Skip to main content
nexeai
All articles

Client references for an AI provider: how to verify them

Nicolas10 min read
Client references for an AI provider: how to verify them

An AI provider sends you a brochure with three logos, two quotes and a two page case study. The layout is polished. Nothing, however, tells you whether those companies exist, whether the project is still running, or whether the figures come from production or from a demonstration. For a small business owner, the client reference has become the main sales argument in this market, and one of the easiest to arrange. Here is a verification method that fits in four questions, one table to fill in and two fifteen minute calls.

Key Takeaways

  • A useful reference names a client you can call, a scope and a go live date. A logo alone proves a commercial relationship, not a result.
  • Three questions are enough to sort them: who uses the tool today, what was the starting point, and what exactly was measured.
  • A figure with no baseline, no period and no scope is worth nothing. Ask for the calculation, not the result.
  • A serious provider arranges the introduction to a client, or explains clearly why it cannot.
  • Without references, a young provider can compensate with a bounded paid pilot and written contractual commitments.

What a client reference proves, and what it does not

A logo proves one thing only: a company signed a contract with that provider. It says nothing about the amount, the duration, the satisfaction, or the state of the project six months later. A quote, meanwhile, is almost always written by the provider and then approved by the client, often at the end of a project that is still fresh. As for the two page case study, it is rebuilt after the fact, with the figures that tell the best story.

What a good reference can prove is narrower, and far more useful: a company comparable to yours, on a comparable use case, obtained a defined result over a defined period, and the system is still running.

That distinction matters because production is precisely where projects die. According to Gartner (July 2024), "at least 30% of generative AI (GenAI) projects will be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value". In other words, a pilot reference proves almost nothing. The only question that matters is whether the project is still running, and who uses it.

The three questions that expose a hollow reference

Ask them in this order, to the provider first, then to the client if you get the introduction.

Who uses the tool today, and how often? An AI agent used by two curious people once a week is not a delivered project. Ask for the number of active users, the department involved and the real frequency of use. A provider who knows their client answers in thirty seconds. A provider reciting their brochure hesitates.

What was the starting point, and how was it measured? A productivity gain assumes a measurement before. If nobody timed the initial process, the announced gain is an impression, not a result. Ask who took the measurement, when, and on what sample.

What changed in the organisation? An AI agent handling inbound requests only works if someone rewrote the procedures, trained the teams and decided who rules on ambiguous cases. A reference that mentions no human work behind the tool is incomplete, or too good to be true.

Element Unverifiable reference Verifiable reference
Client Logo with no company name Legal name, sector, headcount
Contact "We can pass it on" Name and role of a contact
Scope "An AI assistant" Process, volume handled, connected tools
Date None Month of go live
Result "Significant gains" Figure, baseline, period
Status Project described in the past System still in service

Read a case study figure like an auditor

An isolated figure is not proof, it is an argument. Four elements must come with it before it becomes verifiable: a baseline (what value before), a period (over how many months), a scope (how many cases, employees, sites) and a calculation method (who measured, with which tool).

Placing a figure also requires an outside reference point. In a survey of 822 business leaders conducted between September and November 2023, Gartner (July 2024) reports that respondents declared on average "15.8% revenue increase, 15.2% cost savings and 22.6% productivity improvement". These averages cover companies of all sizes and sectors, not French SMEs. They give an order of magnitude: a provider announcing "+50% productivity" is not necessarily dishonest, but they must be able to explain why their result departs that far from an average measured on 822 organisations.

Be wary of figures that do not concern what you are buying. A processing time divided by three on a process you already outsource will bring you nothing. Always ask what the gain translates into inside your own organisation: hours saved, additional cases handled, shorter response time, errors avoided.

How to reach a cited client without burning the relationship

The introduction is the decisive test, and the easiest to arrange. Ask the provider to introduce you to a client whose size and sector resemble yours, not the most prestigious client on the brochure.

Three precautions make the call useful. State the purpose and the length (fifteen minutes, a technical conversation, no sales prospecting on your side). Ask in advance for the client's consent to be cited: a company whose name circulates without authorisation tells you a lot about the provider's rigour. Finally, check the client's existence and real size in a public registry. A reference to a struck off company, or to a headcount very different from the one announced, is a strong signal.

During the call, two questions beat ten. "Would you run this project again with this provider?" and "what went wrong?". The answer to the second is the most informative: a satisfied client who describes a schedule slip and a rewrite of specifications is more credible than a client with nothing to reproach.

The signals that should raise a flag

Some details keep showing up with providers who inflate their reference portfolio. Logos with no company name and no mission date. Quotes signed with a first name and a sector, with no role and no company. Round figures, systematically above market rates, never backed by a calculation. The case study whose scope changes between the brochure and the phone call. Refusal of an introduction on every reference in the name of confidentiality, when the cited client could perfectly well confirm a project without disclosing sensitive data.

Another signal, subtler: references all sitting in a sector very distant from yours, presented as transferable. A successful project in logistics says nothing about your accounting firm, unless the provider explains what transfers. This list overlaps largely with the signs of an unreliable provider that we detail elsewhere.

A provider without references: what replaces the proof

Many young and competent firms do not yet have publishable references. This is common, and not disqualifying in itself, provided the provider offers something other than a promise.

A paid, bounded pilot, with success criteria written before the start, replaces a reference. A technical audit of access and code, run by you or by a third party, replaces declared trust. Precise contractual commitments (deadlines, penalties, code ownership, reversibility procedure, support level) replace logos. Finally, a clear exit clause, with the handover of source code and data, limits the risk if the project goes off track.

These elements are discussed before signing, not after. They overlap with the questions to ask before signing with an AI agency, and they often weigh more than a brochure.

What a reference does not replace

The French market is still young, and AI references in production remain few. According to Insee (July 2026), "en 2025, 18 % des entreprises françaises de 10 salariés ou plus déclarent utiliser au moins une technologie d'intelligence artificielle (IA)", and "plus de la moitié des entreprises utilisant l'IA déclarent être freinées dans leur utilisation par un manque d'expertise". In a market where expertise is scarce, the providers who can deliver it are few, and those who can prove it are fewer still.

A verified reference does not guarantee that your project will succeed. It only guarantees that the provider has already been through a production rollout with a comparable client. It is a filter, not insurance. The rest is decided in the scoping, the quality of the available data and your teams' ability to take over the process.

NexeAI offers an audit and advisory engagement upstream of this kind of decision, as well as the design and deployment of AI agents. We document the scope, the success criteria and the exit conditions before any commitment, and we tell you what can be measured in your context, including when the answer is "not yet".

FAQ

How many references should you ask for?

Two genuinely verified references beat eight logos. Ask for one reference per targeted use case, and at least one in a company your size. If the provider cannot supply any on the precise use case you care about, say so explicitly and adapt the contract accordingly.

Can a provider refuse to give client names?

Yes, and it is sometimes legitimate: a client may forbid publication of their name, or work in a sensitive sector. The refusal must then be explained and compensated, through an anonymised introduction (sector, headcount, use case), a direct conversation without the name on the brochure, or a bounded pilot. A blanket refusal, with no explanation, on every reference, is a negative signal.

Is a case study without figures useless?

No, but it serves one purpose only: understanding the technical scope and how the project unfolded. A case study that describes the connected tools, the volumes handled and the difficulties encountered is instructive even without a quantified result. It becomes suspect when it claims to demonstrate a gain without giving either a baseline or a period.

Should you call every cited client?

No. Pick the reference closest to your situation and dig into it. One reference verified in depth teaches you more about the provider's working methods than five five minute calls. If the provider refuses the introduction on that specific reference, move to the next one and note the refusal.

What if all the references are abroad?

Not disqualifying, but it changes what you need to check: GDPR compliance, data hosting, support language, time zone, and the law that applies in a dispute. Ask where the data is hosted and who answers when there is an incident, before looking at the logos.

Conclusion

Verifying client references requires neither technical expertise nor much time. Three questions to the provider, one table to fill in, one introduction, one fifteen minute call: you will know whether the brochure describes a living project or an arranged story. The core rule fits in one sentence: demand the scope, the date and the method before accepting the result.

If you are preparing an AI project and want an outside view on the references you are being shown, or on what is genuinely measurable in your context, write to us. We will also tell you, when that is the case, that the project is not mature yet.


Share this article

Want to go further?

Let's talk about how AI can apply concretely to your business, in a free first conversation.