From FNOL packet to a set reserve: a claims intake walkthrough

A claim shows up as a pile, not a form: a broker email, an ACORD loss notice, a police report, a repair estimate — four documents in three formats, each holding a different slice of the same event. Someone has to read all of it, key it into the claims system, code the line and the cause, add up the numbers, and set a reserve. InsightXtract does that read in one pass. Below is an actual first-notice-of-loss packet — a commercial fleet collision for Ridgeline Logistics LLC — run end to end into one governed claim record.

The packet

Four files, four shapes — the everyday reality of an FNOL.

• broker_claim_email.pdf — the producer’s notice: insured, policy number, what happened, the other party and carrier.
• acord2_auto_loss_notice.pdf — the ACORD 2 automobile loss notice: claim header plus a schedule of the vehicles involved.
• police_report.pdf — the Austin PD crash report: narrative, at-fault determination, citation status.
• repair_estimate.pdf — the body shop estimate for the insured tractor, line by line.

Step 1 — Classification

Each file is routed to a claims document type. No filename guessing — the classifier reads the content and assigns a role in the claim record.

Those four roles are exactly what the unified claim record knows how to merge — the header from the notices, the narrative and fault from the police report, the numbers from the estimate.

Step 2 — Extraction, with citations

Every field carries the document and page it came from, so a file reviewer can verify without re-opening the packet.

Step 3 — Coding the loss

Free-text narrative is normalized against claims glossaries so the record joins to the policy, routes to the right desk, and rolls up in loss analytics. The raw text stays; the coded value is added alongside.

Field	As written	Coded
cause_of_loss	“rear-ended a car that stopped abruptly”	Collision
line_of_business	“auto liability claim”	Auto Liability
loss_state	“Austin, TX”	TX

Step 4 — The schedules

The ACORD 2 vehicle schedule and the shop estimate come through as typed tables — every row, not a summary.

Vehicles involved ACORD 2 p.1

Owner	Vehicle	Driver	Drivable	Est. damage
Insured	2023 Freightliner Cascadia	James Whitfield	No	$18,400
Third party	2021 Honda Accord	Maria Alvarez	No	$9,250

Repair estimate — insured tractor repair estimate p.1

Line item	Amount
Front bumper assembly R&R	$6,200
Hood replacement	$3,100
Radiator + condenser	$2,400
Body labor (32 hrs)	$5,900
Refinish / paint	$800
Total estimate	$18,400

Step 5 — Deterministic roll-ups

The reserve inputs are computed by rules over the extracted rows — never by asking a model to add. Zero tokens, exact, and reproducible.

The claim record

Everything lands in one structured record — a claim header, coded line and cause, the triage flags, the schedules, and the derived reserve inputs — ready to post to the claims system and set an initial reserve.

Why it sets a reserve — and holds up

• Straight-through triage — a coded line, cause and location plus a clean estimate is enough to auto-set an initial reserve and route to the auto desk; the flags (no injuries, no litigation) keep it out of the escalation queue.
• Present-role consolidation — the record merges only the documents actually in the packet, so a missing police report never leaves a blank masquerading as a fact, and a stray field from an unrelated form never clobbers a real one.
• No fabricated zeros — if a schedule or figure isn’t present, the derived field is skipped, not set to 0. A reserve is never understated by a silent default.
• Versioned & cited — the document types, glossaries and roll-up rules are pinned to a published version, so a record produced today reproduces tomorrow, and every value traces to a document and page.