Claims intake for an insurer

An insurer handles 1.2 million claims a year. That is about 3,300 a day, and up to 30 a minute at peak after storms. Each claim arrives by email with an average of 6 PDF or photo pages. Today an adjuster spends 14 minutes per claim: keying 40 fields into the claims system, checking for fraud signals and writing the first reply to the customer. The customer wants Claude to do the first pass.

Requirements:

• 12 of the 40 fields are critical, such as the policy number, the date of loss and the claimed amount. The customer wants 98% accuracy on them.
• The target is 4 minutes of adjuster time per claim.
• The budget is $0.15 per claim in model costs.
• EU claims must be processed in the EU.
• No payment-related action happens without an adjuster's approval.
• The claims system's API takes 20 writes a second. A processed claim needs 5 writes (create the claim, attach the documents, write the fields, write the fraud flags, save the reply draft).

Assume these prices and sizes:

• Large model: $3 per million input tokens and $15 per million output tokens.
• Small model: $1 per million input tokens and $5 per million output tokens.
• Cached input costs 10% of the normal input price.
• A page image is about 1,600 tokens.

Your tasks:

1. Before you draw anything: how will you show the customer that Claude helps? Give the eval plan, its size and the gates.
2. Design the system: ingestion, parsing, extraction, routing to people, the integration with the claims system, audit and data residency.
3. Show that it fits the budget.
4. How do you roll it out, and what happens on the day a new model version ships?

Write down three things before you look at the guide: the size of your eval set and the gate it must pass, your cost per claim with each step priced, and the one action the model must never take on its own.