Building a Platform for Agentic AI in Clinical Trial Operations

medable · Healthcare · 2026 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Clinical trial operations and agentic AI platforms

Medable wanted to reduce manual clinical-trial operations work by deploying configurable agentic applications for document classification and trial monitoring, while preserving healthcare-grade customer isolation, validation, evaluation, and human oversight.

Constraints:

• clinical trials can take over 10 years and involve global sites, languages, documents, and patient data
• ETMF users must classify and attach metadata to large volumes of trial documents across hundreds of possible classifications
• clinical research associates must monitor data across 13 or more systems
• each customer and clinical trial has bespoke requirements and isolated data environments
• agents must be validatable, evaluated, and trusted before and after deployment in a regulated clinical setting

Architecture

Medable built Agent Studio as a platform layer for internal teams and customers to configure agentic applications. The platform supports model-agnostic agents, RAG knowledge, MCP connectors, workflow functionality, agent skills for context management, multiple triggers, isolated customer environments, versioned draft-to-publish agent development, built-in evaluations, and human-in-the-loop review. ETMF and CRA applications sit on top as multi-agent, agent-powered products.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  agent-studio[Agent Studio] --> model-configuration[Model configuration]
  agent-studio --> rag-knowledge[RAG knowledge]
  agent-studio --> mcp-connectors[MCP connectors]
  agent-studio --> workflow-and-skills-layer[Workflow and skills layer]
  agent-studio --> trigger-layer[Trigger layer]
  agent-studio --> evaluation-harness[Evaluation harness]
  agent-studio --> customer-isolation-layer[Customer isolation layer]
  agent-studio --> etmf-application[ETMF application]
  agent-studio --> cra-application[CRA application]
  rag-knowledge --> etmf-application
  mcp-connectors --> cra-application
  mcp-connectors --> ai-data-layer[AI data layer]
  ai-data-layer --> cra-application
  evaluation-harness --> human-review[Human review]
  human-review --> etmf-application
  human-review --> cra-application

Components

Data flow

A customer or internal product team configures agents in Agent Studio with model parameters, system prompts, relevant knowledge repositories, MCP connectors, workflows, and triggers. For ETMF, document-classification agents use trial-document knowledge and golden datasets to classify documents and metadata, then human reviewers provide trust-building feedback. For CRA workflows, agents connect multiple trial data systems, surface understandable monitoring information, recommend actions, and can take actions with human oversight.

Why this architecture

Medable chose a platform-first architecture because clinical-trial customers have bespoke requirements but share repeated needs: isolated environments, RAG over protocols and documents, connectors to many systems, model/configuration control, workflow triggers, evaluation stage gates, and human oversight. Building Agent Studio lets new solutions reuse these capabilities instead of rebuilding them for each agentic application.

Models used: model-agnostic agents · flagship models · agent skills · RAG systems

Scale

• Medable frames the clinical-trial cycle as taking over 10 years today, with an ambitious goal of reducing operations toward one year.
  “looking to solve the problem that is it takes over 10 years currently to get a drug to market. So where our bread and butter has been the last few years has been in the clinical assessments and e-consent space, where patients are able to complete electronic questionnaires”transcript, stated

  “operation. So that 10 years could be potentially reduced down. Our goal, our ambitious big and hairy goal is one year. And so, essentially like it starts from creation of a protocol, which is the rules of the clinical trial, and actually ensuring that the science team”transcript, stated
• Medable supports clinical sites globally and over 100 languages.
  “>> We are the clinical sites that we support over 100 different languages and we support clinical sites all over the world. >> Oh, wow. Okay, excellent. Yeah, we tend to have a global audience, so I always like to set the scope of what we're talking about. Okay, so now I understand”transcript, stated
• A given customer can upload over 80,000 ETMF documents per year.
  “happened today is there's over 80,000 documents a year uploaded to the system just for a given customer and essentially it takes the users that have to upload these documents into the system around at least 5 minutes per document to assign the associated”transcript, stated
• The ETMF classification task spans over 350 possible classifications.
  “solve, right? There's over 350 classifications that the user has to understand and assign the associated metadata to. So, leveraging AI to solve this problem just made sense where we're able to we are would be able to completely take the humans out of that process of the classification and the”transcript, stated
• Some sponsors or CROs may have over 1,300 clinical research associates.
  “probably sometimes over 1,300 clinical research associates at a given sponsor or or a CRO and the people that are actually funding the trial. And essentially what they what their job is to monitor the data coming through the from the clinical research sites and”transcript, stated
• CRA monitoring spans 13 or more systems.
  “there's about 13 or more different systems where data is coming in through throughout the life of the clinical trial.”transcript, stated
• Single studies can generate tens of thousands of documents per month.
  “the actual scale of the documents. There are single studies that produce tens of thousands of documents per month. This is This is not It's not tens of thousands documents over 10 years. It's tens of thousands documents a month per study. It's a lot of documentation.”transcript, stated
• ETMF evaluation used a golden dataset of over 2,000 documents.
  “take like over 2,000 documents and we're able to run the agent against these 2,000 documents where we knew the correct answer.”transcript, stated

Alternatives considered (roads not taken)

Build one-off agents for each product or customer

Rejected because: Medable's customers and trials differ, but repeated agent capabilities could be reused through a platform.

Limit agents to chat-triggered interactions

Rejected because: Agent Studio supports Teams, Slack, webhooks, and system triggers so agents can be invoked without a human starting a chat.

Trust human review as unquestioned ground truth

Rejected because: The team recognizes human reviewers can be wrong and works with customers to adjudicate agent-human disagreements.

Evaluate only once before launch

Rejected because: Medable runs evaluations as a stage gate and on a regular frequency to detect response drift after deployment.

Patterns

multi-agent-orchestration (agent-platform)

ETMF and CRA are not single agents; they are applications built from ecosystems of agents, sub-agents, front-end experience, and non-agentic databases.

rag-retrieval (knowledge)

Agents attach relevant knowledge repositories and protocol/document context, while the team builds an AI data layer for retrieval across large clinical systems.

tool-use-orchestration (connectors)

Agent Studio connects to external clinical and operational systems through MCP connectors and workflow triggers.

human-in-the-loop-gate (clinical-oversight)

Medable starts classification and recommended-action workflows with human review to build trust, monitor agreement, and adjudicate disagreements.

eval-driven-development (validation)

Agent Studio has built-in evaluations for response quality, tool-order correctness, deployment stage gates, recurring monitoring, and golden-dataset validation.

privacy-preserving-processing (regulated-data)

Medable carries its SaaS isolation model into agents, using per-customer isolated environments where it acts as the data processor.

Tradeoffs

Build a reusable agent platform

Chose Agent Studio as shared platform infrastructure over One-off point solutions for ETMF, CRA, and each customer.

Rationale: A platform lets Medable reuse capabilities across bespoke customer needs while accelerating future agentic products.

Make the platform model-agnostic

Chose Bring-your-own or flagship model configuration over Tying Agent Studio to one model provider.

Rationale: Clinical customers and agent use cases need flexibility to configure models and parameters per agent.

Start with human-in-the-loop deployment

Chose Human review and active monitoring over Immediately removing humans from classification and action workflows.

Rationale: Human review builds trust and produces agreement/disagreement signals while the agent's accuracy is validated.

Use versioned draft-to-publish agent development

Chose Versioned agents with evaluations before publishing over Ad hoc prompt/configuration changes sent directly to users.

Rationale: Clinical-trial agents need traceable configurations and validation stage gates.

Gotchas (what bit them)

Clinical-trial agents cannot be generic out-of-the-box bots.

Root cause: Every trial and customer has bespoke requirements, isolated data, and different operational rules.

Resolution: Use Agent Studio as a configurable platform with isolated customer environments and purpose-built agents.

A human-in-the-loop label is not automatically ground truth.

Root cause: Human document classifiers can be wrong, sometimes discovered years later.

Resolution: Treat disagreements as deltas, work with customers to decide which answer is correct, and feed verified corrections back into the agent.

Large clinical data retrieval is not solved by merely adding connectors.

Root cause: Agents need the right context across large data volumes, multiple systems, MCP design, and context-window limits.

Resolution: Medable is building an AI data layer with a common ontology to align systems for retrieval and tool use.

Agent responses can drift after publication.

Root cause: Model, configuration, and data changes can alter responses over time.

Resolution: Run evaluations before publishing and on a regular frequency after launch.

Evaluation

Medable evaluates at the Agent Studio platform level and at application levels such as ETMF and CRA, using response relevance, tool-order correctness, configuration benchmarking, pre-publish stage gates, recurring monitoring, a golden dataset of known-document classifications, and human agreement/disagreement review.

Metrics: agent response accuracy and relevance · correct tool calls and tool order · performance across multiple configurations · golden dataset accuracy on over 2,000 ETMF documents · human agreement/disagreement with agent recommendations · response changes over time

Outcomes

• Medable deployed Agent Studio as a reusable platform for internal and customer-built agents.
• ETMF uses agentic document classification and metadata assignment with human review and golden-dataset evaluation.
• CRA workflows connect 13 or more data sources, surface data quality and safety information, recommend actions, and can take action with oversight.
• The platform carries forward Medable's customer isolation and data-processor model into agentic deployments.

The case covers clinical operations at global scale: over 100 languages, over 80,000 documents per year for one customer, 350+ classifications, 13+ CRA systems, and a golden evaluation dataset of over 2,000 documents.