Most AI pilots do not fail. They just never leave the pilot.
The demo goes well. The deck gets nods. Someone opens a channel for phase two. Then the calendar fills with other work, and phase two never gets a date. Months later somebody asks what happened to the AI project. The honest answer is that it is still sitting where the demo ended.
That is the pilot scar.
Most people carrying one do not blame the technology. They blame themselves for not asking harder questions at the start. By the time the demo was over, it felt too late to ask.
The questions below are the ones we think would have caught it. Run them against any vendor selling you an AI project. Ours included.
You do not need to understand the technology to use them. You need to understand your own business. And you need plain answers before you sign.
Ownership
Who owns this after the pilot ends, a name, not a department? A pilot with no named owner loses its owner the moment its champion moves on. Ask for the name before you start. Not after the invoice lands.
What happens to it if that person leaves? If the answer involves the vendor, the vendor now owns part of how your business runs. That may be fine. It should be a decision you made on purpose. Not one you find out about during someone's notice period.
The production path
What has to be true for this to run in production, not a demo? A demo runs on clean data and a friendly test case, with someone standing by to fix whatever breaks on camera. Production means your messiest customer and your worst week. Nobody is standing by. Ask exactly what changes between the two.
Who signs off before it touches real customers or real money? If nobody can name that person, the project is not ready to leave the sandbox. However good the demo looked.
What is the plan for when the model or the vendor changes underneath it? Every AI system sits on a model, and models get retired, repriced or replaced. A vendor with no answer here is asking you to bet on their supplier standing still. No supplier stands still.
Verification
What number tells you this is working, and where do you check it? Not a testimonial. Not a feeling that things run smoother. A real figure, in a place you can look at any time. If nobody can point to it, there is nothing to verify. There is only something to believe.
Can someone other than the vendor rebuild that number, or do you take it on trust? A number the vendor makes, reports and explains for you is marketing. A number your own team could rebuild from raw data is evidence. Ask which one you are being offered.
What does failure look like, and who sees it first? Every system fails sometimes. The question is where the failure shows up. On a scorecard you already read, or in a customer complaint weeks later. Ask for the failure case before you ask for the success case.
The exit
What does it cost, in time and effort, to turn this off? Unwinding the project means rebuilding whatever it replaced. If nobody wrote that down, you are carrying a cost nobody has named. Ask for it out loud, before you need it.
What do you keep if you walk away: the data, the process, the people who understand it? A good vendor leaves you with something durable. Data you can export. A process someone wrote down. People who understand how the system worked without the vendor in the room. A pilot that leaves with the vendor was never really yours.
Ask us these too
We wrote this list against ourselves as much as anyone. So here is roughly what you would hear if you put it to us.
A named person runs your function and stays reachable for as long as you are a client. Not a rotating account team.
We do not call a project finished at the demo. We call it finished when it runs on a schedule, unattended. Someone outside this company should be able to read the output and see where it came from.
Every number we report comes with the method that produced it. So you, or anyone you trust, can check our working instead of taking our word for it.
And if the fit turns out to be wrong, we say so. On the first call, or a year in. We would rather do that than keep an unhappy client on a contract.
None of that makes us the right choice for every business. It only means we can survive the same questions we are asking you to put to everyone else.
Run it before you commit
Copy the list into wherever you take meeting notes. Bring it to the next AI pitch you sit through. Ours included.
If a vendor cannot answer most of it plainly, you have learned something worth knowing. Better now than after the pilot quietly stalls.
Would you rather start on your own? The free AI readiness assessment asks seven scored questions and names your two weakest foundations. No email gate. No call.
Would you rather start with a conversation than a checklist? Talk to us. We will walk through your own answers to these questions, on the record, before we ever propose running anything. The call is free.