The industries that employ the most people are getting AI the way you get a weather app.

Take the ONS release of 20 July and set the adoption headline aside. The more revealing question is how the technology gets into a firm in the first place, because the route determines who controls it afterwards.

The survey asked. Firms in manufacturing, in wholesale and retail, in hospitality, and in admin and support most commonly report getting AI through free-to-use software. Firms in construction, in information and communication, and in professional and technical work are more likely to buy it.

The split matters more than it looks. A purchased system arrives with a contract, an owner, a budget line and somebody accountable when it fails. Free software arrives with none of those. It arrives because one person opened a browser.

Now add the training figure. Eleven per cent of businesses with ten or more employees report that more than half their workforce has had AI-related training. Around 40 per cent of medium and larger businesses report integrating AI skills by training or retraining existing staff, which is the most common approach by a distance, but the depth behind it is thin. Hiring people who already have the skills runs at about 2 per cent among the smallest firms and 10 per cent among the largest.

So the pattern in the sectors employing most of Britain is a tool with no contract, no owner and no instruction, used by people who taught themselves.

Three consequences follow, and none of them requires assuming the technology is dangerous.

The first is that a firm cannot see its own exposure. Forty-one per cent of businesses report no barriers at all to adoption in the last three months, which reads as a clean bill of health and may equally mean nobody has looked.

The second is that the capability sits with the individual rather than the organisation. Where staff have taught themselves on free tools, the skill leaves when they do, and the firm has no record of what it lost.

The third is a measurement problem the ONS raises about itself, which is unusual and worth noticing. Software with no market price is hard to place in the national accounts. If a large share of Britain's AI capability arrives free, the standard instruments for valuing capital investment will not see it. The statisticians say they are working towards a recommendation on whether an AI thematic account is feasible by March 2027.

Everything above is measured. What follows is inference, and the join is worth marking: this has the shape of an unmanaged dependency. Not a crisis, and not evidence of harm. A system where the load has moved to a layer with no budget, no owner and no measurement, which is the condition under which small failures stay invisible until they are large ones.

The narrow point for a reader is this. If your employer has never mentioned the tools you use every day, that is not an oversight peculiar to your firm. On these figures it is the ordinary case.