Nasr Khan

12 AI

Agent reliability check

Our AI agent passed most of its tests. Can we launch it, and what will still go wrong?

For CEOs, product leaders and operations heads signing off an agent. Takes about three minutes. Opens with an example.

Your test results

1What you tested

Picking a job fills a typical number of steps. Change it to match your agent.

Each test checks
2In production
$

Refunds, rework, support time and goodwill, per failed task.

3The human check

Seeded test of the reviewers

Mix outputs with known errors into a normal review batch. Count what reviewers catch. Leave the first box at 0 if you haven't run it yet.

What reviewers see

The answer

Not yet

How sure you can be

Where the chain breaks

What the human check catches, each month

Task success as tasks get longer

Best estimateCautious and hopeful ends of the range

What to do next

    Take the result with you

    A PDF report with your inputs, results, charts and method. Or the same as plain text for an email.

    How this tool works

    What it does

    A test pass rate is one number from a small sample. This tool shows the range the real failure rate could sit in, how failures add up when a task has several steps, and how many failures reach customers after your human check.

    It also tests the human check itself. A person in the loop only helps if they have time to look and can spot the problem.

    What you enter

    • Test results. Tests run and passed, whether each test covers a step or a whole task, and steps per task.
    • Production. Tasks a month, the cost of a failure that reaches a customer, and the most failures you can accept.
    • The human check. How many tasks a person checks, how long a check takes, reviewer hours, and a seeded test: known errors planted in a review batch.

    What you get

    • A launch stamp: ready, launch with checks, or not yet.
    • The failure range at the confidence you pick, for steps and for whole tasks.
    • Failures reaching customers each month after the human check, and what they cost.
    • Tests still needed to prove your limit, and a PDF for the launch review.

    The maths, assumptions and sources

    The failure range uses the Wilson score interval, which behaves well for small samples and for rates near zero. For step tests, task success = (1 − step failure rate) ^ steps, assuming steps fail independently. Errors that reach customers = task failure rate × (1 − share checked × share reviewers catch). The share checked falls if reviewers don't have the hours to check every task assigned to them. Without a seeded test, the tool gives the human check no credit, because nobody has measured it.

    Tests needed: the smallest number of tests that would bring the cautious end of the range under your limit, if the agent keeps failing at today's rate.

    • Wilson, E. B. (1927). Probable inference, the law of succession, and statistical inference. Journal of the American Statistical Association, 22(158), 209–212.
    • Gartner (June 2025) expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing costs, unclear value and weak risk controls.
    • AWS, Well-Architected Agentic AI Lens: approval steps fail when reviewers see too much, lack context, or aren't available.
    See answer