Seeing which strawberry is ripe is the part machines have solved. Everything difficult happens in the two seconds after that.

A strawberry is an awkward object to pick. It bruises under a thumb. It grows in a cluster where ripe and unripe sit touching, half of it hidden under leaves, and it has to arrive in a punnet looking saleable a week later. A person does all of this without thinking about it. Getting a machine to do it means solving three problems in sequence, and they are not equally solved.

The first is seeing. Cameras and a trained model judge ripeness from colour and shape. We described this desk's local example, the Cambridge firm Dogtooth, doing exactly that back in June, and it is the part that now works well enough to be dull. The second is reaching, which is harder than it sounds: the arm has to plan a route to one berry through a tangle of other berries, stems and leaves, and arrive without disturbing any of them. The third problem is the one that decides whether you have a picking robot or an expensive way of making jam. Separating fruit from stem and setting it down, at a force low enough that no bruise shows up three days later. Grip is the whole business.

Britain has two visible attempts running. Dogtooth, in Cambridge, sells a machine it prices at £30,000. The company says it picks between 200kg and 300kg a day, that one person can keep twelve of them working, and that about three are needed per hectare, giving an average payback of two years. Those are its own figures, given to prospective buyers, and I have found no independent measurement of any of them. The machine still needs a person to move it between rows and handle the crates. At Tiptree in Essex a different system altogether, built at the University of Essex on a £1.02m grant through Defra's Farming Innovation Programme, is reported to take around two and a half seconds a strawberry and has been adapted to lettuce.

Defra's own review of automation in horticulture is more sober than either. It sorted the field into six groups of technology. Three are ready for wide adoption now: improved production systems, packhouse automation, and field rigs. The review's assessment of those three was that they deliver minimal labour savings, because what they do is make a picker more productive rather than remove the need for one. The three that would replace a picker rather than assist one, including autonomous selective harvesting, were still in development, much of it at prototype. The review also says something the funding announcements tend to leave out: some crops may never be picked by machine at all, because they need a gentleness and a judgment that does not reduce to a rule.

Which matters, because the country is not hedging. The seasonal visa allocation has come down three years running to 42,900 for 2026, and 42,300 people applied in the twelve months to June, closing the gap to about six hundred. The people are being subtracted on a published schedule. The machines are on nobody's schedule.

The fragility here is the ordinary sort rather than the dramatic sort. A farm that has planned around twelve robots and one supervisor has exactly one supervisor. Soft fruit does not wait: a variety comes good over a fortnight and that fortnight is the year. When something breaks in it, the old system had a hundred people in a field who could work slightly harder. The new one has a queue for an engineer.