Five Steps to Governed AI

Chapter Six  /  Step Three

01
02
03
04
05

Calibrate

Run it against real scenarios from your firm and compare the results to what your team would have done.

If it cannot match your team on the hardest five, it is not ready for the other five hundred. The alternative is that the first error reaches a client.

Why it matters

Sample data is not your book of business.

Sample data is clean because someone cleaned it. Your book is not. Untested, the first real exception surfaces in front of a client rather than somewhere you can still change the answer.

What good looks like

Pull real cases from the past year: the exceptions, the unusual account structures, the ones that took someone an afternoon. Run the agent against them. Compare each answer to what your team actually did at the time. Use the hard ones: the easy cases will agree and tell you nothing. Where the two differ, one of them is wrong, and establishing which is the work.

85–95% agreementReadyAgreement with your team. It is ready.
Below 85%Tune itSomething in the spec or the grounding is wrong.
Above 95%InvestigateInvestigate the disagreements. It may be right where your team was inconsistent.

The point of the exercise is not the score. It is the conversation the disagreements start.

Figure 6.1

Your five hardest cases, side by side.

Real expense reports from the past year. The agent’s answer against what the controller actually did.

ScenarioAgentTeamMatch
Standard travel reimbursement$1,240$1,240
Multi-city trip, split billing$4,880$4,880
Conference with client dinner$2,310$2,310
Home office kit, policy changed mid-quarter$3,400$3,400
Client entertainment dinner, two attendees$9,400FLAGGED
AGREEMENT
80%
Below the line. Not ready.
THE ONE THAT CAUGHT IT
The agent applied the per-event limit. The controller applies the per-head limit. That rule was never in the spec, so it was never in the agent. Caught in calibration, not in a reimbursement.

What happened next

  1. The exception went into the spec. The per-head cap applies to all client entertainment. Signed by the controller.
  2. The agent was rebuilt against the amended spec. Not patched. The spec is the source, so the change starts there.
  3. The five cases were re-run. Along with the 195 that already passed.
AGREEMENT, SECOND PASS 100% The loop is the method. A failed test is the step working, not the step failing.

In your firm

The agent matched 200 expense reports to policy. Then it approved the one your controller would have caught in five seconds.

A $9,400 dinner receipt. Policy caps client entertainment at $150 a head, and the receipt listed two attendees. The agent read it as business development and passed it against the $10,000 per-event limit.

200 EXPENSE REPORTS THE AGENT MATCHED
201
What the policy says
Client entertainment: $150 a head. Nothing above VP without CFO pre-approval.
What the agent missed
Per-head math. It checked per-event instead.
What it cost
Reimbursement paid. Flagged three months later, at the next internal review.

Report 201 was never in the calibration set. That is the only reason it was in production.

Skip this step
Edge cases blow up live. The first failure is in front of a client.