When Is an AI Impact Assessment Actually Finished?

Someone asks if Claude can just finish their AI Impact Assessment for them. It's the wrong expectation. Claude can start one. Finishing it takes someone checking the draft against the real system, not just against the description they gave Claude.

Why a polished draft isn't the same as an accurate one

It's tempting to treat a polished-looking assessment as an accurate one. Say someone describes a hiring tool as "matches resumes to keywords," when the real system also scores tone in video interviews. Claude drafts the risk section beautifully from that description, fluent, complete-reading, professionally organized. But the video-scoring risk simply isn't in it, because Claude only ever saw the description it was given. It never saw the actual system.

The three things a working assessment needs

A working AI Impact Assessment needs three things right: what the system actually does, the data feeding it, and who it affects. Claude can write all three sections fluently from whatever information it's handed. What it has no way to do is check that description against the real system on its own.

Watching the stamp change

In the hiring-tool example, the assessment page gets its system and data sections filled in, but the document is stamped "unverified," because the description behind it left out the video scoring entirely. The stamp only changes once the missing fact is added: the data section gets rewritten to include the video scoring, and the assessment moves from "unverified" to "verified," checked against what the system actually does.

Fluent doesn't mean accurate, and rough doesn't mean failed

A fluent, complete-reading assessment doesn't prove it's accurate. Polish is a writing quality, not a fact check. The reverse also holds: a rough, half-filled first draft isn't a failure of the process either. A scaffold that flags exactly what still needs checking is doing its job. An AI Impact Assessment isn't finished when Claude writes it. It's finished when someone checks it against what the system actually does.

Try it yourself

Describe one AI-powered feature you actually use or are building, in two or three sentences. Ask Claude to draft the System and Data sections of an AI Impact Assessment from that description alone. Then add one detail you left out on purpose, a data source, an edge case, a downstream use, and ask it to revise. Compare the two drafts side by side. The second draft is better only because you supplied a fact the first one never had, and that gap is the whole point of the exercise.

Key takeaways

• Claude can draft an AI Impact Assessment fluently, but it can only work from the description it's given, not the actual system.
• A complete-reading, well-written draft is not the same as an accurate one; polish and accuracy are separate qualities.
• The three required components are what the system does, the data feeding it, and who it affects.
• A document should stay marked "unverified" until someone checks the draft against the real system and adds missing facts.
• A rough or incomplete first draft isn't a failure; it's the scaffold correctly flagging what still needs verification.

Who this is for

Teams responsible for AI governance or compliance documentation who use Claude to draft impact assessments, and want a clear standard for when a draft can actually be marked verified rather than just complete-looking.