How do you choose an AI provider when you run a mid-sized company?
The market bundles under one word trades that have nothing in common: agencies doing communications, integrators deploying off-the-shelf tools, vendors selling a licence, independents selling days, and a few teams that actually build bespoke. The right criterion is neither the technology nor the price: it is who answers for the result on a specific process, and what stays in the company once the provider leaves. A CEO who starts by choosing a tool has the order wrong — you start with the process that costs money.
Where do French mid-sized companies actually stand?
Further than people say on intent, less far than people think in practice. Bpifrance Le Lab surveyed 1,209 executives of French companies with more than ten employees between October and December 2024: 43 % have put a strategy in place to integrate AI, but only 26 % use generative AI, 16 % non-generative AI and 10 % both. Roughly a third have adopted it day to day.
The revealing gap is in conviction: 38 % of executives consider AI important or very important to the survival of their business today, rising to 58 % on a three-to-five-year horizon. In other words, most know the subject is coming and do not yet know where to take hold of it.
That is exactly where the choice of a provider is decided: the question is not “which tool”, it is “where to start”.
Agency, integrator, vendor, independent, studio: who does what?
The agency comes from communications and digital. It is good at visibility, content, rebuilding a website; marketing is its natural ground and it is often excellent there. It is not equipped to touch an industrial or financial process.
The integrator deploys and configures existing tools — an augmented office suite, a CRM, an automation platform. The right call when the need is standard, and almost always cheaper than building. Its ceiling is the product it installs.
The vendor sells a licence and a roadmap that is not yours. Excellent when your need is the common need; frustrating as soon as your specificity is precisely what makes your margin.
The independent brings a sharp skill for a set time. Fast, flexible, cheap to engage — but their capacity ends with their calendar, and what they leave ends with their files.
The bespoke build studio works where the other four stop: a process specific to the company, weighing on results, that no market product handles. That is engineering, not advice, and it is judged on what runs in production.
Why do so many AI projects produce no financial effect?
Three causes, always the same. The first is order: a tool is chosen, then a use is sought for it. The useful path is the reverse — start from the two or three processes that consume hours or destroy margin, and build only for those.
The second is adoption. A tool the teams do not use is worth nothing, and usage cannot be decreed from a management committee. It is built with the people who will do the work, during the project, changing the tool when it does not fit — not in a training session once the project is delivered.
The third is the absence of an owner. In a mid-sized company AI often has nobody at its head: the IT director sees a risk, the sales director a gadget, the CEO an urgency without a deadline. With nobody answering for it, the subject stays an agenda item rolled over month after month.
What should you ask before signing?
Which specific process are you working on, and how will the effect be measured? A provider who answers “productivity” has not understood the question.
Who will do the work, and have they done it in live operations before? The distance between a demo and a tool in production is measured in months.
What happens when you leave? Can the teams use it, change it, switch it off? Autonomy of use is the real deliverable.
How long until the first tool is in service? Beyond a few months, the subject will have moved before delivery. We build in two to three months, precisely for that reason.
And a question few CEOs ask: where does my data go? An architecture that keeps data on your side is not a luxury — it is often the condition for the subject to clear the board.
How long, and with what commitment?
A first useful cycle runs two to three months: a few weeks to identify and arbitrate the processes worth treating, then building the tool with the teams who will use it. Neither a six-month audit nor a pilot that goes nowhere.
The rest is a matter of rhythm rather than volume. A few days a month keep the subject alive — arbitrating, unblocking, extending — provided somebody answers for it. That is what a part-time AI lead amounts to: the competence and the accountability, without the full-time seat most mid-sized companies can neither justify nor fill.