The Loss Run That Reads Itself: How AI Turns Claims History Into Underwriting Signal in Minutes

The submission says "loss runs attached." What arrives is four PDFs from three different carriers, one spreadsheet an agency clearly exported by hand, and a scanned fax from 2019 that the insured's brother-in-law "had somewhere." The underwriter opens a fifth tab, starts retyping claim dates, amounts, and reserve columns, and the quote clock, which everyone agreed would be measured in hours, quietly starts burning its entire budget on data entry.

Loss runs are the single most analytically valuable document in commercial underwriting, and also the most hated to process. They are unstructured by design: every carrier formats them differently, every agency interprets "total incurred" a little loosely, and half the time the document that answers the most important question, whether this risk has a pattern, arrives in the format least able to answer it.

There is a better way, and it is not OCR with a prayer. It is an AI document pipeline that understands what a loss run is, reads all of them the way an experienced analyst would, and hands the underwriter a normalized, scored, explained claims history before the first keystroke.

Why Loss Runs Consume Underwriting Time

• Every carrier speaks a different dialect. One file lists claim numbers first, another buries the date of loss in a paragraph, a third reports paid and incurred in alternating columns with no header you can trust. The reader, not the software, has to decide what counts.
• Multiple documents have to become one story. A five-year history usually means three carriers and a coverage-gap question: do these policies overlap, where did one end and the next begin, and who decided the gap was acceptable?
• Open claims hide. The most consequential line on a loss run is often a reserve that is still moving. Manual reviews miss them because they are reported in a different column with a different label on every form.
• The work is invisible in the metrics. Firms measure quote turn time but not the hours buried in transcription, so the bottleneck survives every "process improvement" because nobody ever saw the process.

This is the same structural problem we covered for submission intake: the AI is not doing anything clever with the data until the data exists in one place, in one shape, with its sources intact.

What the AI Actually Does With a Loss Run

The InsuranceClouds AI platform treats a loss run the way it treats every document in the pipeline: identify what it is, extract the meaning rather than the pixels, and map it onto a schema the rest of the system can reason about. Concretely, that means five steps, run in seconds:

• Recognition. The system distinguishes a loss run from an ACORD page, a policy declaration, or a photograph of a loss, so unsorted attachments in a shared inbox stop being a triage project. This builds on the extraction discipline we described in beyond OCR.
• Extraction with confidence. Claim date, incident description, carrier, claim number, status, paid, incurred, reserve, and subrogation are pulled from any layout, scanned or digital, and every field carries a confidence score. Low-confidence values surface for review instead of silently becoming truth.
• Normalization. Three carriers' column vocabularies collapse into one schema: open versus closed gets defined the same way, incurred versus paid stops being ambiguous, and overlapping policy periods stitch into a single chronological record per named insured and location.
• Pattern analysis. Only then does the underwriting value appear. Frequency and severity trends by line of coverage, concentration by cause (slip and falls on one premises do not look scary in three separate PDFs; they look scary in one normalized timeline), open-reserve exposure, and recency weighting that separates yesterday's problem from a five-year-old outlier.
• Explanation. The output is not a black-box score. Every observation cites the claim that supports it: this risk scores higher on premises liability because of three fall-related claims in 24 months, here they are, here are the source pages.

From Parsed History to Priced Decision

Because the normalized loss history lives in the same system of record as the submission itself, it does more than inform a human's opinion. It feeds the risk signals behind submission triage scoring, so a deteriorating claims trend re-ranks a submission's priority without anyone building a report. It flows into program analytics, so loss experience by class, broker, and carrier becomes a standing question the platform answers instead of a quarterly project. And it enriches the quote itself: endorsement and pricing conversations start from "your third fall claim changes the loss history we rated this against," with the evidence attached, which is a far better producer experience than a silent declination.

It also shortens the honest path to a decision. An underwriter who spends twenty minutes reading a clean, cited, normalized history instead of four hours assembling it will decline more risks correctly, and bind more risks correctly, than one working from memory and a spreadsheet.

Where the Judgment Stays With the Underwriter

The AI assembles, normalizes, flags, and explains. It does not decide appetite. A claims pattern that looks like a deterioration might be a single disputed claim with a moving reserve; a spike that looks minor might connect to a litigation trend the underwriter knows from the market. Every output is a draft for expert review, every field traces to its source document, and every decision and its rationale stay on the record, the same audit-ready discipline the platform applies to compliance and carrier review. Automation earns its trust by showing its work, which is also how it earns a seat in an underwriting file.

What Changes When the Loss Run Reads Itself

• Quote turnaround compresses, because the longest step in the workflow, transcription and assembly, shrinks to a review pass measured in minutes.
• Consistency improves, because every submission gets the same five-year, gap-checked, open-reserve-aware analysis no matter who picked it up.
• Producer conversations get evidence, because explanations cite claims instead of opinions, which reduces disputes and improves response quality on the next round.
• The file is defensible, because the loss history you rated, the analysis that supported it, and the reviewer who approved it are one linked record, ready for audit or carrier review.

Loss run analysis is one module of the InsuranceClouds AI platform, wired into the distribution platform where submissions and quotes already live, so the history informs the workflow instead of sitting in an attachment nobody opens twice. See the full document-to-decision lifecycle in our case studies.