The question that comes up in every management meeting — “is AI ready?” — has no answer, and that is what makes it dangerous. Depending on the task you have in mind, the honest answer runs from “for two years now” to “not at all”, with no simple rule to tell which applies. The limits of AI cannot be deduced: they have to be observed.
The 2026 edition of Stanford’s AI Index provides the clearest illustration we have read (Stanford HAI, AI Index 2026). The limits of AI do not form a regular frontier crossed year after year: they form an irregular ridge line, excellent here, absurd three centimetres further on. Hence the purpose of this article: a grid for knowing which side a given task falls on.
The limits of AI fit into a gold medal and a clock
In 2025 a model took a gold medal at the international mathematics olympiad, a test where the world’s best school students regularly fail (AI Index 2026, p. 72). The same family of models reads an analogue clock correctly 50.6% of the time, where humans manage 90.1% (AI Index 2026, p. 96).
The juxtaposition is not an amusing anecdote: it is a methodological warning. A manager who concludes from an impressive demonstration that the tool will manage “the rest” is wrong in both directions. They will overestimate AI on apparently simple tasks — reading a plan, counting objects, comparing two dates — and underestimate it on tasks they assume are out of reach. It is the same gap measured in the article that opens this series, between what is declared and what is deployed.
Agents are progressing fast and still failing
On OSWorld, a test requiring real tasks in a desktop environment — opening a file, filling in a form, chaining applications — the success rate went from about 12% to 66.3% in a year, six points short of the human reference measured at 72.35% (AI Index 2026, p. 73 and 113).
Two readings of that figure are possible, and both are true. The trajectory is spectacular and makes any two-year forecast reckless. And one failure in three is still one failure in three: on a process running a hundred times a day, that means thirty-three manual recoveries, or thirty-three errors if nobody is watching. The question is therefore not the model’s performance, but what its failure costs inside your process.
AI confuses knowing and believing
The phrase AI hallucination covers very different realities, and the report has the merit of separating them. On a benchmark of hard questions, measured rates range from 22% to 94% depending on the model (AI Index 2026, p. 136). On an exercise closer to business use — summarising a supplied document — the same models fall between 1.8% and 5.4%. Quoting the first figure without the second, or the reverse, means telling two opposite stories from the same report.
The most disturbing result is elsewhere. When a false statement is presented to the model as the user’s own belief, its performance collapses: a model that distinguished true from false more than 90% of the time drops to 14.4% (AI Index 2026, p. 138). In office language: the tool tends to agree with you. Anyone who arrives with a preconceived idea and asks for confirmation will get it, and will leave more confident in their error.
This property has a direct consequence for choosing use cases. AI is more reliable the more the expected answer can be checked by someone other than the person who asked. It is also what argues for handing closed decisions to simpler mechanisms, as we showed in not every decision deserves an LLM.
Where the real gains are
The studies gathered this year draw a clear line. Where the task is repetitive, framed and measurable, the gains are substantial: 14% to 15% more tickets handled per hour in customer support, 26% more contributions from equipped developers, around 50% more advertising output per employee in marketing (AI Index 2026, p. 174). Three different metrics, three different studies: they do not add up and do not transpose as they are.
Where the task calls for judgement, context or fine knowledge of what exists, the result degrades, and sometimes reverses. The most instructive study covers experienced developers working on their own code: equipped with AI tools, they took about 19% longer, while being convinced they had been faster (AI Index 2026, p. 219). The result awaits replication. But the gap between perception and measurement is an experience we meet on almost every assignment.
The case of AI in customer service deserves a nuance, because it is the first one people think of and the most often badly framed. The measured gain applies to assisted human agents, not to a bot answering on its own: the tool suggests, the person decides and sends. Substituting instead of assisting changes the nature of the risk — a wrong answer then goes out unseen, and the cost is counted in customers, not minutes.
A four-criteria grid for choosing use cases
From all of the above we draw four questions. They require no technical skill, and are asked before any trial. If a single one gets a negative answer, the use is not forbidden: it needs an explicit safeguard, or it waits.
| Criterion | The question to ask | If the answer is no |
|---|---|---|
| Structured task | Does the expected work fit in an instruction you can write in five lines? | Break the task down before automating anything. |
| Measurable result | Will you be able to say, in a month, whether it worked — in time, volume or errors? | Take the baseline measurement now: afterwards it will be too late. |
| Recoverable error | What happens if the answer is wrong and nobody notices? | Keep the use to internal drafts, never to outputs that commit you. |
| Human validation | Is someone named as reviewer, and do they have the time? | Do not start: a review nobody performs does not exist. |
Three uses pass this grid in almost every small business: sorting and routing incoming requests, preparing a first draft of a document that will be reviewed, and extracting information from structured documents. Three others rarely pass: answering a client directly without review, analysis that grounds an investment decision, and anything touching a regulatory obligation.
Once the uses are chosen, what each person may do with them still has to be written down: that is the purpose of the AI charter. And it remains to look at what those choices do to learning the trade, which we develop in who will train tomorrow’s seniors.
The good news in this edition is not that AI is progressing: it is that its limits are becoming documented, and therefore usable. A business that can name the task, measure the result and name the reviewer does not need to forecast what the tool will manage next year. It will be able to test it in a week, and decide on facts rather than on a demonstration.
Common questions
What are the limits of artificial intelligence?
They do not follow the apparent difficulty of a task. A model can take gold at the mathematics olympiad and read an analogue clock half the time. Limits are established task by task, with a measured trial, never by deduction.
Why does AI hallucinate?
Because it produces the most plausible continuation, not the truest, and has no internal way of signalling its ignorance. Measured rates vary enormously with the exercise: from 1.8% on summarising a supplied document to more than 90% on hard questions.
What are the best AI use cases in a business?
Those meeting four conditions: a task you can describe in five lines, a measurable result, a recoverable error, and a named reviewer. In practice: sorting incoming requests, first drafts of documents that get reviewed, extracting information from structured files.
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. The hallucination rates cited come from two distinct benchmarks, one made of hard questions, the other a summarisation exercise; they are not comparable with each other. 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.