Healthcare · Document Intelligence
Agentic AI that reads tables no algorithm could.
Fortune 10 healthcare enterprise
A multi-agent LLM system — an extraction agent, a standardization agent, and a self-checking feedback loop — that digitizes contract tables too complex for traditional parsing: borderless layouts, merged cells, tables spanning pages.
Healthcare · Document Intelligence
Challenge
What was the challenge?
Extracting data from tables in documents like PDFs is notoriously difficult, and healthcare contracts contain the worst offenders: tables with no visible borders, unusual and inconsistent layouts, and tables that stretch across multiple pages. Existing table-detection algorithms either failed to find them or extracted their contents in a form no downstream system could use.
The consequence was manual work — time-consuming, costly, and error-prone in a field where accuracy is non-negotiable. And even where existing tools extracted something, they did a poor job of converting it into a standard format, which left auditing systems unable to consume the output. The problem wasn't just reading tables; it was producing data other tools could trust.
Approach
How did we approach it?
Instead of building one monolithic model, we designed a division of labor between specialized LLM agents. An Extraction Agent focuses purely on pulling table data out of documents — including open-ended, borderless, and multi-page tables that defeat conventional parsers. A separate Standardization Agent then reorganizes that raw extraction into one consistent, uniform format designed for audits and compliance checks.
Between them sits a feedback loop that compares outputs against a trusted reference dataset and pushes corrections back through the agents, catching incomplete or inconsistent results before they leave the system. Reliability is engineered in, not hoped for.
We were equally deliberate about evaluation. The system is scored on average precision (is what we extracted correct?) and average recall (did we capture everything?), and it additionally uses an LLM as a judge to assess the quality of its own outputs — a self-evaluation layer that flags weak results for another pass.
Delivered
What did we deliver?
- An Extraction Agent that uses large language models to extract table data directly from complex documents — including borderless, open-ended, and multi-page tables.
- A Standardization Agent that converts extracted data into one consistent format ready for audits and compliance checks.
- A feedback loop that validates outputs against trusted reference data and iteratively corrects incomplete or inconsistent results.
- A dual evaluation framework: AP/AR metrics for measurable accuracy and completeness, plus LLM-as-judge scoring for output quality.
Mechanism
How does it work under the hood?
- 01The Extraction Agent processes the document and pulls candidate tables, handling layouts conventional algorithms miss.
- 02The Standardization Agent restructures each extraction into the canonical output format downstream audit tooling expects.
- 03The feedback loop compares results against a trusted dataset; discrepancies trigger correction passes through the agents rather than shipping flawed output.
- 04Average precision and average recall are computed to quantify accuracy and completeness, and an LLM judge scores output quality — low-scoring results are refined before release.
Outcome
What was the outcome?
The proof of concept demonstrated that agentic LLM systems can reliably digitize table layouts that traditional extraction algorithms simply cannot parse — and, critically, deliver the output in a standardized form that audit and compliance tooling can consume directly.
The architecture — specialized agents plus a validating feedback loop plus measurable evaluation — became a reusable blueprint the subsequent agentic projects were built on.
Learnings
Key learnings
Splitting extraction and standardization into separate agents made each one simpler, testable, and independently improvable — division of labor works for AI too.
A feedback loop against trusted reference data converts 'usually right' into 'reliably right' — and is the difference between a demo and a system.
Measure with AP/AR and an LLM judge from day one; agentic systems without an evaluation framework can't prove they're improving.
Related stories
Healthcare · Pharmacy Benefits
Digitizing complex pharmacy contracts with generative AI
An end-to-end GenAI pipeline that reads dense pricing contracts — tables, legal language, and billing rules — and turns them into audit-ready structured data, unlocking roughly $64M in annual savings.
Enterprise · Proposal Operations
Answering RFP questionnaires with agentic RAG
An agentic retrieval-augmented system that answers standard RFP questions instantly from a library of expert-approved responses — and intelligently routes novel questions to the right team.
Work with us