Five years away. AI has entered clinical development. Now he must decide what to trust and what he can defend.
Five years ago, the founder and CEO of a small, successful biotech company disappeared during an Antarctic expedition with his team after a failure left the group stranded and cut off from the outside world.
Five years later, they were rescued.
The company had survived. It had moved forward.
One asset was preparing to enter the clinic. Another had completed Phase I and was moving toward Phase II. The science was familiar. The R&D environment was not.
His team explained what had changed.
AI was no longer something discussed at conferences. It was becoming part of drug development.
“What can AI help us do?” he asked.
The question sounded simple.
A New Kind Of R&D Decision
His first instinct was familiar: evaluate data, challenge assumptions, recognize weak signals, and make decisions when evidence was incomplete.
He knew models could be wrong. He also knew experienced people could be wrong.
Now his team was showing him AI-generated analyses that challenged assumptions he had developed over a career.
AI was being explored across clinical development: analyzing large datasets, identifying clinical patterns, supporting trial design, and helping teams prioritize recruitment strategies. But whether an output was useful depended on the population, the data, the model’s performance, its context of use, and the consequences of acting on it.
The more consequential the decision, the less satisfying the question “Does the model work?” became.
The question became: Can we defend using it for this decision?
He had spent his career making R&D decisions with incomplete information. What was different now was that a machine could generate some of the evidence challenging those decisions.
The Regulators Are Already In The Room
His team showed him what the FDA and EMA were doing.
The FDA’s 2025 draft guidance on the use of artificial intelligence to support regulatory decision-making for drugs and biological products introduced a risk-based framework for assessing the credibility of an AI model for a particular context of use. In January 2026, the FDA and EMA published guiding principles for good AI practice in drug development, emphasizing issues including context of use, data governance, performance assessment, transparency, and human-centric implementation.
The message was not that AI should be avoided. Nor was it another member of the development team.
It was a tool whose credibility depended on what the organization was asking it to do.
The outcome was becoming tangible. The FDA has qualified AIM-NASH, an AI-enabled tool for assessing liver biopsies in clinical trials of treatments for metabolic dysfunction-associated steatohepatitis. The significance was not that AI had replaced clinical judgment. It was that the tool had been evaluated for a specific purpose in a defined context.
The company needed to understand its context of use, its data, its performance, and its limitations.
Ultimately, it needed to understand and stand behind how AI had contributed to a development decision.
That changed the question for the CEO.
If an AI model suggested a different development strategy and the team followed it, who understood why? If the model was wrong, who recognized its limitations? If it produced the right answer for the wrong reasons, would the team know? And if the FDA asked why the decision had been made, could the company explain not only what the model predicted, but why it was appropriate for that particular decision?
For a small biotech, these were practical questions.
There was no internal AI organization.
So the founder went back to his office.
The Whiteboard
He wrote three questions on a whiteboard.
- What is AI telling us?
The first question was about value: Where could AI improve R&D decision-making?
- What does AI not know?
The second question was harder.
A model could generate an output without understanding the biological, clinical, or operational context of a decision. The team needed to understand its context of use, limitations, data, and performance rather than treating it as an answer.
- Who owns the decision?
This was the important question.
AI could inform a decision, challenge it, or force a team to reconsider something that had seemed obvious.
But the organization still had to understand, justify, and stand behind the decision.
AI might change how evidence is generated and interpreted. It did not remove responsibility for the resulting decision.
The New Skill Of R&D Leadership
The founder realized that his challenge was not to decide whether his company should “use AI.” That was becoming less relevant.
His job was to decide how his organization would use it and where its outputs were credible enough to influence consequential decisions.
That meant bringing clinical, biometric, regulatory, medical, and data science expertise together rather than making AI the responsibility of one technical function.
It meant defining, before using a model for a consequential decision, what question it was being asked to answer and what evidence would be required to trust that answer.
It meant considering regulatory engagement early when AI could materially influence a development decision, rather than explaining its use after the fact.
And it meant changing his behavior.
Thirty years of experience remained an asset. But experience could no longer be the final answer just because it was experience. Nor should an AI output become the final answer simply because it came from more data.
He had returned expecting the science to have changed.
What had changed more profoundly was the architecture of decision-making.
His job was no longer to have every answer himself. It was to build an R&D organization capable of knowing when to trust AI, when to challenge it, when to seek additional evidence and when not to use it at all.
He looked again at the three questions on the whiteboard.
What is AI telling us?
What does AI not know?
Who owns the decision?
Five years earlier, he would have started with experience.
Now, he realized, his experience had a new job: not to compete with the machine, but to know when the machine deserved to be heard.
Author
Professor David Adler
Profile:
Crunchbase
Social:
NA
Website:
www.pathora-institute.com
Professor David Adler, MD/PhD, MBA, is an oncology drug development and translational medicine leader with more than 15 years of pharmaceutical and academic experience. He spent a decade in senior roles within Bayer AG’s Global Oncology Clinical Development organization and now serves as Chief Scientific & Medical Officer of the PATHORA Institute of Pathology & Tissue Medicine. He also holds academic appointments at the Hebrew University of Jerusalem, Ben-Gurion University of the Negev, and the University of Bonn.
