A managing director who has been reading the technology press for a year can hold two contradictory beliefs at once: that artificial intelligence is advancing fast enough to make every decision urgent, and that inside their own business nothing has really changed. Both are true. It is precisely that gap which the 2026 edition of the AI Index, published by the Stanford Institute for Human-Centered AI, makes it possible to measure (Stanford HAI, AI Index 2026). That gap is not closed with one more tool: it is closed by AI literacy.
Our reading fits in one sentence: what is missing is no longer the technology but AI literacy — knowing what the tool does, what it does not do, and who decides when it gets things wrong. The first obstacle executives name is neither price nor regulation: it is a gap in knowledge.
Capability is no longer the limiting factor
The report documents a year of technical acceleration. On SWE-bench Verified, a test that requires fixing real bugs in real code repositories, the best models went from about 60% to close to 100% of the human reference level between 2024 and 2025 (AI Index 2026, p. 75). The nuance is worth keeping: this is a ratio to the human level, not an absolute score — on the test itself, the best models plateau around seventy to seventy-five per cent.
That progress comes almost entirely from the private sector: 91.2% of notable models recorded in 2025 were produced by industry, against a handful from academia (AI Index 2026, p. 19). For a business, the consequence is simple. Waiting for the technology to mature is no longer a reasonable waiting strategy: it already is mature for a large number of tasks, and that is not where the difference between two companies in the same sector is decided.
The same acceleration shows on the agent side — systems that chain actions inside software rather than answering a question. On OSWorld, a test that involves completing real tasks in a desktop environment, the success rate went from about 12% to 66.3%, six points short of the human reference measured at 72.35% (AI Index 2026, p. 73 and 113). A third of attempts still fail; but the trajectory, in a single year, leaves little room for the “not mature yet” argument.
Everyone says they have adopted AI
The adoption figures are spectacular, and that is where one should read slowly. 88% of the organisations surveyed report using AI in at least one function, against 78% the year before; 79% report using generative AI (AI Index 2026, p. 193). These figures come from a self-reported McKinsey survey of 1,993 respondents across 105 countries. They say what executives answer on a questionnaire, not what their teams actually do. The report itself calls the results directional rather than comprehensive.
Among the general public, diffusion is real and fast: generative AI reached 53% population adoption in three years, faster than the personal computer or the internet in their early days (AI Index 2026, p. 11). France even ranks fifth among the thirty largest economies, with 44.0% of users in the second half of 2025 — ahead of the United Kingdom and Germany, and well ahead of the United States, twenty-fourth at 28.3% (AI Index 2026, p. 201).
In other words: your staff already use AI. The question is not whether it enters the business, but whether it enters by a decision or through the back door.
That back-door entry has a cost you do not see straight away. Documents end up in personal accounts, answers are checked by nobody, and the business loses track of what was produced with what help. The day a client, an auditor or an insurer asks the question, there is nothing to show — no written rule, no log, no named owner.
Almost nobody has deployed it at scale
This is the most interesting figure in the report, and the least quoted. When the same organisations are asked about their use of agents — systems that carry out tasks end to end, not assistants you query — the most frequent answer, in almost every function, is no use at all. Even in IT and knowledge management, the two most advanced functions, at least two-thirds of respondents report none (AI Index 2026, p. 197). Elsewhere it is worse.
The report adds a correlation that speaks directly to smaller businesses: the companies with the highest revenue are the ones that most often reach the scaling phase. It would be convenient to conclude that size is everything. We think the opposite is true: large organisations do not succeed because they are large, but because they have the functions that size pays for — someone whose job it is to frame, measure and arbitrate. A small business does not have those functions. It can have the decision.
The words matter. “At scale” does not mean “everywhere”: it means a use has left the trial stage, runs on the real flow, has someone answering for it, and produces a result that can be measured. Between a promising trial and that stage there is rarely a technical problem. There is a decision to take about what you agree to hand over, a threshold to set, a procedure to write for doubtful cases. That work cannot be delegated to a supplier.
The bottleneck is not technology, it is AI literacy
Asked what prevents them from implementing responsible AI, executives cite first a gap in knowledge and training, then resources, then regulatory uncertainty (AI Index 2026, p. 145). None of those three obstacles is technological.
| Obstacle cited | Share of executives | What it means in practice |
|---|---|---|
| Knowledge and training gaps | 59% | Teams know neither what the tool can do nor what it cannot. So they use it everywhere, or nowhere. |
| Resource or budget constraints | 48% | The project exists on paper, but nobody has the time to carry it through to a measured result. |
| Regulatory uncertainty | 41% | People wait for clarity. The waiting ends up standing in for a decision nobody actually took. |
These answers come from a survey run by the AI Index team with McKinsey, among business executives, excluding China. They describe a problem of collective competence, not of equipment. And collective competence is not bought with a licence: it is built by running a real process through the tool, watching what breaks, and writing the rule that stops it happening again.
AI literacy is not a training day. It is a state in which everyone, at their own post, can answer three questions: which tasks am I allowed to use it for, how do I recognise a doubtful answer, and who do I flag it to. None of that requires understanding how a model works. And the gap sometimes runs the other way: with 44% of users in the French population, part of your team probably already knows more than the business does.
The gains are real, but not where they are announced
The report gathers several field studies, and their results are not comparable with one another: 14% to 15% more tickets handled per hour in customer support, 26% more contributions from developers equipped with an assistant, around 50% more advertising output per employee in marketing (AI Index 2026, p. 174). Three different metrics, three different contexts. None of them promises your business the same gain.
One study points the other way, and it is the one to read in full: with experienced open-source developers working on their own code, using AI tools lengthened completion time by about 19% — while the participants believed they had been faster (AI Index 2026, p. 219). The result has not been replicated since, so it is not the last word. It says something else: the perception of a gain is not the gain. That is exactly why we ask for a measurement beforehand, not only an impression afterwards.
Finally, the report flags a concern that is not yet an established finding: heavy reliance on AI may slow skill development over the long term (AI Index 2026, p. 174). We devote a whole article to it, because it is the hardest decision a manager has to take, and the most expensive one to take too late.
One gap remains that no tool will close: 73% of the experts surveyed expect a positive effect of AI on how people do their jobs, against 23% of the general public (AI Index 2026, p. 12). Fifty points. When a management team announces an AI project, it speaks from the first group to people who belong to the second.
Where to start when you are a small business
The sequence we apply never starts with a tool. It starts with a process that already exists, that someone carries out every week, and whose cost in time can be stated. Sorting incoming requests, checking supplier documents, preparing a quote: it hardly matters, as long as it is real and measurable.
That process is then broken down step by step, between the mechanical gesture, the reflex decision, the writing and the human judgement. It is that breakdown, not the choice of model, which decides the outcome: we set it out in not every decision deserves an LLM. Before extending, two safeguards are enough: knowing which tasks genuinely lend themselves to it, and writing the rules of the game on one page.
Last comes the question of data, which is not a legal formality at the end of a project. A model placed at the heart of a process sees everything that process handles. Where the computation runs, under which law, with which model: three decisions for the person in charge, and we devote an article in the same series to them. For a first overview, where to start with AI in a small business without exposing your data runs through the approach end to end.
Pace matters as much as method. A first use is framed in a few days, tried for a few weeks on the real flow, and judged against a measurement taken before starting. If no gain appears, you stop: that is a result, not a failure. This discipline, more than the choice of tool, separates the businesses where AI eventually holds a place from those where it stays a subscription line nobody dares cancel.
AI is not plateauing. Organisations, on the other hand, always plateau in the same place: where nobody has the time to learn, or the mandate to decide. That is good news for a small business, because those are two things a small business can settle in a few weeks, with no research budget and no transformation department.
Common questions
How can AI help a business?
On repetitive, bounded tasks, measured gains range from 14% more tickets handled in customer support to around 50% more advertising output in marketing. On tasks that require judgement, results are far less certain, sometimes negative.
How do you introduce AI into a business?
By starting from a real process rather than a tool. Break it down step by step, hand AI the steps that suit it, measure the time saved, and write the usage rule. One accountable person beats a committee.
How many businesses use AI?
88% of the organisations surveyed by McKinsey for the AI Index 2026 report use in at least one function. It is a self-reported survey: use at scale remains rare, with most respondents running no agents in production.
How do you make an AI project safe?
Three safeguards are enough to begin with: a list of approved tools, a list of forbidden data, and human validation on decisions that commit the business. The rest is added with use, once you know what the team actually does with the tool.
Source: Stanford HAI, Artificial Intelligence Index Report 2026, published on 29 June 2026 under a CC BY-ND 4.0 licence. Page numbers refer to the PDF file. Adoption rates come from self-reported surveys — McKinsey, The State of AI in 2025, and the survey run by the AI Index team — not from usage measurements. This article is an analysis by Maeliom; it is neither a translation nor an adaptation of the report, and is neither affiliated with nor endorsed by Stanford University.