01 — Editorial
Why this guide
Each summer, the Stanford Institute for Human-Centered AI publishes the most complete survey the field has. The 2026 edition runs to four hundred and twenty-five pages, it is in English, and its data looks first at the United States. The head of a smaller business will not read a line of it. They will hear about it through press headlines, alarmed or enthusiastic depending on the day, and both versions will ask them to decide something.
We read it in full. This document is not a summary of it: the report’s licence forbids that, and a summary would be of no use anyway. We kept five findings out of the fifteen it highlights, set aside what does not concern a company of twenty to two hundred people, and took a position wherever the figures allow one.
One precaution holds for everything that follows. Many figures are American, and we say so every time: what transposes and what does not is part of the analysis. Adoption rates come from self-reported surveys; they say what executives answer, not what their teams do. Every figure cited carries its page number.
This guide can be read three ways. In five minutes, the five findings are enough. In twenty minutes, the full read gives the reasoning. And if what you want is something to do on Monday, the last two parts — the roadmap and the two tools — stand on their own.
Maeliom
02 — Five findings
Five findings, not fifteen
The report highlights some fifteen takeaways. Only five change anything for a smaller French business, and here they are, with their figures.
The third finding is the one nobody quotes, and it is the most useful. Asked about their use of agents — systems that carry out a task end to end — the same organisations most often answer “no use at all”. Even in IT and knowledge management, the two most advanced functions, at least two-thirds of respondents report none (p. 197). Declared adoption and real deployment do not describe the same world.
02 — Five findings
Two findings people would rather forget
The fourth is about how uneven performance is, and it fits in a juxtaposition. One model took gold at the international mathematics olympiad (p. 72); the best model tested reads an analogue clock correctly 50.6% of the time, where humans manage 90.1% (p. 96). AI is neither superhuman nor stupid: it is uneven, and that unevenness does not follow the apparent difficulty of a task.
The fifth concerns safeguards, and it points the wrong way. The field’s reference database recorded 362 AI-related incidents in 2025, against 233 the year before (p. 132). Over the same period, the share of organisations where a dedicated role leads AI governance rose from 14% to 17% (p. 142): more than eight in ten still have nobody named.
03 — Our reading
France is not lagging behind
This is the position we hold most firmly, because it contradicts a story we hear every week. On the diffusion of artificial intelligence through the population, France ranks fifth among the thirty largest economies in the second half of 2025, with 44.0% of users, up 3.1 points. The United States rank twenty-fourth, at 28.3% (p. 201).
The rest points the same way. France is the European Union’s largest private investor in AI, with $4.36 billion in 2025, fourth worldwide (p. 182). It is Europe’s second contributor of models published since 2018, behind the United Kingdom (p. 336). And it is one of five countries — with Canada, Germany, Italy and Japan — that have endorsed every major international AI governance initiative (p. 151).
None of this says France will produce the next frontier model: infrastructure and capital remain massively concentrated elsewhere, and the guide shows that too. But the story of French lag, imported from an American debate about the model race, leads managers to ask the wrong question. The question is not how to catch up. It is how to give structure to what is already happening, often with no rules and nobody accountable.
This is an extract from the guide. The four other positions, the six manager’s questions and the two tools are in the full document, sent by email.
Common questions
What is the Stanford AI Index?
An annual report published by the Stanford Institute for Human-Centered AI, gathering the available data on artificial intelligence: model performance, adoption, investment, regulation, incidents. The 2026 edition runs to 425 pages and appeared on 29 June 2026.
What should a smaller business take from the AI Index 2026?
That capability is no longer the limiting factor, that declared adoption far exceeds real deployment, and that the obstacles executives cite are knowledge gaps, budget and regulatory uncertainty — none of them technological.
Is this guide a summary of the report?
No. The report is published under a licence that forbids distributing an adapted version. This guide is an original analysis by Maeliom, drawing on the report’s figures, placing them in a French business context and taking a position. The data is sourced page by page.