A London company was built so machines could find new medicines. Its own accounts show what happened to the people

Start with the document, because almost everything written about this company has come from somewhere else.

BenevolentAI was founded in London on a single proposition: that a machine could do the discovering. Not assist a chemist, not speed up a search, but choose the molecule. It floated in Amsterdam in 2022 and it no longer exists in the form it had then. What follows is not from a press release. It is from accounts filed at Companies House and signed off by auditors.

The British parent, BenevolentAI Limited, files a note headed staff numbers and costs. It reports the average number of people employed across each year, split into two categories, research and development, and administration.

In 2020 the group averaged 273 people, 223 of them in research. In 2021, 309, with research at 256. In 2023, 343, with research at 277. In 2024, 221, with research at 170.

There is no figure for 2022. The company was owned by a listed parent registered outside the United Kingdom, and British subsidiaries in that position are exempt from filing group accounts. The numbers went dark in the years the company was most visible to investors, and came back when the listing collapsed. That is worth knowing before anyone builds a trend line.

The averages are one measure. The accounts also give a straight count on the day. At the end of 2023 the group employed 248 people. At the end of 2024 it employed 173. That is 75 people, thirty per cent of the company, inside one calendar year.

Now the part that rules something out.

Divide the wage bill by the number of people and you get what the company paid per head. In 2020 it was about 90,900 pounds. In 2021, 88,800. In 2023, 93,900. In 2024, 89,000.

Five years, a collapse in the middle, and the figure barely moves. So this was not a company shedding its expensive people and keeping the cheap ones, and it was not the reverse either. Whatever happened, it happened roughly evenly across the payroll.

The same discipline applies to who went. Of the 122 posts lost between the 2023 and 2024 averages, 107 were research posts. That sounds like researchers being singled out, and it mostly is not. Research was already 81 per cent of the workforce. An even cut would have taken about 99 research jobs. It took 107. The difference is eight people, and on an administration base of 66, a handful of individuals moves that figure several points. This desk is not going to build an argument on eight people.

Three things this record cannot tell you. It cannot separate the laboratory scientist from the software engineer, because the accounts fold both into research and development. It cannot show the peak, because 2022 is missing and always will be. And the 2023 group figures carry the word unaudited in the accounts themselves, because no group accounts were prepared that year, so the only comparison available to us runs from an unaudited year to an audited one.

What is established, and it is enough. A company built so that machines could find medicines employed 277 researchers in 2023. By the end of the following year the whole group was 173 people. Every figure above is on the public record in Britain and has been sitting there since October last year.