Claims Analyst · Job two

Claims laundry lists: find what matters in a 300-page list.

Brisc reads claims laundry lists, thousands of rows in any cedent layout, and tells you which claims deserve attention. It scores each claim against your own hazard vocabulary and problem jurisdictions, and shows exactly why every flag fired.

Run summary · cedent file 09 4,180 rows scanned · 61 flagged
criticalCLM-2214hazard keyword + venue match
highCLM-1878hazard keyword, exact
mediumCLM-0431synonym match, confirmed
reviewCLM-3067borderline: waiting for a human
Every row stamped. Export carries the config snapshot that scored it.
The position

Know when an insured approaches its attachment point.

The Analyst tracks each insured’s total incurred loss against their layer attachment point and bands it below, approaching, or reached.

How the position is computed

The Analyst resolves which insured a file belongs to, sums values across files with per-claim dedupe, and routes genuinely ambiguous cases to review rather than forcing them into a band.

below approaching reached
The read

How does it read a file it has never seen?

Every cedent lays the document out differently. The Analyst works out what it’s looking at before it counts anything, so nothing fake enters the queue.

Document classification

Claim register, event log, or placement matrix, identified first, so policy wording never becomes a fake claim.

Column-identity mapping

Finds claim number, loss date, claimant, narrative and venue whatever the columns are called.

Row suppression

Totals, preambles and repeated headers are recognised as report artifacts and never counted as claims.

Loss-date normalisation

Twelve date formats to ISO, occurrence ranges detected, merged-cell contamination flagged, the original always kept verbatim.

The screen

Match your vocabulary. Score every claim.

Three-tier keyword matching, exact, synonym and fuzzy, runs against your hazard list behind a two-threshold gate: confident hits auto-accept, borderline hits queue for review, weak ones are ignored.

Venue matching

Plaintiff-friendly venues from your own jurisdiction list. A specific court always counts; a bare state only counts in a venue column.

Risk banding

Keyword severity, confidence, and jurisdiction weights combine into a per-claim score, banded critical to none.

Run summary

Rows scanned, flags by band, keyword hits by category, jurisdiction hits, pending exceptions. Every row stamped.

EXC-1109 Analyst reasoning
“electricuted” accepted

Misspelling of a term on your hazard list. Narrative context supports an electrocution injury. Flagged for review with reasoning attached.

“decrease” rejected

Superficially close to “deceased” but the narrative describes a change in incurred value, not a fatality. Not flagged, and the decision is recorded.

The engine

Deterministic engine. Fail-open AI.

The screening engine is deterministic and explainable. Three AI enhancements sit on top, and all fail open: if a model is unavailable, the run degrades to deterministic behaviour and never blocks ingestion.

Exception adjudication

Borderline near-misses are accepted or rejected with written reasoning. Ambiguous cases are tagged as needing review, never guessed.

One-click enrichment

A plain-language summary of any claim: injury and loss type, parties, venue signal, and a recommended next action.

Nothing overwritten

Originals are kept verbatim. Every AI decision is labelled as such, with the reasoning stored beside the row it touched.

Your list, your tuning

Your keywords. Your jurisdictions. No deployment required.

Per-tenant configuration

Keywords, jurisdictions, thresholds and bands tuned to your book, no deployment required.

Retriage on change

Change the list and re-score an ingested run instantly, with no re-extraction.

An exportable audit pack

A multi-sheet Excel export carrying every flag, every decision, and the exact config snapshot that scored the run.

One Analyst, two jobs

Screened here. Opened next door.

Laundry-list review is job two. Job one reads first notices of loss and opens the claim record.

Common questions

Laundry-list review, answered

Can we use our own keyword list?

Yes, that’s the point. Your hazard vocabulary and your problem jurisdictions are configured per tenant, with no deployment work, and a guard prevents a run against an empty dictionary. The lists stay yours; we never publish or share them.

What happens when the AI isn’t sure?

It stops and asks. Borderline near-misses are adjudicated with written reasoning; genuinely ambiguous cases are tagged as needing review and wait for a human. Nothing ambiguous is forced into a band.

Does hidden or truncated text in a PDF get attributed to the right claim?

Yes. Extraction preserves the link between each value and the row it came from, so text a reader would miss, hidden, truncated or spilling across merged cells, still lands on the correct claim, and the original is kept verbatim.

Can we re-run after changing the list?

Yes, instantly. A configuration change retriages an already-ingested run with no re-extraction, and the export carries the exact config snapshot that scored it, so you always know which list produced which flags.

How is this different from the Bordereaux Analyst?

The laundry-list job screens claims for attention: which rows in a bulk listing deserve a handler’s eyes, and why. The Bordereaux Analyst reconciles premium and claims bordereaux as financial documents, including version comparison across runs. Different questions, different Analysts.

Bring your own laundry list. Watch it flag.

30 minutes, no slide deck. We run a real cedent file, against your vocabulary if you bring it, and show every flag with its reasoning.

Book the walkthrough