An AI proof of concept to find out, within a few weeks, whether your project is worth the investment

Above all, it is for deciding on the facts. It lets you know, before investing:

• whether AI correctly handles the company’s real cases, including the hardest ones;
• whether the available data is sufficient and reliable enough;
• what time or money savings can reasonably be expected;
• how much the solution will cost once in service;
• whether it can connect to existing software;
• whether future users are ready to use it.

It is also a good tool for convincing people: management, a board or investors more readily grant a budget on the basis of measured results.

These four stages often follow one another, but they don’t answer the same question.

And the demo? It is the version you show, often prepared on examples chosen so that everything goes well. It gives an idea of the potential, but says nothing about how the solution will behave with real data.

An AI POC always addresses a precise, measurable question. For example:

• in sales administration: can AI read purchase orders received by email and enter them into the management software with fewer than 5% errors?
• in customer service: can an assistant answer half of routine requests on its own, and pass the others to the right person?
• in finance: can AI spot supplier invoices that don’t match the orders?
• in human resources: can an assistant answer employees’ questions from internal agreements and procedures?
• in operations: can AI estimate a quote or a lead time from the history of projects or orders?

Disappointing POCs often look alike. Here are the most common pitfalls:

• starting without a precise question, and ending up testing “AI in general” rather than your own case;
• forgetting to measure the current situation, and no longer being able to prove the gain at the end;
• testing on data that is too clean, prepared by hand, which doesn’t look like everyday data;
• judging the results “by eye”, without comparing them with objectives set in advance;
• involving users too late, when their comments can no longer be taken into account;
• ignoring the running cost, which can climb quickly when the solution handles thousands of requests.

Most AI POCs last from two to eight weeks, depending on the provider and the complexity of the case. Three factors weigh most on the duration: data availability, the number of software tools the prototype has to access, and how precise the question is. A POC that drags on for several months is often a sign of overly broad scoping. It is then better to split it into several simpler questions.

The price depends mainly on the number of cases tested, the state of the data and the connections needed to existing software. On the market, published prices range from a few thousand to several tens of thousands of euros. The Médiateur des entreprises guide mentions, for this type of experiment, a budget of “a few tens of thousands of euros at most”^[3].

Also think about the running cost once the solution is in service: AI model calls, hosting, human review. A good POC is precisely what lets you estimate it.

Often less than you might think. For many cases, a few dozen to a few hundred real examples are enough to start: documents, customer requests, order histories. What matters is that they reflect the company’s day-to-day reality, with its exceptions and difficult cases.

Before handing them to a provider, three points are worth checking: your rights over this data, the presence of personal data, and where it will be processed. The CNIL offers a checklist on this subject^[2].

It is a perfectly normal outcome, and often a valuable one: it avoids an investment that would not have delivered. A negative POC is still useful if it explains why. Depending on the case, you can then reframe the question, improve the data quality first, choose a simpler solution, sometimes without AI, or drop the project.

It all depends on how it was built. A POC put together in a hurry for a demo usually has to be rebuilt. If it was designed from the start with an architecture able to evolve, a good part can be kept: the business logic, the exchanges with the AI model and the test set. It is a question to ask your provider before you start.

In some cases, yes. In France, the innovation tax credit (crédit d’impôt innovation, CII) lets SMEs recover 20% of certain expenses incurred to design the prototype of a new product, up to €400,000 of expenses a year, until 31 December 2027^[4]. It covers products intended for the market, not tools for purely internal use. Your accountant can check whether your project is eligible.

Bpifrance also offers, under the IA Booster France 2030 programme, support for SMEs and mid-sized companies that can go as far as trialling an AI solution, with part of the costs covered^[5].

Only partly. The EU AI Act does not apply to research, testing and development activities carried out before an AI system is placed on the market or put into service. Real-world testing does not benefit from this exclusion, and other rules, such as the GDPR, continue to apply^[6].

The POC is still the right time to plan ahead. If the future solution falls under high-risk systems, for example in recruitment, credit or education, its obligations will apply from 2 December 2027, the date set by the July 2026 omnibus regulation^[7].

Several agencies and IT services companies offer AI POCs. Their offers differ mainly in duration, price and the follow-up offered after the POC.