Skip to main content

Claude – Claims Data

Working a real claims file with Claude — Participant workspace

Participant workspace · advanced session

Working a real
claims file with Claude

A set of hands-on exercises on a real, fully sanitised claims transactions listing. Work in the file directly, and keep one question in mind throughout: a confident answer is not the same as a correct one.

Download the data fileClaimsTransactionsListing_DEMO_sanitised.xlsx · 54,714 rows

Attach the data file in Claude, then work through whichever exercises suit the room. They stand alone and can be done in any order — you won't need all of them. The aim isn't to trust the tool; it's to learn where it helps and where it needs watching.

Work in

Pairs — the good bits come from comparing what you each got.

Each exercise

Copy the prompt, run it, then talk about what came back.

Stay sceptical

When Claude gives a number, your next move is always "do I believe it?"

01
Ask for the total
Prompt auditing

A simple-looking question. Run it, then decide whether you'd put your name to the answer.

Prompt

Here is a claims transactions file. Tell me the total losses.
Talk about
  • What figure came back — and would you send it to a colleague as "the total losses"?
  • What did the question leave open? (Gross or net, currency, dates, adjustments.)
  • Rewrite the prompt so the answer is one you'd stand behind, and compare with your partner's.
02
Catastrophe share
Questioning data

Get a straight answer to a straight question — then check whether the data let you down.

Prompt

Using this file, what share of claims are catastrophe losses?
Give me the number and the percentage.

Then audit before you trust it

# Question the data before you believe the grouping
Before you answer, audit every category column in this file for
data-quality problems: inconsistent capitalisation, leading/trailing
spaces, near-duplicate labels, and blank or placeholder values like
"<<Unknown>>". List what you find, then give me the corrected grouping.
Talk about
  • Did the corrected number differ from your first answer? By how much?
  • What would you have reported if you'd trusted the first result?
  • What does that tell you about running a grouping before checking the data?
03
Contras and corrections
Questioning data · reconciliation

This file records manual adjustments alongside the system entries. Work out what they actually do.

Prompt

This file has a "System / Manual" column with three values:
System, Manual Contra, and Manual Correction.
1. Show total incurred split by that column.
2. Tell me the net effect of the manual adjustments on the whole file.
3. List the claims that carry at least one manual adjustment.
Talk about
  • What do the manual adjustments do across the whole file, versus on individual claims?
  • If a report only showed the whole-file total, what would a reader miss?
  • Which view would you lead with — book-level or claim-level — and for whom?
04
Build and swap a skill file
Skill files · sharing

Turn the checks you've learned into something reusable, then hand it to another pair and run theirs.

Starter skill file — claims-audit Already converts the dates and runs the category check. Three gaps marked [GROUP TO ADD] for you to complete. Unzip it and you'll find claims-audit/SKILL.md inside, ready to use.
Download SKILL.zip
What to do
  • Read the starter skill with your partner and run it on the file as-is.
  • Fill in the three [GROUP TO ADD] sections: the missing-data rule, the reconciliation from exercise 3, and a currency note.
  • Swap with another pair, run their version, and compare what each skill caught.
  • Talk about: whose wording was clearer, and one line you'd borrow from theirs?
05
Hypotheses and a sceptical test
Ideation · critical thinking

Come up with ideas, test one, then argue against your own result.

Prompt

Brainstorm five hypotheses about what drives large incurred amounts
in this file (e.g. line of business, loss cause, accident year).
Pick one and test it. Then tell me honestly why this analysis
might be misleading given what's actually in the data.
Talk about
  • Which hypothesis did you test, and what did the result suggest?
  • Every reason you can think of to distrust that result.
  • Would you show this to a pricing or reserving colleague? What would you caveat?
06
Break it on purpose
Testing the tool

In your pair, one of you tries to make Claude give a misleading answer about the data; the other designs a prompt that stops it.

Your challenge
  • There's at least one oddity planted in the amounts — something that looks wrong or too neat. Find it.
  • Can the data itself tell you whether it's real, or would you have to go and ask a person?
  • Talk about: the strongest "gotcha" you managed, and the prompt that defended against it.
About this data. Fully sanitised — no real insured, policy, claim or event details remain. Amounts were kept exact so the contra and correction entries still reconcile. Illuminate AI · claims data session.