WEBVTT

1
00:00:08.570 --> 00:00:13.239
Agentic delivery will make
execution faster and more abundant.

2
00:00:13.376 --> 00:00:17.316
Unless the same intelligence
goes into understanding intent,

3
00:00:17.316 --> 00:00:21.211
it will also help us travel much
further in the wrong direction.

4
00:00:21.798 --> 00:00:25.883
It begins with an idea
that isn't yet completely clear.

5
00:00:26.019 --> 00:00:29.452
AI produces a plausible suggestion.

6
00:00:29.591 --> 00:00:34.653
"Yes, that will do." The next
output builds on the first,

7
00:00:34.653 --> 00:00:36.578
then another builds on that.

8
00:00:37.165 --> 00:00:40.624
Nothing looks obviously
wrong in isolation.

9
00:00:40.761 --> 00:00:45.177
But several iterations later there's
a substantial gap between what was wanted

10
00:00:45.177 --> 00:00:47.468
and what has actually been created.

11
00:00:47.605 --> 00:00:50.304
Call it the AI rabbit hole.

12
00:00:51.524 --> 00:00:54.943
The AI has not necessarily failed.

13
00:00:55.080 --> 00:00:58.051
It followed the direction it was given.

14
00:00:58.188 --> 00:01:02.123
A provisional suggestion
became an accepted decision.

15
00:01:02.260 --> 00:01:05.299
AI makes that surprisingly easy.

16
00:01:05.885 --> 00:01:08.829
Transformation has the same problem.

17
00:01:08.966 --> 00:01:13.662
Understanding what is really needed
is often compressed or skipped.

18
00:01:13.799 --> 00:01:18.006
Used carelessly,
AI accelerates every step.

19
00:01:18.143 --> 00:01:23.102
Assumptions become requirements,
requirements become designs, and designs

20
00:01:23.102 --> 00:01:27.622
become working software before anyone
has tested the thinking underneath them.

21
00:01:28.842 --> 00:01:32.411
This isn't an argument
against agentic delivery.

22
00:01:32.547 --> 00:01:35.659
Use AI to shorten development cycles,

23
00:01:35.659 --> 00:01:39.975
generate code, create tests
and produce documentation.

24
00:01:40.562 --> 00:01:45.896
But applying AI only to execution
leaves half the opportunity untouched.

25
00:01:46.033 --> 00:01:48.556
Use it to understand too.

26
00:01:48.693 --> 00:01:54.435
Compare interpretations, expose
contradictions, and test assumptions.

27
00:01:55.022 --> 00:01:58.672
The article behind this
video was written that way.

28
00:01:58.809 --> 00:02:03.723
AI interviewed its author, and challenged
where the position was unclear.

29
00:02:03.859 --> 00:02:08.759
It contributed knowledge without being
allowed to manufacture the point of view.

30
00:02:09.979 --> 00:02:13.724
One customer kept saying
they wanted to go faster.

31
00:02:13.861 --> 00:02:18.712
The obvious response was to take
the familiar sequence of workshops, build,

32
00:02:18.712 --> 00:02:21.737
testing and enablement, and compress it.

33
00:02:22.324 --> 00:02:25.023
That wasn't what they were asking for.

34
00:02:25.159 --> 00:02:29.045
They wanted to start with the
out-of-the-box build, then use testing

35
00:02:29.045 --> 00:02:32.220
and enablement to find
the few changes they needed.

36
00:02:32.807 --> 00:02:36.285
Best practice is whatever's
best for this customer,

37
00:02:36.285 --> 00:02:38.998
with these people, in this situation.

38
00:02:39.134 --> 00:02:43.899
AI is exceptionally good
at reproducing established patterns.

39
00:02:44.036 --> 00:02:45.856
Without critical review,

40
00:02:45.856 --> 00:02:50.552
you get the same best-practice answer
every time, whether or not it fits.

41
00:02:51.772 --> 00:02:53.280
The context that changes

42
00:02:53.280 --> 00:02:58.493
a recommendation is not always captured
in a process document or a data model.

43
00:02:58.630 --> 00:03:01.361
It may be tension between two departments,

44
00:03:01.361 --> 00:03:04.037
the way a particular
leader makes decisions,

45
00:03:04.037 --> 00:03:07.932
or the moment in a meeting
when a sensible suggestion doesn't land.

46
00:03:08.519 --> 00:03:11.666
AI can analyse what was recorded.

47
00:03:11.803 --> 00:03:16.825
It cannot yet be relied on to navigate
every human dynamic around it.

48
00:03:16.962 --> 00:03:22.134
So soft skills do not become less
important as delivery is automated.

49
00:03:22.271 --> 00:03:25.554
They become part of how AI is steered.

50
00:03:26.774 --> 00:03:29.858
The operating model has
to preserve the customer's intent

51
00:03:29.858 --> 00:03:33.921
and the evidence behind each decision
across every project boundary,

52
00:03:33.921 --> 00:03:37.014
instead of resetting
its understanding to zero.

53
00:03:37.601 --> 00:03:42.324
When evidence changes the outcome,
record why and who decided.

54
00:03:42.461 --> 00:03:45.907
Unrecorded drift is moving the goalposts.

55
00:03:46.044 --> 00:03:49.640
Evidence-led, authorised
revision is learning.

56
00:03:50.227 --> 00:03:53.080
Use AI on both sides of delivery,

57
00:03:53.080 --> 00:03:57.754
so faster execution stays connected to
an outcome someone has consciously chosen

58
00:03:57.754 --> 00:03:59.475
and can still defend.
