AI agents changed how fast code gets written. They did not change who is accountable for it. When an auditor asks who approved a change, “the pipeline passed” is not an answer.
The short answer
AI-built changes need two human gates: approval of the spec before work starts, and acceptance of the outcome before it ships. Each gate is signed by a named person, and the record shows who approved what and when, so an auditor can follow it later.
Where approval goes wrong
Most teams already have approvals: a pull-request review, a change ticket, a release sign-off. With agents in the loop, three things go wrong with them.
- Approval comes too late. The first time a person sees the work is the diff. By then, the only choices are to accept a wrong approach or throw away the run.
- Approval is anonymous. A green check from “the team” or a bot does not tell anyone who decided.
- Approval checks the code, not the outcome. A reviewer can confirm the code is clean without knowing whether it does what the business asked.
Gate 1: before anything is built
The first gate is the spec. Before an agent writes a line, a named person approves what will be built: the requirements, the acceptance criteria and the repositories it touches.
This is the cheapest place to catch a mistake. Changing a sentence in a spec costs minutes. Changing the same decision after three agents have built to it costs days.
Gate 2: before anything ships
The second gate is the outcome. The person who owns the business result, not only the engineer who reviewed the code, accepts that the change does what the spec said, with the evidence in front of them: which acceptance criteria passed, and how.
Gate 1 asks “is this what we want?” Gate 2 asks “is this what we got?”
What makes an approval hold up a year later
| Weak and strong approvals | Weak approval | Strong approval |
|---|---|---|
| Who | A team, a bot or a shared account | A named person with the authority to decide |
| What | A diff or a ticket title | A specific version of the spec, or a specific outcome with evidence |
| When | Somewhere in a chat thread | A timestamp on the record |
| Why | Not recorded | The decision and the alternatives it ruled out |
None of this needs a heavy process. It needs the approval to happen in the same place as the work, so the record writes itself.
Your release process still decides
Gates do not replace your pipeline. Code still lands in your repository, passes your checks and ships through your own release process. The gates add the two decisions a pipeline cannot make: what to build, and whether it was built.
In CrewPE
You approve the spec before anything is built. Your business owner accepts the same behavior before anything ships. Every lock records who and when.
See how CrewPE keeps you in controlQuestions teams ask
- What are approval gates for AI-built software?
- Points where a named person must approve before work continues. Two are enough for most changes: approval of the spec before work starts, and acceptance of the outcome before it ships.
- Is a pull-request review enough?
- It checks the code, which still matters. It comes too late to fix a wrong approach cheaply, and it does not confirm the business outcome. A spec gate before and an outcome gate after cover both gaps.
- What should an approval record contain for audit?
- Who approved, in their own name; exactly what they approved, such as a spec version or an outcome with its evidence; when; and the reasoning or alternatives considered for significant decisions.
Koushik Ramani
Co-founder of EnWithAI, working on how AI-built changes get approved, governed and accepted.