A Paris- and Boston-based start-up, backed by several of the world’s largest drugmakers, is betting that its specialized artificial intelligence can solve one of medicine’s costliest problems: the staggering rate at which promising drugs fail.
PARIS & BOSTON, 23 September 2026 – For every ten experimental medicines that enter human testing, nine will never reach the pharmacy shelf. The journey takes more than a decade, and each approved drug carries an average price tag of $2.3 billion — much of it spent on candidates that were doomed from the start.
On Tuesday, a French-American company called ArcaScience said it had built a tool to change that arithmetic.
The company announced the commercial launch of Flow, an artificial intelligence platform designed to help pharmaceutical researchers evaluate the delicate balance between a drug’s benefits and its risks — the calculation that regulators like the Food and Drug Administration and the European Medicines Agency use to decide whether a treatment is worth approving.
ArcaScience said its platform, developed with input from pharmaceutical giants including Sanofi, AstraZeneca, Pfizer and GSK, can outperform the most advanced general-purpose A.I. models by a wide margin on specialized biomedical tasks, recovering 96 percent of known adverse-reaction concepts in an independent benchmark study. The next strongest competitor, Google’s Gemini 3.6 Flash, recovered about 65 percent.
“A promising molecule deserves the strongest possible development strategy,” Romain Clément, the company’s chief executive, said in an interview. “We put millions of benefit–risk configurations within reach of pharma teams, so they can uncover where a candidate can make the greatest difference.”
The announcement reflects a broader shift in the pharmaceutical industry, which has increasingly turned to artificial intelligence to shave years and billions of dollars off drug development. But it also highlights a persistent challenge: general-purpose A.I. models, for all their fluency, are prone to fabricating information — a liability in a field where regulatory compliance is strict and human lives are at stake.
ArcaScience’s answer is a narrower, more controlled approach. Rather than relying on a single large language model, the company’s engine combines 24 specialized A.I. models with a curated evidence base of 100 billion biomedical data points. The company calls the architecture “Exact & Explainable A.I.,” or X-A.I., and says it guarantees that identical queries produce identical, auditable answers — with source-level traceability down to the exact passage in the underlying literature.
That determinism, the company argues, is what separates a tool that can assist a regulatory submission from one that cannot. Flow generates the benefit–risk sections of key regulatory documents, including PSUR/PBRER reports, risk management plans and electronic common technical documents, in hours rather than months.
The company also released the results of a benchmark study evaluating its engine against five frontier A.I. models. The study, which underwent independent qualitative review by the consulting firm inExtenso, found that ArcaScience’s approach was not only more accurate but dramatically cheaper to operate at scale. In a simulated production scenario processing 30 million scientific articles across 16 drug families, the company said its platform cost as little as $0.0162 per article, compared with $1.35 million to $18.53 million for general-purpose model APIs over the same dataset.
“When decisions can shape a drug’s future, A.I. performance has to be measurable,” Mr. Clément said. “Reliability and certainty are non-negotiable in the medical industry.”
ArcaScience, which raised $7 million in September 2025 in a round led by The Moon Venture, says it already counts more than 20 pharmaceutical clients, including Sanofi, AstraZeneca, GSK, Takeda and ICON. Its technology has been used in chronic skin disease research affecting more than 70,000 patients, and during the Covid-19 crisis the French government selected the company to structure and deliver the complete corpus of scientific knowledge on the virus.
Still, the company faces familiar hurdles. Pharmaceutical decision-making is conservative by nature, and persuading regulators to accept A.I.-generated assessments — however traceable — will take time. Competitors, including the largest A.I. laboratories and a growing field of biomedical start-ups, are racing toward the same goal.
But for an industry accustomed to spending billions on failure, even a modest improvement in predicting which drugs will work could prove transformative. ArcaScience is betting that the future of drug development belongs not to the biggest model, but to the most trustworthy one.
The platform is commercially available starting Tuesday for pharmaceutical enterprises, biotech firms and regulatory desks worldwide.
Want to amplify your startup’s story? EU Startup News is your launchpad to reach startup founders, investors, and C-level execs across Europe. Discover our tailored promotional strategies such as Sponsored Articles and Partnerships. Click here to learn more or contact us directly at pitch@eustartup.news. Join us, and let’s make your startup the talk of Europe!
