BenevolentAI found a COVID-19 treatment. Building a business proved harder.

The medicine that gave BenevolentAI one of its clearest successes already existed. Baricitinib was an arthritis drug when the company’s researchers helped identify it as a possible treatment for COVID-19. Their proposal appeared in February 2020, early enough in the pandemic for a useful clue to matter enormously. The task was to spot a connection buried in biology, then give other scientists something worth testing.

That episode explains the attraction of BenevolentAI. Its subsequent history explains the difficulty. An internal eczema candidate disappointed in a clinical trial. Spending was cut. The company entered the public market in 2022 and left in 2025. Today, its website points visitors toward Qor, an intelligence product for people making biopharma business decisions. The scientific insight and the financial struggle belong in the same story.

An old medicine, a new assignment

In the early COVID-19 work, researchers used BenevolentAI’s platform to search for existing drugs that might interfere with infection. Baricitinib emerged as a candidate with a potentially useful combination of biological effects. This was repurposing: asking a medicine to perform a new assignment. BenevolentAI did not invent the original molecule, and its prediction was a hypothesis, not a clinical result.

The evidence came through trials. In ACTT-2, a randomized study involving 1,033 hospitalized patients, baricitinib plus remdesivir produced a median recovery time of seven days, compared with eight days for remdesivir alone. That was a specific result in a specific population. It gave the computational work something an impressive demonstration cannot supply: an observed patient outcome.

In May 2022, the US Food and Drug Administration approved baricitinib for hospitalized adults with COVID-19 requiring specified forms of oxygen or breathing support. The distinction in credit matters. BenevolentAI contributed an early research lead; drug developers, clinicians, trial participants and regulators carried the treatment through later tests. AI earned a place in that chain.

The map behind the molecule

Founded in 2013, BenevolentAI brought together founders Kenneth Mulvany, Michael Brennan and Ivan Griffin around the meeting of technology and life sciences. Its enduring idea is that biomedical knowledge becomes more useful when a computer can connect it. A paper, a gene, a clinical observation and a chemical compound can each illuminate a different part of the same disease.

The company’s knowledge graph organizes relationships among those things. Machine learning tools then help scientists explore the evidence and propose targets. A target is a biological component whose activity might be changed to treat disease. Choosing one is consequential: a beautifully engineered molecule can still disappoint if it acts on the wrong mechanism.

In a 2019 demonstration using glioblastoma, the team showed how its knowledge base brought together disease biology, compounds and clinical trials. Later presentations described tools spanning target assessment, chemistry and repurposing. The appeal is practical. A research team can investigate a connection across different kinds of data instead of relying entirely on what individual specialists have read and remembered.

This is also a company of laboratory work. Its documented discovery operations combined AI with molecular biology, medicinal chemistry and experimental testing in Cambridge. The graph supplies leads; researchers decide what to investigate and how to investigate it.

“We build technology in the service of science”

BenevolentAI · company philosophy

Recursion offers a useful comparison within AI drug discovery. Its founding approach emphasized images of cells and the generation of experimental data at scale. BenevolentAI’s public explanations give greater prominence to connecting biomedical knowledge. Both approaches need evidence beyond the model. BenevolentAI’s distinctive asset is the combination of its accumulated knowledge system and people who understand drug research.

The eczema test that changed the budget

In April 2023, BEN-2293 supplied a sharp reminder of that distinction. The topical candidate was designed to inhibit Trk receptors and treat mild-to-moderate atopic dermatitis. Its Phase IIa trial enrolled 91 adults. Participants received the cream or placebo twice daily for 28 days.

The study met its primary endpoint of safety and tolerability. It missed the secondary efficacy endpoints for itch and eczema severity across the intention-to-treat population. BenevolentAI reported an exploratory signal associated with greater affected body surface area and initially said further investigation was warranted. That qualified observation could not erase the broader efficacy result.

The financial decision followed. In May, BenevolentAI announced a review that prioritized selected assets and new technology products. By its September interim report, the company confirmed no further investment in BEN-2293. This was a concrete change of mind: a candidate that had justified clinical spending no longer justified the next cheque.

Other programs continued. BEN-8744, an oral PDE10 inhibitor being developed for ulcerative colitis, produced positive Phase Ia safety and pharmacokinetic results in healthy volunteers in March 2024. The company said it had reached development-candidate nomination in two years. That demonstrated progress through discovery and early testing; effectiveness in patients remained a separate question.

The customers buy a starting point

BenevolentAI’s commercial customers have included pharmaceutical companies that already possess substantial research operations. AstraZeneca began a target identification collaboration in 2019, initially in chronic kidney disease and idiopathic pulmonary fibrosis. The relationship expanded in 2022 to heart failure and systemic lupus erythematosus. In June 2024, AstraZeneca added a lupus target that it had experimentally validated to its discovery portfolio.

For that buyer, the product is a better-supported starting point for research. A selected target can feed an existing development organization. It is valuable without being a finished treatment, and its eventual medical value remains subject to the experiments ahead.

Merck KGaA, the German science and technology company, signed a different collaboration in September 2023. BenevolentAI would help identify and develop small molecules, initially against three targets across oncology, neurology and immunology. The work combined computational chemistry tools with laboratory capabilities and aimed to deliver candidates for further development.

The agreement’s headline value was up to $594 million, plus potential tiered royalties. The upfront payment was described only as a low double-digit million-dollar sum. Most of the headline opportunity depended on future milestones. Counting the entire ceiling as funding would turn a conditional contract into imaginary cash.

The structure also explains the business model. A partner pays for work today and may pay again as a program advances. BenevolentAI can participate in future success through milestones and royalties, while its internally developed assets offer another route: licensing a candidate to someone prepared to finance the next stage.

The price of keeping the experiment going

The company financed its ambition with substantial outside capital. A 2018 round raised $115 million at a $2 billion pre-money valuation. Temasek invested $90 million in 2019. BenevolentAI’s 2022 full-year results recorded €225 million, or £186.8 million, in gross proceeds from its business combination and listing. Those are dated financing events, not a present-day price tag.

Discovery remained expensive. For the first half of 2024, BenevolentAI reported £2.8 million in revenue and a £32.3 million operating loss. Research and development spending was £19.9 million. These figures describe the whole business for six months; they do not reveal the cost of finding one target or running the eczema trial.

In December 2024, with Mulvany back as executive chairman, the company announced another overhaul. It would concentrate investment on technology, reduce early drug development expenditure and make workforce changes. Management also argued that maintaining a public listing diverted resources from core activities. Its projected runway into 2027 was a forecast attached to those changes.

Shareholders approved the merger underpinning its private-company structure on March 12, 2025. The original entity merged into Osaka Holdings, which converted into a public limited company under Luxembourg law and adopted the BenevolentAI name. The shares were delisted the following day. Public limited legal form did not mean continued stock-exchange trading.

Qor moves the question into the boardroom

The current website puts Qor in front of biopharma decision makers and life science consultancies. Its advertised work includes monitoring competitive and commercial developments, producing strategic briefings, and letting users question the reasoning or explore scenarios. Prospective customers are invited to book a demonstration.

This changes the immediate job to be done. A drug discovery scientist might ask which protein deserves an experiment. A biotech executive might ask how an asset compares with competitors or which developments should affect a partnering discussion. Qor’s pitch addresses the second set of questions, with evidence attached to the briefing.

A useful demonstration would begin with a live business decision. Give the product an asset and a concrete question, then inspect the supporting material, challenge assumptions and see whether new information changes the recommendation. That is an evaluation method suggested by the advertised features, not an independently established performance claim.

The same limits apply to the underlying intelligence: incomplete coverage, stale information or weak evidence can spoil a confident-looking answer. Executive judgment still matters, especially where the decision depends on information that is private or has never been measured.

The useful habit is to invite a rebuttal

BenevolentAI has publicly celebrated collaboration across disciplines. Its 2022 Innovation Week brought the team together in London to work on shared projects. That gives a specific glimpse of its working culture before later reductions. It also points to a habit other organizations can copy: put the people building the model alongside those equipped to challenge its conclusions.

The larger lesson is to decide what evidence would stop a project before becoming attached to it. Baricitinib shows the value of finding a testable connection. BEN-2293 shows why a plausible connection still needs a stopping rule. Qor now asks customers to trust the company with decisions closer to the boardroom. The interesting test is whether it helps them ask a better question before they sign the next cheque.