Your team has sat through the AI talk. Maybe two. Everyone agrees it matters. Nobody has bought anything, and nobody can say what the first purchase should even be.
You are not behind. You are the worldwide majority, and the last year put numbers on both ends of the problem.
The answer in 60 seconds
The famous numbers describe the delivery end: roughly 95% of generative AI pilots show little or no measurable P&L impact (MIT’s much-debated 2025 estimate), and 42% of companies abandoned most of their AI initiatives in 2025, up from 17% a year earlier (S&P Global). France just published the rarer number - the buying end of the funnel: after one year of a national programme that reached more than 35,000 companies with AI awareness sessions, the programme’s own next step - a subsidised AI diagnostic - had been engaged by 70 companies, against a target of 1,000 (French Ministry of the Economy, 3 September 2026). Companies either don’t buy, or buy something vague and abandon it. Awareness worked. The first step didn’t.
The pattern is not French. It is what happens whenever the step after “now you know” is either free content or a large, vague commitment - with nothing purchasable in between. A workable first AI project has five properties:
- It targets one process you run every week, not “the company”.
- Its scope is fixed before you sign - what is in, what is out.
- Its price is known before it starts - a quote, not a day rate that runs.
- It has a dated deliverable, weeks away, not quarters.
- It defines what “worked” means - a pass/fail test agreed in advance.
The rest of this article shows the evidence, then walks through the five criteria, three classic first projects, and how to judge the result - including when a public or subsidised programme is the right buy.
In this article
- What the official numbers actually say
- Why awareness doesn’t trigger a purchase
- The official ladder - and where it stops
- Five criteria of a real first step
- Three classic first projects
- How to judge that it worked
- When a subsidised programme is the right buy
1. What the official numbers actually say
France’s « Osez l’IA » plan launched on 1 July 2025 to push AI adoption across small and mid-sized companies. Its one-year report, published 3 September 2026, is unusually honest raw material, because it counts both ends of the funnel:
| Programme tier | One-year result |
|---|---|
| Companies reached by awareness sessions (615 “AI ambassadors”, 20 regions) | 35,000+ |
| AI solution vendors referenced in the official catalogue | 88 |
| Documented use cases on the public platform | 54 |
| Subsidised AI diagnostics engaged (target: 1,000) | 70 |
The government’s own framing of the next phase says the quiet part aloud: the priority is now moving « de la sensibilisation à l’action » - from awareness to action. Trade press was blunter, calling the first year “a modest start”, and quoting the AI minister’s acknowledgement that many companies still believe AI is not for them (LeMagIT, 4 September 2026).
Two honest caveats before building on these numbers. “Made aware” means reached by the ambassador network - a talk, a workshop, a meeting - not a qualified buyer. And the 70 counts one specific programme: companies that bought AI work privately, outside the plan, are invisible here. So the number does not prove that almost nobody bought anything. What it does show is narrower and more useful: when the step after a free talk is a large generic diagnostic, almost nobody takes it. One company in 500 did.
It helps to set this next to the failure statistics your board has already seen. MIT NANDA’s “GenAI Divide” report estimated in 2025 that about 95% of enterprise GenAI pilots deliver little or no measurable P&L impact - a widely quoted and widely criticised figure, best treated as directional. S&P Global Market Intelligence found that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% a year earlier, scrapping nearly half of their proofs of concept before production. Those numbers measure what happens after a purchase - usually a large, loosely specified one. The French data measures what happens before: offered a big generic first step, companies simply decline to take it. Two symptoms, one disease: the missing object is a small, well-specified first purchase.
2. Why awareness doesn’t trigger a purchase
An awareness session is designed to answer “what is AI and why should I care”. It is structurally unable to answer the three questions a buyer actually has on the way out:
- What should we buy first? “Use cases exist in your sector” is not an answer. A catalogue of 54 examples and 88 vendors transfers the research work to you.
- What should it cost? Without a reference price, every quote feels either suspiciously cheap or unjustifiably expensive, so the safe decision is no decision.
- How will we know it worked? Nobody wants to be the person who spent the budget on something whose success can’t be demonstrated.
If your company has been “in the AI conversation” for a year with nothing launched, the diagnosis is rarely scepticism or lack of budget. It is that nobody converted the interest into a purchasable first step: an object with a scope, a price, and a test. Left unanswered, these three questions produce exactly the two failure modes the statistics describe - no purchase at all, or a vaguely scoped pilot that drifts into pilot purgatory because nobody defined what it was supposed to prove. That conversion is work, and this article is that work.
3. The official ladder - and where it stops
The French programme is worth examining even from outside France, because its structure is typical of how institutions - governments, but also large vendors and consultancies - build AI adoption ladders:
| Rung | What it is | What it costs | Who it fits |
|---|---|---|---|
| Awareness session | A talk by an accredited ambassador | Free | Everyone |
| Self-serve resources | Use-case platform, vendor catalogue | Free | Everyone with time to research |
| Subsidised diagnostic | 8 days of consulting: technical review, use-case identification, prioritisation | €10,000 (excl. VAT), 40% state-funded - ~€6,000 out of pocket | Companies with 10 to 2,000 employees |
| Accelerator | 18-month programme: 12-day diagnostic plus two 13-day modules, 46% funded | Substantial | Companies with 50+ employees and €8M+ revenue |
Details are on the official programme page; figures may evolve.
Look at the gap. Between “free talk” and “€10,000 of general-purpose analysis”, there is nothing - no small, concrete, provable first project. Companies under 10 employees are not even eligible for the diagnostic rung. And the diagnostic itself delivers analysis: a report and a prioritised list, which still leaves the actual first purchase ahead of you.
This is not a criticism of subsidised diagnostics - for some companies they are exactly right (see chapter 7). It is an observation about ladder design: the missing rung is the small first project, and if the institutions won’t build it, you have to specify it yourself.
4. Five criteria of a real first step
Use these five criteria on any proposal - from an agency, a freelancer, a software vendor, or an internal champion. They are provider-agnostic on purpose.
1. One process, not “the company”. The project targets a single workflow you run every week: quotes, meeting follow-ups, support replies, supplier document intake. If the proposal talks about “your AI transformation”, it is a strategy engagement wearing a project’s clothes.
2. Fixed scope, written down. What is in, what is out, what the project explicitly does not attempt. A scope that will be “refined during discovery” is a day-rate engagement with no edges.
3. Price known before signature. A fixed quote for the defined scope. Day rates are legitimate for open-ended work - which is precisely what a first project should not be.
4. A dated deliverable, weeks away. Something running, in your hands, on a calendar date. First projects measured in quarters are pilots that will be renamed “phase one” when the quarter ends.
5. A pass/fail test agreed before you pay. Written down with numbers in it: “the assistant drafts replies for these four request types and the team keeps using it in week four”, or “the meeting workflow produces action lists the project lead accepts without editing at least 8 times out of 10”. If a provider resists defining what failure would look like, that is information.
A proposal that meets all five can still fail - but it fails visibly, cheaply, and with lessons. That is what a first step is for.
5. Three classic first projects
Three shapes come up again and again as sound first purchases. All three fit the five criteria; which one fits you depends on where the pain is.
Training on your real tasks - not generic training. If your team has licences but shallow usage, a half-day to one-day working session built on your actual documents and workflows installs habits that a generic webinar never will. The test is behavioural: are people applying it to their own work two weeks later? In the EU, there is a regulatory tailwind - Article 4 of the AI Act has required organisations using AI to ensure adequate AI literacy of their staff since February 2025 - but buy training for the habits, and let the compliance documentation be a by-product. One small thing that comes up in almost every developer session: teams on Claude, Codex or Cursor subscriptions rarely know how close they are to a usage limit until they hit it, which is why we built Ration, a free open-source menu-bar meter.
One automated workflow. Pick a repetitive flow with a visible weekly cost and automate it end to end - the classic example is meeting notes to assigned actions. The deliverable is a running workflow, not a slide about workflows. This is also the shape where “hours saved per month” is easiest to measure honestly.
One narrow audit. Not a general état des lieux - an audit of one named problem: “why does quoting take four days?”, “can an assistant handle our tier-1 support?”. The deliverable is a decision you can act on, with a recommended (and priced) next step. If an audit’s conclusion is “it depends”, the audit was scoped wrong.
What these have in common: each produces evidence - kept habits, measured hours, a decided question. What none of them requires: an AI strategy, a data lake, a committee, or an 18-month commitment. For a sense of how project shape drives cost structure, see what a software engagement actually costs.
6. How to judge that it worked
Decide the verdict mechanism before the project starts, and keep it small:
- Measure against the pre-agreed test, at a pre-agreed date. Thirty to ninety days after delivery, run the pass/fail test from criterion 5. Not vibes - the numbers you wrote down.
- Count usage, not sentiment. “The team likes it” is weaker than “the team still uses it without being reminded”. Adoption after the novelty fades is the honest signal.
- Time is the cleanest unit. Hours per month, before and after, on the one process in scope. Resist annualised ROI extrapolations from week one.
- A clean negative is a success of the method. “We tested an assistant on tier-1 support; accuracy wasn’t there; we documented why and stopped” is a cheap, valuable outcome - and it beats a zombie pilot that survives because nobody defined failure.
If the first project passes, you now have something no awareness session provides: internal proof, a team that has shipped with AI once, and a shortlist of next candidates. That is the point where broader questions - access, permissions, what goes to production - become real; our pilot-to-production governance checklist covers that stage.
7. When a subsidised programme is the right buy
Fairness requires the reverse case. A large subsidised diagnostic is a reasonable purchase when:
- you are big enough that “one process” is the wrong altitude - hundreds of employees, several plausible AI sites, and the real question is sequencing;
- there is genuine internal disagreement about where to start, and an external, state-framed report has political value inside the company;
- the subsidy covers work you were going to buy anyway at market price.
If you are in France, the programme page lists the current mechanisms, and France Num’s practical guide is a sound free starting point for very small companies. One discipline transfers to any government scheme, French or not: name the step you will take after the report, before you commission the report. A diagnostic with no committed next step tends to become the shelf it is printed on.
Conclusion: the first step is chosen small
The French numbers put a figure on something many leaders feel privately: being informed about AI and knowing what to do next are different states, and the second one doesn’t arrive by attending more talks. It arrives when someone converts interest into a purchasable object - one process, fixed scope, known price, dated deliverable, agreed proof.
You don’t need a strategy to take that step. You need a shortlist of tasks your team repeats every week, and the discipline to buy one small thing whose failure you could afford and whose success you could demonstrate.
You know the tasks. What’s missing is the first step.
Bring one process - the quotes, the meeting follow-ups, the support inbox. In a 30-minute scoping call we will tell you honestly whether it is a training case, an automation case, a narrow audit - or not worth buying yet. Fixed quote if there is something to do; a straight “not yet” if there isn’t. We run our own operations this way - here is what that looks like in practice.
If the answer is training, that is the scope of our corporate AI training; if it is a workflow, it is our n8n AI Automation service.
Frequently asked questions
No - the dependency mostly runs the other way. One delivered project produces the evidence a sensible strategy is built from. Strategy-first sequencing is how organisations end up with a document and no habit.
No. The Diagnostic Data IA is priced at €10,000 excluding VAT for 8 days of consulting, with 40% covered by the France 2030 programme - roughly €6,000 remains with the company, and eligibility starts at 10 employees. Figures current as of September 2026; check the official page before relying on them.
There is no universal number, and be wary of anyone who opens with one. The honest anchor: it should be a fraction of the €10,000 the French state considers a diagnostic worth - because a first project buys a working result on one process, not a general analysis. Insist on a fixed quote against a fixed scope, and compare quotes on the deliverable and the proof test, not the day rate.
If you kept the scope small and defined the test in advance, a failure costs one contained budget and yields a documented reason - which is more than most companies get from a year of meetings about AI. The expensive failure mode is not the small project that misses; it is the open-ended pilot that can’t.
Sources and limits
- French Ministry of the Economy press release, 3 September 2026; DGE - plan Osez l’IA programme page; LeMagIT, 4 September 2026; France Num; MIT NANDA “The GenAI Divide”, August 2025, via Fortune; S&P Global Market Intelligence 2025 survey, via CIO Dive.
- Facts checked on 11 September 2026. Programme prices and eligibility rules are volatile; verify them on the official pages before acting.
- The 70-diagnostics figure counts one public mechanism and says nothing about private AI purchases; the government’s claim that adoption “tripled in two years” is its own and was not independently verified.
- The MIT 95% figure is a contested estimate of pilots’ P&L impact and the S&P Global 42% counts abandoned initiatives - different populations and definitions; they are directional context, not one number.