Skip to content
FocalIDFocalID

Human in the loop: where a person should approve an AI agent’s work

An AI agent that acts in your systems should not act alone on everything. “Human in the loop” means a person approves, corrects or takes over at defined points. The skill is choosing those points well: too few and mistakes reach customers, too many and the agent saves nobody any time.

By the FocalID team in Colombo, , 2 min read

Two people reviewing notes on paper beside open laptops.

Sort actions by what a mistake costs

List everything the agent can do, then sort each action by two things: how costly a mistake would be and how easily it can be undone.

  • Let the agent act alone on cheap, reversible steps: looking up an order, drafting a reply, tagging a ticket, creating a draft record.
  • Ask for approval on steps that move money or stock, or that a customer will see as final: refunds, credit notes, discounts, confirmed orders.
  • Keep with a person anything with legal or regulatory weight: credit decisions, contract terms, filings and disciplinary matters.

Design the hand-over properly

An approval step only works if the person can decide quickly. When the agent hands over, it should show what was asked, what it found, what it proposes and why. The person should be able to approve, edit or reject in one step, in a tool they already use, such as email, WhatsApp or the system’s own task list.

A hand-over that dumps a raw conversation on someone and makes them start again is worse than no agent at all.

Tell the agent when to stop

Apart from planned approvals, the agent needs rules for stopping on its own. Useful ones are:

  • The request does not match any task it was built for.
  • Required data is missing or two sources disagree.
  • The amount is above a set limit.
  • The customer asks for a person, or is clearly upset.
  • It has tried a step twice and failed.

Widen the agent’s freedom with evidence

Start strict. In the first weeks, have a person approve most actions, and record how often the agent’s proposal is accepted unchanged. When one type of action is approved without edits almost every time, let the agent do that one alone and keep sampling it. Where people often edit the proposal, improve the agent before loosening anything.

Keep a log of every step either way. If something goes wrong, you need to see what the agent read, what it decided and who approved it.

Common questions

Does human in the loop mean a person checks everything?

No. It means a person is involved at chosen points: approvals for risky actions, and hand-overs when the agent is unsure. Routine, reversible steps run on their own.

Will approvals cancel out the time saved?

Not if the hand-over is well designed. Approving a prepared proposal takes seconds, while doing the task from scratch takes minutes.

Can approval rules change later?

Yes. They should be settings, not code, so you can tighten or loosen them as you see how the agent performs.

Related