AI Loss Run Analysis from Loss Runs to Decision Ready Underwriting

Nomad Data
November 12, 2026
At Nomad Data we help you automate document heavy processes in your business and find the right data to address any business problem. Learn how you can unlock insights by querying thousands of documents and uncover the exact internal or external data you need in minutes.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

A clean loss summary can still leave an underwriter with important unanswered questions.

Are the figures current? Is a policy year missing? Does one large claim explain the total, or are smaller losses accumulating at one location? If losses have increased, has the business grown even faster?

Extracting the numbers correctly is essential. But those numbers still need to be organized, compared, verified, and understood in the context of the account.

For a small business with a handful of claims, that work may be straightforward. For a large insured, loss history can span hundreds of pages across carriers, policy periods, coverages, locations, and reporting formats. Accurate extraction leaves a substantial analytical task: reconciling versions, calculating statistics, identifying outliers, examining changes over time, and connecting the findings to the rest of the submission.

That is where AI loss run analysis and underwriting automation can deliver more value. The objective is not to replace underwriting judgment. It is to give underwriters a clear, verifiable view of the loss history, with material patterns, data limitations, and unresolved questions already surfaced.

When the preparation is done well, the underwriter can spend less time assembling the first usable view of the account and more time assessing what the evidence means for risk selection, pricing, terms, and follow-up.

What Makes Loss Run Analysis Decision Ready

A decision-ready loss history helps an underwriter answer three questions: What does the evidence show? Where did the information come from? What still needs review?

It should organize claims into useful views, preserve the context behind the figures, and make material gaps visible. An underwriter reviewing a total should be able to identify the claims included in it and verify the underlying amounts without rebuilding the analysis.

Separate Source Facts Calculations and Interpretation

A useful output distinguishes source facts from derived calculations and analytical observations. A carrier-reported incurred amount is a source fact. An average calculated across selected claims is a derived statistic. A statement explaining why a pattern deserves attention is interpretation. Each has a role, and each needs support.

That distinction prevents a polished narrative from obscuring how the conclusion was reached. The underwriter should be able to see which figures came directly from the loss run, which assumptions shaped a calculation, and which observations require professional judgment.

A decision-ready output commonly includes:

• A loss summary by policy period, coverage, location, or other relevant segment.

• Claim counts, paid amounts, reserves, total incurred, average cost, and maximum loss where the records support those calculations.

• Large-loss and recurring-loss views that identify the claims driving the account experience.

• Citations or source links that allow material figures and observations to be verified quickly.

• A focused list of missing information, conflicts, stale valuations, and questions for follow-up.

Make Data Gaps and Uncertainty Visible

Decision-ready information does not hide uncertainty. If a policy year is missing, a valuation date is unclear, a location cannot be assigned reliably, or a financial field is blank, the output should say so. Those limitations affect what comparisons the underwriter can reasonably make.

A blank reserve field is not the same as a reported reserve of zero. A missing policy year is not the same as a documented year with no losses. A claim that appears in two reports may be a duplicate, a later valuation, or a related but separate matter. Automation should preserve these distinctions instead of forcing every record into a falsely complete structure.

What to Automate in Loss Run Analysis

Loss runs often convey similar information in very different ways. Carrier layouts vary. Column names change. Claim descriptions may be abbreviated. Tables may continue across pages, and scanned documents may make it difficult to associate an amount with the correct claim.

Extract and Standardize Core Claim Data

According to Brad Schneider, CEO of Nomad Data:

“Extracting claim data correctly is step one.”

That first step deserves serious attention in any evaluation. An error in digitization can carry through every calculation and summary that follows. A transposed amount, misread claim status, or value assigned to the wrong row can change totals and create a misleading pattern.

Once information is accurately captured, automation can organize it into a consistent structure. Common fields include claim identifiers, dates of loss and report, policy periods, coverage, locations, cause or nature of loss, claimant descriptions, status, paid amounts, reserves, recoveries, and total incurred where available.

The goal is not to erase the source language. It is to create a usable common view while retaining the original labels and records needed for verification. This is especially important when carriers or TPAs use different definitions, expense categories, or status codes.

This challenge is part of the broader insurance document automation problem: documents arrive in inconsistent formats, and useful automation must handle the variation found in real underwriting work.

Detect Duplicates Versions and Missing Periods

The same claim may appear in overlapping reports or successive valuations. Treating every appearance as a separate loss can inflate both counts and totals. Automation can identify potential matches using claim numbers, dates, parties, descriptions, and financial values, but the possible duplicate should remain reviewable.

Version awareness matters as well. A later report may reflect a reserve increase, a payment, a reopened claim, or a closure. The workflow should identify the newest usable valuation while retaining earlier versions when development is relevant to the analysis.

Completeness checks should look beyond blank cells. They should test whether expected policy periods are present, whether the reports cover comparable time spans, whether valuation dates are visible, and whether the submission includes the exposure information needed to interpret frequency and severity.

Make Loss Metrics Easy to Verify

Useful loss analytics begin with clear definitions. Paid, reserved, and incurred amounts serve different purposes. Before aggregating them, the workflow needs to preserve how the source labels those figures and make any differences in treatment visible.

Valuation dates matter just as much. Two loss runs may cover adjacent policy years but reflect different reporting dates. One period may have had considerably more time for claims to develop. An apparent improvement or deterioration may therefore require qualification before it supports an underwriting conclusion.

A decision-ready summary should show which reports were used, their valuation dates, and how displayed totals were calculated. If an aggregate excludes claims with missing amounts, the exclusion should be explicit. If a reported total differs from the sum of extracted claim rows, the discrepancy should be flagged.

The practical test is simple: can the underwriter follow a material figure back through the calculation to the supporting record? Source-linked analysis across ACORD forms, loss runs, and underwriting submissions makes verification part of the workflow rather than a separate reconstruction exercise.

Use Loss Run Statistics to Reveal Hidden Patterns

Account totals matter, but they rarely tell the whole story. Underwriters also need to understand how losses are distributed, what is changing, and which claims or locations are driving the result.

Analyze Frequency Severity and Large Losses

Depending on the records available, useful views can include claim counts, total paid and incurred losses, average cost per claim, maximum loss, open versus closed claims, and the share of total losses attributable to the largest claims. Breaking these measures down by policy period, location, coverage, cause of loss, or claim type can reveal patterns that account-level totals obscure.

Consider an illustrative account whose total incurred losses remain flat across two comparable periods. At one site, claims of the same type increase from four to eight, while average incurred cost declines. The account total conceals that change. A location-level view brings it into focus.

The pattern still needs interpretation. Did the site become busier? Were more employees working there? Did reporting practices change? Are the claims associated with a particular operation or activity? The statistics give the underwriter a specific issue to examine. They do not establish its cause.

The same discipline applies to large losses. A single severe claim can dominate an average and obscure the experience across the remaining claims. Showing the maximum, median or average where appropriate, and the broader distribution helps the reviewer understand what is driving the numbers.

Compare Loss Trends With Exposure Growth

One of the clearest ways to misread a loss history is to evaluate growth in losses without considering growth in exposure. A company may report increasing losses over several years, but the submission may also show that the business has expanded substantially.

With suitable, comparable exposure information, that account may present a different picture from the one suggested by loss totals alone. The relevant exposure measure depends on the business and coverage. Employee count may provide context in one situation; payroll, mileage, vehicle count, sales, units, square footage, or another measure may be more appropriate in another.

The periods and scope must align. A higher claim count does not by itself establish a higher claim rate. Equally, a growing business does not automatically explain away a worsening loss pattern. Changes in operations, business mix, locations, controls, and claim severity may still matter.

This is why loss history should be examined alongside applications, supplemental questionnaires, exposure schedules, broker correspondence, statements of values, and other supporting records. Cross-document analysis can show whether the context needed to interpret a trend is already present elsewhere in the submission.

The same principle applies to underwriting submission triage: the useful output is not a stack of extracted fields, but a supported view of the account that reconciles the documents and makes gaps visible.

Turn Loss Patterns into Focused Follow Up

A useful analytical output helps the underwriter identify what to investigate next. For the location with an increase from four to eight similar claims, the output should identify the relevant periods, show the underlying claims, and indicate whether the submission contains comparable exposure information.

That supports focused questions: Did activity at the site increase? Did operations change? Do the claims describe a recurring cause? Were corrective measures introduced, and when? Are open reserves concentrated in a small number of files? Has the account changed carriers, TPAs, or reporting practices?

The full submission may already answer some of those questions. A supplemental form may describe a new production line. A broker email may explain an acquisition. Supporting records may document a change in loss controls. Bringing those sources into the same analysis can reduce unnecessary broker follow-up and give the underwriter a more complete starting point.

Present Facts Without Filling Gaps with Speculation

Where the information is absent, the output should identify the gap plainly. It should not supply a plausible story to complete the narrative.

According to Brad Schneider, CEO of Nomad Data:

“AI should present the facts. It should not make anything up. It should not speculate.”

That boundary is essential for trustworthy underwriting automation. The system can explain that claim frequency increased while payroll data is missing. It can show that several similar incidents occurred at one location. It can point to an unanswered question. It should not invent the operational change, control failure, or causal explanation behind the pattern.

External research can add context when the workflow calls for it, but every finding should be identified by source and period, and its relationship to the insured should be clear. Research should not be presented as resolving a gap when the available evidence does not support that conclusion.

How Doc Chat Supports Decision Ready Underwriting

At Nomad Data, Doc Chat supports loss run processing alongside broader submission review. It can help teams extract and normalize loss information, generate statistics and structured summaries, compare records across documents, and answer underwriting questions with citations back to the supporting material.

The value comes from connecting loss history to the account, not treating the loss run as an isolated spreadsheet. An underwriter can examine a frequency increase, ask whether exposure changed, locate the supporting application or broker explanation, and identify what remains unresolved.

Choose Conversational or Preset Analysis

Underwriting teams do not all want the same output or interaction. Some want to begin with an accurately digitized loss history and ask questions conversationally. An underwriter might request the largest losses, compare average incurred amounts across years, investigate claims associated with one location, or isolate open claims above a threshold.

Others want a repeatable analysis generated automatically for every submission. Through a preset architecture, a carrier can define the statistics, comparisons, deviations, questions, and output structure it wants produced as part of the workflow. Underwriters can then ask follow-up questions when an account warrants closer examination.

A standard analytical starting point can improve consistency, while conversational exploration allows the reviewer to pursue the particular features of the account. The carrier should define which comparisons are useful, what deserves a flag, and how the information should be presented based on its lines of business, guidelines, and review practices.

Keep Risk Interpretation and the Decision with the Underwriter

Automation can prepare and explain the evidence. The underwriter remains responsible for deciding what those findings mean for the account. That includes determining whether the history is sufficiently complete and comparable, whether operational changes or controls alter the assessment, whether an apparent exception matters under the carrier’s guidelines, and what further information is needed.

According to Brad Schneider, CEO of Nomad Data:

“This division of work gives expertise more room to operate.”

Instead of spending hours constructing the first useful view of the file, the underwriter can begin with that view and focus on the risks, inconsistencies, and business questions that require judgment.

How to Evaluate AI Loss Run Analysis

A credible pilot starts with the accuracy of the loss run data, but it should not end there. The test should reflect the formats, document lengths, valuation practices, and difficult cases the underwriting team actually receives. Then the evaluation should follow the data through to the completed work product.

Use Representative Test Files

The test set should include clean native files, scanned loss runs, multi-carrier histories, overlapping valuations, missing periods, difficult tables, open and closed claims, and submissions with relevant exposure data elsewhere in the packet. It should also include examples where the correct conclusion is that the information is incomplete or not comparable.

Reviewers should define expected outputs and material error categories before the pilot begins. A small formatting difference is not equivalent to a missing severe claim, incorrect total, unsupported calculation, or unrecognized policy-year gap. Weighting consequential errors more heavily produces a more realistic measure of underwriting usefulness.

Measure the Complete Underwriting Workflow

The evaluation should ask: Are extracted fields correct? Are calculations reproducible? Are comparisons appropriately qualified? Can reviewers verify sources quickly? Are material deviations and missing periods surfaced? Does the output match the team’s preferred format? How much correction or reformatting is required?

Operational measures should include preparation time, correction rates, verification effort, broker follow-up, time to a supported assessment, and adoption by the underwriters expected to use the output. Follow-up volume needs interpretation. Better analysis may uncover a material gap that previously went unnoticed. A necessary broker question can represent an improvement, while repeated requests caused by poor preparation represent avoidable work.

As Nomad Data’s discussion of AI evaluation emphasizes, value should be measured across the complete task. Extraction accuracy is necessary, but the stronger question is whether the workflow produces a faster, more reliable, and more defensible underwriting assessment.

For underwriting leaders, the goal is a shorter path from incoming loss runs to an assessment the underwriter can explain and support. Accurate digitization makes that possible. Statistics and contextual analysis make it useful. Traceability makes it reviewable. Professional judgment turns the evidence into a decision.

Check out Doc Chat to see how complete-file analysis, loss run normalization, source-linked answers, and carrier-defined outputs can support faster and more defensible underwriting workflows.

Learn More

FAQs

What is AI loss run analysis?
What information should be extracted from a loss run?
How does loss run automation help underwriters?
Can AI determine whether an account is a good underwriting risk?
How should carriers verify AI-generated loss run summaries?
How does Doc Chat support loss run review?