> ## Documentation Index
> Fetch the complete documentation index at: https://docs.adopt.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Human-in-the-loop (HITL) review

> The review gate that sits between an agent's work and anything becoming final — the single most important concept in Adopt.

If you take away one idea from this knowledge base, make it this one. Every piece of work an agent prepares passes through a human before it counts.

## What it is

Every piece of work an agent prepares passes through a **human-in-the-loop review gate** before it's treated as final. The agent prepares; a person approves. This is structural — **it is not a setting you can turn off.**

## Why it exists

Finance and accounting work carries professional and regulatory responsibility. A prepared workpaper, a filed return, a reconciliation — these need a qualified person to stand behind them. The HITL gate makes that explicit and auditable: there is always a named reviewer and a timestamp on every approval.

## How it works

<Steps>
  <Step title="A run finishes">
    An agent run completes and produces output.
  </Step>

  <Step title="It lands in the review queue">
    The output lands in the review queue for that engagement, marked **action needed**, and the assigned reviewers are notified.
  </Step>

  <Step title="A reviewer decides">
    A reviewer opens it, checks the work against its cited sources, and either **approves** it (it becomes final) or **sends it back** with notes (it returns for correction).
  </Step>

  <Step title="The decision is recorded">
    The decision, the reviewer's identity, and the time are all recorded in the audit trail.
  </Step>
</Steps>

## What a gate looks like in practice

Gates aren't only an end-of-run yes/no — an agent can pause *mid-run* at a specific decision point. In a Revenue Flux run, for example, the agent completes its variance analysis and then opens a **Controller exceptions gate**: it lists the exceptions it couldn't resolve on its own (say, an unfamiliar vendor in Data Costs) with the supporting detail, and waits. Only once a controller clears those exceptions does the agent go on to draft commentary and build the workbook. The gate is a checkpoint *inside* the workflow, not a rubber stamp at the end.

<Info>
  A gate can be conditional — always, only on low-confidence items, or only on flagged exceptions — so routine work flows through while genuinely uncertain items stop for a human. Builders configure this; the gate itself is always present.
</Info>

## Who reviews

Reviewers are assigned **per engagement** (per process), not per client. R\&D reviewers and transfer-pricing reviewers for the same client can be different people, each seeing only their own queue. This lets review responsibility follow expertise.

## For end users

Most of your time in Adopt is spent at this gate. See [Reviewing and approving work](/end-user-guide/reviewing-and-approving) for the day-to-day.

## For builders

You don't wire the gate up — it's always present. What you control is who the reviewers are, and (in Advanced setup) how routing and escalation behave. See [Instructions and skills](/builder-guide/instructions-and-skills) and [Notifications](/end-user-guide/notifications).

## Next steps

1. [Reviewing and approving work](/end-user-guide/reviewing-and-approving)
2. [The object model](/core-concepts/the-object-model)
3. [Notifications](/end-user-guide/notifications)