Running AI in production is everything that happens after the demo works: knowing what the system really gets right, deciding which work it may do alone, sending the rest to people who actually check it, and noticing when quality slips. The work is mostly measurement, and the measurement has to be independent of the system being measured.
The core method starts with refusing proxies. An approval rate, a thumbs-up count or an agent's own "done" says how a process behaved, not whether the output was right. With a 99% approval rate, a blind sample of 200 approved items can still find 17 wrong, about 1 in 12. So the first number I want is a blind audit, reported as counts with a range rather than a bare percentage.
From there, confidence has to be earned. A score built from verifiable evidence, checked against a reliability table, decides what is automated and what goes to a person, with a threshold set for each action type from the harm a missed error causes. Disagreement between model families is a hand-off signal. Reviewers are part of the system too, so they are measured: review screens make people choose an answer instead of approving one, the proposed label is hidden on a slice of items, and seeded known-wrong items give each reviewer a catch rate with its range.
The same discipline carries into release and operation. Automation is gated on the lower bound of a range, not a lucky week. Zero-failure claims are sized honestly: 73, 189 or 381 clean items for 95%, 98% or 99%. Releases are gated on evidence that the tests actually ran, and every rate on a dashboard carries its width.
The essays below cover each part of that loop: what an approval number hides, how to put a range on a rate, how to pick the metric that matches what a mistake does next, how to design the review screen and when an agent should hand off.
Start here
- A 99 percent approval rate is not an accuracy number
- Design the evidence review screen before automating the decision
- A rate is a belief with a width
- Counting honestly: ranges, zero failures, paired changes
- Pick the metric by what a mistake does next
- A lucky week is not a better model
- The release gate is the product
- Design agent supervision around delivery, intervention and bounded work
- Separate probabilistic development from deterministic shipping
Questions this topic answers
- What does a human approval rate actually measure, and what should replace it?
- When should an agent stop and pass work to a person?
- How do you know whether human reviewers catch errors?
- How many clean results do you need before letting an action run unreviewed?
- How do you keep a production AI system from drifting quietly after launch?
If you are running AI systems past the demo stage, connect with me on LinkedIn to compare notes and discuss.