FNOL severity
Claims
First-notice photographs and text contradict the adjuster's opening reserve, and the gap is only found at closure.
- Cycle time
- Reserve accuracy
- Reopen rate
Industries
Insurance
Accelerate claims. Detect fraud. Improve underwriting. Reserve with confidence — on numbers that survive an actuary, an auditor and a regulator in the same week.
The problem
The book is decided one case at a time, and nobody can see it whole
An insurer makes the same handful of decisions hundreds of times a day. What reserve to open. Whether this claim looks like the last one that turned out to be organised. Whether the risk just bound sits inside appetite. Whether the loss that came in last month should have been ceded.
Each of those decisions is made by somebody with part of the evidence. The photographs are in the adjuster's file, the exposure is in the cat model, the fee schedule is in a system the claims handler does not open, and the treaty is a PDF. The decision is not wrong because the people are wrong — it is wrong because assembling the evidence takes longer than the decision is allowed to take.
That is the shape of problem a model is genuinely good at, and it is why the answer has to be explainable rather than merely accurate. A severity score with no contributing factors does not survive a file review, and a fraud flag with no evidence trail does not survive a complaint.
Use cases
Where models earn their place
Each of these is a catalogue entry with a trained reference implementation behind it — and each is one you can watch run in the console at /substrate.
Organised claim rings
Fraud
Claims that share a repair shop, a clinic and a phone number across postcodes — invisible one file at a time, obvious as a graph.
- Fraud loss
- Referral precision
- Investigator hours
Missed subrogation
Recovery
Third-party liability sits in the loss report in plain language and is never pursued, because nobody read the report as data.
- Recovery rate
- Leakage
- Time to referral
Underwriting drift
Portfolio
Bound risk quietly exceeds the appetite the cat model was calibrated against, one acceptable exception at a time.
- Exposure vs appetite
- Rate adequacy
- Consistency
Medical bill leakage
Workers' compensation
Billed codes exceed the fee schedule across hundreds of lines, at a volume no human review will ever reach.
- Paid vs schedule
- Leakage recovered
- Review coverage
Catastrophe response
Event
A storm track crosses thousands of in-force policies, and the question is which ones to contact first, today.
- Time to first contact
- Reserve accuracy
- Adjuster allocation
Connected to
The systems your book already lives in
Read at query time through governed connectors. None of it is copied into a vendor database — the platform holds metadata, your data stays in your systems.
Policy administration
Claims & FNOL
Adjuster documents
Telematics & IoT
Bureau & rating data
Finance & general ledger
Medical & clinical records
Reinsurance treaties
Governance
The ground you have to stand on
Use cases arrive with the frameworks they were designed against already mapped, which is what shortens a model-validation cycle rather than lengthening it:
- IFRS 17 and Solvency II for reserving, capital and the numbers that carry an actuarial signature
- NAIC model bulletins on the use of AI systems by insurers
- EU AI Act, NIST AI RMF and ISO 42001 for the governance of the model estate itself
- GDPR and PDPA for the personal and medical data these models necessarily read
Fairness controls — approval-rate and error-rate parity — are standard on scoring models rather than a roadmap item, and every scoring model carries reason codes natively. That matters anywhere a declined claim or a rated-up renewal has to be explained to the person on the other end of it.
Worked designs
How we would build it
Reference deployments: a problem shape, the design we would propose for it, and what we would expect to be measured afterwards.
Talk to us
Start with one decision
Bring us one claims or underwriting decision you make hundreds of times a month, and we will tell you what the evidence behind it actually looks like.