AI training in companies. 80% lead to nothing, here is why

A statistic to open the ball. 57% of French companies have been deploying generative AI use cases for over a year (Usine Digitale, 2025), but only 1% reach AI maturity (McKinsey State of AI, 2025). Between the two sits a graveyard. And in that graveyard there are mostly AI training PowerPoints. This essay defends a simple, unloved and probably unpopular thesis among classic training providers. Generic AI training is dead. It confuses awareness and adoption, it sells well because it is fundable, and it produces operational emptiness. Here is what replaces it.

Why does generic AI training produce nothing anymore in 2026?

Generic AI training produces nothing anymore because it was designed for a world where AI was a curiosity topic, not a production tool. When 71% of companies already use AI (McKinsey State of AI, 2025), explaining to an executive committee what an LLM is amounts to training a baker to recognize flour. The ROI of an AI awareness session in 2026 is statistically indistinguishable from zero.

The problem comes from a structural confusion between three objects the profession treats as synonyms. Awareness (understanding), acculturation (being able to talk about AI) and adoption (integrating AI into a measurable business process). The market sells 1 while claiming to deliver 3. Because AI is an OPCO/CPF-fundable topic with a Qualiopi certification attached, the offer massified on the least demanding format. The collective acculturation day. The result. Employees walk out happy, HR ticks a box, and zero business frictions are solved.

At Arkange we watch this movie too often. The typical scenario on the market. An organization trains several dozen employees in generative AI with a certified provider. Six months later, the internal verdict fits on two lines. Four or five people use ChatGPT daily, none has changed their working process, and shadow AI has tripled because nobody set a framework. The training was perfectly executed. The operational result looks like a Polaroid underwater.

The thesis in one sentence

The AI training that works in 2026 is not a product. It is an adoption mechanism anchored in a named business friction, with an operational deliverable at the exit, and a usage measurement device in the following 100 days. Everything else is joyful pedagogy subsidized by the OPCO.

This thesis has four uncomfortable consequences. The format matters more than the content, the trainer’s profile matters more than their certification, the measurement horizon matters more than post-session satisfaction, and the sponsor matters more than the budget. Let us take them one by one.

First sub-argument. The format matters more than the content

The format of AI training determines 80% of the adoption result. The same pedagogical content delivered in a 3-hour plenary, in a 2-day workshop on business use cases, or in 1-to-1 sparring on a real project produces three radically different adoption curves, while the slide deck is identical. This is documented by Stanford Digital Economy Lab (Brynjolfsson et al., 2026, study of 51 successful deployments). Organizations that combine training with simultaneous operational deployment have a production go-live rate 4.3 times higher.

The alternative that works at training scale. Co-building learning paths by job family with measurable individual deliverables. An HR learner does not take the same AI training as a wealth management learner, and the assessment is done on a prompt asset built for their future role, not on a multiple-choice quiz. On large-scale deployments in vocational education or continuing training, this segmentation by job family radically changes the post-training usage rate.

The structural limit of plenary formats, observed at scale across the market. A clean presentation, demos that work, thanks at the end of the session, and six months later a measured adoption rate close to zero. No use case triggered, no budget unlocked, no pilot launched. This is not a facilitation quality question, it is a format construction question. Arkange replaced that format with the IAkhathon. 48 hours, mixed teams, one prototype per pair at the end. The project trigger rate at 100 days is radically different.

Second sub-argument. The trainer’s profile matters more than their certification

A Qualiopi AI trainer who has never deployed an agent in production tells things that are correct but useless. The Qualiopi certification (which Arkange holds, to be clear, because our clients need OPCO funding) guarantees a pedagogical process, not field experience. And AI in companies is a subject where the gap between theory and real deployment is immense. According to the a16z Enterprise CIO Survey 2025, 37% of large-company CIOs now run 5 or more AI models in parallel, a subject no trainer without operational experience can treat seriously.

In a demanding industrial environment, the typical request is not classic AI training but a Right Hand. A senior consultant embedded in the teams to transfer method and architecture during the deployment. Formally, it is training. But the deliverable is not a certificate. It is an agent in production and an internal team able to operate it.

Direct implication for a leader. Before signing an AI training quote, ask the provider how many agents their team put into production over the last twelve months. If the answer is vague, it is pedagogy. If the answer is precise and quantified, it is operations. The difference is financially non-trivial.

Third sub-argument. The measurement horizon matters more than post-session satisfaction

Post-training AI satisfaction is a misleading indicator. Short sessions usually get the best scores because they are the most comfortable. Little friction, a lot of enthusiasm. The real indicator is the usage rate at 100 days, meaning how many participants actually use the tools in their business process three months after the session. At Arkange, on the Workshop trainings delivered in 2025, the delta between day 1 satisfaction (94%) and active usage at day 100 (38%) illustrates the problem exactly.

That is why we structure every significant AI training with the 10/30/100 method. 10 days to produce a first concrete deliverable with the participant, 30 days to measure usage adjustment in real conditions, 100 days to document the ROI. This time frame applies to every internal AI program we instrument, not because it is marketing, but because it is the only way to separate useful training from a financial product.

On an AI training program deployed in a professional services firm, the right KPI followed by management is not the post-session satisfaction rate, but the number of case files processed with AI assistance at day 100. The training is judged successful on that KPI, and only on that KPI.

Fourth sub-argument. The sponsor matters more than the budget

An OPCO-funded AI training without a real executive sponsor is an administrative expense. The AI training that produces adoption has a named operational sponsor, who carries an identified business friction, who unlocks the organizational arbitrations and who is accountable for the result. It is rare. It is uncomfortable. It is the only reliable predictor.

On an executive AI acculturation program in the local public sector, the right sponsor is not HR (who funds), it is the transformation directorate, with a clear mandate. Identify 3 business processes to automate within the 6 months following the training. This sponsor-friction-deliverable alignment is what separates a profitable training budget from a tracked training budget.

The Edelman-LinkedIn B2B Thought Leadership 2024 study quantifies this dynamic differently but converges. 63% of C-suite leaders say quality thought leadership influenced a purchase decision, which means the decision is made at the top, and training without a sponsor at the top is an effort without a lever.

Direct recommendation by profile

For an executive committee or general management

Do not order a 3-hour collective AI awareness session. It is ineffective and it is demonstrated. Order instead a short executive masterclass (90 minutes maximum) followed immediately by a 1-to-1 Sparring Partner with each executive committee member on a business friction they carry. ADA cycle applied. Audit of the portfolio, selection of a priority friction, operational deliverable within 30 days. Realistic budget. Between €15,000 and €40,000, OPCO-fundable on the training part.

For operational teams (HR, finance, ops, expertise)

No generic AI training. An IAkhathon or Workshop format on real business use cases, with individual deliverables (prompts, mini-agents, automations) built during the session. Usage measurement at 30 and 100 days mandatory. Eligible for OPCO/CPF funding through Qualiopi. This is the format that moves teams (accounting, finance, services) from an exploratory stance to an adoption stance.

For technical experts (CIOs, data, dev)

A custom plan with an embedded Right Hand for 2 to 4 sprints. No training in the traditional sense. A transfer of method and architecture during a real deployment. This is the format that produces an autonomous internal team able to operate the agents without permanent external dependency, particularly suited to demanding industrial environments.

What it changes for you, concretely

If you have an AI training budget planned for next quarter, ask the provider three questions before signing. How many AI agents has your team put into production this year? What is the expected operational deliverable for each participant at the end of the session? How do you measure usage at 100 days? If the three answers are not quantified and precise, you are buying awareness dressed up as transformation. And the generic AI training market, in 2026, is dying of exactly that confusion.

AI is not bought, it is operated. AI training is not consumed, it is deployed.