Loss runs from every prior carrier, set out by policy year before anyone prices the risk.

Today someone on the desk opens each run, works out which policy it covers, and copies the figures onto your loss summary sheet: claim counts, paid, reserved and incurred by policy year, the open claims, every loss above the large-loss line your underwriters set, and the date each run was valued. Years with no run, or a cover note that disagrees with the runs, become questions for the retail agent.

In the proof of concept, AI reads every run in the file and drafts that summary, with each figure traced to the page and line it came from, plus a list of what is missing or doesn't reconcile. It doesn't price, decline or ask anyone for anything. Your reviewer, the person whose loss summaries your underwriters trust most, puts each draft next to the one your desk actually built.

What arrives depends partly on the prior carrier's state. Illinois has an insurer send the first named insured three policy years of loss information within 30 days of a request, straight to the producer if the insured asks in writing, and each open claim shows its date, a description, payments and reserves; workers' comp sits outside that rule. Connecticut, since 2012, has every commercial insurer send a summary of claims within 30 days of a written request, reaching back up to four years and giving each policy number, coverage period, claim count, paid losses and date of loss. Louisiana puts up to three years of loss runs inside a commercial nonrenewal notice, but its rule leaves surplus lines policies out. One file can mix a statutory summary with a full claim listing, and the draft labels each.

Totals mean different things on different runs. Incurred is paid plus reserves for the period, and as IRMI defines it, it ordinarily leaves out losses incurred but not reported, so the newest policy year always looks lighter than it will end up. Some carriers fold allocated loss adjustment expense, meaning the outside defense counsel, experts and investigators billed to a claim, into paid and reserved; others give it a column of its own or leave it off. Under a professional liability policy with defense within limits, that expense also uses up the limit. The draft keeps indemnity and expense apart wherever the run allows it, and notes where it doesn't.

Putting a claim in the right year takes care. Under an occurrence policy, a claim belongs to the year the loss happened. A claims-made policy, the usual form for professional and management liability, is triggered by a claim made during the policy period, subject to any retroactive date, so the claim belongs to the year it was reported. A run that sorts claims-made losses by date of loss, or a renewal that changed the policy number partway through the history, can place a claim in the wrong year or count it twice. The draft groups each claim by the trigger of its policy and lists any claim that shows up on two runs.

Then the valuation date. A run is a snapshot, and open claims keep moving after it. Where workers' comp runs are in the file, for example behind an umbrella that sits over employers liability, the experience mod values the same claims on its own calendar. NCCI's unit reports value a policy's losses 18 months after it took effect and then every 12 months, and a policy counts toward the mod only if it began between 21 and 57 months before the rating effective date. California, New York and a few other states run their own plans. So one comp claim can carry two figures, and the draft prints each run's valuation date beside every year it summarizes.

Your reviewer will look first at what a summary can hide: a policy year with no run and no reason given, a run valued earlier than the market you're approaching will accept, a claim closed without payment that reopened on a later run, one loss split across two claim numbers, and a cover note claiming no losses where a run lists some. Every figure points back to its source page.

Once your reviewer says yes, the summaries run on new submissions for about 30 days.

An underwriter reads every summary before it is relied on. The AI never prices, quotes or declines a risk, never asks a retailer, an insured or a carrier for a run, and never writes into your management system. A licensed person on your staff decides what the history means and what goes back to the retail agent.

The same reviewer scores the live month against the criteria agreed for the test.