Implementation
How to Pilot AI Without Betting the Company
Most AI projects in small and mid-size businesses fail the same way: too big, too vague, and nobody agreed in advance what success looks like. Having run operating businesses for 30 years, I'd never adopt a tool any other way than the way I'd try a new foreman — on one job first, watched closely, judged by results. Here's that approach, formalized.
Step 1: One slice, not the business
Pick a single, bounded piece of the operation: after-hours calls only. One intake queue. One building's resident inquiries. One estimator's document review. The slice should be small enough that a failure costs you nothing but the pilot, and real enough that a win is undeniable. "Transform our operations" is not a pilot; "answer every call after 5pm" is.
Step 2: Capture the baseline first
Before anything goes live, measure the slice as it runs today: calls answered, response times, hours spent, jobs booked — whatever number the pilot is supposed to move. Thirty days of baseline is ideal; even two weeks beats none. Without a baseline, the post-pilot conversation is opinions. With one, it's arithmetic.
Step 3: Install a kill switch
Decide, in writing, how the pilot gets turned off and what happens when it is. For an AI phone agent, that's call forwarding back to the old setup — a five-minute change. For document automation, it's the manual process everyone still remembers. Knowing you can revert in minutes is what lets your team relax enough to give the pilot a fair run. It also keeps any vendor honest: if they can't tell you the rollback plan, walk.
Step 4: Set the number that decides — in advance
Agree before launch on the single metric and threshold that means "expand" versus "stop": booked jobs up 15%. Response time under one minute. Five office hours a week recovered. Write it down with a date. When the date arrives, the number decides — not the demo, not the salesperson, not sunk cost. Some pilots should die; a framework that can't kill a pilot can't be trusted when it says expand.
The 60-day shape: weeks 1–2, baseline and setup. Weeks 3–4, live on the slice with daily spot-checks. Weeks 5–8, run clean and let the data accumulate. Day 60, the number decides. Total exposure: one slice of one process, fully reversible, for two months.
Run AI this way and the worst case is a cheap, fast "no" backed by your own data. The best case is a proven win you can expand with confidence — because you already know exactly what it's worth.