The FDA Just Handed Industry a Rulebook on AI in Trials. Now Comes the Hard Part.

Two years ago, a pharma sponsor asked the FDA whether AI could be used to identify patients most likely to respond to a therapy. The agency’s answer, per Shashidhar Joshi’s public reading of the regulatory arc, amounted to: “It depends, and we’ll figure that out together.” That answer held for years. Then January 2025 arrived, and the FDA published its draft guidance “Considerations for the Use of AI to Support Regulatory Decision-Making for Drug and Biological Products,” introducing a seven-step, risk-based credibility framework. Twelve months later, the FDA and EMA issued joint “Guiding Principles for Good AI Practice in Drug Development,” the first coordinated transatlantic regulatory position on AI in pharma. The ambiguity, as Joshi puts it bluntly, is gone.

But gone how? That is where the LinkedIn chatter this week gets interesting, and where the surface-level celebration of regulatory clarity runs into something more complicated.

The frameworks exist. The question practitioners are wrestling with is whether they enable genuine AI integration or codify a posture of managed caution that will age poorly against the pace of model development.

The Data Layer Doctrine

Read the RTCT comment letters from Novartis, Lilly, Bayer, and Daiichi Sankyo together, as Amit Patel did this week, and a shared doctrine emerges that most AI coverage has missed entirely. These are not letters arguing for autonomous AI systems. Daiichi was explicit: AI should not autonomously make dose escalation or safety decisions. Every letter drew the same line between AI that supports qualified clinical judgment and AI that replaces it. What the sponsors want underneath that line is aggressive: automated reconciliation, coding consistency, discrepancy detection, safety surveillance. The back office, not the bridge.

Patel‘s read is that the sponsors who lead AI adoption will not be the ones with the most sophisticated models. They will be the ones whose data operations are modern enough to move in real time.

That reframe matters more than it might appear. The industry conversation about AI in trials has been dominated for three years by patient recruitment algorithms, adaptive protocol engines, and biomarker prediction models. Sophisticated, visible, headline-worthy. But per the comment letters, the near-term battleground is cleaner data delivered faster to humans who still make the calls. Lilly’s formulation, cited by Patel, is worth sitting with: the FDA should select for the most mature oversight of AI, not the most mature AI.

A brilliant model with weak governance does not clear this bar. A well-governed model doing unglamorous work does.

A Rulebook Written in Pencil

Here is the counterintuitive read on the FDA/EMA joint principles: the regulatory frameworks published in 2025 and 2026 are not the final word on AI in trials. They are the opening bid from institutions that have received, per Joshi’s post, over 800 AI-component submissions since 2016 and now have nearly a decade of hands-on experience evaluating them. That is not a small sample. The FDA has seen what happens when sponsors deploy models without documented data lineage, without change control, without evidence of human review beyond a checkbox. The seven-step credibility framework and the joint principles are less an innovation agenda than a distillation of what has gone wrong.

Which raises an uncomfortable question for the sponsors celebrating regulatory clarity: if these frameworks were written partly in response to observed failures, how many current AI deployments in active trials would fail to meet them today?

The answer, given what the comment letters reveal about current data infrastructure, is probably more than sponsors would prefer to say publicly. Lilly and Bayer independently proposed a DMF-style pathway, per Patel’s analysis, that would let AI vendors disclose model documentation directly to the FDA with sponsors referencing it. The parallel to the Drug Master File mechanism is deliberate. It signals that sponsors recognize their own AI vendors may hold documentation that sponsors themselves cannot fully produce on demand. That structural gap between deploying AI and governing it is precisely the gap the FDA’s credibility framework is designed to expose.

Silvio Galea, writing from his position as Chief AI Officer this week, puts the operating principle cleanly: “Close enough doesn’t work. In clinical trials, AI needs to be grounded in verified data, informed by deep expertise, and designed to support, not replace, human judgment.” That framing aligns with the FDA/EMA posture. It also describes an implementation challenge that most trial sponsors are still in the early stages of meeting.

Who Actually Writes the Standard

The more systemic question underneath all of this is about institutional authorship. The FDA and EMA published the frameworks. But the comment letters from Novartis, Lilly, Bayer, and Daiichi were not passive responses to finished policy. They were active attempts to shape what “mature AI oversight” looks like in practice, including the DMF-style vendor pathway that neither agency had proposed on its own. Sponsors are not waiting for regulators to define the standard and then complying. They are submitting detailed operational models and asking regulators to ratify them.

That dynamic is not new in drug development. The negotiation between sponsor operational reality and regulatory principle has always shaped guidance over time. What is new is the pace. The FDA’s January 2025 draft guidance and the January 2026 joint principles arrived within twelve months of each other, a compression that suggests the agency is trying to keep frameworks current with deployment rather than publishing once-a-decade position papers. The 800-plus AI-component submissions since 2016 created the empirical base. The comment letters are now adding the operational layer.

Joshi’s framing of this moment is precise: this is no longer a “watch this space” conversation. It is a compliance conversation. Sponsors who have been treating AI governance as a future-state planning exercise now face concrete requirements: specify how each AI model is used in the study protocol, identify and mitigate risks, validate continuously as data environments shift, and integrate data scientists with clinical leads throughout the trial rather than at handoff points. Those are not aspirational principles. They are checkboxes with audit trails attached.

The DMF-style vendor pathway proposal from Lilly and Bayer deserves particular scrutiny as the FDA evaluates it. If adopted, it would create a new layer of regulatory relationship between the agency and AI vendors that currently sits entirely within the sponsor-vendor contract. That structural shift would give the FDA direct visibility into model documentation it currently must request through sponsors, and it would put AI vendors into a compliance posture they have not historically occupied. Whether the agency moves on that proposal will signal more about the future of AI governance in trials than any principles document published so far.

The sponsors filing comment letters are not asking the FDA to slow down. They are asking for a governance architecture they can actually build. The FDA’s next move, whether it adopts the DMF-style pathway or holds the existing sponsor-accountability model, will determine whether the 2025-2026 frameworks become an operational foundation or a compliance floor that the industry immediately starts negotiating its way around.

References

1. Amit Patel, LinkedIn post on RTCT comment letters from Novartis, Lilly, Bayer, and Daiichi Sankyo, 2025
2. Shashidhar Joshi, LinkedIn post on FDA draft guidance (January 2025) and FDA/EMA Joint Principles (January 2026), 2025
3. Silvio Galea, LinkedIn post on responsible AI in clinical research, 2025