Short answer
Human-in-the-loop is a design pattern where an AI system prepares work and a person approves, edits or rejects it before it takes effect: drafted replies queued for a one-click send, extracted invoice fields shown for confirmation. Review is placed where a mistake would be expensive, rather than on every output the system produces.
Also called: HITL, Human review, Human oversight
The design question is which decisions to gate, and reversibility is a better test than importance. Sending an internal summary to the wrong channel is embarrassing and fixable. Issuing a refund, quoting a price, or emailing a client on the company’s behalf is not easily undone, so those belong behind an approval. Tiering this way keeps the volume of review small enough that people actually read what they approve.
The pattern fails when approval becomes a formality. A queue of two hundred items with an approve-all button trains reviewers to click without reading, which is worse than no review because it manufactures a record of oversight that never happened. Showing the source passage beside the answer, surfacing only the uncertain cases, and keeping the queue short are what keep the human step meaningful. Coverage should then be reduced deliberately, as measured accuracy justifies it, rather than eroding through fatigue.
Common questions
Does human review cancel out the time saved?
Not usually, because checking is faster than producing. Reading a drafted reply and pressing send takes a fraction of the time it takes to write one from scratch, and confirming four extracted fields is quicker than typing twelve. The saving shrinks only when everything is reviewed with equal care regardless of risk.
Where this comes up in our work
Related terms
AI Agent
An AI agent is a system that uses a language model to decide which actions to take, then takes them through connected tools: reading a database, sending an email, updating a booking record.
Guardrails (AI)
Guardrails are the controls that constrain what an AI system may say and do: topic limits, validation of output against an expected structure, blocked actions, spending caps, and filters on both the incoming request and the generated response.
Hallucination (AI)
A hallucination is output that is fluent, plausible and false: an invented citation, a policy the company never wrote, a confidently wrong figure.
Workflow Automation
Workflow automation is software that runs a defined sequence of business steps without a person moving data between systems: a submitted form creates a record, notifies the right team, files the attachment and schedules a follow-up.
Reading definitions because you are scoping a project? Skip ahead and just ask.