BramForgeLabs / Products / Insurance AI Radar / Preview

Free sample · Use Case 1.4 · Domain 01

APS Retrieval & Clinical Summarization

The canonical insurance worked example. Same orchestration as the IT playbook’s knowledge-base pattern, applied to a life-underwriting file. The agent does not issue, decline, or rate. It produces a cited clinical packet a human underwriter can use.

Radar: Pilot · ★ High Value · 🏦 Regulation-heavy · POC: Medium · Pattern: Retrieve → Synthesize → Verify → Escalate

1. Problem & business context

Attending Physician Statements arrive as PDFs, faxes, and portal dumps. An underwriter spends 20–45 minutes per APS extracting diagnoses, dates, medications, unresolved investigations, and contradictions against the application. Backlogs sit in evidence-procurement, not in judgment. The waste is reading, not deciding.

A summarization agent that cites page and source, flags contradictions, and refuses to invent labs is a legitimate production candidate. A summarization agent that writes a rating class is not — that is Use Case 1.5, and it stays human-gated.

2. Success metrics

MetricPOC targetWhy it matters
Time-to-first-clinical-packet< 5 minThe operational win
Citation coverage100% of findings have page/sourceWithout this, it is theatre
Hallucinated finding rate0 on the gold setAn invented finding is a model-risk incident
Underwriter accept / minor-edit≥ 70%If they rewrite it, you built a chatbot
Escalation rate15–30%Too low means the verifier is asleep

3–5. Architecture, roles, orchestration

Four agents, one orchestrator, no write-back to the policy admin system in v1.

APS PDF → Retrieval (page-map, source hashes) → Synthesis (JSON packet + narrative) → Verification (grounding, purpose-limit, prohibited fields) → high confidence: UW workbench (read-only) / else: Escalation (reason codes + immutable trace)

AgentResponsibilitySuccess criteria
RetrievalFind the passages that matterGold-set recall ≥ 0.9 on diagnosis/med/date entities
SynthesisPacket from retrieved context onlyZero claims without a span; schema-valid
VerificationGrounding, safety, purpose-limitation100% of findings re-found in chunks
EscalationHuman UW queue + traceReconstructable six months later

Prohibited fields on this packet: rating_class, table_rating, decline_reason, issue_decision. If ops asks “can we just auto-issue clean files?”, that is a different system with a different E-23 inventory line — not a new column here.

If the packet would be the sole basis of an automated disposition, Verification must escalate under Law 25 s.12.1. This use case is designed so that never happens: the underwriter still decides.

6. Data & purpose

In: APS with page map, e-App coded impairments, UW-manual excerpts as a retrieval set. Optional labs/Rx/MIB as separately cited channels. Out of the context window: neighbour cases, marketing features, unrelated advisor notes. Purpose: new-business underwriting only. Using the same APS later for persistency is a purpose change — blocked.

8. Failure modes (the ones that kill the project)

9. POC that counts (2 weeks)

One product, one language, PDF APS only, 30-file gold set, read-only packet view. Demo: side-by-side original page highlight, packet finding, verifier pass. If a stakeholder cannot see the span, the POC has failed even if the prose is fluent.

11. How to build — the short version

  1. Define ClinicalPacket schema first. The narrative is a rendering of the JSON.
  2. Page-aware retrieval (OCR → pages → chunks). Heading-aware split beats naive 512-token windows.
  3. Synthesis may only emit schema fields. Empty span → drop the finding.
  4. Verifier is a different prompt and, where cost allows, a different model. Same-model self-check is not verification.
  5. Immutable trace: request id, model ids, chunk hashes, verdicts.
  6. Stop. Do not add rating.

Concrete path: 42-year-old applicant, e-App discloses elevated cholesterol. APS page 3: lipid panel. APS page 9: “follow-up with oncology 2024, records requested.” Synthesis emits hyperlipidemia (page 3) and unresolved oncology (page 9) plus contradiction vs e-App. Escalation routes to UW with reason contradiction — not to a decline engine. That single path is the product. Fluency is not.

This is one of 39 cells. The Radar’s job is to tell you which 13 should not be built as autonomous agents yet — and to hand you this spec for the ones that should.

Back to Insurance AI Radar →  ·  How specs are written →