What is a proof of concept?
It answers one question and stops. Keep using it past that point and you are running production traffic on something nobody built for the job.
Proof of concept (PoC)
PoC
A proof of concept is a small build that tests whether an idea can technically work. Conditions are chosen to make the test fast and cheap. It answers one narrow question, then gets thrown away, rather than kept and grown into the product.
The question a PoC answers is narrow on purpose. Can a model classify these documents accurately enough? Will this API return data fast enough? It is not asking whether customers want the feature, or whether the system holds up under real traffic. Those are different questions, and a PoC was never built to answer them.
Because the question is narrow, the build can cut corners everywhere else. Authentication can be a stub. Error handling can be missing. The data can be a clean sample somebody picked, not the messy export a real customer will upload. That is not sloppiness. It is the whole point of a PoC: spend as little as possible to get one honest answer.
The pilot trap starts here
MIT's 2025 State of AI in Business report found 95% of enterprise AI pilots showed no measurable profit or revenue impact. A common cause sits in the gap between a PoC and a production system. The proof of concept worked, so someone pointed real users at it and called that shipping.
A PoC that answered its one question honestly was never tested against the questions production asks. Nobody checked what happens on input nobody anticipated. No one is named to answer when it stops responding, or when ten people use it at once instead of one. None of that was in scope, because none of it was the question.
Skipping the rebuild feels efficient in the moment. It looks like time saved. In reality, the cost just moves to a later date nobody chose. Usually that date is the first serious outage, when the price is higher and everyone is watching.
- One questionCan this work, technically, at all.
- Picked conditionsClean data, no concurrent users, no edge cases.
- Answer reachedYes or no. The PoC's job is done.
- Decision pointRebuild for production, or discard.
- Rebuilt systemAuth, monitoring, error handling, ownership added.
Skip the decision point and the PoC keeps running with none of the last step done.
A proof of concept against the things it gets mistaken for
| Criterion | What people build instead | What a PoC actually is |
|---|---|---|
| A prototype | Shows what the product will look and feel like, usually to a person, not to real data. | Tests a technical question, usually with real data, and is not meant to be looked at. |
| A minimum viable product | The first real version of the product, built to be used and kept if it works. | Answers one question, then is discarded whether the answer is yes or no. |
| A pilot | Runs with a limited group of real users over weeks, testing adoption and value. | Runs once, briefly, against a technical question with no user rollout attached. |
Common questions
01What is the difference between a PoC and an MVP?
A proof of concept answers one technical question and is then thrown away, whatever the answer was. A minimum viable product is the first real version of the product, built to be kept and extended if it tests well with users. A PoC settles whether something is technically possible. An MVP, built separately once that answer is yes, tests whether people want it.
02Can a PoC be promoted straight to production?
Not without a rebuild. A PoC skips authentication, monitoring, error handling and load testing on purpose, because none of that was needed to answer its one question. Production needs all four. Promoting a PoC without adding them is how a working demo becomes an outage nobody planned for.
03How long should a proof of concept take?
As long as it takes to answer the one question honestly, and no longer. A PoC that keeps growing new features has stopped being a proof of concept. It has turned into an unplanned MVP, usually without the production-grade foundations an MVP needs.
Related terms
- What happens after the AI pilot →The four things a pilot skips that production needs before it can carry real traffic.
- AI prototype to production →Where we rebuild a proof of concept into a system built to carry real users.
- Vibe-coded prototype vs. production build →The same gap in a different shape: a build that impressed a room against one that survives one.
