Skip to main content

Self-assessment · two minutes

Is your AI checkable?

Ten plain questions about an AI system you run, or are about to buy. Answer honestly and you get a score out of 20 and the three places to start.

It runs in your browser. Your answers are not stored, sent or tracked.

  1. 01Is there a written statement of what the AI is for, and what it must not do?

    One page anyone in the team can read. Without it, nobody can say whether it works.

  2. 02Do you have a set of real past cases with known right answers to test it against?

    Even 50 to 200 of your own examples beats any amount of demo.

  3. 03Were pass marks agreed before the system was built or bought?

    For example: flags at least 95% of the invoices your team would reject.

  4. 04Do the tests run again every time the prompt, model or data changes?

    Models and prompts drift. A test that ran once at launch is a memory, not a check.

  5. 05Can you see where each answer came from: the source passage, record or rule?

    If staff cannot check an answer in seconds, they either trust it blindly or ignore it.

  6. 06Does a named person approve any decision that matters before it takes effect?

    Payments, emails to customers, anything about a person. The AI drafts; a person decides.

  7. 07Is there a record of what the AI did, when, and who approved it?

    An audit trail you can read afterwards, not just application logs.

  8. 08Is there a cap on what it can spend or send, and does it stop when the cap is hit?

    Usage caps and send limits stop one bad loop becoming an expensive day.

  9. 09Do you know how it fails, with examples, not just how often it succeeds?

    The wrong answers are the useful part. Someone should have read them.

  10. 10Could someone outside the team re-run your headline accuracy figure and get the same result?

    If the number cannot be reproduced, it is a claim, not a result.

Progress

0 of 10 answered

Your score appears here when all ten are answered.

How we make our own builds checkable: tests agreed in week 1. Investing in an AI company? Use the AI due diligence checklist.

    We use cookies.

    Is your AI checkable? A ten-question self-assessment · vocabotics