Building AI Assistants Clinical Trial Teams Can Trust: Q&A with E.K. Koh, Suvoda

As artificial intelligence (AI) assistants move from answering questions to taking actions inside clinical trial systems, the stakes rise sharply: an incorrect answer about drug supply, a visit window, or a participant’s status can carry safety, regulatory, and privacy consequences.

To explore this further, Applied Clinical Trials recently caught up with E.K. Koh, chief product officer at Suvoda, on what makes an AI assistant reliable enough for a regulated environment, how architecture can protect the study blind, and which everyday trial tasks are the best fit for AI assistance right now.

ACT: What separates an AI assistant that clinical trial teams can actually trust from one that creates more risk than it removes?

Koh: In clinical trials, the bar is higher than in many settings, and so a trustworthy AI assistant is one that provides reliable, validated information clinical teams can use to make informed decisions. Study teams are adopting AI hoping to speed up repetitive and time-intensive work, but the risk is that people act on an incorrect answer. An error about drug supply, a visit window, or a participant’s status can carry health and safety implications for participants, along with regulatory compliance and privacy consequences.

Several characteristics reduce that risk. A reliable AI assistant for clinical trials should answer only the questions it has been designed and tested to answer, saying “I don’t know” when a question falls outside that scope. Accuracy and reliability must be built into the system from the outset, rather than treated as features to improve after deployment. Its answers should come from tested skills that work with trial data it is authorized to access, supporting consistent, repeatable responses. It should operate within the roles and permissions already established in the trial systems, and its activity should be fully auditable to support regulatory compliance and data integrity. Without these safeguards, teams must check every response themselves, and the time savings disappear.

For that reason, sponsors evaluating any vendor should ask how the assistant’s responses are tested and validated, what data they are based on, and how the system manages questions outside its defined capabilities.

ACT: How is AI built on curated clinical knowledge different from a general-purpose model, and why does that distinction matter in a trial environment?

Koh: A general-purpose AI model is trained to respond to a very wide range of requests, and it generates each answer as it goes. That flexibility is useful for many tasks, but it makes the model’s behavior difficult to test in advance.

An assistant built on curated clinical knowledge takes a different approach: conversational AI can work to understand the user’s question and map it to the appropriate topic area. Subject matter experts define step-by-step instructions for how specific questions should be addressed. Those instructions become skills that are tested and validated like other trial software. The assistant then uses these predefined skills and authorized study data to generate a response. The system needs to be tightly constrained, reducing the risk of unsupported or fabricated responses.

Sofia, the AI assistant that my team developed at Suvoda, is an example of how this process can work in practice. It covers more than 40 topics across randomization, drug supply, and patient-reported outcomes, and each has been through our full software development lifecycle. Writing and testing each skill takes sustained effort, so we expand deliberately, guided by what customers have told us over the years about the information they need and the actions they want to take quickly.

Coverage may be narrower than a general-purpose model, but in a regulated environment, reliability, validation, and auditability within a defined scope matter more than the ability to answer every possible question.

ACT: When an AI assistant in a clinical trial moves from answering questions to taking actions, what safeguards must be in place?

Koh: Clinical trial data is submitted to regulators such as FDA and EMA, so an action taken in a trial system has to be appropriately authorized, controlled, and traceable to the responsible user. Errors can have significant consequences for study operations, data integrity, and participant safety, making appropriate safeguards essential.

Taking actions is where an assistant may save the most time. Deactivating all of the sites in a country, for example, would ordinarily mean working through several screens and steps, and an assistant can carry it out from a single request. Because an action like that has consequences, three safeguards need to be in place.

Before acting, the assistant should summarize exactly what it intends to do and require the user’s confirmation. Each action should be attributed to the user who requested it and captured it in an auditable record, so a reviewer can determine who did what and when. The assistant should also carry no more authority than the user, meaning it can perform only the actions that person is already permitted to perform.

Sofia follows this model. The user confirms each action and remains accountable for the result.

ACT: How do you make sure an AI assistant respects roles, permissions, and the study blind without undermining what makes it useful in trials?

Koh: The architecture itself has to protect the study blind. If protection depends on the AI assistant exercising good judgment, prompting or system behavior could create an unacceptable risk of exposing restricted information. In a blinded study, the design must enforce blinding protections at the system level, rather than relying on the assistant to decide what information is appropriate to disclose.

The most reliable principle is that the assistant has no independent access of its own. It can reach only the information the signed-in user is already authorized to see under their role, and it applies the same permissions that govern direct use of the system. Sofia is built on that principle. It operates within the Suvoda Platform, inside a governed framework that protects the study blind and creates the audit trail clinical trials require.

Limiting the assistant to each user’s existing access does not make it less useful, because people already do their work within those roles. A site coordinator gets fast answers about their site, a drug supply manager gets answers about supply, and each can ask in their preferred language. Sofia is part of the platform, so there is no separate product to learn and no new login if the trial is using multiple systems like RTSM and eCOA. Importantly, Suvoda applies a zero-data-retention policy to the underlying AI model and does not use customer data to train public models.

ACT: What everyday trial tasks are the best fit for AI assistants right now, and what changes for the people doing them?

Koh: The strongest use cases today are the frequent, time-sensitive tasks or inquiries that take several actions, clicks, or screens to complete. Study teams track drug inventory, review shipment details, monitor visit schedules, and gauge how well patients are keeping up with their questionnaires across global sites. A drug supply manager, for example, may need to know the inventory held at a depot or which lots are released for a country. The answers exist in the system, but accessing and assembling that information can require navigating multiple screens.

With an AI assistant, the user asks in plain language, in their preferred language, and receives an answer drawn from live study data, which can also be presented in a table or chart. The same conversation can support an authorized next step, such as processing a drug supply order, with user confirmation. Sites work across many systems, so the time saved on navigation adds up quickly. As monitoring capabilities are added, I expect some routine oversight tasks to be delegated to agents, with people continuing to review and decide.

I’ve seen in our early adopter trials that Sofia can reduce the administrative burden on routine work so that people can focus on the parts of a trial that call for their expertise, judgment, and decision making. Ultimately, the value of these assistants isn’t simply how quickly they can answer questions or complete tasks. It’s whether they can help clinical teams work more efficiently while maintaining the accuracy, oversight, and auditability that clinical research requires. When those elements come together, technology can support the people responsible for moving trials forward to help bring new therapies to patients.

E.K. Koh is chief product officer at Suvoda, where he leads product strategy, roadmap, engineering, and delivery. He is an experienced product executive who has scaled product organizations across startups, public companies, and heavily regulated industries. Sofia, Suvoda’s AI assistant, is now available for all new clinical trials using Suvoda RTSM/IRT or Suvoda eCOA.