Most AI pilots die the same way: they get built by someone who understands the model but not the workflow, handed to someone who understands the workflow but not the model, and nobody owns getting it from "works in a demo" to "runs on live data every day."
The pilots that ship are the ones built next to the person who will actually use them — the acquisitions lead, the analyst, the ops manager — not built in isolation and thrown over the wall. That's slower on day one and much faster by week six, because there's no re-education phase where the tool has to be explained to the person who was supposed to want it.
If a pilot has been "almost done" for more than a month, the problem usually isn't the model. It's that nobody who understands the day-to-day workflow has been in the room since kickoff.
The handoff gap
A pilot has two halves. One is the model: prompts, data pipelines, evaluation. The other is the workflow: who uses the output, when, in what format, and what they do when it is wrong. Most teams staff the first half and assume the second will sort itself out.
It doesn’t. The builder ships something that works on the sample data. The intended user tries it on a real case, finds it misses a step they consider obvious, and goes back to the spreadsheet. Nobody is at fault, and nobody owns the fix.
Five signs a pilot is stalling
- It has been almost done for more than a month.
- It runs on sample data and has never run on live data.
- The person it was built for has not used it in the last two weeks.
- Nobody can say which number it is supposed to move, or what that number was before the pilot started.
- Updates come from the builder and never from the user.
One of these is a warning. Three means the pilot will not ship without a change in who is in the room.
What the pilots that ship have in common
- 01One owner on the business side.A named person whose own work gets better when the system works, and who can decide how it should behave.
- 02Built next to the user.The builder watches the real workflow and the user sees early versions often. Surprises show up in week two instead of week ten.
- 03A definition of done that involves live data.Done means it runs on real cases on a schedule. A successful demo is a milestone on the way there.
- 04A number measured before and after.Hours per task, turnaround time, cost per run. Without a baseline nobody can tell whether the pilot worked.
- 05A plan for after launch.Someone trained to run it, someone responsible for keeping it running, and a clear answer to who owns the code.
What that looks like in practice
When we built a comp analysis agent for Axia Partners, we built it side by side with the firm’s acquisitions lead so it would match how the team actually underwrites. The agent runs the same steps their analysts ran. A comp that took 80 analyst hours per park now comes back the same day, and the agent is being extended into underwriting.
The model was not the hard part of that project. Matching the workflow was, and that only happened because the person who owned the workflow was involved from kickoff.
How to restart a stalled pilot
If you have a pilot stuck in the almost-done state, three moves usually get it going again:
- 01Name the owner.Pick the person whose work the pilot is meant to change and give them the authority to say what done looks like.
- 02Cut the scope to one workflow.Drop everything that is not needed for one real case to run end to end on live data.
- 03Put builder and user in the same room every week.Short sessions on real cases, until the user would rather use the tool than the old process.
If the pilot still does not move, it may have been the wrong project to begin with. How to choose your first AI initiative covers a better filter.
This is also why our process ends with an upskill step. We hand the system over and train your team to run it, extend it, and trust it. A pilot that nobody on your side can run has not shipped.


