AI Proof of Concept: How to Scope One Before You Build

Most AI projects that stall do not fail at the model. They fail at the scope. The team builds a demo that looks great, nobody agreed what "good enough" meant, and three months later the question "should we put this into production?" still has no answer. An AI proof of concept exists to answer that one question, quickly and with your own data. This is how we scope one before anyone writes code.
Here is an example from our own work. An event management company we work with had managers spending a large part of their week answering suppliers' organizational questions: lighting, screens and the other details that have to be right before a booth goes up. We did not start with "an AI assistant for operations". We started with one job, supplier conversations, and one measure, the hours managers spent on them. The agent cut that time by 30%, about 3 to 4 hours a week per manager in high season. That result was easy to judge because the scope made it measurable.
What an AI proof of concept is, and what it is not
A proof of concept answers one question: can AI do this specific job well enough, with your data, at a cost that makes sense? It is not a small version of the final product.
Proof of concept | Prototype | Pilot | MVP | |
|---|---|---|---|---|
Answers | Can AI do this job well enough? | Will people use it this way? | Does it hold up in daily work? | Will customers pay for it? |
Users | Your team, on real data | A few test users | A small group of real users | Real customers |
Built to last? | No | No | Close to it | Yes |
Data | Real cases, small sample | Sample or mock data | Live data | Live data |
Ends with | A go or stop decision | Design feedback | A decision to roll out | A product in the market |
The difference matters for budget. If a proof of concept quietly turns into an MVP, you pay for production quality before you know the idea works.
When you need a proof of concept, and when you don't
Run one when there is a real unknown that could sink the project: whether your data is good enough, whether the model can do the job reliably, or whether it will work inside your systems. Skip it, or do something else first, when:
The goal is still vague. "Use AI to improve customer experience" cannot pass or fail. Start by picking the job.
A simple rule, script or ready-made tool would do the job. Use AI only where it beats the simpler option.
The data does not exist yet. Collecting and cleaning it comes first.
Step 1: Write down the one question it must answer
Before anything else, write the question in one sentence. "Can an agent answer 70% of supplier questions without a manager?" is a question. "Explore how AI can help operations" is not.
If the team cannot agree on the sentence, that is your first finding. It is much cheaper to learn it now than after a month of building.
Step 2: Pick one job and one success metric
A good first job happens often, follows rules a person could write down, and has an outcome you can check. Then choose one number that says whether it worked:
Hours saved per person per week.
Share of cases handled without a person.
Accuracy against answers a person checked.
Cost per task compared with doing it by hand.
Measure that number today, before the proof of concept starts. Without a baseline, "it feels faster" is the only result you will get.
Pick a business number, not only a model score. An accuracy figure tells your engineers how the model performs; it does not tell you whether the job got cheaper or faster. If you track accuracy, track it next to one of the business numbers above.
Step 3: Check the data before you write code
Most of the effort in custom AI development is not the model. It is finding, cleaning and connecting the data the model needs. Before you commit, check:
Where the data lives, and whether you can get it out of each system.
Whether it is complete enough for the job: real cases, not only the tidy ones.
Who owns it, and whether you are allowed to send it to an AI provider.
Generative AI development, such as drafting replies or summarizing documents, needs less data preparation than a model trained on your history. It still needs real examples to test against.
Step 4: Set the limits
A proof of concept needs a fence around it. Agree in advance:
A time box. Usually 2 to 4 weeks. If the question cannot be answered in that time, the scope is too big.
Which systems it can touch. Read-only access where possible.
What is out of scope. Integrations, user interface polish and edge cases usually wait for later.
A budget for running it, including the cost of model calls on real volumes.
Step 5: Agree the rules for data and access
A proof of concept often gets more access than it needs, because "it's only a test". That is how sensitive data ends up in a tool nobody reviewed. Decide which data the proof of concept may use, who can see the results, and whether your AI provider can train on your data. On our projects, client code and data are never used to train public AI models.
If your company has no AI usage policy yet, this is a good moment to write one. Our AI governance and compliance services cover that, along with frameworks such as the EU AI Act and ISO 42001.
Step 6: Decide what "go" and "stop" mean before you start
Write the decision rule next to the question from Step 1. For example: "If the agent handles at least 60% of supplier questions correctly in a two-week test, we plan the production build. If it handles less than 40%, we stop. In between, we run one more round with better data."
A stop is a good result. It means you spent weeks, not a year, learning that this job is not ready for AI.
Stop, too, if the data you need is not available, if the people who would use it do not need it, or if running it in production would cost more than it saves.
Why most AI proofs of concept never reach production
A proof of concept can pass and still go nowhere. These are the reasons we see most often, and all of them are decided during scoping:
It was tested on clean, hand-picked examples. Real data is messier, and the result does not hold.
Nobody agreed on a success measure, so nobody can say it passed.
It was judged on model accuracy alone. Budgets are approved on business results: hours saved, cases handled, cost per task.
No one owns it after the demo. Production needs a named person accountable for how it runs.
There is no path into the systems people use every day, and no budget for running it at real volumes.
Access and data rules were left for later, and the security review stops it.
The fix is to write the production questions into the scope on day one: who will own it, which system it will live in, and what it may cost to run.
A one-page scope you can copy
Field | What to write |
|---|---|
Question | One sentence the proof of concept must answer |
Job | The one task the AI will do, with an example |
Success metric | One number, with today's baseline |
Data | Sources, owner, and permission to use them |
Limits | Time box, systems, budget, what is out of scope |
Rules | Who can see what; provider data terms |
Decision | Go, stop or one more round, and the thresholds for each |
Owner | The person who makes the call at the end |
After a successful proof of concept
A working proof of concept is the start of the build, not the end. Production needs what the test skipped: access controls, logging of every decision, testing on edge cases, and a person accountable for what the AI does. Our guide on how to build an AI agent covers those steps.
This is where senior engineers matter most. AI speeds up routine work; across our projects we measure 30-40% less of it. But someone has to decide what is safe to ship. With a senior-led team, the first production deployment usually comes 4-8 weeks after the start. See how we work on our AI-enhanced software development page.
FAQ
How long should an AI proof of concept take?
Most published guides put it at 2 to 4 weeks, and up to about 8 when the test includes real integrations. Set a time box before you start. If the question cannot be answered within it, narrow the question rather than extending the project.
What is the difference between an AI proof of concept and an MVP?
A proof of concept tests whether AI can do a job well enough, on your data, for your team. An MVP tests whether customers will use and pay for a product. A proof of concept is not built to last; an MVP is.
What if the proof of concept fails?
Then it did its job. You learned, at low cost, that this task or this data is not ready. Check whether the cause was the data, the scope or the model, and decide whether a narrower question is worth one more round.
Do we need our own data for a generative AI proof of concept?
You need real examples to test against, even if the model is not trained on them. A proof of concept judged only on made-up examples tells you little about how it will perform on your work.
What is the difference between a proof of concept and a pilot?
A proof of concept tests whether AI can do the job, usually on a small set of real cases with your own team. A pilot puts a working version into real users' daily work, to see whether it holds up and whether they adopt it. Run the pilot only after the proof of concept passes.
When should you skip a proof of concept?
When the goal is still vague, such as "use AI to improve customer experience": pick the job and the measure first. Also skip it when a simple rule, a script or a ready-made tool would do the job. A proof of concept should test something you genuinely don't know.
Planning an AI proof of concept? Tell us the job you want AI to do, and we will help you scope it.
