1
00:00:00,06 --> 00:00:01,05
- [Instructor] How should humans

2
00:00:01,05 --> 00:00:03,06
and your AI agent collaborate?

3
00:00:03,06 --> 00:00:05,02
That's the last critical pillar

4
00:00:05,02 --> 00:00:07,06
to moving your agent into production,

5
00:00:07,06 --> 00:00:09,06
and this is all about finding the right balance

6
00:00:09,06 --> 00:00:12,05
between autonomy and oversight.

7
00:00:12,05 --> 00:00:14,02
Why does it matter so much?

8
00:00:14,02 --> 00:00:16,06
Well, because if you give your agent too much freedom,

9
00:00:16,06 --> 00:00:19,07
you risk mistakes or unexpected actions.

10
00:00:19,07 --> 00:00:20,06
Give it too little,

11
00:00:20,06 --> 00:00:23,08
and it's just a fancy chatbot with little value.

12
00:00:23,08 --> 00:00:25,03
The key is designing the interface

13
00:00:25,03 --> 00:00:27,00
so that humans stay in control,

14
00:00:27,00 --> 00:00:30,02
while the agent does the heavy lifting.

15
00:00:30,02 --> 00:00:31,09
You'll hear two common patterns.

16
00:00:31,09 --> 00:00:34,02
The first one is human in the loop.

17
00:00:34,02 --> 00:00:36,08
The human improves or edits the agent's output

18
00:00:36,08 --> 00:00:39,00
before it moves forward.

19
00:00:39,00 --> 00:00:41,02
The other one is AI in the loop.

20
00:00:41,02 --> 00:00:44,04
This is a more augmented approach where the agent assists

21
00:00:44,04 --> 00:00:47,05
or automates a human-led process.

22
00:00:47,05 --> 00:00:48,08
Both types are valid.

23
00:00:48,08 --> 00:00:51,02
Which one you choose really depends on the risk

24
00:00:51,02 --> 00:00:53,04
and the context of the task.

25
00:00:53,04 --> 00:00:57,02
In our presales agent, we use a human in the loop setup.

26
00:00:57,02 --> 00:01:00,00
The agent prepares the quote and drafts the email,

27
00:01:00,00 --> 00:01:02,08
but the salesperson reviews it before sending it out,

28
00:01:02,08 --> 00:01:04,07
and that's the right balance in this case.

29
00:01:04,07 --> 00:01:08,02
Automation helps, but humans stay accountable.

30
00:01:08,02 --> 00:01:09,02
Keep a human in the loop

31
00:01:09,02 --> 00:01:11,07
whenever the decision has real consequences,

32
00:01:11,07 --> 00:01:15,05
like pricing, compliance, or communication with customers.

33
00:01:15,05 --> 00:01:19,02
AI agents can suggest or prepare, but humans in this case,

34
00:01:19,02 --> 00:01:21,03
decide and approve.

35
00:01:21,03 --> 00:01:23,05
On the other hand, full autonomy works great

36
00:01:23,05 --> 00:01:26,01
for repetitive or low-risk tasks,

37
00:01:26,01 --> 00:01:28,03
like fetching data, summarizing notes,

38
00:01:28,03 --> 00:01:30,03
or preparing internal drafts.

39
00:01:30,03 --> 00:01:32,07
That's where agents shine.

40
00:01:32,07 --> 00:01:36,04
A well-designed interface makes those boundaries obvious.

41
00:01:36,04 --> 00:01:39,01
In NAM, you can use the human in the loop nodes

42
00:01:39,01 --> 00:01:42,02
that allows you to send messages and have humans respond

43
00:01:42,02 --> 00:01:45,00
before the agent continues with a workflow.

44
00:01:45,00 --> 00:01:47,07
This allows you to let humans step in when needed,

45
00:01:47,07 --> 00:01:50,06
and always make it easy to pause, edit,

46
00:01:50,06 --> 00:01:53,07
or override the agent's actions.

47
00:01:53,07 --> 00:01:55,09
My tip, design with a high level

48
00:01:55,09 --> 00:01:58,04
of human-in-the-loop involvement at first.

49
00:01:58,04 --> 00:02:00,04
Build your workflow so a person reviews

50
00:02:00,04 --> 00:02:03,04
each critical output, like quotes, emails, or decisions,

51
00:02:03,04 --> 00:02:05,00
before they go out.

52
00:02:05,00 --> 00:02:08,01
And once the agent performs reliably and consistently,

53
00:02:08,01 --> 00:02:11,06
you can start to remove or automate those approval steps.

54
00:02:11,06 --> 00:02:14,05
It's the fastest way to scale without losing trust,

55
00:02:14,05 --> 00:02:16,00
and it keeps your team confident

56
00:02:16,00 --> 00:02:19,00
as the agent takes on more responsibility.