How to Run a 30-Day AI Pilot That Ends in a Decision

Most AI pilots do not fail. They drift. Nobody set a finish line, so the pilot is extended, then extended again, until it quietly becomes "something a few people use."

The book's 30-Day AI Pilot is built to prevent that. Each week ends with a deliverable and an exit check that must pass before the next week starts.

Before Day 1: the groundwork

Thirty days is realistic only once the groundwork exists: a map of the process, a baseline, a redesigned workflow with its human checkpoints marked, an approved tool list and data classes, and a way to measure. Build it once; later pilots inherit it.

Week 1 - Select and baseline

Pick one process and one slice of it - routine returns rather than all returns - so the cases are similar and the results mean something. Name an owner, a sponsor, and five to twelve daily users.

Then write the pilot charter and set every line before any result exists, because a line drawn after you see the numbers tends to land just below them:

• a stop rule with continue, revise, and stop lines;
• a quality and risk limit - for example, no errors passed to a customer;
• the sponsor's outcome;
• an adoption line, plus the fallback rate;
• the counted window, skipping the first two live days while people learn.

Exit check: a baseline covering a normal cycle of the work.

Week 2 - Prototype

Build the simplest design that works, then test it on real past cases before anyone uses it live. The test set should include an input with no valid answer, a case the AI cannot classify, a hostile instruction, and a forced failure that proves the fallback works.

Exit check: the test set passes and a named person - not the builder - signs the go-live.

Week 3 - Controlled use

Named users run the workflow on live work, behind the gate. Log every issue.

Exit check: no open high-risk issue.

Week 4 - Measure and decide

Compare with the baseline and apply the stop rule you wrote in Week 1. Keep hours and cash separate: recovered time is capacity, not savings, unless payroll changes, a hire is avoided, or throughput rises in a way you can measure.

Exit check: a signed decision - standardize, revise and rerun once, or stop. Extending is not a decision.

A first pilot is meant to teach

Even a pilot that clearly helps can fail its own payback test at measured cost. That is not a waste. A first pilot builds the template, the rules, and the measurement habit that every later pilot inherits.

The AI Project Charter and ROI Calculator from the book are free downloads on the Resources page.