Pfizer to Deploy Edison's Kosmos AI Scientist Across R&D Pipeline

Edison Scientific will deploy its Kosmos AI scientist across Pfizer's biopharmaceutical R&D pipeline, marking the San Francisco company's second major pharma deployment in five months. The agreement, announced Oct. 29, covers internal drug-development data in oncology, internal medicine, immunology and other therapeutic areas.

Edison, spun out of the nonprofit research lab FutureHouse in 2025, has raised $70 million from Spark Capital, Triatomic Capital and other investors. Its first comparable deal came May 19, when Incyte agreed to embed Kosmos across discovery and development work, starting with target discovery, validation and translational biology.

Sam Rodriques, Edison's CEO, said the platform works "across the full R&D life cycle from target identification through clinical development and filing." In the Pfizer collaboration, Kosmos "will develop integrated R&D world models by combining access to publicly available and proprietary data sources from Pfizer's internal drug development efforts," Rodriques said in written responses to R&D World.

"Pfizer has spent decades generating data about what works and what doesn't work in human biology," Rodriques said in the announcement. "That experience serves as the foundation for Kosmos, ensuring lessons from each program are maximized to inform other programs, with every experiment and clinical readout strengthening the entire system."

Edison's forward-deployed scientists and engineers will work with Pfizer research teams to build Kosmos into specific workflows, according to the announcement. No financial terms of the Pfizer agreement were disclosed.

How accurate is the AI scientist?

The evidence base comes from the Kosmos paper posted to arXiv in November 2025. Independent scientists judged 79.4% of statements in Kosmos reports to be accurate. The breakdown matters for R&D managers weighing where to trust the system:

• 85.5% of data-analysis statements were reproducible.
• 82.1% of literature-review statements held up against primary sources.
• 57.9% of synthesis statements were accurate — meaning over four in ten failed scrutiny.

Each run lasted up to 12 hours, executed an average of 42,000 lines of code and read 1,500 papers, the paper reported.

Rodriques acknowledged the weak spot. Asked how partner scientists can separate data-supported findings from interpretation, he pointed to Kosmos's citation system: "All statements are cited back to executed code, source data, or literature, even including specific pages and figures associated with the claim." He also noted the paper's results "were published one year ago" and that "LLM capabilities have advanced dramatically in the last few months." That claim of improvement is unverified by independent measurement.

Where do the 75% and $99 million figures come from?

Edison's Sept. 30 case study on work with Population Health Partners (PHP), a New York investment firm that creates companies and develops drugs, projects a roughly 75% cut in team time needed to prepare an IND, plus about $99 million in recovered peak sales per program.

The caveats are substantial. Rodriques said the 75% figure "is the projected time compression PHP expects in its second program with Kosmos" — a projection, not a measured result. It extrapolates from drafting and review speedups in a first stage involving one PHP program preparing for a Phase 2b study.

The measured results from that first stage include concrete numbers. Kosmos produced a first acceptable draft of a Nonclinical Overview in 4.2 hours, against roughly 100 hours of expert time. It found 98% of source-data errors Edison itself planted in a draft document — a test the vendor designed. Kosmos also drafted roughly 70% of a Pre-IND Briefing Meeting Request that PHP then edited and submitted to FDA.

The $99 million figure breaks down into three projected gains per program:

• $42 million from cutting IND preparation from four months to one.
• $32 million from cutting protocol amendment turnaround from 12 weeks to three.
• $25 million from halving the probability of amendments and nearly eliminating clinical holds.

All three components are projections based on PHP's own four-month baseline, not outcomes from completed filings. The case study was produced and published by Edison, the vendor benefiting from the numbers.

Beyond pharma workflows

Rodriques said Edison co-builds agents around each partner's tasks. At PHP, agents produce IND drafts using the firm's established rubrics, and one agent collaborates with experts inside Microsoft Word through comments and tracked changes — "the research environment where subject matter experts are often most comfortable."

For Pfizer, the open questions are whether Kosmos's synthesis accuracy improves on real proprietary data, and whether projected time compressions survive contact with regulated filings. The deployment's results, once Pfizer programs run through it, will provide the first independent test.