Sanofi is moving generative artificial intelligence from a general office tool into the specialised work of drug development. The pharmaceutical group, OpenAI and Formation Bio announced a collaboration on 21 May 2024 to build purpose-made software, tuned models and AI agents for research and development. Fortune examined the plan eight days later through an interview with Sanofi chief executive Paul Hudson.

The arrangement matters because each participant supplies a different scarce input. Sanofi brings proprietary pharmaceutical data and a large development pipeline. OpenAI provides model capabilities, fine-tuning knowledge and technical resources. Formation Bio contributes engineering and a platform built around clinical development. Their task is not to place a public chatbot beside scientists, but to design tools for particular decisions and documents.

A partnership built around complementary assets

Sanofi, headquartered in France, wants to become a biopharmaceutical company powered by AI at scale. The companies' announcement said their custom systems would cover the drug-development lifecycle. It did not identify a medicine produced by the collaboration or report a validated clinical result.

Formation Bio has experience building software around trials and developing licensed drug assets. OpenAI, based in the United States, contributes a general modelling foundation. Sanofi supplies the domain context needed to make that foundation useful. This division of labour illustrates why valuable industrial AI often depends less on model access alone than on data rights, workflow design and expert validation.

What the partners said they would combine

  • proprietary scientific and operational data from a global pharmaceutical pipeline;
  • models that can be fine-tuned for narrowly defined research and development tasks;
  • engineering systems capable of fitting those models into regulated workflows;
  • drug-development and clinical-trial experience for testing whether outputs are useful.

The cost of failure defines the opportunity

Hudson told Fortune that developing a drug costs Sanofi between $3 billion and $4 billion and that 80% fail in phase-one clinical trials. These were the chief executive's estimates, not audited results of the new partnership. His commercial objective is to identify weak candidates earlier so capital and scientific effort can move toward projects with better prospects.

That ambition is economically important even if AI improves only selected stages. A development portfolio contains many expensive gates: selecting a biological target, designing a candidate, running laboratory studies, recruiting trial participants, monitoring safety and preparing regulatory evidence. Earlier rejection of an unsuitable candidate may be valuable, but a confident algorithmic prediction still has to survive experimental and clinical scrutiny.

A text-free three-lane physical infographic moves document cards, molecular candidates and patient figures through repeated validation gates while rejected options enter a separate tray
Purpose-built models may reduce search and drafting work; they do not remove the evidence gates that protect patients and determine approval.

Documents offer the earliest practical test

Hudson expected the first outputs by the end of 2024, most likely initial drafts of documents for the US Food and Drug Administration. That is a more bounded task than inventing a medicine. A system can organise existing evidence, assemble a draft and flag missing material while qualified teams remain responsible for accuracy, interpretation and submission.

This near-term use also provides a disciplined evaluation path. Managers can measure time saved, correction rates, omitted evidence, traceability and reviewer confidence. If a model cannot handle controlled documentation reliably, broader claims about molecular design or patient selection deserve even more caution.

Data access creates both advantage and obligation

Proprietary data can make a model more relevant because it reflects compounds, experiments and operating decisions that are absent from public material. The same data can include commercially sensitive research and information subject to strict controls. Access therefore needs defined purposes, technical separation, retention rules, monitoring and clear accountability for every downstream use.

Quality is equally important. Historical decisions may contain inconsistent terminology, incomplete records or biases in which patients entered trials. Training on a large archive does not automatically correct those weaknesses. The partners need evaluation sets that represent the intended task and experts able to recognise plausible but unsupported output.

Operational AI is the real experiment

Sanofi's plan differs from buying a broad assistant licence. Hudson told Fortune that the collaboration was about operationalising large language models. That means placing a model inside a controlled sequence where inputs, permissions, review and escalation are explicit. The difficult work is organisational as well as technical: scientists, clinicians, regulatory specialists, data teams and software engineers must agree on what the system may do.

The partnership was an investment in capability, not proof of a faster approved drug. Its progress should be judged through specific outcomes: reproducible predictions, fewer avoidable iterations, faster preparation of accurate documents and well-governed use of sensitive data. If those measures improve without weakening evidence standards, generative AI could become useful infrastructure for pharmaceutical development rather than another layer of experimentation.