Automated Clinical Document Generation Platform for Pharmaceutical R&D

abbvie · Healthcare · 2024 · source Agentic: yes · promotional bias: low · 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 and regulatory document generation

AbbVie needed to automate hundreds of regulated clinical and R&D document types without turning generative AI into an unchecked document author or one-off point solution.

Constraints:

• clinical study startup alone has 87 documents per study
• documents are regulated and require consistent accuracy, writing style, and formatting
• the platform must keep humans in charge and support human review
• source data must be fresh, secure, and drawn from many internal repositories and systems
• the platform must scale beyond one document type to hundreds of document types

Architecture

AbbVie built GAIA as a generative AI document-creation platform rather than a single-document automation. Users with role-based access select a document type and criteria in an intuitive UI. A document orchestrator defines the order of operations for each document, an enterprise prompt library shares prompt assets, an API hub pulls real-time master data from document repositories, literature, data lakes, management systems, LLMs, and prompt libraries, and AWS serverless infrastructure runs the platform. GAIA generates whole documents or sections, retains outputs temporarily, supports regeneration, and is evolving with agents, context engineering, vectors, graphs, and human-feedback loops.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  gaia-user-interface["gaia-user-interface"] --> document-orchestrator["document-orchestrator"]
  document-orchestrator["document-orchestrator"] --> enterprise-prompt-library["enterprise-prompt-library"]
  document-orchestrator["document-orchestrator"] --> feature-catalog["feature-catalog"]
  document-orchestrator["document-orchestrator"] --> api-hub-integration-layer["api-hub-integration-layer"]
  api-hub-integration-layer["api-hub-integration-layer"] --> source-data-systems["source-data-systems"]
  enterprise-prompt-library["enterprise-prompt-library"] --> document-generation-llms["document-generation-llms"]
  document-orchestrator["document-orchestrator"] --> document-generation-llms["document-generation-llms"]
  enterprise-prompt-library["enterprise-prompt-library"] --> agentic-catalog["agentic-catalog"]
  feature-catalog["feature-catalog"] --> developer-teams["developer-teams"]
  agentic-catalog["agentic-catalog"] --> developer-teams["developer-teams"]
  aws-serverless-runtime["aws-serverless-runtime"] --> document-orchestrator["document-orchestrator"]
  document-orchestrator["document-orchestrator"] --> temporary-document-retention["temporary-document-retention"]
  gaia-user-interface["gaia-user-interface"] --> human-reviewers["human-reviewers"]
  human-reviewers["human-reviewers"] --> feedback-learning-loop["feedback-learning-loop"]
  feedback-learning-loop["feedback-learning-loop"] --> enterprise-prompt-library["enterprise-prompt-library"]
  feedback-learning-loop["feedback-learning-loop"] --> document-generation-llms["document-generation-llms"]

Components

Data flow

A permitted user logs into GAIA, selects a document type and AI-enabled sections, and provides criteria through the prompt-based interface. The document orchestrator configures the document-specific sequence, pulls fresh data through the API hub, invokes reusable platform features and prompts, processes the content with AI, generates a first draft or section, and makes it available for download or regeneration. Business users review quality, style, and formatting, while human feedback informs future draft generation and prompt improvement.

Why this architecture

AbbVie chose a platform because automating one document would not scale to the hundreds of clinical and R&D documents required. The document orchestrator, prompt library, feature catalog, and API hub create reusable building blocks, while role-based access, temporary retention, fresh master-data pulls, and human-in-the-loop review address regulated life-sciences constraints. Context engineering and feedback loops are added because LLM-only generation plateaus around 70-80% insight.

Models used: GAIA · AWS serverless infrastructure · AWS Bedrock · LLMs · agents · vectors · graphs

Scale

• Clinical study startup alone requires 87 documents per study.
  “clinical study startup alone, we have 87 documents that are part of that for every clinical study.”transcript, stated
• GAIA automated 26 document types by the end of the calendar year.
  “by the end of this calendar year, we'll have automated 26 documents types in our platform.”transcript, stated
• The 26 automated document types are on track to save 20,000 hours per year.
  “That is on track to save us 20,000 hours per year for those 26 documents.”transcript, stated
• AbbVie targets well over 350 document types by 2030.
  “We're looking by 2030 to automate well over 350 documents”transcript, stated
• The 2030 target is over 115,000 hours saved.
  “save over 115,000 hours from our teams.”transcript, stated
• Manual effort can be reduced by up to 90% per document.
  “reduce most of the manual effort for these documents by up to 90% per document right now.”transcript, stated

Alternatives considered (roads not taken)

Automate one document type as a point solution

Rejected because: The team expected hundreds or thousands of documents, so one CSR-style solution would not scale.

Use GAIA as a document repository

Rejected because: The platform pulls master data in real time and uses temporary retention so generated drafts stay fresh and governed.

Expect LLMs alone to extract all needed insight

Rejected because: The team saw a 70-80% plateau and moved toward context engineering with vectors, graphs, and deeper context methods.

Patterns

human-in-the-loop-gate (regulated-workflow)

GAIA is explicitly human-in-the-loop, with users reviewing generated regulated documents.

tool-use-orchestration (workflow)

A document orchestrator configures each document and defines the operation order for generation.

prompt-versioning (prompt-management)

An enterprise prompt library shares prompt assets across GAIA and other organizational agents.

agent-memory (learning-loop)

The platform is being designed as self-learning from human feedback to improve draft generation and prompts.

data-quality-loop (data)

GAIA pulls master data in real time and uses freshness/retention decisions to improve document quality.

Tradeoffs

Build a platform

Chose reusable GAIA document platform over single-document automation.

Rationale: A platform supports repeated document automation over time and avoids rebuilding point solutions for each regulated document type.

Pull data in real time

Chose fresh master-data access over turning GAIA into a long-lived document repository.

Rationale: Fresh data and temporary retention support quality, regeneration, and governance.

Measure automatable percentage and human-in-loop remainder

Chose agile value reassessment over fixed upfront savings assumptions.

Rationale: Initial assumptions were not fully baked, so value realization needed reassessment after seeing what could really be automated.

Gotchas (what bit them)

A point solution would not scale to the document universe.

Root cause: Clinical and R&D operations require hundreds of document types, not only one CSR-style document.

Resolution: Build a reusable platform with orchestrator, prompt library, feature catalog, and integration layer.

LLM-only drafting hit an insight plateau.

Root cause: Clinical documents need deeper context than the LLM alone could surface.

Resolution: Add context engineering with vectors, graphs, and other methods to include deeper insights.

Automated regulated documents can drift in tone, accuracy, and formatting.

Root cause: Different document teams have specific style expectations and regulatory quality needs.

Resolution: Define accuracy, writing style, and formatting early with business participation and sprint demos.

Evaluation

AbbVie evaluates GAIA through value realization, automatable document percentage, human-in-the-loop remainder, quality of generated sections, accuracy/style/formatting definitions, hours saved, and document-type automation progress.

Metrics: 26 document types automated · 20,000 hours saved per year target for 26 documents · 350+ document types by 2030 · 115,000+ hours saved target · up to 90% manual effort reduction per document · quality of generated document sections · human-in-loop effort remaining

Outcomes

• GAIA automated 26 document types by the end of the reported calendar year
• the platform is projected to save 20,000 hours per year for those 26 documents
• AbbVie targets well over 350 document types and over 115,000 hours saved by 2030
• manual effort is reduced by up to 90% per document
• the platform supports whole-document and section-level generation with human review

The strongest scale facts are 87 startup documents per clinical study, 26 automated document types, 20,000 annual hours saved for those 26 types, a 350+ document target by 2030, and over 115,000 annual hours saved target.