How to Run a 30-Day AI Pilot
This piece is adapted from Chapter 9 of Copilot at Work, my free guide for teams adopting Microsoft 365 Copilot. The plan below works for any AI tool; the guide runs it specifically for Copilot, with the scripts and checklists included. Download it free.
Most AI pilots don't end. They trail off, undecided, which means the organization pays full cost and receives no verdict. This plan exists so that can't happen: one team, thirty days, and a decision at the end that someone says out loud.
It's deliberately built for one team rather than the whole organization, because the fastest way to a successful company-wide rollout is a small, undeniable win that other teams ask to copy. Evidence recruits better than mandates.
Before day one: choose the workflows
The owner (a named person with authority over the team's workflows, and if you don't have one of those, stop and get one) picks two or three specific workflows. Not "email" but "the Monday status report." Not "meetings" but "the weekly ops sync and its follow-ups." Good candidates happen at least weekly, visibly annoy the people who do them, and involve reading unstructured stuff and turning it into structured action.
Resist choosing the most valuable workflow. Choose the most provable one: frequent, measurable, low blast radius if the first week is rough. The point of round one is a clean verdict.
And do the thing everyone skips: put the day-31 decision meeting on the calendar now, before anyone knows the answer. The plan enforces itself from that one appointment.
Week one: baseline and setup
Capture the before-numbers. For each chosen workflow, count what it currently costs: hours per week of assembly, days of cycle time, number of interruptions. Precision is not required; honesty is. A simple shared sheet, filled in by the people who do the work, becomes the most important artifact of the month, because without a before, no after can prove anything.
Then set expectations in one conversation: the AI is an assistant, not an oracle; first answers are drafts; you own what you send; anything feeding a decision gets checked; share what works. Teams that hear that paragraph calibrate in a week. Teams that don't spend a month oscillating between "this is magic" and "this is useless."
Train light: a handful of prompt patterns demonstrated on the team's actual workflows, not a prompt library nobody will browse under deadline.
Week two: run both paths, watch closely
The team uses the AI on the chosen workflows daily while the old way still exists as a safety net. The owner's job is presence, not enforcement: two ten-minute check-ins, wins shared specifically, flops diagnosed out loud so failures become technique instead of verdicts.
The owner's other job this week: draft the retirement list. Based on what's actually working, write down what stops next week, with dates.
Week three: retire the old way
This is the week that separates real adoptions from tool trials. The old path closes for the chosen workflows. The report the AI now assembles is no longer built by hand; the human role becomes review and send. The meeting whose only output was status transfer gets cancelled or halved. The old template is formally done.
Retirement will surface real problems, and that's the design. Where the new way genuinely can't carry the load, the owner adjusts scope honestly rather than quietly reopening the old path for everything, because two paths is the stall condition. A team that feels nothing in week three retired nothing.
Week four: measure and decide
Rerun the week-one counts. Same sheet, same units. Then the owner writes a one-page verdict: the numbers before and after, what the team says, and the decision. There are four exits, and only three are allowed:
Expand. It worked and the numbers show it. Extend to the next workflows, and let the one-pager recruit the next team.
Fix. Real value, identifiable blockers. Adjust and run another cycle.
Kill. The workflows didn't repay the effort. Retire the experiment honestly and bank the finding. A clean kill in thirty days is dramatically cheaper than an ambiguous fade over a year.
Drift is the forbidden one: no decision, full cost, no verdict. Day 31 exists so it can't happen, and it was on the calendar before the pilot started.
All three allowed exits are wins, which is the emotional core of the whole plan: designed this way, the pilot cannot waste your money, because even its failure produces the one thing stalled pilots never deliver. An answer. The math on what the alternative costs is worth reading if you need to make that case to someone else, and the three questions to answer before you even start are the natural prequel to this plan.
The full version, with the team scripts, the checklist, and the measurement chapter that follows it, is in Copilot at Work, free for you and your team.
Rosemarie Withee has spent thirteen years helping operations teams get real work out of their software, first Microsoft 365, now AI. She’s written six books for Wiley and builds AI products at Portal Integrators.