Document AI — from PDF claim parsing to rule-based decisions and audit trails

9:00 a.m. in the VeridianLife claims room

Insurance claims that arrived over the weekend are waiting to be processed. Manual review of every PDF is slow, but trusting every extracted value could let a misread amount or missing diagnosis code drive an automated decision.

Today you are part of VeridianLife's claims-operations team, deciding whether automation may continue for CLM-2025-006 or the claim should be handed to a person with evidence.

Which claims can move automatically, and which should be handed to a person with evidence?

Decision criteria for this mission

• Perspectives in play — Processing automation structures repeated documents quickly, while claims review checks whether the extracted values are trustworthy and whether the handoff reason will travel with the case.
• Decision constraints — CLM-2025-006 has an amount and diagnosis code, but its confidence is 0.52, below the 0.6 automation threshold. Having a value and trusting that value are separate decisions.
• Completion signals — Separate the source document, intake metadata, and structured result; record ESCALATED with its reason; then recheck the handoff through claimant-policy relationships and the full decision distribution.

Workshop goals

By the end of the Workshop, you will be able to explain four decisions:

• Distinguish values read from the document from values known at intake.
• Decide how far a structured result can be trusted for automated handling.
• Trace the reason automation stopped through data, relationships, and the operations view.
• Explain the complete solution from the opening operations problem to the finished handoff.

From one received claim to an explained result

One claim passes through five scenes: intake, extraction, rule decision, human handoff, and operations review. After an evidence check, the last chapter returns to the opening problem and explains what the flow solved.

flowchart LR
    accTitle: Five scenes, an evidence check, and a retrospective for one insurance claim
    accDescr: Inspect the claim PDF and intake metadata, judge the structured result, route the claim to automation or human review, verify operations evidence, check the boundary, and explain the solved problem.
    receive["1. Claim intake<br/>Find the first clue"] --> extract["2. Structure the document<br/>Choose trusted values"]
    extract --> decide["3. Apply ordered rules<br/>Handle or hand off"]
    decide --> trace["4. Trace relationships<br/>Recover context"]
    trace --> operate["5. Review operations<br/>Choose the next action"]
    operate --> check["6. Evidence quiz<br/>Check the boundary"]
    check --> reflect["7. Retrospective<br/>Explain the outcome"]

Each chapter resolves one question. Follow the core practice first, and open Deep dive only when you need implementation detail or the full inventory.

Prerequisites

• An analyst or engineer account with D.Hub Portal access (Editor or higher)
• About 43 KB of download space for one ZIP
• 100 minutes of practice time

You do not need a terminal, Python, or a clone of dhub2-examples. If scenario import is new to you, complete Import and tour a complete hands-on scenario first.