Operations · 8 min read · Aug 12, 2026

Agree the pass mark before you build.

Evaluation is the least glamorous part of an AI build and the one that most reliably decides whether it reaches production. A pilot without an agreed pass mark cannot conclude — it can only continue, and eventually be abandoned for reasons nobody can quite articulate.

Why pilots run forever

A pilot with no defined success criterion has no state in which it is finished. Output is reviewed, opinions are formed, some are positive and some are not, and the decision defaults to whoever is most senior or most tired.

The fix is unglamorous and it has to happen before the build: write down what the system must get right, on which cases, at what rate, agreed by the people who will judge it. That single artefact is what allows a pilot to conclude in either direction — which is the point, because a pilot that cannot fail was never an evaluation.

If nobody can state the number that would make this a success, the project has no way to end. Agree it while it is still cheap to disagree.

Where test cases come from

Invented test cases produce misleading confidence, because people invent tidy examples. Real evaluation sets come from real history, and a good one has four ingredients:

  • Ordinary cases. The routine bulk. Establishes baseline competence and is the easiest to gather.
  • Historical exceptions. Cases that previously required a human decision — this is where systems actually fail.
  • Known-hard cases. The ones experienced staff can name from memory. Ask them; they will produce a list in minutes.
  • Adversarial cases. Malformed input, contradictory data, missing fields, and content designed to mislead.
Four sources of evaluation cases: ordinary cases for baseline competence, historical exceptions where systems actually fail, known-hard cases that experienced staff can name, and adversarial cases with malformed or contradictory input.
Four sources. The second carries most of the value and is most often missing.

The second category carries most of the value and is the most commonly missing. A system evaluated only on routine cases will pass evaluation and fail deployment, because production is where the exceptions live. If you take one thing from this: go and get the exceptions.

Measure what failure costs, not just how often

A single accuracy percentage hides the thing that matters. Two systems at 92% can be very differently deployable, because the shape of the 8% differs.

QuestionWhy it matters more than accuracy
Which way does it fail?Missing something is not the same as inventing something
Does it know it is unsure?A flagged low-confidence case is a handled case
What does one failure cost?Decides how much review the process needs
Is failure detectable downstream?Silent errors are the expensive kind
Four questions that matter more than an accuracy percentage: which direction the system fails in, whether it signals its own uncertainty, what a single failure costs, and whether failure is detectable downstream.
Four questions a single accuracy percentage hides.

The second row is the most useful property to design for. A system that is right 85% of the time and reliably flags its uncertainty is more deployable than one that is right 95% of the time with uniform confidence — because the first can be operated with a review queue and the second cannot be operated at all without reviewing everything.

The human baseline nobody measures

Evaluations are usually run against an implicit assumption that the current manual process is correct. It is not, and comparing against a perfect standard that does not exist sets an unreachable bar.

Have people redo a sample of historical cases without seeing the original answers. You will find disagreement between them, and disagreement between the same person on different days. That number is the honest benchmark, and it is frequently lower than everyone expects. It also identifies genuinely ambiguous cases — which should be routed to a human by design rather than counted as model failures.

Evaluation is not a phase

The most common structural mistake is treating evaluation as a gate before launch rather than a permanent fixture. Model providers update models. Behaviour changes on upgrade, including on changes described as improvements. A system validated against the previous version has not been validated.

Keep the evaluation set runnable on demand and run it after every model change, every prompt change and every retrieval change, plus on a schedule. It should be the cheapest thing in the project to execute — if running it is a two-day manual exercise, it will not be run, and the system will drift unobserved.

The same set doubles as a regression suite. Fixing one failure mode frequently introduces another, and without a stored set of cases that were previously passing, nobody notices until it reaches a customer.

Dealing with this in your own group?

We answer scoping questions before there's a contract in sight — including the ones about cost and data handling.

Questions

Short answers,
in full.

The questions this article gets asked most, answered so each one stands on its own.

Talk to us
Why do AI pilots never reach a decision?

Because no pass mark was agreed before the build, so there is no state in which the pilot is finished. Output gets reviewed, opinions form, and the decision defaults to whoever is most senior or most tired. Writing down what the system must get right, on which cases, at what rate — agreed by whoever will judge it — is what lets a pilot conclude in either direction.

Where should AI test cases come from?

Real history, not invention — people invent tidy examples. A good set mixes ordinary cases for baseline competence, historical exceptions that previously needed a human decision, known-hard cases that experienced staff can name from memory, and adversarial cases with malformed or contradictory input. The exceptions carry most of the value and are the most commonly missing.

Is accuracy the right metric for an AI system?

On its own, no. Two systems at the same accuracy can be very differently deployable depending on which way they fail, whether they signal uncertainty, what a single failure costs, and whether failure is detectable downstream. A system that is right 85% of the time and reliably flags uncertainty is more operable than one right 95% of the time with uniform confidence.

How often should AI systems be re-evaluated?

After every model, prompt or retrieval change, plus on a schedule. Providers update models and behaviour shifts on upgrade, including on changes described as improvements — a system validated against the previous version has not been validated. Keep the evaluation set cheap to run; if it takes two days of manual work it will not be run and the system will drift unobserved.