Discovery Gains Are Real – But the Development Opportunity Is Still Largely Untapped

The first session featured three pharma R&D executives — Don Bergstrom (President of R&D, Relay Therapeutics), Fiona Marshall (President of Biomedical Research, Novartis), and Andrew Plump (President of R&D, Takeda) — discussing how their organizations are deploying AI. As expected, a lot of the conversation centered on discovery — AI-driven target identification, molecular design, binding prediction. This is well-trodden territory by now, and the investments are enormous.

But then the conversation shifted, and this is where it got interesting.

The panelists acknowledged the real gains AI has already delivered in discovery — shorter timelines and new molecular entities that are now in late-stage trials. What they kept coming back to was the opportunity that remains: clinical development, and specifically the enrollment bottleneck that derails timelines on nearly every program regardless of how good the molecule is.

Clinical trial enrollment is consistently one of the longest and least predictable phases of drug development. It’s where timelines slip, budgets balloon, and competitive advantages erode. Don Bergstrom from Relay was candid about this — he described the difficulty of finding patients even for their own oncology studies, and mentioned one program where they’re actively employing AI-driven patient identification approaches.

So why hasn’t this been solved already?

The Missing Infrastructure: Site Networks

After the panel, I had a side conversation with Bergstrom where I pressed on this point. The AI tools for patient identification and feasibility are impressive — but the patients are going to be recruited at sites, and the data lives in site systems. Given the time pressure to get a trial started, how do you actually get the infrastructure in place fast enough?

His answer was sharp: Site Networks

Specifically, he pointed to HCA Healthcare and Sarah Cannon Research Institute (SCRI), which had acquired US Oncology — an oncology CRO and site network. That combination gave them a pre-built infrastructure: sites that were already enabled, data systems that were already connected, and an operational layer that was ready to activate when a new trial needed to start. They didn’t have to build it from scratch under the gun of a trial timeline. The network was already there.

That’s the insight that stuck with me from Monday. The bottleneck isn’t the AI. It’s the absence of site networks that are wired up and ready for AI to work through. And the companies that have figured this out — like HCA/SCRI with US Oncology — are the ones that can actually translate AI-powered feasibility into enrolled patients.

At Graticule, this is exactly the problem we’re working to solve — and I want to be clear about what “solving it” actually requires, because it’s easy to think the answer is just better technology.

It’s not. Technology is important, but the foundation of a functioning site network is governance, trust, and alignment on how to simplify doing business together. It’s contracts and data-sharing agreements. It’s standardization across sites so that a sponsor doesn’t have to negotiate bespoke terms with every institution. It’s deciding, up front, what a pharma company is allowed to do with the network — and making those rules clear enough that the network can actually move at speed.

Here’s a concrete example. Every pharma company we talk to wants the same thing: give us patient counts from 30 large health systems, instantly, diving into clinical notes to refine the feasibility of recruitment. The technology to do this exists today. We can run NLP across unstructured EHR data and return refined cohort counts in hours. But the governance to allow it? That’s the bottleneck. Can you deploy self-service feasibility tools before you’ve contracted the sites? Who has access? What data can they see? Under what terms? These are governance questions, not technology questions — and they have to be answered before the first query runs.

What makes this harder is the organizational disconnect inside pharma itself. Today, most companies can run these kinds of deep feasibility queries in their real-world evidence groups — they have access to large de-identified datasets and the analytical tools to interrogate them. But those RWE teams sit in a completely different part of the organization from the clinical trials groups who actually need the answers. The data that could accelerate site selection and enrollment lives behind an organizational wall, not a technical one.

Building a site network that works for clinical trials means solving all of these problems together — the technology, yes, but also the governance frameworks, the contracting models, and the operational standardization that let sites and sponsors work together efficiently and repeatedly, not just on a one-off basis. That’s what we’re building at Graticule: not just a data platform, but the enablement layer that makes a network of health systems ready to participate in AI-powered clinical development across therapeutic areas.

A Different Kind of Investment — And a Different Kind of Strategy

Here’s what most pharma companies haven’t yet internalized: the investment profile for AI in development is fundamentally different from AI in discovery.

In discovery, AI is a tool you can buy and feed data. You acquire platforms that crunch chemistry, predict binding, and optimize molecules. It’s capital-intensive, but it’s a familiar investment model — license the software, hire the computational team, point it at your target library.

Development doesn’t work that way. You can’t just feed data into a patient identification algorithm if the data isn’t accessible at the sites where patients are seen. You can’t accelerate site feasibility if the sites aren’t organized to respond. The hard part in clinical AI isn’t the model — it’s that patients are recruited at sites, the data lives in site systems, and the sites themselves aren’t wired for this.

This means the investment isn’t in tools. It’s in building and fostering networks — or fundamentally changing your operating model to select networks of sites that are organizing for efficiency, rather than defaulting to the traditional model of picking the best KOLs or the sites that historically enrolled the most patients. That’s a sea change in clinical development strategy. It requires rethinking not just how you run a trial, but how you build out the operations infrastructure that makes trials possible in the first place. And here’s the paradox: the enemy in clinical trials is time, and time is also the reason this infrastructure never gets built. Trials are always a hurry-up offense. The moment a program enters the clinic, the pressure is relentless — enrollment targets, competitive timelines, patent clocks. There’s never a “good time” to pause and invest in network readiness.

We see this at Graticule every day. A pharma company comes to us with a trial and says: here’s our list of sites — are they in your network? That’s backwards. Or we hear it from the site side: I have all these pharma companies knocking on my door, and every one of them wants to work differently. The sites are overwhelmed, the data systems aren’t interoperable, and each sponsor is reinventing the wheel.

The companies that break out of this cycle will treat site network development the way a football team treats the offseason. You don’t build the playbook during the game. You set up the plays in the downtime — investing continuously in the network, the data infrastructure, and the site relationships — so that when the next program enters the clinic, you’re ready to execute from day one. The investment has to happen before the trial, not as a scramble once it starts.