AI won’t fix clinical trials until we redesign how decisions are made

Imagine it's 2032… 

 …a major pharmaceutical company has just tripled the number of treatments approved by the FDA in a single year. At 70% of the cost.

That outcome did not come from a breakthrough in molecule design. It came from something less visible: a complete reinvention of how clinical development decisions are made.

All development activity now flows through a connected decision environment that combines continuous simulation with a network of specialised AI agents. Before a trial ever reaches a patient, thousands of alternative futures have been explored. Different designs, populations, doses, endpoints, timelines, operational constraints. Decisions are rehearsed and stress-tested in silico, while there is still time to change them.

Research agents continuously scan historical trial data, real-world evidence, and emerging literature, proposing protocol designs grounded in what has and has not worked before. Operations agents monitor recruitment, site performance, and data quality in real time, surfacing adjustments before timelines slip. Regulatory agents assemble draft submissions and flag inconsistencies long before they reach health authorities. Every agent operates within strict policy bounds. Every recommendation is explainable. Every action is logged and auditable.

The effect has been a step change. Assets reach Phase III readout nearly 40% faster. Trials that once took six months to plan are submitted for regulatory approval in a day. Late-stage failures driven by avoidable design and operational errors have collapsed.

But the most profound change is organisational. There are no longer separate global functions for clinical development, clinical operations, biostatistics, safety, and regulatory affairs handing work off in sequence. Phases I, II, and III are no longer treated as rigid gates. Entire layers of coordination, synthesis, and documentation work have vanished. Small, highly skilled teams oversee development portfolios that once required thousands of people. Headcount has fallen significantly, but capability has increased.

This might feel like a far-off vision. But the building blocks - from simulation powered trial designs to protocol drafting agents - are all proven. So the question is what it actually takes to get there. 

Why better tools alone won’t fix the problem

There is a principle from software engineering that holds: an organisation's systems will mirror its own communication structure. 

Consider how a study design decision are made today. Clinical development drafts initial protocol concepts. Biostatistics shapes the statistical design, endpoints, and sample size. Clinical operations assesses feasibility. Regulatory reviews against agency expectations. Medical affairs may weigh in on clinical relevance.

Bolting on better tools without changing the structure they sit inside produces a slightly faster version of the same system. As my colleague Steph Skeet has written, the result is like putting a Ferrari engine on bicycle wheels: a supercharged discovery engine feeding into a 1990s execution system.

No matter how powerful the science, point solutions that accelerate one stage cannot overcome the system as a whole. Each function optimises for its own constraints. The time taken to run a trial is still constrained by information hand offs and sequential manual iteration between teams of people. Scaling amplifies the number of people required and the complexity of the trade offs. And by the time a protocol is locked, it reflects a series of sequential compromises that no single person designed, and no single person would have chosen. 

Tom Oliver, Head of Product at Faculty, describes this clearly in his recent blog: this is one of the most consequential decisions in a trial, one that locks in cost structure, risk profile, and timeline - and the logic behind it is essentially invisible.

How to get there: organising around the decision 

Now consider what changes when you reorganise around the decision rather than the function. 

For study design, work that was spread across functions, assembled sequentially, is now orchestrated around a single point of commitment. Technology designed to support the decision meeting simulates and optimises across all design parameters at once. High-bandwidth discussions and evidence-driven trade-offs are now instantaneously possible. Entire layers of coordination shrink or disappear. Smaller, highly skilled teams now oversee many trials in parallel. The study design process collapses into a single day. 

Capturing this decision in technology means the logic behind each commitment can be captured, traced, and learned from. What was known, what was chosen, what followed. Decisions stop being invisible and become reusable, compounding assets – so that each trial becomes better and faster to design than the last. 

And this is just what it might mean for study design. 

Realising this vision, however, cannot be delegated to a data science team or resolved by selecting the right vendor. It requires development leaders to ask whether the functional boundaries and processes they have inherited are fit for purpose for the current moment. A moment where AI progress means we have an opportunity to radically increase the number and speed of new treatments brought to patients. 

This is a big change. But one others are already on the way to realising.

Our approach to getting there in a way that is incrementally valuable and collectively transformative is one we’ve written about in our 10 lessons book with Shreeram Aradhye, President, Development & CMO at Novartis. The starting point is not asking what AI or agents can do. It is mapping the decisions that drive the KPIs you care about—and then working backwards to design everything around them: the AI, the data, the processes, and the people.

At Faculty, this is what we are working on with some of the world’s largest pharmaceutical companies. With Faculty Frontier™ and the scale that comes with being part of Accenture, we support not just the technology, but the organisational transformation required to realise the value.

The organisations that reach the 2032 scenario will not be the ones that spent the most on AI. They will be the ones that were willing to treat the decision-making process itself, and the organisational structure around it, as the thing that needed to be rebuilt.