AI agents in 2026: what is actually happening

Two figures about the same thing. A YouGov survey for IONOS, fielded between January and March 2026, puts the share of Spanish SMEs already using AI tools day to day at 41%. Spain’s national statistics office, in its official ICT survey, records 21.1% among companies with ten or more employees and 13.4% among those with fewer than ten.

They do not contradict each other: they measure different things. The survey captures self-perception across a sample of 514 responses; the official statistic records systematic use. Between “someone at the company uses ChatGPT” and “AI sits inside a process” there is a wide gap, and that gap is exactly where most projects die.

This year was announced as the year of autonomous agents. It pays to read the original sources rather than the headlines. Microsoft’s report on 2026 trends, published in December 2025, is far more cautious than what gets attributed to it: it describes agents as “digital coworkers” acting under human direction, and its second trend is not about autonomy but about safeguards. Their security lead sums it up by saying an agent joining the workforce needs controls comparable to those of an employee.

Gartner puts numbers on both sides. On one hand, it forecasts that 40% of enterprise applications will embed task-specific agents by the end of 2026, up from under 5% in 2025. On the other, it estimates that more than 40% of agentic AI projects will be cancelled before the end of 2027: runaway costs, unclear business value and insufficient risk controls.

A third figure explains much of the noise. Gartner calls it “agent washing”: of the thousands of vendors claiming to sell agents, it reckons only around 130 genuinely do. The rest is, at best, a chatbot with a new label.

McKinsey points the same way from the customer side: 88% of organisations report regular AI use, but no more than 10% have scaled agents in any specific function. Plenty of pilots, little production.

For a company considering the move, the useful question is not whether to adopt agents. It is which specific process is worth automating. And that question is answered by looking at the problem, not at the tool catalogue.

Our criterion is simple: an agent makes sense when the task is repetitive, its output can be verified, and someone is there to supervise it. If any of the three is missing, what you get is a pilot that quietly dies within six months.

2026 is not turning out to be the year of autonomous agents. It is the year agents move from demo to production with a person watching, and the year the market starts telling the real thing from the label. Less spectacular, far more useful.