← BACK TO INSIGHTS
AI AUTOMATION FOR SMB OPERATIONS BY INDUSTRY9 MIN READ

Insurance Agency Loss Run Review Automation

How loss run automation works, step by step

S
Hana Rahman
NOVEMBER 16, 2026
AI AUTOMATION FOR SMB OPERATIONS BY INDUSTRY
SANSA LAB NOTES · 03

Loss run review automation turns inconsistent carrier documents into claims data an agency can use for renewals and underwriting. Its value is a controlled process that extracts, standardizes, checks, and delivers that data without making agents re-key it. The best place for an agency to start with AI is work it can check for accuracy and whose results improve the work that follows.

What a loss run report actually contains and why manual review breaks down

A loss run is an official carrier-issued record of a commercial policyholder’s claims history. It typically covers three to five years and identifies the named insured, policy number, policy period, carrier, and coverage type. Each claim adds a claim number, date and cause of loss, status, paid losses, outstanding reserves, and total incurred. Those amounts are not interchangeable. Paid losses represent money already disbursed, reserves estimate future payments, and total incurred is their sum.

The valuation date establishes when those amounts and statuses were last updated. An open claim can change between valuations, so a report with a stale date may understate or otherwise misrepresent the current exposure. A closed claim and an open claim with the same paid amount do not tell the same underwriting story.

Layouts, field names, terminology, and date conventions vary, while scanned reports may prevent even a basic search or copy-and-paste. Running subtotals can be mistaken for additional claims, inflating the account’s losses. More staff can read more documents, but cannot remove these structural differences. After interpreting each report, reviewers must re-key the results into spreadsheets or underwriting systems, introducing a second opportunity for error.

Why loss run review earns its keep as the first workflow to automate

Having to re-key those results makes loss runs a good first project, and the work that relies on them adds to the case. Submission intake puts the start of an account into a structured form, while loss run extraction does the same for its history. FurtherAI identifies both as foundations for underwriting workflows, including document-heavy operations already handled by agents at carriers, MGAs, and brokers in 2026. Renewal scoring, risk pricing, and carrier recommendations all need claims history they can use. Unreliable inputs do not become reliable because the next system is clever.

Compared with broader decision-making applications, loss run review is bounded, document-heavy, and governed by checks an agency can understand. The totals either reconcile or they do not. OutcomeCatalyst places loss run extraction among the underwriting automation stages that pay back early, together with submission intake and appetite triage. That makes it a defensible starting point before expanding into other workflows, rather than a task to postpone until an ambitious AI program is complete.

The agency can test the proposition directly with its own submissions, including poorly formatted reports, and compare extracted totals with manually keyed results. There is no need to wait for a long calibration period to discover whether a subtotal was counted twice. The process has five steps, namely ingestion, parsing, normalization, validation, and output.

Step one: ingesting documents regardless of format

Ingestion begins before the system extracts a single claim. It must accept the PDFs, scanned images, and spreadsheets carriers send without asking agency staff to sort or reformat them first. A scanned loss run from a regional carrier is a useful example. Because the page is an image rather than readable text, optical character recognition, or OCR, must make its contents readable before extraction can begin.

AI document parsing then works over that readable content, interpreting the document rather than relying only on pre-identified field positions. This allows it to handle unfamiliar layouts without requiring a template for every carrier’s exact format. OCR supplies the text from which parsing builds the claims record.

The entry route also determines how much work disappears. FurtherAI describes agents that open broker emails and attachments, classify documents, and extract their data. Agencies also receive loss runs through carrier downloads into their agency management system, or AMS. An ingestion layer needs to connect to those pathways as well as accept manual uploads. Otherwise, staff still have to move documents between systems, and some documents stay outside the automated workflow.

Step two: extracting claim fields from unstructured text

Once the document is readable, extraction turns its contents into labeled fields. The system identifies each claim’s number, date of loss, status, paid losses, reserves, total incurred, and cause wherever they appear in the carrier’s layout. This is the line-by-line reading work a reviewer previously performed. The system must work out what the text means as well as find it beside a familiar label.

For example, one carrier may use “total incurred” where another uses “gross paid plus reserve.” Extraction must work out how those labels relate to the amounts beneath them. It must also distinguish claim rows from running subtotals, because a derived total is not another loss. A parser that misses that hierarchy can produce neatly structured, thoroughly wrong data.

Status requires equal care because an open claim with a large reserve signals different exposure from a closed claim with the same paid amount. Extracting dollars while dropping status strips away part of the underwriting meaning. Well-designed systems also attach confidence levels to extracted fields and surface uncertain readings. Opspring uses this approach in its ACORD auto-fill engine, marking fields in yellow when it is less sure it has read them correctly. For loss run extraction, the relevant feature is the same distinction between a confidently read value and one that needs a person’s attention.

Step three: normalizing data across carrier formats into a common schema

After extraction, the agency still has data in each carrier’s own format. Normalization maps those values into a common schema so claims can be compared across carriers and years. Dates are a straightforward example. A report using MM/DD/YYYY and one using YYYY-MM-DD need a consistent date representation before the agency can reliably analyze claim frequency over time.

Cause-of-loss descriptions need a shared taxonomy as well. One carrier’s “slip and fall” and another’s “general liability — bodily injury — premises” must map to a consistent category if the agency wants to count comparable claims across its book. Financial conventions require similar attention, particularly the distinction between an incremental reserve and total incurred. Normalization puts those amounts on a consistent basis.

At this point, AMS integration becomes an agency decision. Opspring describes bi-directional integration with systems including Applied Epic, AMS360, HawkSoft, EZLynx, and QQCatalyst. Writing normalized claim data back makes it part of the account record. Leaving it in an export means someone must re-import or re-enter it for subsequent work. A cleaner spreadsheet is useful, but it is not the same as a connected workflow.

Step four: validation, flagging anomalies, and keeping a human in the loop

Validation checks whether the structured record makes sense before anyone relies on it. Paid losses plus reserves should reconcile with reported total incurred, and discrepancies should route the document to a reviewer. Missing valuation dates, claims without a status, and open claims with zero reserves also warrant attention. These are signals to investigate before concluding that an error has occurred.

AI hallucinations can create errors and omissions exposure. Errors in claims data can affect pricing, renewal terms, and carrier submissions. The process must expose uncertainty and require review because the model can make mistakes. A system that systematically flags ambiguous fields and inconsistent totals is safer than a manual process in which those errors pass unnoticed.

Human sign-off is therefore part of production design. Opspring describes licensed professionals retaining authority at critical steps before outputs reach a client or carrier. FurtherAI likewise gives each agent a defined point where a person approves, adjusts, or overrides its work. For loss run review, edits and approvals should also be logged for E&O documentation, extending the audit-trail principle Opspring applies to ACORD auto-fill. The human checkpoint makes clear who is responsible for the work between uploading a PDF and submitting a number.

Step five: generating usable outputs that connect to the renewal workflow

Outputs can include structured summaries, spreadsheet exports, and data sent directly to underwriting or analytics systems. They should go where agents need them for the renewal and submission workflow. A correct record that still needs manual reconstruction has not completed the job.

InsuranceAgency.AI describes its Account Manager AI agent pulling claims data from carrier downloads, reviewing loss history, and flagging patterns before the renewal conversation. That history feeds a broader renewal summary that covers what has changed since last year, possible coverage gaps, coverage that may no longer be needed, and carriers that may better fit the current risk profile. The company also describes automatically flagging accounts at risk of not renewing for retention outreach, turning analysis into a reason for producer action.

OutcomeCatalyst evaluates underwriting automation through underwriter minutes per submission touched, quote turnaround time, hit ratio, and underwriting expense ratio. These measures connect document processing to operational results rather than treating extraction speed as the whole achievement. AMS write-back completes the loop by preserving normalized history in the account record. The first cycle saves extraction work, while each subsequent cycle starts from a cleaner baseline.

What the time recovered actually buys a small agency

For a small agency, that cleaner baseline expands capacity. Manual renewal preparation forces choices about which accounts receive a thorough review and which receive a quick glance and a renewal notice. InsuranceAgency.AI identifies this uneven attention as a consequence of no one watching the whole book at once. There is a shortage of hours in which to apply that care.

Systematic claims review and renewal preparation make outreach possible beyond the accounts a producer remembers to call. Contact can happen before the carrier’s rate-increase letter. Producers can spend more of their time on relationships, complex commercial placements, and renewals flagged as at risk. FurtherAI describes the corresponding underwriting shift as moving from doing busywork to managing agents, allowing the same team to handle more accounts without adding headcount.

SANSA looks at how the team could use this extra capacity before it starts the build. Its two-week diagnosis identifies the highest-leverage AI opportunities, often loss run processing and renewal preparation, and the first production system is scoped around that bottleneck. The intended progression is from recovered capacity to new business or service lines the team could not previously staff. This practical deployment model lacks a quantified before-and-after result for a named agency.

What to look for before choosing a loss run automation system

For an agency ready to act, the evaluation should follow the failure modes. Test the system on the agency’s own loss runs, including scans, unusual layouts from regional carriers, and multi-year reports that have subtotal rows. Compare the extracted values and totals against manually keyed results. Clean PDFs are the easy examination paper. The awkward reports show whether the system can do the work it is meant to take over.

Does the system write normalized claims history into the AMS, or stop at a spreadsheet export? Review checkpoints deserve the same scrutiny, including whether uncertain fields are surfaced and whether a person signs off before output reaches a client or carrier. Edits and approvals should appear in an audit trail for E&O documentation. Fully automated delivery without those controls is a governance risk.

Opspring describes a pilot timeline of one to three weeks, while SANSA’s two-week diagnostic is intended to identify the bottleneck worth addressing first. Those timelines cover distinct stages short of a completed deployment. Loss run automation can be operational in weeks, and a six-month proposal for this specific workflow suggests a broader scope than the problem requires. The useful endpoint is a working production system that starts improving the next renewal cycle.

Want results like these for your business?

Book a free 30-minute consultation. A focused conversation, no commitment.

Talk to Us
RELATED INSIGHTS