Human review of AI agent work: a process that leaves proof

Why review needs a process

A 2026 field study of 802 developers found AI doubled the changes per engineer and also doubled the review load per reviewer, with AI-written changes taking about 20 % longer from first review to merge. Review is where AI productivity goes to be taxed, and it is the part most teams don't measure or even assign.

Clients are asking too. In a 2026 survey of professional services, 85 % of clients said firms should disclose AI use on their matters, and few firms do. A record of human review is the strongest answer.

Three roles

• Requester: the person who hands the work to the agent and writes the brief. Stays responsible on the task.
• Reviewer: the person who checks the result. Often the requester, sometimes a senior for client-facing work.
• Approver: the person who moves the task to done. For client work, never the agent itself.

A review process in six steps

1. Put every piece of agent work on a task, with acceptance criteria in the description or checklist.
2. Hand it to the agent. Watch the session, answer its questions and stop it if it drifts.
3. When the agent finishes, the task goes to In review, not Done.
4. The reviewer checks it against the criteria, tracks their review time on the task and either sends it back or approves it.
5. Only a person moves the task to Done. Sent-back tasks count as rework.
6. At the end of the period, report agent hours, review hours, rework and approvals.

What to keep as evidence

• The request, the agent's progress notes and its result, on the task and its thread
• Who moved the task to done, and when
• The reviewer's time on the task
• How often the task was sent back
• A report per client and period, frozen when you share it

How Hourtick supports it

Agent outcomes shows, for each finished task, the agent's time and cost, the people's briefing and review time, the number of times it was sent back and who moved it to done, a person or an agent. An AI work report turns this into a read-only page for a client: people's and agents' hours, which agents and models, and for each task whether a person reviewed or approved it. It has no rates, costs, margins or notes, is frozen when shared with a fingerprint of its contents, and can be revoked.

Agents can also be limited so review is realistic: scope to chosen projects, read-only mode, and monthly budgets so work doesn't pile up faster than people can check it. Hourtick measures tasks and agents and doesn't score people.