Agentic AI for Clinical Trials: Medable’s Roadmap to Speed and Scale

In this in-depth Q&A, we sit with Andrew Mackinnon to explore Medable’s agentic AI platform and its first agents for clinical trials. We discuss how the Agent Studio, featuring CRA and TMF agents, is engineered to cut trial timelines, reduce administrative burden, and unlock parallel execution across larger networks. The conversation delves into practical design principles, governance models, regulatory considerations, and implementation strategies to balance rapid AI-enabled progress with rigorous data integrity and compliance. No conclusions—just the insights shaping the next era of trial automation.

Moe: I’ve framed Medable as agentic AI to speed timelines from years to under a year. What human bottlenecks still slow trials, and how do CRA/TMF agents fix them?

Andrew: Absolutely. I’m not talking about replacing humans; I’m augmenting them. The CRA agent doesn’t merely tick boxes; it analyzes across systems to surface root problems—missing data, misaligned protocols, or out-of-sequence documents. It can write queries into the EDC, chase missing items by email, or surface issues for human triage. The real value is shrinking the micro “white space” where nothing happens, so people can shift time toward planning, interpretation, and strategy while agents handle reconciliation, data gathering, and routine chasing. This enables more trials in parallel—think of a CRA able to manage 50-100 sites instead of 5-10—by augmenting scarce CRA capacity with an agentic workforce. We’re building a scalable asset funnel, not just speeding a single trial.

The approach also addresses asset throughput. By expanding capacity, sponsors can push more assets through the funnel concurrently, which is crucial as discovery systems return more candidates. The agents’ impact isn’t simply faster task completion; it’s enabling a broader, faster pipeline that unlocks parallelism. If you only add humans, you’ll hit a ceiling; with agents, you can handle more sites, more data streams, and more complex correlations without a linear headcount explosion. That’s how we start to turn years-long timelines into a multi-year-to-months reality, while preserving quality and oversight.

Moe: TMF is often seen as a documentation hurdle, yet it’s where audit risk concentrates. What resistance have you seen automating TMF workflows, and how did that shape your design?

Andrew: The TMF agent centers on a guiding principle: can we move a document from inception to the correct TMF location with the right metadata and protocol references, ideally without human touch? The goal is a humanless pathway that keeps TMF inspection-ready—complete, current, and auditable. Historically, filing into TMF was a Friday-afternoon chore that bled into the next week, month, or even Christmas downtime, amplifying risk and stress. Automating this workflow reduces that cycle time, improves audit readiness, and cuts dull administrative workload. Resistance in TMF is relatively modest because the payoff is tangible: higher quality, lower risk, and less cognitive burden. For companies processing north of a million documents a year, automation can dramatically reduce resource needs. TMF serves as a pragmatic gateway, proving automation’s value and creating room to extend to other domains.

From a change-management lens, we’re cognizant that fear of automation persists; however, when teams see near-real-time, inspection-ready TMF improvements, the conversation shifts. The TMF domain is a strong proof point for broader automation, because it directly touches quality and compliance. In practice, TMF automation also offers a clean, measurable ROI: faster issue identification, fewer late-stage corrections, and more predictable audits. As adoption grows, TMF becomes the anchor that demonstrates governance, reliability, and the potential of agentic workflows across the trial lifecycle.

Moe: You emphasize human-in-the-loop checkpoints for validation. How do you decide where automation stops and where humans must remain? Were there debates about speed vs. trust?

Andrew: We designed an autonomy dial—organizations can tune how autonomous an agent is, depending on risk tolerance and regulatory expectations. For the TMF example, the agent provides a confidence score (for instance, 94% certainty on classification). If confidence falls below a threshold (say, 80%), we could require a human double-check. If confidence is high but the classification is ambiguous (e.g., 95% one way, 92% another), governance rules decide whether a human resolves the ambiguity. Audits and inspections remain non-negotiable; we can support these with inspection agents and human oversight where needed. Because agents are probabilistic rather than deterministic, validation becomes ongoing governance: validate initially, launch, and continuously monitor performance and drift. Regulators will increasingly offer guidance, so we’ll adjust the autonomy level as those expectations evolve. The debates aren’t about stopping speed—it’s about ensuring trust, traceability, and defensible decisions in every workflow, especially where compliance risk is high.

Moe: From a market perspective, how will agentic layers reshape sponsor–CRO dynamics, given system-agnostic integration and cross-stack reach?

Andrew: The Model Context Protocol (MCP) approach makes integration lightweight and broadly applicable. Agents can connect across most organizations’ tech stacks, reducing barriers for adoption, particularly for smaller CROs. That means smaller players can expand capabilities without proportional headcount growth, leveling the playing field against larger integrators. At the same time, tech-focused, agent-centric CROs can become more powerful, offering scalable capabilities beyond their headcount. Sponsors and CROs benefit from faster onboarding of new workflows and more flexible, scalable trial operations. The broader implication is a more competitive, agile ecosystem where the success of trials depends not only on science but on how effectively teams adopt, govern, and leverage agentic AI across diverse environments.

Moe: If 95% of TMF documents are still processed manually, what legacy mental models keep manual reconciliation alive, and how will automation shift that thinking?

Andrew: The core bias is the belief that humans are inherently better at these tasks than machines. That assumption hasn’t kept pace with AI capabilities that handle millions of documents with many classifications. Deterministic, rule-based filing can’t scale to millions of cross-referenced documents, versions, and protocols. Agents use probabilistic, context-aware reasoning and continuous learning, enabling more nuanced classification and routing. As organizations observe tangible gains in quality and efficiency, the old mental model loses credibility. TMF thus becomes a compelling proof point for broader automation, and as adoption grows, teams realize automation doesn’t compromise accuracy—it enhances it. This mindset shift is essential to unlock parallel trial operations and broader agent deployment across the workflow.

Moe: With several agents rolling out, how do you prevent cognitive overload and ensure adoption remains clear? What signals indicate success or risk?

Andrew: Our rollout is deliberately structured around implementation science and change management. We start with a proof-of-value phase tied to explicit ROI KPIs, then scale in measured steps to solve high-impact problems first. Clear success signals include measurable ROI, faster issue resolution, fewer manual touchpoints, and positive user feedback. Negative indicators—rising cognitive load, governance bottlenecks, or unclear benefits—trigger recalibration: narrow scope, adjust governance, or reduce autonomy. We maintain a patient, evidence-driven approach so stakeholders see value quickly and can scale without overwhelming teams. The aim is to balance speed with governance, ensuring adoption builds confidence rather than fatigue.

Moe: Finally, how will Medable align rapid AI progress with regulatory demands and data integrity?

Andrew: The balance is speed plus governance. We embed continuous validation, audit trails, and governance into deployment, starting with TMF and then extending to other domains. This approach demonstrates steady gains in inspection readiness and data quality while staying aligned with evolving regulatory guidance. We’re prepared to adapt validation strategies as expectations evolve, relying on ongoing monitoring, traceability, and risk-based governance. The objective is faster trials without compromising integrity or compliance, delivering outcomes that regulators can trust and organizations can scale with confidence.