What is human in the loop (HITL)?
The line between an agent that prepares an action and an agent that takes it.
Human in the loop (HITL) — Human in the loop (HITL) is a control pattern where an automated system prepares an action and a person reviews, edits or approves it before it takes effect. It is used where actions are irreversible, external or reputationally costly.
Why it matters
Not every automated action needs a person to sign off. Tagging a lead or summarizing a call carries little risk if it is wrong. But sending an email, changing a deal's stage, or deleting a contact does, because the cost of a mistake is external, visible or hard to undo. HITL is the pattern that puts a person between the model's proposal and the action for exactly that class of work.
For a sales team, this means an agent can draft a hundred follow-up emails overnight, but none of them leave the building until someone has looked at them. For an agency running several client instances, it means the same review discipline applies everywhere, regardless of which client's data the agent is touching. The pattern only works if it is implemented once, not per feature: a single queue, a single log, a single owner. In SalesCrew this is implemented as per-agent modes (off, draft, review, auto) with external sends defaulting to review.
HITL as a sequence
- 1
The agent decides an action is needed
Based on new activity, it works out what should happen next.
- 2
It prepares the action, not the effect
A drafted reply, a proposed stage change: written, but not yet live.
- 3
It is queued for review
The proposal sits in one place with the evidence behind it.
- 4
A person approves, edits or rejects it
The action only takes effect after this step, or is corrected first.
- 5
The outcome is logged
What was proposed, what was decided, and by whom, all become part of the record.
The mistake to watch for
Questions
- How is human in the loop different from human on the loop?
- Human in the loop means a person reviews an action before it takes effect, so nothing external happens without a decision first. Human on the loop means a person monitors after the fact and can intervene, but the system acts on its own by default. HITL is the stricter, slower pattern.
- Does every AI action need a human in the loop?
- No. Low-risk, internal, reversible actions, such as tagging, scoring or summarizing, are usually left to run without review, since a wrong one costs little to fix. HITL is reserved for actions that are external, hard to undo, or carry reputational cost.
- Where does human in the loop get implemented in practice?
- As a single approval queue and audit log that every agent's proposed actions pass through, not as a separate confirmation dialog bolted onto each feature. One queue is what makes review consistent and auditable across a whole system.