"What is an AI pilot?" Owners ask it when a vendor proposes one, when a board member says the company should run one, or when a manager asks for permission to try a tool and nobody knows what to answer. The word is used for everything from an afternoon of experimenting to a six-month program with a steering committee, which is why it is hard to say yes or no to. This page gives it one meaning an owner can use and a short list of what to decide before the first day.

I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them. I have watched a great many pilots, of systems and of processes, and the ones that were worth running had one thing in common: someone had written down, before the start, what would count as success and who would decide. I am a lawyer by training, which makes me insist on that page. Now, at 013 Labs, I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies, and the pilot is where most of those plans begin.

The short answer

An AI pilot is a small, time-boxed trial of one tool on one real job, with a written test of success and a named person who decides at the end whether it stops, changes or grows. Each part matters. Small means a few people and a limited slice of the work, not the whole company. Time-boxed means a date on which it ends, whatever has happened. One real job means actual work from your business, not a demonstration with sample material. The written test means you can say in one sentence, before the start, what would make you keep going. And the named decider means the end of the pilot is a decision, not a drift.

A pilot is not a license to try everything, a purchase in disguise, or a way of postponing a choice. If it has no end date and no test, it is not a pilot. It is a tool that has quietly arrived.

What a pilot is not

It is not an experiment on your own time. People trying a tool on a Sunday is useful, and the page on shadow AI explains how to find out what is already being tried. A pilot is the version the company chose, with a boundary around it.

It is not a demonstration. A vendor's demonstration shows the tool at its best, on material chosen to flatter it. A pilot runs the tool on the messy work your people actually do, on a normal Tuesday.

It is not a rollout with a nicer name. If the plan is to switch the tool on for everyone and see what happens, that is a rollout. A pilot can fail cheaply because it was small. A rollout that fails has already touched every customer.

It is not a project. A project builds something. A pilot answers a question: does this tool, on this job, with these people, do what we hoped, at a cost and risk we can accept? Everything in the pilot exists to answer that question honestly.

Why pilots stall

When owners and executives talk about pilots that went nowhere, the causes are rarely about the technology. They are about the pilot.

Nobody owned it. A pilot with a vendor champion and an enthusiastic manager, but no named person who answers for it, ends when the enthusiasm does.

There was no test. Without a sentence written beforehand, every result can be read as a success by the people who like the tool and a failure by the people who do not, and the argument never ends.

The job was wrong. A pilot on a task nobody does often, or on one where a mistake costs a customer, teaches little. The page on what an AI use case is explains how to choose a job worth testing, and the page on picking your first AI project has five tests for it.

It never ended. The date passed, the trial continued by default, and a year later the tool is in daily use with no decision ever having been made.

It ended without a decision. The pilot closed with a slide deck and a round of thanks. Nobody said stop, change or grow, so nothing followed.

The three decisions before day one

The same three decisions that sit under every AI plan sit under every pilot, only smaller.

Who owns it. One named person, with a half-page written scope: what they may approve alone, what comes to the owner, and the date they report. Usually it is the manager of the team whose job is being tested, not the IT provider. The page on who owns AI in a company is about choosing that person.

What information stays out. One sentence the pilot team can repeat. For most companies: customer lists, prices, payroll, anything about an employee, drawings and recipes, and anything under a non-disclosure agreement do not go into a tool the company has not approved, on an account it does not hold. A pilot is the time when people are most eager and least careful, so the sentence matters most here. The page on whether free tools are safe for business use covers the account and data side.

What the tool is never allowed to do alone. In a pilot this list is short and firm: nothing leaves the company, changes a figure or touches a customer until a named person has looked. The page on human in the loop is the long version of that check.

If you cannot write the three decisions in a few lines each, the company is not ready to start. That is a useful thing to learn before spending anything.

The one page of a pilot

A pilot needs a single page, kept where the owner can find it. Six lines is enough.

The job. One sentence: the work being tested, who does it today, and how often.

The tool and the account. Which tool, on whose account, with which settings. Company accounts only, never someone's personal login.

The people. The few who use it, the one who owns it, the one who decides.

The test of success. One sentence, written now, in words an outsider could check. Choose something you can see with your own eyes: the first draft is ready the same day instead of the next, the weekly report no longer needs an evening, the person who checks finds the output usable with only small changes. Decide in advance what you will count and how you will count it, and write the figure in your own words, not a vendor's.

The end date. A day on the calendar, usually a few weeks away, and not moved except by the decider in writing.

The decision. At the end the decider chooses one of three: stop, change or grow. Stop means the tool does not stay and the lesson is written down. Change means one thing is altered and the trial runs again for a short, fixed time. Grow means the tool moves to a larger slice of the job, with the same checks.

Check yourself

Seven yes or no questions. Each no is a decision still to make.

Mostly no? That is the normal starting point for an owner-led company. The AI readiness scorecard asks ten questions like these across the whole business and shows which of the three decisions is missing.

How long, how big, how much

An owner's first questions are usually about size, and the honest answers are short.

How long: long enough to see a normal cycle of the job, and no longer. If the job repeats daily, a few weeks shows it. If it happens monthly, a few weeks shows one cycle only, and the pilot should be about a job that repeats more often. A pilot that needs a year to show anything is the wrong pilot.

How big: a few people, one team, one slice of the work. The aim is a result you can believe, not coverage. A small pilot done carefully tells you more than a large one done loosely.

How much: the cost is mostly the time of the people who use it and the person who checks, plus whatever the tool costs. Count the time. The price of the tool is rarely the larger number, and a pilot that looks free because the tool is free has simply hidden the cost in someone's week.

What the end looks like

At the end, three sentences should be possible. What we tested. What we saw. What we decided.

If the answer is stop, say so and write down why. A stopped pilot that taught the company something is a good outcome, and an owner who rewards that honesty gets honest pilots afterwards.

If the answer is change, name the single change and set a short new date. Changing five things at once teaches nothing.

If the answer is grow, grow by one step: the next team, or the next slice of the job, with the same checks and the same page. Growth is a new pilot with a larger boundary, not a rollout.

Whatever the answer, update the list of tools and who owns each. A pilot that ends without that update has left a tool in the building nobody is looking after.

Picture it going wrong

A made-up example, to show the shape. An eighty-person regional building-products wholesaler lets its customer service team try a tool for drafting replies to order enquiries. Nobody writes anything down. The vendor's contact calls it a pilot, the team calls it a trial, the owner assumes it is a test. It runs for four months. Replies are faster, and people like it. Then a customer receives a reply promising a delivery date the warehouse had never agreed to, and the owner asks what the tool has been doing and for how long. Nobody can say who owned it, what was fed into it, what was supposed to happen at the end, or whether the faster replies were worth anything.

Nothing in that story is a technology failure. The tool did what tools do. The failure is that the pilot had no owner, no test, no end date and no check before a reply left the building. With one page written, the trial runs for three weeks with one team, a named supervisor reads every reply that mentions a date, the result is judged against a sentence written on the first day, and on the last day the owner decides: change one thing, run two more weeks, and then grow.

Mistakes that repeat

Frequently asked questions

What is an AI pilot?

A small, time-boxed trial of one tool on one real job, with a written test of success and a named person who decides at the end whether to stop, change or grow. It answers a question about a tool in your business. It is not a rollout, a demonstration or an open-ended experiment.

Do I need an AI pilot before I use AI?

Not for every use. A tool people already use for low-stakes work, with the data rule and a named owner in place, may not need a formal pilot. A pilot earns its place when the job touches customers, money or records, when the tool costs real money, or when more than a few people would use it. If a vendor says you need one to buy, ask what question it answers.

How long should an AI pilot run?

Long enough to see a normal cycle of the job, usually a few weeks, with the end date written on the first day. Extend it only by a written decision from the person who decides, and for a stated reason, never by drift.

Who should run an AI pilot in a mid-sized company?

The manager of the team whose job is being tested, with the owner as the decider or a named executive in their place. Not the IT provider alone and not the vendor. The person who runs it must be the one who feels the result in their week and has the standing to stop it.

Should the vendor run the pilot for us?

A vendor can support it, set up the account and answer questions. The choice of job, the test of success and the decision at the end belong to you, because a vendor who writes the test usually passes it. Keep your own page, in your own words.

What should be in the success test?

One sentence, written before the start, that an outsider could check. Something observable: how soon the first draft is ready, how much the person who checks has to change, whether a weekly task still needs an evening. Add the cost of the checking time. If you cannot write the sentence, the job is not ready to pilot.

What happens after a pilot succeeds?

Grow by one step, with the same checks, and update the written list of tools and owners. A successful pilot becomes a normal part of how the job is done only when the owner, the data rule and the check come with it. The page on whether your insurer will ask about AI is a good reason to have those on paper by then.

Not sure where you stand? Take the 3-minute AI readiness scorecard.