Human in the loop
A design in which a named person approves, corrects or can stop an agent's actions at defined points, not after the fact.
Human in the loop is a design in which a person has a real opportunity to approve, correct or stop what an agent does. The word real matters. A review that happens after an action has already reached a customer, or a reviewer who cannot see the reasoning, is not a control.
The design question is not whether a person is involved but where, how often and with what information. Too many approval points make the agent useless. Too few make the person a rubber stamp.
In practice
I classify each action the agent can take. Reads and drafts usually need no approval. Writes to internal systems usually need a check. Anything that leaves the organisation, moves money or changes a customer record needs an explicit approval from a named person before it runs. The approver sees the proposed action, the evidence the agent used and the rule that triggered the review.
I also measure the approval rate. If a reviewer approves almost every proposal without a change, the review point is either too early or the agent is good enough to move the threshold. Either way, the decision should be deliberate.
Example
[Composite] A procurement agent drafted purchase requests. Reviewers approved almost every draft. The team moved low-value, in-policy requests to a sampled review, keeping full review for anything above a set amount or outside the approved vendor list. The approval queue fell by most of its volume, and the sampled reviews caught two policy errors in the first month.
Common mistakes
- Putting a person at every step, which hides the real risk behind routine clicks.
- Showing the approver only the final proposal, not the evidence.
- Assuming a person in the loop removes the need for evaluation. The reviewer also needs measuring.
Related terms
Cite this definition
Human in the loop. Mayur Patil, mayurpatil.ai, updated 2026-11-01. https://mayurpatil.ai/glossary/human-in-the-loop/