Building Agent Studio: How Medable Is Using Agentic AI to Accelerate Clinical Trials

Overview

• Speaker: Teresa Torres with guests Luke Bates, Jen, Piera Matthews, and Matt Schoolfield
• Channel: Product Talk
• Main topic: Building agentic AI systems and agent platforms at Medable to accelerate clinical trials
• Purpose: To share how Medable built Agent Studio and utilized agentic AI to tackle complex clinical trial operations, reduce drug development timelines, and handle massive data volumes while maintaining regulatory compliance. Teresa Torres hosts a panel discussion with Medable team members Luke Bates, Jen, Piera Matthews, and Matt Schoolfield about Agent Studio, their platform for building agentic AI products. They discuss the massive administrative, regulatory, and data challenges in clinical trials, how they transitioned from legacy systems to agent-driven workflows, the importance of human-in-the-loop design, data ontologies, and evaluation frameworks for AI agents.

Topic Map

Introduction to Medable and Agent Studio

• Explanation: Panelists introduce themselves and Medable's mission to bring effective therapies to patients faster by streamlining clinical trials.
• Key claims:
  • It currently takes over 10 years to bring a drug to market.
  • Medable aims to reduce clinical trial timelines down to one year.
  • Agent Studio is a platform allowing both internal teams and customers to configure and build agentic products.
• Examples:
  • Moving from 10-year clinical trials to ambitious 1-year goals.
• Terminology:
  • Agent Studio
  • Agentic-powered solutions
  • ETMF (Electronic Trial Master File)
  • CRA (Clinical Research Associate)
• Why it matters: Setting the scope on how agentic AI addresses deep operational bottlenecks in the multi-billion-dollar clinical trial industry.

Products Built on the Agent Platform

• Explanation: Detailed overview of the specific products built on top of Agent Studio, including ETMF, CRA assistants, and protocol adherence tools.
• Key claims:
  • ETMF handles electronic trial master file documents, allowing FDA auditors to quickly find and review trial records.
  • ETMF processes over 80,000 documents a year per customer, requiring hours of metadata classification.
  • CRA agents monitor clinical research site data across 13+ different legacy systems.
• Examples:
  • Uploading documents to ETMF and automatically assigning metadata classifications.
  • CRAs navigating 13 different data source systems to monitor patient safety risks.
• Terminology:
  • ETMF
  • CRA
  • EDC (Electronic Data Capture)
  • Ontology
• Why it matters: Illustrates real-world applications of agentic workflows solving data fragmentation and regulatory compliance.

Platform Architecture and Deployment Patterns

• Explanation: Explaining the technical architecture of Agent Studio, supporting agnostic models, rag knowledge, MCP connectors, workflow orchestration, and flexible deployment models.
• Key claims:
  • The platform allows common users and engineers to configure agents using agnostic model selection.
  • Supports multiple deployment patterns including top-level agent platforms and forward-deployed engineering support.
  • Enables webhook and non-human triggers for event-based agent execution.
• Examples:
  • Connecting agents to different data systems using MCP connectors.
  • Triggering agents via Microsoft Teams, Slack, or webhooks without human intervention.
• Terminology:
  • Agnostic models
  • Rag knowledge
  • MCP connectors
  • Workflow functionality
  • Agent skills
  • Context windows
• Why it matters: Provides technical depth on how enterprise software companies can scale agentic solutions across diverse customer stacks.

Data Layers, Ontologies, and Context Management

• Explanation: Discussing the challenges of messy data, data layers, ontologies, and managing context windows across disparate clinical trial systems.
• Key claims:
  • Data fragmentation across 13+ systems creates heavy cognitive load for humans.
  • Building an AI data layer aligns all systems into a common ontology to simplify querying and recommendations.
  • Managing context windows and pre-filtering data prevents LLM hallucinations and latency.
• Examples:
  • Mapping disparate clinical terms like 'participant' across 13 systems to a unified ontology.
  • Pre-filtering tool lists using smaller sub-models before calling heavy LLMs.
• Terminology:
  • Data layer
  • Ontology
  • Vectorize
  • Sub-model
  • Context window
  • GCP (Good Clinical Practice)
• Why it matters: Addresses the fundamental data engineering hurdles required to make enterprise agents accurate and reliable.

Human-in-the-Loop and Evaluation Frameworks

• Explanation: Exploring the role of human oversight, benchmarking, and evaluation mechanisms to ensure agent outputs remain accurate and compliant.
• Key claims:
  • Human-in-the-loop ensures trust and accuracy in high-stakes clinical trial decisions.
  • Evaluations test whether agents achieve expected intents, follow design specifications, and adhere to GXP standards.
  • Benchmarking configurations against test evidence ensures consistent agent performance.
• Examples:
  • Reviewing agent classification corrections against human inputs to calculate deltas.
  • Benchmarking different GPT and Claude configurations for accuracy.
• Terminology:
  • Evaluation harness
  • Benchmarking
  • GXP compliant
  • Delta
  • Ground truth
• Why it matters: Crucial for regulated industries where mistakes can invalidate clinical trials or compromise patient safety.

Key Points

Solving Cognitive Load Limits

• Explanation: Human workers in clinical trials face massive cognitive loads, reading 200+ page protocols and navigating 13 disparate systems.
• Evidence: Clinical trials involve thousands of documents per month per study and strict regulatory oversight.
• Practical implication: Agentic AI acts as a connective tissue, absorbing the heavy data-processing burden and providing actionable recommendations.

Platform Approach to Agent Building

• Explanation: Medable provides both a platform and an experience, enabling internal teams and customers to build custom agents on top of core capabilities.
• Evidence: A customer built a full Microsoft Teams agent ecosystem in a 70-minute session using Medable connectors.
• Practical implication: Platform flexibility accelerates product delivery cycles and empowers non-technical users to configure domain-specific agents.

The Necessity of Data Ontologies

• Explanation: Without a unified ontology, agents struggle to query fragmented data sources effectively.
• Evidence: Different clinical systems use different terminology for the same concepts (e.g., participant vs. patient).
• Practical implication: Investing in an AI data layer that maps data to a common ontology drastically improves agent accuracy.

Frameworks, Models & Processes

Agent Studio Platform Model

• How it works: Provides foundational agent primitives (models, MCP connectors, workflow, triggers) that power both custom applications and customer-built agents.
• Components:
  • Agnostic model support
  • MCP connectors for data systems
  • Workflow functionality
  • Agent skills
  • Multiple trigger types (chat, webhook, integrations)
• When to use: When scaling AI agent development across multiple enterprise workflows and custom customer use cases.

Human-in-the-Loop Evaluation Framework

• How it works: Combines automated agent outputs with human review loops to measure deltas, evaluate correctness, and continuously improve agent accuracy.
• Components:
  • Ground truth comparison
  • Automated benchmarking against configurations
  • Human review and correction logging
  • Iterative prompt and model tuning
• When to use: In highly regulated environments like clinical trials where error rates must be strictly monitored and minimized.

Examples & Case Studies

A customer with business-side background onboarded with Medable for 40 minutes, stepped away for a month, and returned having built their own ecosystem of Microsoft Teams agents using Medable connectors.

• Illustrates: The usability and power of a well-designed agent platform.
• Lesson: When foundational primitives and connectors are well-built, non-technical domain experts can rapidly build complex agent workflows.

Processing over 2,000 to 80,000 clinical documents through ETMF agents to assign metadata classifications.

• Illustrates: Automating heavy manual administrative overhead in clinical trials.
• Lesson: Agents drastically reduce the time auditors spend searching for trial records.

Actionable Takeaways

• Immediate:
  • Recognize the distinction between raw LLM prompting and structured agent platforms.
  • Map out data fragmentation before attempting multi-system agent workflows.
  • Implement evaluation harnesses early to track agent accuracy.
• Strategic:
  • Adopt a platform approach to agent development to empower both internal teams and customers.
  • Focus AI adoption on reducing administrative operational burdens in high-friction industries.
  • Treat human-in-the-loop workflows as a continuous improvement loop rather than a static check.
• Questions to investigate:
  • How do you effectively manage context window limitations when agents span 13+ external systems?
  • What is the optimal balance between autonomous agent execution and human oversight in GXP-compliant environments?
  • How can ontologies be dynamically updated as new data structures emerge?

Claims Worth Verifying

• It currently takes over 10 years to bring a drug to market. (Statistical / Industry norm)
• Single clinical studies can produce tens of thousands of documents per month. (Operational statistic)

Notable Quotes

"Welcome to Just Now Possible with Teresa Torres." (at 0:04) "We're a SaaS platform company from the very beginning. We've always solved our solutions from a platform approach." (at 27:30)

Compressed Summary

• Medable's Agent Studio enables internal teams and customers to build agentic clinical trial workflows.
• Clinical trials suffer from heavy administrative burdens and fragmented data across 13+ disparate systems.
• A unified AI data layer and ontology are critical for accurate multi-system agent querying.
• Human-in-the-loop evaluation frameworks ensure GXP compliance and build trust in regulated environments.
• Keywords: agentic ai, clinical trials, agent platform, ontology, evaluation framework
• Core insight: Building successful enterprise AI agents requires a robust platform architecture, unified data ontologies, and rigorous human-in-the-loop evaluation frameworks.