MassBio is partnering with the Arthur D. Little team on November 2 to host 2026 R&D Reimagined – Techbio Solutions, the second in a series of events designed to connect biotech and pharma R&D leadership with the technology innovators and solution providers transforming the life sciences with new tools and services.
There is a pattern playing out across life sciences that is worth naming directly.
A company gets serious about AI. A budget is approved, a platform is selected, and domain experts are brought in to define use cases. But months later, the teams who were supposed to benefit from the work are still struggling to make it useful.
The problem is rarely that scientists or clinicians are left out. It is what happens when they say the solution does not work.
AI in drug development is a domain challenge. Whether any given application is actually useful depends on understanding the work itself. Is the output scientifically defensible? Does it account for the regulatory pathway being pursued? Can it distinguish between meaningful evidence and adjacent information? Can it handle the judgment calls a person makes in the real workflow?
Domain experts are asked these questions constantly. But when the underlying AI cannot meet their requirements, the response is often to engineer around the gap. More integrations, additional retrieval layers, workflow redesigns, further customizations. The problem gets bent to fit the solution.
Better implementation solves some problems, but it cannot compensate for a fundamental mismatch between what an AI can do and what the work requires. If a model cannot reliably account for scientific context, regulatory nuance, or the kind of expert judgment a use case demands, a better integration is not the answer. It may be a different model, a different approach, or honest acknowledgment that the use case is not yet ready for AI.
Without actively making that assessment, organizations drift into treating AI limitations as implementation problems. Worse, they begin redesigning expert workflows around the constraints of a tool.
This is particularly consequential in regulatory affairs and clinical development. Regulatory professionals spend careers learning to distinguish evidence that will withstand scrutiny from evidence that appears compelling on the surface. Clinical development leaders carry similar judgment about study design, data quality, and benefit-risk. That kind of knowledge is built through lessons from programs that succeed, programs that fail, and interactions with agencies.
AI can support elements of that work. Generating a plausible answer, however, is not the same as knowing whether the answer is appropriate in context. Expert judgment remains essential to determining what should even be asked, what constitutes sufficient evidence, and when an output should not be trusted.
None of this diminishes the role of IT. Deployment, cybersecurity, data architecture, access controls, and enterprise scalability are genuinely hard problems that require serious technical expertise. The critical question is knowing when a challenge is one of implementation and when the AI itself is simply not capable of supporting the work as required. That line can only be recognized by people who understand what the work is actually for.
This is why the role of domain experts cannot end at defining requirements or testing a pilot. They need real authority, including the ability to say a model is not good enough, an approach needs to change, or a use case should not move forward. That is not to say technology should not be used, it is the appropriate exercise of professional judgment in a context where the stakes are high and the failure modes are not always obvious.
What this also requires is greater technical fluency within scientific and clinical functions. Not as engineers, but as professionals who understand enough about AI to challenge assumptions, recognize limitations, and distinguish between something that can be fixed and something the AI fundamentally cannot do well.
The organizations that get this right will not be those that have deployed AI into the most workflows. They will be the ones that stayed honest about where AI genuinely improves expert work, where a different approach is warranted, and where human judgment still sets the standard.
AI should adapt to the work. Not the other way around.
About the Author
Craig Wylie, Managing Partner, Head of Arthur D. Little USA
Craig Wylie is Managing Partner and Head of Arthur D. Little USA and a member of the firm’s global Healthcare & Life Sciences Practice. With 30 years of experience in life sciences and a background in computer science, he advises pharmaceutical and medical device companies on quality and regulatory operations, clinical trial effectiveness, manufacturing compliance, and operational transformation.
His combination of deep life sciences expertise and technology experience has placed him at the forefront of AI innovation in biopharma, helping organizations apply emerging technologies to complex regulatory, quality, and operational challenges. Craig has supported many of the world’s leading life sciences companies through major transformations and remediation efforts, with a focus on making quality and compliance an enabler of performance and innovation.
