Chapter 22 of 36 · ~1 min

Human-in-the-Loop

Some decisions should not be automated, because the stakes are high, because the rules are not yet known, or because a person is accountable for them. Human-in-the-loop design puts a person at a chosen point in the process: to approve an action, to correct a draft, to resolve an ambiguity, or to take over entirely.

The design question is not whether to include people but when the agent should ask rather than proceed. Ask too often and you have built a slow form. Ask too rarely and you have built a liability. The answer is usually different for different steps, and it should change as trust is earned or lost.

A human checkpointModel draftsa reply, an actionPerson decidesapprove, or send backAction runsonly on this pathFeedbackwhat should changeapprovednot yetthe model revisesThe decision is recorded next to the draft, which is what makes the approval auditable later.
The consequential action exists only on the approved path.

Experiment

Live model

The model drafts a reply. Nothing is sent until you approve, or send it back with a note.

Runs against a live model through Chatterfly's server. Your text is sent to the model provider and not stored. Nothing is sent anywhere; the point is that the send step cannot run without your decision.

Big question

Which decisions in your process must a named person own?