30–90 day AI pilot: how a public institution can move from an idea to measurable evidence

An AI pilot is not a demonstration designed to prove that the technology works.

It should test whether a specific solution improves a specific problem for specific users in a specific context.

The first step is to formulate a hypothesis.

For example: “A document assistant grounded in our approved corpus can reduce average procedure-search time without increasing the error rate beyond the accepted threshold.” This changes the logic of the project.

It forces the institution to define a baseline, a user group, a corpus, quality criteria and an expected decision.

A 30–90 day pilot can follow four periods.

Day 0–15: scope.

Define the problem, users, sponsor, perimeter, authorized sources, risks, baseline and success criteria.

Day 15–30: build and test.

Configure the prototype, establish access controls, prepare a test set, verify output quality and train initial users.

Day 31–60: operate in controlled conditions.

Collect metrics, track errors, observe adoption, document incidents and correct what needs to be corrected.

Day 61–90: evaluate and decide.

Compare results with the baseline, analyze value, risk, cost and maintainability, then take an explicit decision.

The final decision may be: stop, redesign and retest, extend, scale gradually or prepare broader deployment.

The pilot should also create durable assets: an Evidence Pack, error register, test set, user documentation, decisions and recommendations.

Without these, the institution learns very little and may repeat the same experiment a few months later.

The most important discipline is to avoid changing success criteria during the pilot simply to make the project look successful.

If the expected value is not demonstrated, the report should say so.

A useful pilot may end in a stop decision.

BEXTIO considers an experiment successful when it supports a better-informed decision, not only when it leads to deployment.