DSME Global Links
DSME Global Links
Strategy

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.

Muhammad Dayyan·Founder & CEO·October 8, 2025·6 min read

"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. 1

    Proof of concept

    Is this possible with our data? Weeks. Throwaway code. Output is a written finding, not an asset.

  2. 2

    Pilot

    Will people use it, and does it create value? Real users, narrow scope, reversible.

  3. 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

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.

M
Written by
Muhammad Dayyan
Founder & CEO, DSME Global Links