The Pilot-to-Production Checklist
Most AI pilots never reach production — Q2 2026 data put the failure share at 69%. The 31% that ship are not luckier. They exhibit signals you can check for in advance. Here are twelve of them.
Stage One: Align — Do You Know What This Is For?
The hold-up fits in one sentence.
If describing what is slowing things down takes a paragraph, the diagnosis is not done and the build will inherit the confusion.
A named person carries the cost.
Initiatives where "the team" owns the gap stall; initiatives where a specific operator feels it weekly ship.
The win condition is observable.
"Numbers we can point to" survives contact with reality. "Transformation" does not.
Stage Two: Meet — Do You Have a Plan or a Vibe?
The scope is written down with a price and a timeline.
Fixed scope is not bureaucracy; it is the only known cure for the change-order spiral.
Someone with operational context made the data-model call.
The single most predictive signal on this list. Schemas decided implicitly by an AI model are the leading cause of month-seven rebuilds.
The tool count went down during planning.
Teams that add tools in planning are avoiding a decision. Teams that cut them made one.
The pilots that die do not die in production. They die in planning, and production is just where the body is found.
Stage Three: Coach — Is the Build Being Steered?
Something ships every week.
Working increments, not status meetings. A build you can only evaluate at the end is a build you cannot steer.
Drift gets named early.
"This is not what we pictured" said in week three costs nothing. Said at handoff, it costs the project.
Failure modes are designed, not discovered.
Fallback chains, confidence thresholds, and vendor-API abstraction exist before launch.
Stages Four and Five: Lead and Reflect — Will It Survive Contact?
Production load was simulated before users arrived.
Demo volume proves nothing. The systems moving $125M+ a month were architected for that number before they saw it.
The team that uses it helped shape it.
Floor-staff adoption is engineered in collaboration, not mandated at rollout.
Someone owns it after launch.
Monitored, maintained, and improved weekly — or orphaned at handoff. Production is a tense, not a milestone.
The Proof Strip
69% of pilots fail to reach production (Digital Applied, Q2 2026). 12 signals, checkable before a dollar is committed. 90 seconds to score your own initiative.
Scoring Yourself Honestly
| Signals present | Where you likely stand | The move |
|---|---|---|
| 9–12 | Production-bound | Protect the cadence; do not add scope |
| 5–8 | Coach stage — drifting or stalling | Name the drift this week, in writing |
| 0–4 | Align stage wearing a pilot costume | Stop building; start diagnosing |
Frequently Asked
Can a struggling pilot be saved?
Usually — if the failure is direction, not effort. Mid-build course correction is a defined engagement shape, and it starts with an honest read of the current state, not a rebuild pitch.
Is this checklist a substitute for a full assessment?
It is the pattern; the assessment is the instrument. The free Paddock20 Diagnostic asks eleven questions built on these signals, scores the result live, and hands you a two-page PDF — the gap between where you think you are and where your answers put you, whether we ever talk or not.