1
00:00:02,000 --> 00:00:07,489
You are leading an airline during a difficult
morning. Bad weather is approaching. Several

2
00:00:07,489 --> 00:00:09,136
aircraft are arriving late.

3
00:00:09,436 --> 00:00:14,576
Customers want answers, and your operations team
wants to know where to intervene.

4
00:00:15,667 --> 00:00:21,256
Three requests land on your desk. First: which
flights are most likely to depart late?

5
00:00:21,556 --> 00:00:27,487
Second: how should we explain the disruption to
passengers? Third: which passengers qualify

6
00:00:27,487 --> 00:00:30,486
for a meal voucher under our published policy?

7
00:00:31,583 --> 00:00:37,473
All three requests involve information. All
three might appear in an AI proposal. But they

8
00:00:37,473 --> 00:00:39,501
require different capabilities.

9
00:00:39,801 --> 00:00:45,289
Your task is to select the approach that fits
the decision, rather than choose a technology

10
00:00:45,289 --> 00:00:47,761
first and search for a problem afterward.

11
00:00:48,833 --> 00:00:52,085
For the first request, imagine a forecasting
analyst.

12
00:00:52,385 --> 00:00:57,607
The analyst studies previous flights, weather
conditions, turnaround times, and other

13
00:00:57,607 --> 00:01:01,232
relevant evidence. The output is an estimate of
delay risk.

14
00:01:01,532 --> 00:01:06,722
This is a common role for classical machine
learning: learning patterns from examples to

15
00:01:06,722 --> 00:01:09,435
produce a prediction, classification, ranking,

16
00:01:09,655 --> 00:01:11,312
or numerical estimate.

17
00:01:12,375 --> 00:01:15,928
For the second request, imagine a communications
adviser.

18
00:01:16,228 --> 00:01:21,719
The adviser receives verified facts and drafts a
helpful explanation in language the

19
00:01:21,719 --> 00:01:23,353
passenger can understand.

20
00:01:23,653 --> 00:01:28,995
This is a common role for generative AI:
producing content, including summaries,

21
00:01:28,995 --> 00:01:30,664
explanations, and drafts.

22
00:01:30,964 --> 00:01:36,382
The business evaluates whether that content is
accurate, useful, appropriate, and efficient

23
00:01:36,382 --> 00:01:36,978
to review.

24
00:01:38,042 --> 00:01:42,533
For the third request, imagine a supervisor
applying a clear policy.

25
00:01:42,833 --> 00:01:48,277
If the qualifying conditions are satisfied, the
passenger receives the specified benefit.

26
00:01:48,577 --> 00:01:54,435
That may require ordinary rules-based
automation. Learning from examples adds little
value

27
00:01:54,435 --> 00:01:57,233
when the organization already knows exactly

28
00:01:57,453 --> 00:02:01,938
which rule should apply and can express the
relevant conditions reliably.

29
00:02:03,000 --> 00:02:07,532
Your first strategic responsibility is to
identify the work product.

30
00:02:07,832 --> 00:02:13,401
Are you asking for an estimate, new content, or
consistent application of a known rule?

31
00:02:13,701 --> 00:02:18,130
Different parts of the same customer journey may
require different answers.

32
00:02:18,430 --> 00:02:23,639
A single conversational interface can conceal
several distinct business capabilities

33
00:02:23,639 --> 00:02:24,507
underneath it.

34
00:02:25,583 --> 00:02:31,685
These categories are useful, but they are not
absolute scientific walls. Generative AI is

35
00:02:31,685 --> 00:02:34,221
usually built using machine learning.

36
00:02:34,521 --> 00:02:39,946
It can also classify information. Predictive
systems can work with text and images.

37
00:02:40,246 --> 00:02:45,996
We distinguish the approaches because business
tasks create different evaluation, cost, and

38
00:02:45,996 --> 00:02:47,323
control requirements,

39
00:02:47,543 --> 00:02:50,690
not because the labels establish rigid
boundaries.

40
00:02:51,750 --> 00:02:57,407
Consider a beautifully written passenger message
stating that a flight will leave in twenty

41
00:02:57,407 --> 00:02:59,832
minutes. The message sounds reassuring.

42
00:03:00,132 --> 00:03:04,736
But if nobody has confirmed the departure time,
its fluency creates risk.

43
00:03:05,036 --> 00:03:09,385
The system has produced plausible language
without sufficient evidence.

44
00:03:09,685 --> 00:03:15,460
Customers may act on that statement even though
the underlying information does not justify

45
00:03:15,460 --> 00:03:15,650
it.

46
00:03:16,708 --> 00:03:22,613
Now consider an accurate delay forecast that
reaches no employee who can change the outcome.

47
00:03:22,913 --> 00:03:29,096
Its predictive performance may be impressive,
but its business value is limited. Capability

48
00:03:29,096 --> 00:03:30,998
alone does not create value.

49
00:03:31,298 --> 00:03:36,686
Someone must use the output within a workable
process, with enough time and authority to

50
00:03:36,686 --> 00:03:38,338
take an appropriate action.

51
00:03:39,417 --> 00:03:44,321
This brings us to a second distinction:
prediction is not explanation of cause.

52
00:03:44,621 --> 00:03:50,366
A model might identify a relationship between
operating conditions and delays. That

53
00:03:50,366 --> 00:03:53,343
relationship can help prioritize attention.

54
00:03:53,643 --> 00:03:57,823
It does not automatically prove which
intervention will prevent a delay.

55
00:03:58,123 --> 00:04:02,992
Business teams must evaluate the proposed
intervention as well as the prediction.

56
00:04:04,042 --> 00:04:09,372
Similarly, a language model can write a
convincing explanation of a prediction without

57
00:04:09,372 --> 00:04:12,161
access to the model's actual reasoning or the

58
00:04:12,381 --> 00:04:18,645
operational facts. Treat an unsupported
explanation as another generated claim requiring

59
00:04:18,645 --> 00:04:19,570
verification.

60
00:04:19,870 --> 00:04:25,483
Ask what evidence supports it, what uncertainty
remains, and whether a person could mistake

61
00:04:25,483 --> 00:04:27,950
a suggestion for a confirmed conclusion.

62
00:04:29,000 --> 00:04:34,456
As a strategist, ask four questions before
discussing products. What output does the

63
00:04:34,456 --> 00:04:37,704
business need? What evidence supports that
output?

64
00:04:38,004 --> 00:04:41,895
What happens if it is wrong? And who has
authority to act on it?

65
00:04:42,195 --> 00:04:47,854
Those questions establish the strategic decision
boundary and make the investment discussion

66
00:04:47,854 --> 00:04:48,715
more concrete.

67
00:04:49,792 --> 00:04:55,213
Let us move from the airport to Harborline
Distribution, a fictional company supplying

68
00:04:55,213 --> 00:04:57,230
equipment to business customers.

69
00:04:57,530 --> 00:05:03,090
Some orders miss their promised delivery
windows, and account managers spend too much
time

70
00:05:03,090 --> 00:05:04,758
preparing customer updates.

71
00:05:05,058 --> 00:05:11,583
The chief operating officer proposes one
generative AI assistant to solve both problems.

72
00:05:12,667 --> 00:05:18,030
The proposal is attractive because employees
could ask questions in ordinary language.

73
00:05:18,330 --> 00:05:24,044
But a convenient interface does not establish
that one underlying approach should perform

74
00:05:24,044 --> 00:05:24,750
every task.

75
00:05:25,050 --> 00:05:30,704
Harborline first separates the need to estimate
delivery risk from the need to explain

76
00:05:30,704 --> 00:05:33,005
verified information to a customer.

77
00:05:34,083 --> 00:05:40,276
Harborline has historical records describing
orders, shipment movements, promised dates, and

78
00:05:40,276 --> 00:05:41,689
actual arrival dates.

79
00:05:41,989 --> 00:05:47,913
If those records are reliable and
representative, a predictive model could
identify orders

80
00:05:47,913 --> 00:05:49,887
at increased risk of lateness.

81
00:05:50,187 --> 00:05:55,367
The relevant business question is whether early
warnings help employees prevent service

82
00:05:55,367 --> 00:05:55,902
failures,

83
00:05:56,122 --> 00:05:59,191
not simply whether the model produces a score.

84
00:06:00,250 --> 00:06:05,930
A model that raises too many unnecessary
warnings can overwhelm the operations team.

85
00:06:06,230 --> 00:06:11,746
A model that misses important failures can leave
customers without support. Both mistakes

86
00:06:11,746 --> 00:06:14,102
matter, but their consequences differ.

87
00:06:14,402 --> 00:06:19,641
Missing a delivery to a critical customer may be
much more costly than unnecessarily

88
00:06:19,641 --> 00:06:21,513
checking an ordinary shipment.

89
00:06:22,583 --> 00:06:24,657
Now examine communication.

90
00:06:24,957 --> 00:06:30,338
Account managers gather order information,
review customer history, locate relevant

91
00:06:30,338 --> 00:06:32,737
policies, and compose an explanation.

92
00:06:33,037 --> 00:06:37,432
Generative AI could prepare a first draft using
approved information.

93
00:06:37,732 --> 00:06:43,235
It might summarize the situation, adapt the
tone, and suggest an appropriate next step.

94
00:06:43,535 --> 00:06:49,293
It should not manufacture a confirmed arrival
date simply because the customer expects one.

95
00:06:50,333 --> 00:06:55,923
Harborline separates three responsibilities. The
predictive component estimates delivery

96
00:06:55,923 --> 00:06:56,241
risk.

97
00:06:56,541 --> 00:07:01,967
Authoritative business records establish what is
currently known. The generative component

98
00:07:01,967 --> 00:07:04,680
prepares a message grounded in those records.

99
00:07:04,980 --> 00:07:08,514
A manager approves sensitive promises or
exceptions.

100
00:07:08,814 --> 00:07:14,204
Connecting these capabilities does not give any
component unlimited authority over the

101
00:07:14,204 --> 00:07:14,643
others.

102
00:07:15,708 --> 00:07:21,448
Notice the difference between saying a shipment
has elevated delay risk and saying it will

103
00:07:21,448 --> 00:07:24,063
arrive tomorrow. The first is a forecast.

104
00:07:24,363 --> 00:07:29,832
The second sounds like an operational
commitment. Passing information between systems
must

105
00:07:29,832 --> 00:07:31,412
preserve that distinction.

106
00:07:31,712 --> 00:07:37,149
A draft should clearly distinguish a confirmed
event, an estimate, and a proposed action.

107
00:07:38,208 --> 00:07:43,741
Before selecting a solution, examine the data.
Historical delivery records may contain

108
00:07:43,741 --> 00:07:44,642
missing dates.

109
00:07:44,942 --> 00:07:50,603
Customer-service notes may reflect inconsistent
practices. An important supplier may have

110
00:07:50,603 --> 00:07:52,384
changed its operating model.

111
00:07:52,684 --> 00:07:57,999
A large collection of records is not
automatically a useful collection of evidence,

112
00:07:58,219 --> 00:08:02,374
and more records do not necessarily correct a
systematic gap.

113
00:08:03,417 --> 00:08:09,169
Training means learning patterns from examples.
Using the resulting model on a new case is

114
00:08:09,169 --> 00:08:10,639
often called inference.

115
00:08:10,939 --> 00:08:16,636
If training examples do not reflect current
operations, predictions may become less useful.

116
00:08:16,936 --> 00:08:22,431
If recorded outcomes are unreliable, the model
may learn from mistakes in the records.

117
00:08:22,731 --> 00:08:28,009
Business owners must understand these
limitations without implementing the technical

118
00:08:28,009 --> 00:08:28,512
process.

119
00:08:29,583 --> 00:08:35,222
A general-purpose model may have broad language
capability without knowing Harborline's

120
00:08:35,222 --> 00:08:37,879
current policies or customer commitments.

121
00:08:38,179 --> 00:08:44,226
One response is retrieval-augmented generation,
often shortened to RAG. Think of giving the

122
00:08:44,226 --> 00:08:47,083
communications adviser a library assistant.

123
00:08:47,383 --> 00:08:53,388
Before the adviser writes, the assistant finds
relevant approved documents. The adviser uses

124
00:08:53,388 --> 00:08:55,737
that material to prepare a response.

125
00:08:56,792 --> 00:09:02,636
Amazon Bedrock Knowledge Bases supports this
retrieval-and-generation pattern, including

126
00:09:02,636 --> 00:09:05,891
organizational information and source
references.

127
00:09:06,191 --> 00:09:11,278
Retrieval can improve relevance and support
verification, but it does not guarantee

128
00:09:11,278 --> 00:09:12,014
correctness.

129
00:09:12,314 --> 00:09:17,762
If the retrieved policy is outdated, the
response may repeat outdated information.

130
00:09:18,061 --> 00:09:22,893
The organization still owns document quality,
appropriate access, and review.

131
00:09:23,958 --> 00:09:30,408
Do not confuse retrieval with fine-tuning.
Fine-tuning adapts model behavior using
examples.

132
00:09:30,708 --> 00:09:35,857
It is not automatically the best way to keep
frequently changing facts current.

133
00:09:36,157 --> 00:09:41,966
For Harborline's changing delivery policies,
access to approved current information is the

134
00:09:41,966 --> 00:09:43,386
immediate requirement.

135
00:09:43,686 --> 00:09:49,304
Choose an adaptation approach according to the
specific quality problem you are trying to

136
00:09:49,304 --> 00:09:49,682
solve.

137
00:09:50,750 --> 00:09:56,256
At a high level, Amazon SageMaker AI is an
option for custom predictive or other

138
00:09:56,256 --> 00:09:58,252
machine-learning development.

139
00:09:58,552 --> 00:10:04,813
Amazon Bedrock is an option for generative AI
applications using available foundation

140
00:10:04,813 --> 00:10:05,329
models.

141
00:10:05,629 --> 00:10:09,456
These are strategic starting points, not
exclusive boundaries.

142
00:10:09,756 --> 00:10:14,925
Managed services still require customer
decisions about data, quality, operating

143
00:10:14,925 --> 00:10:17,768
responsibility, and acceptable business use.

144
00:10:18,833 --> 00:10:24,248
Harborline could consider Amazon Q Business for
internal knowledge assistance if its

145
00:10:24,248 --> 00:10:26,181
capabilities fit the workflow.

146
00:10:26,481 --> 00:10:28,972
Product fit requires evaluation.

147
00:10:29,272 --> 00:10:34,632
A product supporting internal knowledge work is
not automatically suitable for every

148
00:10:34,632 --> 00:10:36,163
customer-facing process.

149
00:10:37,250 --> 00:10:43,290
Amazon Bedrock Guardrails can help detect or
filter selected undesirable content and

150
00:10:43,290 --> 00:10:44,872
sensitive information.

151
00:10:45,172 --> 00:10:51,173
Supported grounding checks can help identify
responses that diverge from supplied sources.

152
00:10:51,173 --> 00:10:53,774
These controls reduce particular risks.

153
00:10:54,074 --> 00:10:59,807
They do not replace access management, business
authorization, or evaluation of whether the

154
00:10:59,807 --> 00:11:02,454
response is suitable for its intended use.

155
00:11:03,500 --> 00:11:08,815
A content filter and a permission check answer
different questions. One concerns what the

156
00:11:08,815 --> 00:11:09,531
system says.

157
00:11:09,831 --> 00:11:13,384
The other concerns what information the user may
receive.

158
00:11:13,684 --> 00:11:18,903
A manager's approval answers another question:
whether the business should make a particular

159
00:11:18,903 --> 00:11:19,527
commitment.

160
00:11:19,827 --> 00:11:24,800
Strong governance preserves these separate
responsibilities instead of assuming one

161
00:11:24,800 --> 00:11:26,537
safeguard performs every job.

162
00:11:27,583 --> 00:11:30,111
Now bring the decision back to economics.

163
00:11:30,411 --> 00:11:36,215
A presentation promises that generated messages
will save every account manager several

164
00:11:36,215 --> 00:11:38,751
hours each week. That is a hypothesis.

165
00:11:39,051 --> 00:11:44,837
Harborline must test actual time saved after
checking facts, correcting drafts, and handling

166
00:11:44,837 --> 00:11:45,528
exceptions.

167
00:11:45,828 --> 00:11:51,995
Time released may create capacity, reduce
overtime, or improve response times; it is not

168
00:11:51,995 --> 00:11:53,747
automatically cash saved.

169
00:11:54,792 --> 00:11:59,776
For predictive machine learning, measure whether
employees can act on a warning in time.

170
00:12:00,076 --> 00:12:05,846
For generative AI, measure whether a draft is
useful, supported, and faster to approve than

171
00:12:05,846 --> 00:12:07,178
writing from scratch.

172
00:12:07,478 --> 00:12:10,098
Avoid one broad AI accuracy score.

173
00:12:10,398 --> 00:12:16,063
Track delivery-risk performance, message
quality, human review effort, and customer
outcomes

174
00:12:16,063 --> 00:12:16,740
separately.

175
00:12:17,792 --> 00:12:19,434
Consider recurring cost.

176
00:12:19,734 --> 00:12:25,326
Long conversations, extensive background
material, repeated requests, and unnecessary

177
00:12:25,326 --> 00:12:27,498
retries can increase model usage.

178
00:12:27,798 --> 00:12:33,548
Predictive models have ongoing costs too,
including operation, information maintenance,

179
00:12:33,548 --> 00:12:35,134
evaluation, and updates.

180
00:12:35,434 --> 00:12:39,638
There is no universal rule that one category is
always cheaper.

181
00:12:39,938 --> 00:12:44,786
Compare the cost of completing the business task
at the required quality and volume.

182
00:12:45,833 --> 00:12:47,672
Begin with a bounded pilot.

183
00:12:47,972 --> 00:12:53,418
Establish the current baseline: missed
commitments, message preparation time,
unnecessary

184
00:12:53,418 --> 00:12:55,254
interventions, and complaints.

185
00:12:55,554 --> 00:13:00,956
Evaluate representative cases, including
incomplete records and policy exceptions.

186
00:13:01,256 --> 00:13:06,491
Test predictions on cases not used for learning.
Check generated messages against

187
00:13:06,491 --> 00:13:07,912
authoritative records.

188
00:13:08,212 --> 00:13:13,658
Use named owners, spending limits, and a
practical fallback to the existing process.

189
00:13:14,708 --> 00:13:20,327
Now apply that reasoning. A distributor has
reliable delivery history, confirmed shipment

190
00:13:20,327 --> 00:13:22,284
records, and approved policies.

191
00:13:22,584 --> 00:13:28,190
It wants a limited pilot to predict delays and
prepare customer updates without inventing

192
00:13:28,190 --> 00:13:28,946
commitments.

193
00:13:29,246 --> 00:13:35,271
Option A uses one general-purpose generative
model for everything and measures employee

194
00:13:35,271 --> 00:13:36,171
satisfaction.

195
00:13:36,471 --> 00:13:42,005
Option B separates predictive risk assessment
from evidence-grounded drafting and reviews

196
00:13:42,005 --> 00:13:42,752
commitments.

197
00:13:43,792 --> 00:13:49,356
Option C uses a predictive risk score alone to
establish a confirmed delivery date.

198
00:13:49,656 --> 00:13:55,649
Option D fine-tunes on historical emails and
sends updates automatically once the writing

199
00:13:55,649 --> 00:13:56,861
matches the brand.

200
00:13:57,161 --> 00:14:00,570
Pause and select the approach that satisfies the
requirements.

201
00:14:00,870 --> 00:14:06,339
Consider evidence, evaluation, and authority
rather than the apparent simplicity of the

202
00:14:06,339 --> 00:14:06,905
proposal.

203
00:14:07,958 --> 00:14:09,557
The best answer is B.

204
00:14:09,857 --> 00:14:15,129
It matches prediction to the risk task and
generation to communication, while preserving
the

205
00:14:15,129 --> 00:14:17,536
distinction between a forecast and a fact.

206
00:14:17,836 --> 00:14:23,956
A fails because satisfaction does not establish
predictive or factual quality. C turns an

207
00:14:23,956 --> 00:14:26,706
estimate into an unsupported commitment.

208
00:14:27,006 --> 00:14:31,978
D addresses style while leaving current facts
and authorization unresolved.

209
00:14:33,042 --> 00:14:38,883
Read the requested outcome and constraints.
Identify what must be predicted, generated,

210
00:14:38,883 --> 00:14:41,099
verified, and governed by a rule.

211
00:14:41,399 --> 00:14:47,017
The analyst estimates. The adviser communicates.
The policy manual governs explicit

212
00:14:47,017 --> 00:14:47,897
entitlements.

213
00:14:48,197 --> 00:14:53,550
A strong AI strategy gives each the appropriate
job, connects their work carefully, and

214
00:14:53,550 --> 00:14:55,949
keeps accountability with the business.