Expert evaluation tiles representing private clinical development AI testing and review.

Life sciences · · 8 min read

How to evaluate AI agents for clinical trials

Acumen Editorial · Updated November 20, 2026

AI agents have reached clinical operations. On November 18, 2026, Oracle announced Oracle Life Sciences AI Agents for early adopters. Oracle says they monitor study performance, flag operational issues, suggest likely causes of KPI variances and recommend actions that follow the trial protocol. Oracle plans to extend them to clinical data management, trial design and post-market work.

The pressure behind this is real. Two days later, Reuters reported that scientists at several large drugmakers say AI has cut the time to get new molecules to the clinic by as much as half. Faster discovery pushes more candidates into trials, and trials are where most drugs fail: nearly 9 in 10, according to PhRMA figures cited by Reuters.

So the next bottleneck is clinical development. If you lead clinical operations, data management, biostatistics or AI in R&D, here is how we would test trial agents before they touch a live study.

What do AI agents actually do in clinical trials?

The jobs are getting specific. Based on Oracle's announcement and current research, the common agent roles are:

  • Portfolio and study monitoring agents that track enrollment, site performance and KPIs, and flag studies drifting off plan.
  • Root-cause agents that suggest why a metric moved, such as slow enrollment at a group of sites.
  • Action-recommendation agents that propose next steps within the protocol and route them to the right team.
  • Data management agents that check incoming data, raise queries and reconcile sources.
  • Trial design agents that propose design elements such as dose-finding approaches, endpoints or eligibility criteria.

Each one fails in a different way. A monitoring agent that misses a slow site costs weeks. A design agent that proposes the wrong dose-finding approach can shape an entire program. One accuracy score cannot cover both.

Does FDA's AI credibility framework apply to trial agents?

Sometimes, and the line matters. FDA's January 2025 draft guidance on AI to support regulatory decision-making for drugs sets out a seven-step credibility framework. It is still a draft. The guidance ties credibility to a defined context of use, meaning how the model answers a specific question.

Model risk has two parts in the framework, per Troutman Pepper's analysis. Model influence is how much the output drives the decision compared with other evidence. Decision consequence is the harm if the output is wrong.

The scope is the part buyers miss. The draft excludes drug discovery and operational efficiencies such as workflows and resource allocation. It covers models that affect patient safety, drug quality or the reliability of study results.

That means the same agent can sit on either side of the line. A dashboard that ranks site performance is likely operational. An agent that recommends protocol actions, flags safety signals or cleans data feeding the primary endpoint can affect the reliability of results. Plan for the stricter case.

Why aren't public benchmarks or vendor demos enough?

Because trial work depends on context that is not written down. TrialDesignBench, presented at a NeurIPS 2026 workshop, tested five frontier models on 617 Phase I and I/II oncology dose-finding trials. Each trial was paired with an expert-validated reference design reflecting FDA's Project Optimus. None of the five models matched the reference standard.

The authors traced the gap to two things. Models could not infer the design determinants that trial records leave out. They also lacked structured, decision-tree reasoning. Supplying the missing determinants raised balanced accuracy by 0.22, and adding structured reasoning raised it by 0.24.

The lesson for buyers: the expert knowledge that is not in the source documents is what moves performance. A demo on a clean sample study will not show you that gap. Oracle's announcement includes no performance metrics, which is common for launches. You need your own evidence.

How should you build the evaluation set?

Build it privately, from your own studies, and score each agent on its own job. Here is the checklist we use:

  • Define the context of use for each agent. Write down the question it answers, its data sources, what it may recommend and what it must never do.
  • Rate model influence and decision consequence. Use the result to set how many cases, how many graders and how strict the pass threshold should be.
  • Use real, messy inputs. Past enrollment curves, site histories, protocol deviations, query logs and amendments, with identifiers removed.
  • Seed known problems. Plant issues you already know the answer to, such as a site with a known data quality problem or a deviation pattern.
  • Write reference answers and rubrics. Score correctness, protocol fidelity, reasoning and actionability as separate criteria.
  • Match graders to the task. Clinical operations leads, monitors, clinical data managers, biostatisticians and clinical pharmacologists, depending on the agent.
  • Grade blind and adjudicate. At least two graders per item, disagreements resolved and agreement reported.
  • Document everything. Your plan, results and deviations become the credibility record if the agent ever falls in scope.
  • Rerun on every change. New model version, new data source, protocol amendment or new study type means a new run.

Which metrics matter for each agent?

Use the metric that matches the failure that hurts most.

AgentFailure that mattersPrimary metricLikely FDA scopeWho grades
Study monitoringMissed off-plan study or siteRecall on seeded issues, lead timeUsually operationalClinical operations leads
Root-causePlausible but wrong explanationExpert agreement on top causeUsually operationalClinical operations, monitors
Action recommendationAction that breaks protocolProtocol-fidelity error rateCan affect resultsClinical operations, regulatory
Data managementMissed discrepancy in key dataQuery recall and precisionCan affect resultsClinical data managers
Trial designWrong design for the questionAgreement with expert reference designCan affect safety and resultsBiostatisticians, clinical pharmacologists

For monitoring and data agents, recall matters more than precision. A false alert costs a reviewer a few minutes. A missed one can cost a database lock date. For design agents, grade the reasoning, not only the final answer, since TrialDesignBench shows reasoning structure drives the score.

What should you ask a clinical trial AI vendor?

Ask for the evidence, not the summary:

1. What context of use was each agent tested for, and on what cases?

2. Who graded the outputs, and what were their roles and credentials?

3. What are the recall and error rates on seeded issues, by agent?

4. How does the agent behave when the protocol does not answer the question?

5. Can we run our own private evaluation set before rollout and after each update?

6. What audit trail exists for each recommendation, approval and action?

Human review helps, and Oracle says its agents are "designed to operate with human oversight." But a reviewer approving a fast stream of confident recommendations needs to know how often they are wrong. Only evaluation tells you that.

How Acumen helps

We build the evaluation and training data behind life sciences AI. Our life sciences model evaluation service supplies private evaluation sets written and graded by practicing scientists and clinical development specialists. That includes reference answers, rubrics, seeded-issue cases and regression runs on every release. Every item is reviewed and adjudicated.

Use it to test a vendor's agent, to tune your own, or both. For how lab-facing AI scores on raw experimental data, see our breakdown of BioBasicsBench.

Request a sample dataset Schedule a conversation

FAQ

What are AI agents for clinical trials?

They are AI systems that complete trial tasks with some autonomy instead of only answering prompts. Current examples monitor study and site performance, suggest causes for KPI changes, recommend actions within the protocol, check incoming data and propose design elements. Vendors such as Oracle describe them as working under human review, with people approving actions before they run.

Does FDA's AI guidance cover clinical trial operations agents?

It depends on the use. FDA's January 2025 draft guidance excludes operational efficiencies such as workflows and resource allocation. It covers AI that affects patient safety, drug quality or the reliability of study results. An agent that only ranks sites is likely out of scope. One that cleans endpoint data or recommends protocol actions may be in scope.

What is a context of use for an AI model?

FDA defines context of use as how an AI model is used to address a certain question of interest. In practice, it means writing down the exact question the agent answers, its inputs, its role in the decision and its limits. The context of use sets the risk level and how much evidence the evaluation needs.

How do you test an AI agent that monitors clinical trials?

Build a private test set from your past studies with identifiers removed. Seed known problems, such as slow-enrolling sites or deviation patterns, and measure how many the agent catches and how early. Have clinical operations leads grade its explanations and recommendations blind against a rubric. Rerun the set after every model or data change.

Can LLMs design clinical trials yet?

Not reliably on their own. TrialDesignBench tested five frontier models on 617 Phase I and I/II oncology dose-finding trials against expert-validated reference designs. None matched the reference standard. Adding missing design determinants and structured reasoning improved scores, which suggests expert input remains the main lever for trial design tasks.

Who should grade clinical trial AI outputs?

Match the grader to the job. Clinical operations leads and monitors judge study and site recommendations. Clinical data managers judge queries and reconciliation. Biostatisticians and clinical pharmacologists judge design proposals such as dose-finding. Use at least two graders per item, resolve disagreements by adjudication and report agreement rates alongside scores.

From the Acumen archive. Explore expert data and Healthcare & Pharma.

← Back to Insights