Early-stage biotech’s AI advantage is a window that closes. Sol Babani and Derk Arts on how to use it.

At the AWS Life Sciences Symposium the previous week, Solomon Babani, Founder and CEO of Symbiosis Advisors, sat through presentation after presentation on AI in clinical development. His takeaway: “There’s a lot of retrofit going on.” Companies were taking existing processes, existing workflows, and asking how to bolt AI onto them. He left mostly disappointed.

That framing opened a live conversation on April 15 between Sol and Derk Arts, MD, PhD, CEO and founder of Castor, on what AI strategy actually looks like for early-stage biotech clinical trials teams. An hour, no slides, no scripted pitch. The session covered why most AI guidance misses the point, where the FDA’s guidance framework actually helps, and what separates the teams that will get this right from the ones that won’t.

The distinction is not budget, team size, or therapeutic area. It’s whether you’re building or retrofitting. Early-stage biotech can build. Most of pharma can’t.

The greenfield window is real

Big pharma is stuck. Studies designed years ago, processes embedded in SOPs that cost more to change than to work around, and organizational immune systems that turn almost any innovation into a retrofit. Early-stage biotech teams are in a completely different position.

“I think it would be almost foolish not to entertain how to modernize your approach. You have the opportunity, you don’t have to worry about undoing legacy systems, legacies, processes, procedures, and all of that, and look for a more efficient way.”Sol Babani, Founder and CEO of Symbiosis Advisors (14:29)

That window closes the moment a team builds a first study on manual processes. The longer you wait, the more you are retrofitting too. Derk noted that Castor itself is working to remove manual data entry from the process entirely, describing how Castor Catalyst handles source document review and data capture with human oversight on the output rather than at the data entry stage. The goal, he said, is to create a digital trail that makes monitoring a remote and efficient process rather than a site visit.

FDA guidance: useful signal, limited map

Derk’s read on the current FDA AI guidance was specific: positive direction, weak on practical detail. His estimate was that about ninety-eight percent of the innovation actually happening in this space involves teams using existing models, not building their own. The guidance focuses heavily on model development. That maps to a fraction of what most teams are actually doing.

Sol’s answer: stop waiting for detailed guidance and start with a structured decision framework instead. Identify the business problem. Define the specific AI use case. Work through validation requirements and data quality. Estimate cost and timeline. Define what success looks like before you start. That sequence matters. Skipping any step is where implementations go wrong.

Build versus buy: the compliance variable

For data collection and electronic data capture infrastructure, Sol’s position was unambiguous: in a 21 CFR Part 11 environment, building your own compliance layer introduces validation risk that established vendors have already absorbed. The cost gap between buying sophisticated tools and building from scratch has collapsed. Most biotech teams don’t need to build. They need to choose well.

Human oversight is a methodology, not a checkbox

Sol was clear that human-in-the-loop oversight is not just a regulatory compliance requirement. It is an active design choice that shapes how a team builds its AI workflows from day one. Derk put the practical risk plainly:

“It just produces believable bullshit if you’re not careful. When you do it properly, it doesn’t do that. But if you don’t do it properly, it probably will.”Derk Arts, MD, PhD, CEO and founder of Castor (39:58)

The right setup requires a well-structured data source, a paid model with appropriate data protection agreements, and a documented review workflow. The shortcut through free public models is not a starting point. It is a risk that looks like one.

Derk also framed the financial stakes of getting AI right across the industry, noting that failed oncology studies represent billions of dollars in lost investment annually, as context for why modernizing clinical trial infrastructure matters beyond individual efficiency gains.

What’s in the full recording

The session covered more ground than the highlights above capture. Specific moments worth finding:

Sol walks through exactly why CRO economics are a structural obstacle to faster AI adoption in monitoring. A large part of CRO revenue is tied to monitoring visits, and he explains where the conflict actually sits and what biotechs can do to address it in negotiations. (26:00)

Derk breaks down what he calls the four-stakeholder problem in clinical AI: the vendor, the CRO, the sponsor, and the site, each with different definitions of what “saving money” means and different incentives to move toward or resist change. (29:04)

A detailed exchange on the buy versus build decision, including Derk’s specific reasoning for why building 21 CFR Part 11 compliant infrastructure against an existing EDC system would be “quite risky” for most biotech teams, and where Sol draws the line between cases that justify building versus buying. (22:13)

Sol’s five-step decision framework, presented in full with specifics on validation approach, training data requirements, timeline and cost estimation, and how to define what a successful implementation looks like at the end. (37:34)

Frequently asked questions