How to review AI-written changes without reading the diff

On many teams the PM's review of an AI-written change is a thumbs-up on a pull request they didn't really read. That isn't review. Here is a loop that lets a product owner judge a change by evidence instead of code.

1. Start from an outcome, not a task

Write each piece of work as an outcome script, one issue with four parts:

• Outcome: what a user can do afterwards, in one or two sentences.
• Personas and states: which users matter (admin, new user, suspended tenant).
• Acceptance journeys: three to six steps per persona that prove the outcome.
• Out of scope: what the agent must not touch.

The agent turns the journeys into tests before it writes code. Your review then starts with a question you can answer: did the journeys pass?

2. Ask for evidence first

A good pull request description leads with a plain-English summary of what a user can now do, then evidence: test results, a browser trace, screenshots of the app with realistic data. Technical notes come last. If the evidence is missing, send it back before anyone reads code.

3. Try it as the people it's for

Open the live preview and log in as each persona in the outcome script. Does the suspended tenant see a clear message or a crash? Does a new user understand the empty state? This is where realistic seeded data matters; an empty preview passes everything.

4. Pin feedback where you see it

Instead of describing a problem in chat, pin a comment on the exact element, while logged in as the persona who saw it. The pin records the page, the element and the persona, which is everything an agent needs to fix it.

5. Close the loop

When you're done, send the pins back to the agent in one batch. It fixes them, redeploys the preview and reports what it changed. Aim for most pins fixed in a single round; if they aren't, the outcome script probably needs to be clearer.

Merging stays human. The agent proposes, the evidence convinces, a person decides.

Where Adaptl fits

Adaptl's Outcome Workspace gives each change its outcome script, a Living Preview seeded with your personas, and the audit overlay. Hitting "Complete audit" bundles your pins and notes, checks each one against your design tokens, architecture boundaries and critical journeys, and fixes them in the workspace. The tech lead then signs off on a plain-English impact card instead of the diff, and the merged change becomes an entry in your product lineage.