Agentic AI Best Practices – Designing Humans in the Loop (HITL)

AI
1

Human-in-the-loop is not just a safety net — it is an architecture decision.

In agentic AI, the real question is not simply “Should a human be involved?”
It is “Where should human judgment live?”

This article introduces a practical Agent–Case boundary pattern for enterprise AI:

  • The Agent prepares, reasons, recommends, and evaluates confidence.
  • Deterministic rules decide whether work can continue automatically.
  • The Case owns formal human work when SLA, approval, reassignment, auditability, or exception handling is required.

The core best practice:

AI prepares. Rules govern. Cases own human work.

This pattern helps make agentic AI more predictable, auditable, and production-ready.

2

Agreed @laucf ! Very nice article :slight_smile:

3

Thanks @laucf!
One thing I often wonder about in the HITL Assisted pattern is: how do you keep the human in their critical Judge role, and how do you avoid them just rubber-stamping everything the agent has suggested without a real review?

Humans can at times be too thrusting, especially if the agent’s output is high quality 99 out of a 100 times. Which patterns have you seen to make sure they are an active judge in the process?
We’ve all seen those videos of people reading books or sleeping while their not-yet-fully-self-driving car zipped along the highway :slight_smile:

4

@Tim_Straatsma Great question. I think the answer is to

Look at the model performance & the human bias!

If a reviewer agrees with the agent 99.9% of the time, that may indicate one of two things:

  1. The agent is genuinely performing extremely well.
  2. The human has drifted from “judge” to "rubber stamp."

But there is an opposite failure mode as well:

Humans repeatedly override the agent, not because the recommendation is wrong, but because they distrust the system, feel a loss of control, or are concerned about their job.

The interesting metric is therefore not approval rate, but:

  • Override frequency
  • Types of overrides (in which situation, by whom)
  • Time spent reviewing
  • Variance between reviewers (distance between agent and human output)
  • Correlation between confidence scores and override decisions
  • Outcomes of overrides (was the human or the agent ultimately correct?)

Together these measures help identify both automation bias (“the AI is usually right, so I stop thinking critically”) and displacement bias (“I override because I do not trust the AI or feel threatened by it”), which can be equally damaging.

In mature implementations, HITL becomes a monitored control rather than simply a workflow step.

This necessitates to collect this data in the CaseType to be able to measure.

The goal is not just to have a human in the loop, but to continuously validate that human judgment is still being exercised appropriately—neither blindly accepting nor reflexively rejecting the agent’s recommendations.

The self-driving car analogy is a good one: the challenge is not only preventing people from falling asleep at the wheel, but also preventing them from grabbing the wheel unnecessarily every few minutes because they don’t trust the system. Both behaviors are signals that the human–AI partnership is not yet calibrated correctly.

5

Great article and key to predictability