Contract Intelligence System: AI-Powered Contract Q&A for Healthcare

The Challenge

BEFORE STATE

Dozens of Contracts. Critical Terms Buried. No Way to See Across the Portfolio.

A typical mid-size healthcare organization maintains 10 to 50 active contracts at any time, payer-provider agreements, TPA arrangements, EDI clearinghouse deals, Medicare Advantage plans, and more. Each is a unique, highly technical document, often 20 to 80 pages, covering reimbursement rates, prior authorization rules, stop-loss thresholds, renewal terms, and compliance obligations. The challenge was never storing these documents. It was knowing what was inside them.

Answering a simple question like "which contracts require prior authorization for inpatient admissions?" meant manually searching through stacks of PDFs, hours of work, prone to error. Prior auth rules sat buried in exhibits, so claims were denied and revenue leaked from missed provisions. Stop-loss thresholds were scattered across TPA contracts, creating unexpected financial exposure at year-end. Expiry dates were tracked in spreadsheets, so contracts auto-renewed unintentionally. With no cross-contract visibility, the organization missed negotiation leverage and benchmark gaps, and every compliance audit meant re-reading PDFs for days rather than hours.

The Approach

PHASE 1

Discovery and Portfolio Mapping

Axiant mapped the active contract portfolio: twelve healthcare contracts spanning eight contract types, covering payer-provider relationships, TPA administration, EDI clearinghouse agreements, Medicare and Medicaid programs, and value-based arrangements. Discovery confirmed that the core problem was not document storage but retrieval, the organization had the contracts but could not reliably surface what was inside them.

Each contract carried provisions that directly affected revenue, compliance, and patient care, yet those provisions were locked inside lengthy PDFs with no consistent structure across documents. Similar questions were answered differently depending on who read which contract, and no single system understood the portfolio end to end.

PHASE 2

Standardization and Design

Before any system was built, Axiant defined the structured intelligence layer. Every contract would be auto-tagged at ingestion with boolean flags, stop-loss, volume discounts, auto-renews, prior authorization, and network type, enabling instant filtering at the metadata level without a semantic search call.

The design separated two responsibilities deliberately: a structured metadata store for precision on dates, flags, and parties, and a semantic search engine for understanding the dense, technical language inside each document. Every query would benefit from both, structured-data precision and semantic document comprehension, so that exact figures and nuanced provisions surfaced together.

PHASE 3

Build and Configuration

The solution was built around three layers, an AI orchestration layer, a semantic search engine, and a structured metadata store, coordinated by a contract AI agent. Four specialized retrieval tools handle distinct query shapes: semantic search across documents, lookup by contract name, combined filter-and-search, and full portfolio listing. Session memory retains conversation context so multi-turn analysis holds across a line of questioning.

The critical architectural innovation is the per-document search loop. Rather than running a single global search, the system fetches the matching contract list from the metadata store, fires one targeted search per document, and merges the deduplicated results. Every contract is evaluated individually, guaranteeing that contracts with brief but relevant mentions surface alongside those with dense, prominent sections. A scheduled daily job drives expiry alerts at configurable 90, 60, and 30-day thresholds.

PHASE 4

Validation and Testing

The system was validated against a 30-question test suite across four difficulty tiers: single-document retrieval, cross-contract comparison, cross-document synthesis, and expert-level edge cases, including deliberate hallucination traps designed to provoke confident but false answers. The system achieved a 100% pass rate. A known earlier-version miss, prior authorization for inpatient admissions returning only some of the qualifying contracts, was resolved by the per-document search loop, which correctly surfaced all four.

PHASE 5

Optimization and Roadmap

With validation complete, optimization focuses on extending coverage and routing efficiency: structured extraction of procedure codes with authorization flags at ingestion, routing flag-based queries directly to the metadata store to bypass semantic search entirely, side-by-side provision comparison between named contracts, and amendment tracking on re-ingest. Notification delivery for the expiry workflow is schema-ready and pending configuration.

The Results

AFTER STATE

Hours of PDF Review Become Seconds. Every Answer Cited to Source.

The contract portfolio became a conversational knowledge base. Questions that previously took hours of manual PDF review now return in under three seconds, cited to the source document. Prior authorization compliance is complete because every contract is evaluated, not just the ones a given staffer happens to read. Expiry tracking moved from spreadsheets to automated alerts, and new contracts are auto-classified and searchable in under sixty seconds. Across a 30-question validation suite spanning single-document retrieval, cross-contract comparison, synthesis, and hallucination traps, the system passed every case.

What Is Next

The next phase extends the system from retrieval into structured contract intelligence: extracting CPT and DRG codes with authorization flags at ingestion, routing flag-based queries directly to the metadata store, side-by-side provision diffs between two named contracts, and version history with change detection on re-ingest. Email and chat notification delivery for the expiry workflow is schema-ready and will be documented as it is configured.