you.

274
00:16:43,787 --> 00:16:45,760
The record is five lines every morning.

275
00:16:46,352 --> 00:16:50,474
It writes down what it read and
when each source was last updated,

276
00:16:51,066 --> 00:16:54,375
then which rule it applied
and what that rule flagged,

277
00:16:54,700 --> 00:16:56,859
and finally who looked at the result.

278
00:16:57,184 --> 00:16:58,184
Keep those.

279
00:16:58,508 --> 00:17:02,339
Regulated work keeps this kind
of log for at least six months,

280
00:17:02,711 --> 00:17:06,646
and that is a fair floor for a
decision somebody will ask you about

281
00:17:06,681 --> 00:17:07,681
in March.

282
00:17:08,063 --> 00:17:10,153
Some of this you get without asking.

283
00:17:10,745 --> 00:17:16,272
Inside Microsoft three sixty five, Purview
is already logging Copilot and agent

284
00:17:16,306 --> 00:17:21,392
activity, including what was accessed,
and it holds that for a hundred and

285
00:17:21,484 --> 00:17:23,551
eighty days on the standard tier.

286
00:17:24,236 --> 00:17:27,278
That is the floor underneath
you rather than your record,

287
00:17:27,603 --> 00:17:31,772
and your record is those five lines,
in language a person can read.

288
00:17:32,371 --> 00:17:37,038
Then the cost, and the number that
matters is not what a run costs.

289
00:17:37,956 --> 00:17:39,791
It is what a useful run costs.

290
00:17:40,464 --> 00:17:44,248
If one morning in three needs
fixing before you can use it,

291
00:17:44,747 --> 00:17:48,870
you are paying for three mornings
and your own time to get one page

292
00:17:48,904 --> 00:17:49,886
you trust.

293
00:17:49,926 --> 00:17:52,457
Copilot will at least quote you a price.

294
00:17:52,967 --> 00:17:56,903
In Copilot Studio an agent
action costs five credits,

295
00:17:57,495 --> 00:18:02,197
grounding in your own tenant's data costs
ten, and two hundred dollars a month

296
00:18:02,789 --> 00:18:05,448
buys twenty-five thousand
credits for the whole tenant.

297
00:18:06,249 --> 00:18:10,846
So one morning run that takes
a handful of actions is cents.

298
00:18:11,649 --> 00:18:16,826
Five people running theirs against
eleven sources every day of launch week

299
00:18:17,104 --> 00:18:20,031
is a number you can put
in front of a sponsor.

300
00:18:20,622 --> 00:18:25,231
And the credits expire at the
end of the month, so what you are

301
00:18:25,277 --> 00:18:28,993
buying is capacity, not a balance.

302
00:18:29,585 --> 00:18:33,996
All of these limits move, so
check them the week you build,

303
00:18:34,507 --> 00:18:36,184
not the month you planned it.

304
00:18:36,784 --> 00:18:39,965
And decide how you will know
whether any of this worked,

305
00:18:40,337 --> 00:18:41,509
before you turn it on.

306
00:18:42,020 --> 00:18:46,084
Pick one number you already care
about, like how long the go,

307
00:18:46,409 --> 00:18:50,994
no go takes, or how many surprises
show up in the first week after

308
00:18:51,075 --> 00:18:54,373
launch, or how many times
you chase the same answer.

309
00:18:54,744 --> 00:18:57,623
Write it down now, while
it is still the old way,

310
00:18:57,995 --> 00:19:01,884
because once this is running you
cannot measure what it replaced.

311
00:19:02,314 --> 00:19:05,413
And do not ask people
afterwards whether it helped.

312
00:19:06,005 --> 00:19:07,457
There is a study I think about.

313
00:19:07,782 --> 00:19:11,845
Experienced developers were
given AI help on real work.

314
00:19:12,437 --> 00:19:15,560
Going in, they expected to
be about a quarter faster,

315
00:19:15,990 --> 00:19:19,461
and when it was measured they
were nineteen percent slower.

316
00:19:19,972 --> 00:19:23,432
And afterwards they still
believed they had been faster.

317
00:19:24,024 --> 00:19:27,751
These things feel like they are
helping even on the days they are not,

318
00:19:28,343 --> 00:19:30,466
which is why you write
the number down first.

319
00:19:31,066 --> 00:19:32,250
None of this is new.

320
00:19:33,051 --> 00:19:36,604
If you work somewhere with
a formal change process,

321
00:19:37,196 --> 00:19:42,757
the tooling there probably already
scores how risky a release is and routes

322
00:19:42,804 --> 00:19:45,091
the risky ones to a person.

323
00:19:45,776 --> 00:19:49,352
What is new is that you can
build the small version yourself,

324
00:19:49,677 --> 00:19:53,148
this week, for one question
that belongs to you.

325
00:19:53,578 --> 00:19:55,017
Be clear about the limits.

326
00:19:55,389 --> 00:19:59,429
It does not decide anything, and it
will not stop a launch that somebody

327
00:19:59,475 --> 00:20:00,927
has already decided to make.

328
00:20:01,519 --> 00:20:06,488
What you get is a page that shows
where every number came from and

329
00:20:06,534 --> 00:20:10,853
when it was last touched, and somewhere
sensible to start looking on Tuesday

330
00:20:10,911 --> 00:20:11,911
morning.

331
00:20:12,008 --> 00:20:14,446
So here is the whole thing in one picture.

332
00:20:15,247 --> 00:20:17,674
Your eleven sources sit on the left.

333
00:20:18,475 --> 00:20:21,609
You named every one of
them, and it can only read,

334
00:20:22,202 --> 00:20:23,202
never write.

335
00:20:23,781 --> 00:20:28,773
Your rules sit in the middle, with
the thresholds you set back when you

336
00:20:28,843 --> 00:20:29,853
planned the launch.

337
00:20:30,538 --> 00:20:33,138
One of those rules does most of the work.

338
00:20:33,568 --> 00:20:38,374
A yellow that arrives without a
plan, an owner and a date comes back

339
00:20:38,421 --> 00:20:39,421
to you in red.

340
00:20:39,953 --> 00:20:43,425
Below the rules is the
split that keeps it steady.

341
00:20:43,936 --> 00:20:47,651
Ordinary code does the
counting and the comparing,

342
00:20:48,022 --> 00:20:50,635
so Tuesday's answer matches Monday's.

343
00:20:51,320 --> 00:20:55,639
The model only gets used on the written
part, the comments and the notes

344
00:20:55,662 --> 00:20:57,925
that somebody typed in their own words.

345
00:20:58,525 --> 00:21:03,749
And that split is where most of
your bill goes, or does not go.

346
00:21:04,400 --> 00:21:08,138
You pay a model for what you
send it and what it sends back.

347
00:21:08,707 --> 00:21:13,548
Eleven sources on a launch is a
lot of words, and most of those

348
00:21:13,594 --> 00:21:16,299
words never needed judgment at all.

349
00:21:16,961 --> 00:21:23,080
Counting open defects is not judgment,
and neither is comparing eighty-seven

350
00:21:23,300 --> 00:21:29,140
against ninety, or checking whether a
date falls inside the last seven days.

351
00:21:29,906 --> 00:21:34,399
If the model does that work, you are
paying to push all eleven documents

352
00:21:34,434 --> 00:21:38,648
through it, every morning,
for the whole of launch week.

353
00:21:39,241 --> 00:21:43,292
If ordinary code does it, the
model never sees most of them.

354
00:21:43,977 --> 00:21:48,958
It gets the numbers already worked out,
and the few passages where somebody

355
00:21:49,167 --> 00:21:50,990
actually wrote something worth reading.

356
00:21:51,930 --> 00:21:55,680
Same page at the end, a
fraction of the input.

357
00:21:56,621 --> 00:22:01,416
In Copilot Studio every one of those
model steps is an action you are

358
00:22:01,462 --> 00:22:04,515
billed for, so the
fewer of them there are,

359
00:22:05,316 --> 00:22:06,790
the further your credits go.

360
00:22:07,592 --> 00:22:11,609
And the part you pay least for
is the part that cannot wobble on

361
00:22:11,678 --> 00:22:12,577
you.

362
00:22:12,617 --> 00:22:15,264
Out on the right is the page itself.

363
00:22:15,670 --> 00:22:21,150
You get three places to start, in
order, each one with the reason it

364
00:22:21,197 --> 00:22:25,690
is there and the date its
source was last touched.

365
00:22:26,096 --> 00:22:29,242
It runs on a clock, every
morning of launch week.

366
00:22:29,742 --> 00:22:33,074
It runs under its own
account with its own access,

367
00:22:33,248 --> 00:22:36,081
so your colleague gets
the same answer you do.

368
00:22:36,510 --> 00:22:38,774
Then it stops and drafts to you.

369
00:22:39,459 --> 00:22:44,034
The approval still carries a person's
name, and it waits until that person

370
00:22:44,092 --> 00:22:44,876
answers.

371
00:22:44,916 --> 00:22:48,260
And along the bottom, the five
lines it writes down every morning.

372
00:22:48,689 --> 00:22:53,252
Next to them, the one number you
chose to watch before you switched any

373
00:22:53,299 --> 00:22:54,299
of this on.

374
00:22:54,689 --> 00:22:59,252
If you have ever pasted your notes
into a chatbot and asked what you

375
00:22:59,287 --> 00:23:02,363
were missing, you are already at step one.

376
00:23:03,141 --> 00:23:08,273
Next time we build one of these live,
inside the tools you already have,

377
00:23:08,586 --> 00:23:11,257
and we go further than we went today.

378
00:23:11,930 --> 00:23:17,491
We fix the prompt that works beautifully
for you and badly for everyone else.

379
00:23:18,281 --> 00:23:22,983
A real M C P server goes
behind it, and we work out who

380
00:23:23,076 --> 00:23:25,525
owns that door before anybody opens it.

381
00:23:26,106 --> 00:23:29,670
And we move the arithmetic out
of the model and into code,

382
00:23:30,181 --> 00:23:34,581
which is where most of a token bill
goes, and the cheapest thing on

383
00:23:34,627 --> 00:23:36,264
this list to fix.

384
00:23:36,856 --> 00:23:39,814
I am Guillermo, and this is my AI clone.