Dear reader,
On 27 October Anthropic said it had bought Mithril Medicine, a company with one full-time employee (Anthropic). It did not disclose the price. Its founder, Cody Tranbarger, has been developing treatments for two rare forms of muscular dystrophy. He used Claude to organise contracts, logistics and experiments with 11 external labs, and handed it many other parts of the company’s operations. Mithril took one possible treatment from an idea to promising results in mice in less than 18 months. Anthropic expects human trials to start as soon as late 2028.
This is Anthropic’s account of a company it bought partly to show what its model can do, and the work is still preclinical. It is also the clearest example so far of something I see more and more: a very small team that runs its work around AI gets done in weeks what a large company does in quarters.
What Claude did at Mithril
The technical report Mithril published on 26 October gives more detail (report). Under the team’s direction, Claude searched the literature, turned the contract labs’ reports and instrument data into tables, re-analysed and plotted the data, and drafted text and figures. In one week it turned a 341-page data package on 113 antibody constructs into one plan for five contract research organisations. Claude’s first answers came in a median of 96 seconds; the labs took a median of 8.5 hours to reply. Claude also made mistakes. The authors examined six: Claude caught five, one of them only after a lab asked a question, and three reached a lab in emails the team had edited and sent. The decisions stayed with the people. The authors conclude that “the size of a screen need not be limited by the size of the team that directs it.”
Most of the lab work was done by the 11 suppliers. The founder retained control of the drug programme: what to test, with whom, and what to do with the results.
Pleiades and Enveda
The same Anthropic post describes two more cases. The Column Group, a venture investor in biotech, is setting up Pleiades, a company that will use Claude “to run many more—and more complex—drug programs than is usual for a small team”. At Enveda, where chemists had spent up to 75% of their time on paperwork, Claude reviewed 1,061 pages of manufacturing records. It found all 11 problems the human reviewers had found and seven they had missed, and review time fell from over 25 hours to under three. It also wrote the first draft of a regulatory document in four hours; by hand that took about three months. A first draft is not a filed document.
Mithril is not the only one
In June 2025 Wix bought Base44, an app-building platform, for $80m (TechCrunch). Maor Shlomo had started it as a side project six months earlier. At the sale it had nine people, Shlomo and eight employees, and 250,000 users, and in May it made $189,000 of profit.
Lovable, a Stockholm company founded in late 2024 that lets people build software by describing it, reached $400m in annual recurring revenue in February 2026 with 146 full-time employees (TechCrunch). That is about $2.7m of revenue per employee. In August it raised $400m at a valuation of $13.3bn and said employees at nearly two-thirds of Fortune 500 companies use it, with teams at Adidas, NVIDIA and Deutsche Telekom building internal software on it (Lovable).
For comparison, Cognizant, one of the large IT service firms that build and run software for banks and insurers, reported $5.66bn of revenue for the third quarter with 359,900 employees (Cognizant). On an annual basis that is about $63,000 per employee. Lovable’s revenue per employee is roughly forty times higher, although the ratio cannot separate the effect of AI from the difference between a software platform and a services firm. Cognizant’s chief executive, Ravi Kumar S, said more than half of the company’s code is now AI-assisted, and its headcount grew by 10,100 over the year.
The comparison has a limit. Building a new product from a clean slate is much easier than using AI to automate processes that run on the legacy systems of a bank or an insurer, where they are often complex, poorly documented and regulated. That is the work Cognizant does, and AI does not make it simple.
Large companies: many projects, few numbers
On 6 October Evident published its yearly index of AI in 50 large banks (Evident). The average score rose 26% over the year, the fastest rise since the index started in 2023. But only 12% of the AI use cases it tracks report an effect on operational measures, and about 1% report concrete financial returns. Twelve banks now report a realised or expected return on their AI activities as a whole, up from eight a year ago.
On 27 October the Regional Chamber of Commerce in Katowice published a survey of 134 Polish organisations, written by Anna Majkowska and Joanna Tokar (RIG Katowice). 94% of them saw AI as an opportunity. 16% said AI was part of the company’s wider direction; 39% used it in selected areas and 19% were at the stage of individual tests or pilots. 85% said they were open to testing new solutions, but only 37% rated highly their ability to put changes in place quickly, and 35% had a named person or team in charge of the change. The sample is small and self-selected, and fewer than a quarter of the organisations had 250 or more staff. “Implementing a tool does not yet mean that a company can make use of what it offers,” Majkowska said.
Some large banks do keep up
Size alone does not decide the outcome. DBS, Singapore’s largest bank, says that in 2025 it ran more than 2,000 models in over 430 use cases, and that its data analytics and AI work produced about SGD 1bn of economic value (DBS). That is the bank’s own figure and covers much more than generative AI, and the bank puts it down to early and sustained investment in data and technology. In the Evident index the leaders hardly change: every bank in this year’s top 10 has been there before, six of them every year, with JPMorgan Chase first. The gap is opening between banks as much as between banks and start-ups.
What I heard at IBM TechXchange
From 26 to 29 October I was at IBM TechXchange, at sessions on scaling AI from pilot to production, auditable credit decisions and agent security. I did not hear much that was new. My main impression was that many large companies talk at length about how much they are doing with AI, and the results they can show are much smaller than the talk.
Why small teams are faster
In my view large legacy companies are, and will remain, far behind the new small teams that are built around AI from the start. Headcount matters less than it did: Mithril had one person and 11 suppliers, Base44 had nine people and Lovable has 146.
The reasons are in the examples above. In a small team the person who decides is also the person who does the work, and AI takes over the part of the work that grows with the size of the job. The Mithril report says it plainly: Claude did the reading, calculation and drafting, the judgements stayed with the people, and the pace was set by the labs’ replies and the team’s own decisions. In my view the need to agree each change between the business owner, IT, risk, procurement and suppliers is a large part of why big companies move more slowly.
The Polish survey also points to unclear responsibility for change: only 35% of organisations had a named person or team in charge of it; in 37% responsibility was spread across several people, and in 28% nobody formally had it.
Legacy systems add work. A small team starts from a clean slate. A bank or an insurer has to fit AI into legacy systems and processes that are often undocumented and regulated, which is why the Cognizant comparison has a limit. DBS shows that it can be done, after years of building its data foundations. The Evident index shows where most banks are: plenty of use cases, and about one in a hundred with a financial result that the bank is willing to publish.
Briefing
On 28 October, KRiBSI, the commission that supervises AI in Poland, held its first meeting, on the day Articles 8–18 and Chapters 3–5, 8 and 9 of the Polish act on AI systems came into force (Dziennik Ustaw). Electronic forms for complaints and individual opinions have been open since 29 October, and the commission’s first explanations will cover Article 50 of the AI Act: labelling AI-generated content and telling users when they deal with AI (KRiBSI). A deputy chair of KNF sits on the commission. A firm can now face a complaint before any guidance is out.
On 30 October, Pedro Machado of the ECB’s Supervisory Board said that “around 85% of the banks use generative AI already”, and that frontier models “shorten the time between a vulnerability being discovered and exploited”. He called fast patching “the biggest challenge for the banks today” and said the ECB would assess the cyber-resilience action plans banks were due to submit by the end of October (ECB). Issue 71 covered the same gap between finding software flaws and fixing them.
On 29 October, Mistral published the weights of Mistral Large 4 under a modified MIT licence. Companies can download it and run it on their own infrastructure; only firms selling the model as a service need a separate commercial licence. Through Mistral’s API it costs $0.70 per million input tokens and $2.10 per million output tokens (Mistral). Mistral has not yet published its technical report. A bank running it on its own servers would not be affected by US export restrictions like those that made Anthropic switch off Fable 5 in June.
Summary
On 27 October Anthropic said it had bought Mithril Medicine, a one-person company that ran a drug programme through 11 external labs. Lovable reached $400m of annual revenue with 146 employees. In banks, Evident finds about 1% of AI use cases with published financial returns, and in a Polish survey only 37% of organisations think they can put a change in place quickly.
In my view large legacy companies are, and will remain, far behind small teams that work with AI from the start. Mithril’s example suggests that AI lets a small team coordinate a much larger programme without building an equivalent internal workforce.
Stay balanced, Krzysztof
Krzysztof Goworek is founder of Quintant — AI advisory that gets enterprises from experiment to production value.