Classification done right: intent over form, and a signal every app can use

Every insurance operation has a mailroom problem. Submissions, claims, legal filings, broker threads, loss runs, cyber reports, medical records, and spreadsheets pour in through email, portals, and scans — heterogeneous, mislabeled, and often bundled. The very first decision, before any value can be extracted, is what is this document, and how urgent is it? Get that wrong and a served summons sits in an FNOL queue for a week; a Department of Insurance inquiry gets triaged as routine correspondence; a three-year loss run gets read as a single active claim.

InsightXtract treats classification as a first-class, agentic step — not a filename regex or a layout template match. It reads the messy reality (OCR text, scanned PDFs, email bodies, legal filings, claim forms, spreadsheets) and assigns exactly one business class that captures the document’s intent: an obligation, a notification, or a data artifact. And it returns that verdict with the evidence that drove it — so every downstream system can trust it, and act on it.

The core principle: classify by meaning, not by layout

The design rests on a simple, hard-won rule: a document is what it does, not what it looks like. A lawsuit pasted into an ACORD form or forwarded inside a broker email is still a lawsuit. So InsightXtract never decides on template alone. Four ideas make that work in practice:

• Intent over form. Classify what the document does in the claims/underwriting lifecycle, not its shape.
• Obligation beats description. A document that creates a legal or regulatory obligation (a filing, a regulator order, an attorney demand) outranks one that merely reports an event (an FNOL, an incident, an investigation).
• Noise is stripped before meaning is judged. Forwarding headers, signatures, disclaimers, and confidentiality footers are discarded — unless that text is the operative content.
• Single-label, severity-ranked. Documents carry many signals at once. The system always returns one class, chosen by a deterministic severity ladder — never by model whim.

What comes back: an evidence-bearing verdict

Classification doesn’t hand back a bare label. Every document gets a structured, auditable verdict — the payload that downstream systems consume:

Field	What it carries	Why it’s there
class / code	The one business class, e.g. legal_proceeding	The routing key every other app reads.
priority	Severity band P1–P7	Drives queue placement and SLA/urgency.
confidence	Float 0.0–1.0	Below threshold → auto-flag for human review.
reason	One or two sentences, grounded in the document	Human-readable justification for reviewers.
evidence	1–3 quoted spans that drove the choice	Citations — makes the decision defensible.
runner_up	The second-best class + its confidence	Shows what it almost was — review context.
applied_rules	Which override rules fired (e.g. R2)	Explains why it beat a plausible alternative.
ignored_noise	What was discarded before judging	Transparency on the noise-stripping step.

The taxonomy: 23 classes across 7 severity bands

Classes are ranked P1 (highest — an active legal obligation) to P7 (lowest — operational and catch-all). When several classes are defensible, the resolver picks the highest band, then breaks ties with the override rules.

Band	Class	Code	What it means
P1	Legal Proceeding	legal_proceeding	A filed or served legal action against the insured/carrier.
P2	Regulatory Action	regulatory_action	A notice, inquiry, or order from a government/regulatory body.
P2	Attorney Demand	attorney_demand	Counsel asserting representation, settlement, or a legal threat.
P3	Claim Assertion	claim_assertion	A formal claim asserting liability/damages against the insured.
P4	First Notice of Loss	first_notice_of_loss	First party reporting that a loss or event has occurred.
P4	Circumstance Notice	circumstance_notice	Precautionary notice of a potential future claim.
P4	Claim Acknowledgement	claim_acknowledgement	Carrier-side acknowledgement / adjuster assignment / receipt.
P4	ACORD Form	acord_form	A standard ACORD form with no stronger embedded meaning.
P5	Cyber Incident	cyber_incident	Notification of a cyber/privacy event reported for coverage.
P5	Forensic Investigation	forensic_investigation	DFIR / root-cause / engineering / inspection findings.
P5	Authority Report	authority_report	Police report, FIR, fire-marshal, regulator inspection.
P6	Coverage Position	coverage_position	Reservation of rights, coverage opinion/denial, claim handling.
P6	Policy Contract	policy_contract	Policy wording, declarations, endorsements, schedules of forms.
P6	Loss Run	loss_run	Historical claims/loss-experience report (data table).
P6	Statement of Values	statement_of_values	Property/asset schedule with values (SOV).
P6	Exposure Schedule	exposure_schedule	Rating-basis data: payroll, sales, vehicle/driver/fleet, locations.
P6	Underwriting Submission	underwriting_submission	Application / submission / quote-request package.
P6	Financial Record	financial_record	Invoices, loss calculations, payment/settlement, proof of loss.
P6	Medical Record	medical_record	Medical reports, injury records, treatment summaries, IME.
P7	Supporting Evidence	supporting_evidence	Exhibit-only attachments: screenshots, logs, photos, exhibits.
P7	Broker/Insured Comm.	broker_insured_comm	Operational correspondence with no operative content.
P7	Other	other	Identifiable document that matches no class.
P7	Unknown	unknown	Insufficient, empty, or corrupted content.

Note that data artifacts are first-class: loss runs, SOVs, exposure schedules, and submissions are explicit underwriting-data classes — never dumped into “other.”

How the verdict is reached: normalize → route → vote → resolve

Classification runs as a small agentic subgraph. Cheap steps do recall; an LLM does judgment on a shortlist; and a deterministic resolver — no LLM — makes the final call so the outcome is repeatable and explainable.

• Normalize. Strip email chains, headers, signatures, disclaimers, and footers — unless that text is the operative content — and segment bundled attachments so each is judged on its own.
• Route. A cheap, high-recall pass (keyword matcher + embeddings) narrows 23 classes down to a shortlist of candidate families, so the expensive step stays small.
• Skill voting. Each shortlisted skill — a focused detector for one family — runs in parallel and returns {applies, best_class, confidence, evidence, disqualified_by}. Adding a class means adding a skill entry, not rewriting a giant prompt.
• Resolve. A deterministic resolver takes the candidates above a confidence floor, picks the highest severity band, and applies the override rules. Crucially, the rules can only escalate severity — never silently downgrade — because escalation is the safe direction for an insurer’s obligations.

A few worked examples

Document (after noise-strip)	Almost…	Winner	Why
ACORD 25 PDF containing a served Summons	acord_form	P1 legal_proceeding	R1 + R2 — a filed action beats the form.
AG letter about a data breach affecting residents	cyber_incident	P2 regulatory_action	R4 — regulator beats incident.
Insured email: “ransomware hit us last night”	first_notice_of_loss	P5 cyber_incident	R5 — no third-party demand; more specific.
3-year spreadsheet of claims with paid/reserved	claim_assertion	P6 loss_run	R10 — historical data, not an active claim.

Why this is the most reusable signal in your stack

Here is the part that matters beyond the extraction engine. Classification isn’t a private, internal step — it produces a clean, coded, cited verdict that any system can consume. One classification call becomes the routing brain for the whole operation:

Consuming system	What it does with the verdict
Claims triage / workflow	The priority band sets the queue and SLA. legal_proceeding and attorney_demand jump straight to litigation/supervisor with a suit-response clock; an authority_report attaches to the right claim.
Compliance / regulatory	A regulatory_action starts the statutory response and breach-notification clocks automatically — the class is the trigger that a deadline exists.
Underwriting workbench	Data classes route to the right place: underwriting_submission, loss_run, statement_of_values, and exposure_schedule each open the matching extraction agent and populate the submission record.
Extraction pipeline	The winning class’s YAML spec is the extraction plan — classification tells the agent exactly which fields and tables to pull, so extraction is targeted, not generic.
Document management / DMS	Files itself under the right record and folder, coded consistently — no manual indexing, and searchable by business class.
RPA / straight-through bots	Bots switch on the code to pick the next action — acknowledge, assign, decline, or escalate — without bespoke per-document logic.
Policy administration	policy_contract and coverage_position route to policy/coverage systems instead of the claims queue.
Analytics / portfolio	Consistent classes make the inbound mix measurable — severity trend, litigation frequency, submission volume — across the whole book, over time.

How other apps actually get the signal

The plumbing is deliberately simple, so integration is a day, not a quarter:

• A shared vocabulary. Each class is a real DocumentClass record. Your systems key off stable codes, not free text — so “lawsuit” means the same thing everywhere.
• The verdict travels with the document. The winning class is written to the document record and the full JSON verdict (evidence + applied rules) is persisted alongside it — available to any consumer, and defensible in an audit.
• API & events. Classify on ingest and read the verdict over the API, or subscribe to the classification event so a PAS, claims system, mailroom, or bot reacts the moment a document lands.
• Extensible without redeploys. Add a class in the taxonomy config and seed a DocumentClass row — the prompts render from the taxonomy, so new classes flow through to every consumer without a code change.

Why it matters to the business

• The right document reaches the right desk in seconds. Severity-ranked routing means a served summons never waits behind routine mail — the most costly misses (P1–P3) are exactly the ones the ladder protects.
• Deadlines start themselves. When a regulator notice or a legal filing is recognized on arrival, the compliance and suit-response clocks begin automatically instead of when someone happens to open the file.
• Every decision is defensible. The evidence spans and applied rules make each single-label choice reviewable by claims and legal — and auditable for exams and disputes.
• One signal, many systems. Build the classification capability once; claims, underwriting, compliance, DMS, analytics, and your bots all consume the same verdict. That is leverage no per-app rule engine gives you.