Global pharmaceutical giants are moving artificial intelligence from back-office work such as handling regulatory filings into the core of molecular design and drug candidate screening. Senior scientists at GSK, Bristol Myers Squibb, Novartis, Eli Lilly, Novo Nordisk, AstraZeneca and Sanofi say AI has cut the time needed to develop new molecules and advance them to clinical trials by as much as half this year.

AstraZeneca executive Sharon Barr summed up the shift in one sentence: “There is no part of the process from drug discovery to development that does not use AI.”

The results disclosed in this round are concentrated at the very front end of the R&D chain. Fiona Marshall, president of Novartis Biomedical Research, said the average time from project launch to identifying a candidate drug has been compressed from about three years to 18 months to two years, and AI tools also helped the company pinpoint five priority kidney disease treatment targets. Novo Nordisk said AI has cut the number of iterations in designing, producing and testing new molecules from about eight rounds to three, with each round shortened from eight weeks to four.

Sanofi offered a more concrete case: using AI-assisted methods to turn early biology research into clinical decisions allowed the company to skip a dose-selection study, putting the development of its asthma candidate lunsekimig at least a year ahead of schedule. Robert Plenge, Bristol Myers Squibb’s chief research officer, said that without the company’s AI systems, one of its sickle cell anemia candidates would never have been developed. He said the next priority is connecting models with laboratory automation systems to produce higher-quality molecules faster.

CompanyDisclosed AI R&D resultsNovartisTime from project launch to candidate drug cut from about 3 years to 18 months to 2 years; 5 kidney disease targets identifiedNovo NordiskNew molecule iterations cut from about 8 rounds to 3, each round from 8 weeks to 4 weeksSanofiSkipped a dose-selection study; lunsekimig development at least 1 year aheadBristol Myers SquibbA sickle cell anemia candidate would not have been developed without AI systemsAstraZenecaAI used across the entire process from discovery to development

Note: The above are figures publicly disclosed by the companies, and their methodologies are not entirely consistent.

Matt Truppo, global head of computational and AI strategy in Sanofi’s R&D division, expects AI to shorten the early discovery phase by 20% to 25% across the company’s more than 100 drug development programs over the next two years, and possibly by half within five years. But he cautioned that replicating such efficiency gains in therapeutic areas where the company has less proprietary data will take time.

Whether efficiency gains translate into higher approval success rates remains unknown. Historically, nearly 90% of drugs developed with traditional methods ultimately fail clinical trials; statistics from the US Pharmaceutical Research and Manufacturers of America (PhRMA) show that a new drug typically takes 10 to 15 years from research to market and can cost billions of dollars. Yet the number of AI-assisted candidates entering clinical trials has grown only modestly — as of March this year, 69 companies had 129 such drugs in clinical stages, up from about 117 eight months earlier, with most still in early development.

Waqas Haque, an oncology researcher at the University of Chicago who helped compile the data, expects the number of AI-assisted preclinical drugs and those entering clinical trials to rise markedly over the next one to three years. But he was blunt: “I don’t know whether the success rate of these drugs will be higher than that of non-AI-developed drugs, and I don’t think they will win approval faster.” He noted that bottlenecks in human trials and regulatory review will continue to constrain how quickly drugs reach the market.

Aliza Apple, head of Eli Lilly’s TuneLab, is similarly cautious. TuneLab promotes collaboration by opening some of Lilly’s internal AI models to small biotechnology companies. She argued that claims by Anthropic CEO Dario Amodei and the founder of Alphabet’s Isomorphic Labs that AI can cure all diseases overstate the technology’s current capabilities. “If they can cure one disease first, that would of course be great. Once they’ve done that, we can talk about other diseases.”

The AI companies’ own moves are reshaping the industry landscape. Anthropic completed an initial public offering this week at a $1.5 trillion valuation, has opened its own laboratory and announced several investment plans in recent months, launched AI software for scientific research called Claude Science, and invited Novartis CEO Vas Narasimhan onto its board. Drugmakers including Novo Nordisk are already using its models and Claude Science in R&D. Speaking on Nasdaq about his father, who died of illness, Amodei said clinical trials and experiments still take time, but “I believe AI can significantly accelerate these processes. We hope that in the future we can reduce such tragedies: families losing out on a chance at a cure simply because they were a few years too late.”

Luca Giani, a senior principal at venture capital firm Alumni Ventures, said Anthropic and other frontier AI labs are scouting pharmaceutical acquisition targets, “They want more data, and they want to acquire more companies,” adding that this is “just the beginning.”

Traditional drugmakers are divided on the matter. Apple sees such AI companies as both allies and potential rivals: “Early on, of course we saw them as vendors. To be fair, that’s how they positioned themselves at the time. I think we increasingly see them as partners now.” She added that a year ago they had no related R&D programs underway, and no one can say what things will look like a year from now. GSK Chief Scientific Officer Tony Wood, by contrast, sees Anthropic as “another competitor,” while Bristol Myers Squibb and Novo Nordisk lean toward viewing it as a partner.

The ceiling on collaboration is data. Both Novo Nordisk and Sanofi said they will not provide proprietary data to frontier AI labs for model training. “Without data to train the models, the whole system quickly breaks down,” Truppo said. Some pharmaceutical scientists argue that decades of accumulated R&D experience and data are precisely the moat that traditional drugmakers hold over AI newcomers. Marshall said the company has always had large amounts of data and can now use it in more advanced ways.

For investors, this round of disclosures offers two windows for assessing the value of AI in drugmaking: whether shorter early-stage R&D cycles can be replicated in more therapeutic areas, and whether AI-assisted candidates can beat the historical benchmark for success in clinical trials. The former determines how far the cost curve can fall; the latter determines whether revenue can deliver. With the pace of regulatory review unchanged, the real boost AI gives the pharmaceutical industry’s financial performance will still have to await the verdict of clinical data.