Proof of concept, pilot, production: knowing which one you're funding
Three stages with different goals, budgets and exit criteria. Confusing them is why so many AI demos never ship.
"We built a proof of concept and now we need to productionise it" is one of the most expensive sentences in enterprise AI — usually because the PoC was built to impress rather than to answer a question.
Three stages, three questions
Each has its own budget, exit criteria and definition of done. Confusing them is why demos stall.
- 1
Proof of concept
Is this possible with our data? Weeks. Throwaway code. Output is a written finding, not an asset.
- 2
Pilot
Will people use it, and does it create value? Real users, narrow scope, reversible.
- 3
Production
Can we run this reliably? Monitoring, support, security, cost control — a build, not a promotion.
A proof of concept answers one question
Is this technically possible with our data? Weeks, not months. Throwaway code is fine. The output is a decision and a written finding, not an asset. A PoC that becomes the foundation of the product was not a PoC.
A pilot tests whether people will use it
Real users, real data, limited scope, production-grade enough to be trusted but narrow enough to be reversible. The question is adoption and value, not feasibility — you settled that already.
Production is a different discipline
Reliability, monitoring, support, security review, cost control, documentation. The model may be identical to the pilot's; almost everything around it is new work. Budget for it as a build, not as a promotion.
Write the exit criteria first
Each stage should have a defined kill condition agreed before it starts:
- PoC: if accuracy on our real data is below X, we stop
- Pilot: if fewer than Y% of users adopt it in six weeks, we stop
- Production: if unit cost exceeds Z per transaction, we redesign
Write the kill condition before you start
Proof of concept
- Stage
- Weeks, throwaway code
- Stop if
- Accuracy on real data is below the agreed bar
Pilot
- Stage
- Real users, narrow scope
- Stop if
- Adoption stays under target after six weeks
Production
- Stage
- Reliability, support, security
- Stop if
- Unit cost per transaction exceeds the ceiling
| Stage | Stop if | |
|---|---|---|
| Proof of concept | Weeks, throwaway code | Accuracy on real data is below the agreed bar |
| Pilot | Real users, narrow scope | Adoption stays under target after six weeks |
| Production | Reliability, support, security | Unit cost per transaction exceeds the ceiling |
Stopping is a good outcome
The organisations that get the most from AI are the ones that kill weak ideas cheaply and early. A PoC that concludes 'not with this data' has done its job and saved a great deal of money.