The lab Prentis has existed since April this year. It has a little over 25 employees. And it is negotiating to raise 100 million dollars (around 92 million euros) at an estimated valuation of one billion dollars. Four months of existence, a billion in value - that is maths that works in exactly one industry on earth.

What Prentis is building is not another conversational model. The company trains its models by watching office workers click through documents and systems, in order to build agents that will operate the computer themselves and finish the job. Insurance claims processing. Customs refunds. Precisely those tasks where a human today searches folders and enters data into three different systems.

The founder and chief executive is Ritankar Das, aged 31. The youngest recipient of Berkeley's University Medal in more than a century - he graduated at 18 with a double major, then a master's at Oxford, then an abandoned doctorate at Cambridge. Since 2014 he has run a holding company called Titan, which he describes as a deliberate return to the old Berkshire Hathaway model - a holding financed from its own sales rather than from outside funds.

The other two names on the business card carry more weight. Reid Hoffman, co-founder of LinkedIn and partner at Greylock, an early investor in OpenAI, the man who stepped down from Microsoft's board last month after nearly a decade. And Mark Pincus, founder of Zynga, who now runs an investment firm with Hoffman as an adviser. For both of them, Prentis is a side project. For the valuation, their names are half the argument.

The figures the company puts on the table itself

According to people familiar with the negotiations, Prentis has already signed contracts worth up to 50 million dollars with several clients - a healthcare management organisation, a manufacturer, and producers of goods and clothing. Investor materials project an annual revenue run rate of around 75 million dollars by the third quarter of this year.

Here it is worth reading the small print, because the company wrote it. Those figures, its own presentation states, are an estimated annual value calculated on the basis of a contracted fee of 20 percent of the savings - not recognised revenue - and are "outcome-dependent and subject to final execution". In other words: if the software saves money, there will be revenue. If it does not, the figure is a wish on paper. That is not fraud, it is a normal way of calculating in this industry - but it is not the same as money in the account, and headlines routinely treat it as though it were.

The company claims its Hive-32B model outperforms OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6 on two computer-use benchmarks. It also claims it costs about ten times less per task than the leading models because it is smaller and cheaper to run. None of those results has been independently verified.

The market it is entering is not empty. Anthropic, OpenAI and Mira Murati's lab are working on the same thing, and this year Anthropic bought the Seattle startup Vercept outright, kept the founders and shut down its product. When the giants buy people and switch off products, that is a signal of how seriously they take this field.

The work these agents automate - data entry, claims processing, administrative searching - is exactly what the region sells as a service to foreign clients. This is not a distant American venture capital story. If a billion dollars is flowing into a machine that does that work more cheaply, the question coming is not whether, but when.