There are two ways a supermarket gets its AI, and only one of them employs anybody

In July the Office for National Statistics asked British firms how artificial intelligence reaches them. In wholesale and retail the most common answer was free-to-use software. I wrote about that on 4 August and said those industries were getting AI the way you get a weather app.

Then Tesco and Sainsbury's published their year-end filings, and I have something to add to it.

The occasion is February, and the reason to look now is that the documents only arrived in April. In the last week of February both companies announced changes to their office teams, reported at the time as AI reaching the back office. Tesco published its annual report on 15 April. Sainsbury's published preliminary results on 23 April.

Start with the mechanism, because there are two ways a company acquires this capability and they are not versions of one thing.

The first route is to download a general tool and use it on top of whatever the business already does. There is no contract, no owner, no budget line and no team. The capability sits in whoever happens to be using it, and it leaves the company the day they do.

The second route is to build it inside the business. That means systems trained on the company's own data, engineers to keep them running, and a department with a name and a headcount. The first route is a subscription. The second is an employer.

The filings say which route these two took. Sainsbury's records that every food product now runs on a machine learning forecasting platform, and that it has set up an AI Centre of Excellence. Tesco's chief executive wrote that the company has doubled the size of its tech team over the past six years. Neither of those is a downloaded tool.

Which is where February matters. Sainsbury's put around 300 roles at risk, the majority in technology and data. Tesco removed around 180 head office roles and created around 250. At two of the largest employers in British retail, the part of the business being resized is the part that builds the capability, not the part it was supposed to replace.

Now the correction to my own piece. The ONS counts firms, not people. Counted by firms, retail overwhelmingly downloads, and that finding stands. Counted by employees, retail is dominated by a handful of companies that build, and Tesco alone reported around 342,000 people worldwide at the end of February. The sector average is accurate and it describes almost nobody working for the sector's biggest employers. I did not draw that distinction, and it changes what the number means.

The systems reading is concentration. If this capability lives inside a small number of in-house teams, then the AI industry, as a destination for a range planner whose job has been consolidated, is measured in thousands of posts rather than hundreds of thousands. And those teams are now being resized themselves.

Simplifying slightly, because neither company publishes a split between office and store headcount: we can see which door the capability came in through. We cannot yet see who left by the other one. What would change my reading is a filing showing office numbers falling while the technology team grows. Nobody has published one.