By the end of this you will be able to read any loss run in five minutes, turn a stack of carriers' reports into one clean table with AI, and catch the totals it gets wrong.
Five years of history. Three carriers. Three formats. One is a scanned fax.
The underwriter wants a loss summary by Friday.
This is one of the best jobs you can hand to AI, as long as you know what a loss run is saying first. So let's start there.
What a loss run is
A loss run is the carrier's report of the claims on a policy over a period of time. It's the account's claims history, and underwriters use it to price the risk. ICW Group's explainer on what a loss run is and how it can help or hurt your premium puts the carrier's view plainly.
Most loss runs carry the same core facts, in wildly different layouts:
| Field | What it tells you |
|---|---|
| Valuation date | The date the numbers were valued. Everything is "as of" this day |
| Policy number and period | Which policy year the claim belongs to |
| Date of loss and date reported | When it happened, and how fast it was reported |
| Claim description | What happened, often in a few cryptic words |
| Status | Open or closed |
| Paid | What the carrier has paid so far |
| Reserve | What the carrier still expects to pay |
| Incurred | Paid plus reserve |
| Expense | Adjusting and legal costs, sometimes split out, sometimes included |
If paid, reserves and incurred feel slippery, Milliman's beginner's guide to the casualty actuarial language defines them well.
How to read a loss run in five minutes
- Valuation date. How old is this? Reserves move after it. A stale report can make an open claim look small.
- Policy periods. Do you have every year the submission asks for, from every carrier? A missing year is a question the underwriter will ask.
- Open claims. Read these first. Big reserves on an open claim worry an underwriter more than a closed claim that paid out.
- Incurred by year. Is it trending up or down? One bad year with a fix behind it reads very differently from a steady climb.
- Large losses. Anything that stands out needs a sentence: what happened and what the client changed.
- Gaps. A year with no losses should come with a no-loss letter or a loss run that says so, not silence.
That's it. Five minutes, and now you know what the AI should produce.
Turn five carriers into one table
This is where AI earns its seat. Each carrier names its columns differently. Workers comp loss runs often split medical, indemnity and expense. Some reports include expense in incurred and some don't.
You want one table.
Attached are loss runs for [CLIENT NAME] from [CARRIER A], [CARRIER B]
and [CARRIER C].
Make one table with these columns:
Carrier | Policy period | Valuation date | Claim number | Date of loss |
Description | Line | Status | Paid | Reserve | Incurred | Expense included?
Rules:
- One row per claim. Copy amounts exactly as printed.
- If a column is missing on a report, write NOT SHOWN. Do not calculate it.
- Flag any claim that appears on more than one report.
Then compute, using code, totals by policy year: claim count, paid,
reserve, incurred. Show the totals each carrier printed next to yours.
What should come back? Something like this, for a made-up account:
| Policy year | Claims | Paid | Reserve | Incurred | Open |
|---|---|---|---|---|---|
| 2022 to 2023 | 2 | 14,200 | 0 | 14,200 | 0 |
| 2023 to 2024 | 1 | 3,100 | 0 | 3,100 | 0 |
| 2024 to 2025 | 3 | 22,750 | 41,000 | 63,750 | 1 |
One glance and you know where the conversation with the underwriter will go: the open claim in the latest year. That is the claim to get a status update on before you submit.
The last line of the prompt is your safety net. If the model's total for a year doesn't match the carrier's printed total, something got dropped or misread.
Prompts for the rest of the job
Once the table is right, AI drafts the words around it.
Using the loss table above, draft a short loss summary for an
underwriter: one paragraph per policy year, plain language.
For every open claim and every claim over [AMOUNT], add a line
for what happened and a placeholder [CORRECTIVE ACTION] for me to fill.
Do not invent causes or corrective actions.
List the questions I should ask the client about open claims and
large losses in this table, so the underwriter gets answers before
they ask.
Where AI slips on loss runs
- Scanned reports. A faxed loss run turns 8 into 3 and 1 into 7. Spot check every row on a scanned page.
- Expense in or out. Mixing reports that include expense with ones that don't overstates or understates incurred.
- Duplicates. The same claim appears on two valuations of the same report. Keep the latest.
- Arithmetic in prose. Totals typed in a chat reply can be wrong. Make it calculate with code and show its work; OpenAI's file uploads FAQ covers how ChatGPT handles spreadsheets and data files.
- Sensitive details. Workers comp claims describe injuries. Use a business plan with training off, and keep only what the submission needs. See is ChatGPT safe for insurance client data.
Put it to work in the submission
The loss table feeds the prior carrier and loss history section of the commercial application. Our ACORD 125 guide covers that section, and insurance submission intake automation shows where loss runs sit in the whole packet.
A clean summary also does something quieter. It tells the underwriter your agency knows the account. That is worth more than any one prompt.
When you'd rather not chase them at all
Half the work of loss runs is getting them: requesting from each carrier, following up, filing them to the account. That's work Salience does for agencies. Salience requests the loss runs from each carrier, follows up until they arrive, and prepares the summary, and nothing goes to an underwriter until you approve it.
The rest of the task guides are on the AI by task shelf, including AI for insurance renewals, where these loss runs end up next.
Give us a call
Hey, all the tech is great, and innovation is great. There's a lot of responsibility and liability across the entire space, and a lot of people are saying the same words that mean different things. Give us a call. We plug in specifically to your processes and take care of the GDS first to unlock you and bring home the bacon.
Call +1 (916) 866-7867 or book a walkthrough.
Would you rather approve this than type it?
Forward the request or the prior application. Salience prepares the form, flags what's missing, and nothing goes out until you approve it.
Frequently asked questions
How do you read a loss run?
Check the valuation date first, then the policy periods covered. Then read each claim's date of loss, status, paid amount and reserve. Paid plus reserves is the incurred amount, and open claims with large reserves are what an underwriter will ask about.
Can AI analyze loss runs?
Yes, for the mechanical part. ChatGPT or Claude can pull claims from several carriers' loss runs into one table and total them by year. Have it compute totals with code, then check them against each carrier's own totals.
What is a valuation date on a loss run?
It is the date the figures were valued, the as of date. Paid amounts and reserves change after it, so an old valuation date can understate open claims. Underwriters generally want recent loss runs.
What is the difference between paid and incurred on a loss run?
Paid is what the carrier has already paid on a claim. Incurred is paid plus the reserve still held for what the carrier expects to pay. On closed claims the two usually match; on open claims incurred is higher.
How do I summarize loss runs for a submission?
Put every carrier's claims into one table by policy year with claim counts, paid, reserves and incurred, then write a short narrative for any large or open claim covering what happened and what the client changed afterward.