StackLift
All terms

Glossary

Proof of Concept (PoC)

A small, time-boxed build that tests whether one risky technical idea can work at all.

A proof of concept (PoC) is a small experiment that tests whether a specific technical idea is feasible before a full build. It answers one question, such as whether a model can read your documents accurately, and is usually discarded or rebuilt afterward.

A PoC is not a product and not meant for end users. It has a clear question, a short time box and a pass or fail rule agreed in advance.

It differs from a prototype, which tests a design with users, and from an MVP, which tests demand with real users. Running one early narrows the uncertainty in a software estimate.

Start in 60 seconds

Put this to work on your project.

Get an AI-assisted, architect-reviewed scope, cost range, and timeline. Free, no login.

The complete guide

Proof of Concept (PoC): the complete guide

Updated November 18, 2026 ยท Published by , Founder, StackLift AI. Written with AI assistance; how we make content.

Short answer

A proof of concept (PoC) is a small, time-boxed build that tests whether one risky technical idea can work at all. It answers a feasibility question, such as whether a model can read your documents accurately or two systems can exchange data, and it is usually thrown away or rebuilt before the real product is made.

What is a proof of concept in software?

A proof of concept tests one assumption: can this be done? It is not a product and it is not meant for users. It is an experiment with a clear question, a short deadline and a result you can read as yes, no or not yet.

The idea is older than software. Engineers and scientists have long built a rough version of the risky part to find out whether the whole plan is sound before paying for the rest. In software, that rough version is code, a script, a notebook or a thin slice of a system, aimed at the single point that could sink the project.

The output of a PoC is knowledge, not software. What you keep is the answer, the evidence behind it and a clearer sense of the risks. The code is often discarded.

Proof of concept vs. prototype vs. MVP: what is the difference?

The three are often mixed up because each is an early, partial version of something. They answer different questions, and the usual order is proof of concept, then prototype, then MVP.

Three early-stage builds and the question each one answers
BuildQuestion it answersWho sees itWhat happens to it
Proof of conceptIs this technically possible?The team and sponsorsUsually discarded or rebuilt
PrototypeWill people understand and want to use this design?Test users and stakeholdersInforms the design; may be replaced
MVPWill real users adopt it and pay or return?Real usersGrows into the product

The side-by-side pages go deeper: proof of concept vs. prototype and prototype vs. MVP. Not every project needs all three. Many skip straight to the stage that matches their biggest unknown.

When does a proof of concept make sense?

Run one when a single technical unknown could sink the project and you cannot answer it by reading or asking. The unknown should be specific and testable.

  • A language model must extract data from messy documents, and you do not know if accuracy will be good enough.
  • Two systems must exchange data and the vendor's API may not support what you need.
  • A workflow depends on a new technique, such as RAG, that your team has not used on your data.
  • A performance or scale limit decides whether the design is viable.
  • A legacy system must be read or written in a way nobody has tried.
  • A device, sensor or hardware connection must work before software around it makes sense.

If nothing technical is in doubt, skip it. A PoC for a well-understood web app only delays the build.

How do you run a proof of concept?

  1. Write the question as a sentence that can be answered yes or no.
  2. Define success before you start: what result counts as a pass, measured on real or realistic data.
  3. Set a time box and a small team, and cut everything that does not serve the question.
  4. Build the thinnest thing that can answer it. Skip polish, login screens and edge cases.
  5. Test against real samples, including awkward ones, and record what failed as well as what worked.
  6. Write a short decision note: pass, fail or change the approach, with the evidence and the next step.
  7. Decide what to keep: findings, test data and learnings carry forward; the code usually does not.

A time box protects you from the PoC growing into a project. If the question is not answered by the deadline, that is itself a finding: the problem is harder than assumed.

What are examples of a proof of concept?

These illustrate the kind of question a PoC asks. They are typical patterns, not results from any named organization.

Example questions and pass rules
ProjectPoC questionExample pass rule
Invoice processing agentCan a model extract the required fields from our real supplier invoices?Fields are correct on a set of real invoices at the level the finance team agrees is workable
Internal knowledge searchDoes retrieval over our documents return the right passage for typical questions?The right source appears in the top results for most of a prepared question list
CRM and ERP syncCan we read and write orders through the vendor's API without duplicates?A sample batch syncs both ways with no duplicates or lost fields
Voice assistantDoes speech recognition handle our domain terms and accents?Key terms are recognized well enough in recordings from real calls
Legacy modernizationCan we extract business rules from the old code or database?A chosen module's rules are reproduced and match the old system's outputs

In each case the rule is set before the test, using the people who will rely on the result. For AI work, see the AI agent evaluation checklist.

Why are proofs of concept common in AI projects?

AI features depend on how a model behaves on your data, which you cannot know from a vendor's demo. A PoC tests that behavior early, when changing direction is cheap.

  • Accuracy: is the model right often enough on your real inputs?
  • Cost: what does a typical task cost in model usage? Current prices are on the LLM pricing page.
  • Speed: is the response fast enough for the workflow?
  • Data access: can the system reach the data it needs, safely?
  • Failure behavior: what happens when the model is wrong, and is that survivable?
  • Approach: would RAG, prompt engineering or fine-tuning fit best?

A PoC proves something can work in a lab. It does not prove it will hold up at volume with all the edge cases, which is why a production build still needs testing, guardrails and monitoring.

What should you agree before starting a proof of concept?

Most failed PoCs fail in the setup, not the build. A short written agreement before work starts prevents the usual arguments about what the result means.

  1. The single question being tested, in one sentence.
  2. The pass rule, with a threshold the business owner has accepted.
  3. The data to be used, who provides it and any privacy limits on it.
  4. The time box and the people involved.
  5. What is out of scope, written down, such as login, design and performance at scale.
  6. Who decides at the end, and what each outcome leads to.
  7. Where results and test data will be stored, so findings outlive the code.

The AI project discovery checklist covers similar ground at project level, and the discovery phase is where PoC candidates are usually identified.

How do you measure a proof of concept result?

Use the rule you set at the start, applied to realistic data. Report the numbers or counts without rounding in your favor, and show the failures, not just the successes.

  • Accuracy or correctness on a fixed set of real cases.
  • Coverage: how many cases the approach could handle at all.
  • Speed per task, compared with what the workflow needs.
  • Cost per task, such as model usage or compute.
  • Reliability: does it give the same result when repeated?
  • Effort to maintain: how much hand-tuning did the result depend on?

What goes wrong with proofs of concept?

  • No pass or fail rule, so the result is argued about afterward.
  • Using clean sample data that hides the problems real data will show.
  • Treating the PoC as the first version of the product and shipping it. Shortcuts that are fine for an experiment become technical debt.
  • Widening the question until the PoC becomes a small project with no end.
  • Judging only the happy path and ignoring cost, speed and failure handling.
  • Running one when the real risk is demand or usability, which a prototype or MVP tests better.
  • Letting a vendor run the PoC with their own data, which proves their demo and not your case.

If a PoC already exists and the team wants to ship it, the vibe-coded app rescue page describes how to audit and harden experimental code.

What happens after a proof of concept?

The decision note leads to one of three paths. Each has a next step, and the evidence from the PoC feeds the plan.

PoC outcomes and next steps
OutcomeMeaningNext step
PassThe idea is feasible on the tested dataMove to design or a prototype, then scope the build; the PoC findings go into the estimate
PartialIt works in some cases or needs a different approachNarrow the scope, change the method or run a second, focused PoC
FailIt does not work well enough or costs too muchStop, or rethink the problem; you saved the cost of the full build

A failed PoC is not wasted. It cost a small amount to avoid a large mistake. Record why it failed so the same idea is not retried blindly later.

What does a proof of concept mean for cost and timeline?

A PoC is meant to be small and short, and its cost follows from the question it answers. A narrow question on available data is quick. A broad question with hard-to-get data, many systems or a need for special hardware takes longer.

  • How narrow the question is.
  • Whether real data is available and clean enough to test with.
  • How many systems or APIs must be touched.
  • Access: security approvals and credentials often take longer than the build.
  • How much measurement and evaluation is needed to trust the result.

A PoC lowers the uncertainty in a software estimate, which is a reason to run one before committing to a large budget. For context on full-project ranges, see the AI app cost guide and run the free estimate.

How does StackLift approach a proof of concept?

Every engagement starts with structured discovery and an architect-reviewed scope, which identifies the technical unknowns worth testing. The AI Scope Sprint is a 3-day scoping plan with detailed scope, user stories, tech stack, and a cost and timeline range.

Where a risky assumption remains, it is named in the scope so it can be tested first. Once feasibility is clear, the plan moves to an MVP Build Sprint or another path from the engagement ladder, with CI/CD, observability, security review and rollback in place from sprint one.

Questions people ask

What does PoC stand for?

PoC stands for proof of concept. In software it means a small experiment built to test whether one technical idea is feasible before a full build.

How long should a proof of concept take?

As short as the question allows. Set a time box up front. Many are measured in days or a few weeks, and a 3-day AI Scope Sprint can help define what to test first.

Is a PoC the same as a pilot?

No. A PoC tests feasibility, usually inside the team. A pilot runs a working solution with a small group of real users to learn how it performs in daily use.

Can a proof of concept become the final product?

It should not. PoC code is written for speed and learning, often without security, tests or error handling. Rebuild the proven idea properly, or audit and harden the code before relying on it.

Do I always need a PoC before building software?

No. Skip it when the technology is familiar and the risks are known. Use one when a specific technical unknown could change the cost or the decision to proceed.

Who should approve the success criteria?

The people who will rely on the outcome: the business owner who needs the result and the engineer who can say what is realistic. Agree the rule in writing before the test begins.

What should a PoC report contain?

The question, the success rule, how it was tested, the results with examples of failures, cost and speed observations, the decision and the recommended next step.

Who should run a proof of concept?

A small team that includes someone who can build the technical part and someone from the business who owns the question. Keeping the team small and the question narrow keeps the time box realistic. An outside partner can help when you lack the skills, as long as the test uses your data and your pass rule.