Insights

Why legacy insurance data cannot train an AI model: what was lost and what to do about it

Ten years of PDFs is not ten years of data: where the information in a legacy book was lost, what can still be recovered, and how to stop losing it.

Updated 13 November 2026 · Written and reviewed by the Soldek team

Most insurance AI initiatives start with an optimistic inventory: years of policies, thousands of claims, surely enough to learn from. Then the data scientist asks for the application answers and discovers they are in proposal forms, scanned, attached to a policy record that stores the premium but not what produced it. The project becomes an extraction project, the budget goes to document processing, and the model never gets built. This is not bad luck. It is what legacy systems do to data.

Where the information went

The application
Captured as a form, stored as a document. The system has 'proposal received'; the facts in it were read once by an underwriter and never entered as fields.
The rating
Done in a spreadsheet beside the system, or in a rating module whose tables were replaced without history. The premium is stored; the factors are not.
The decision
The policy exists, so it was accepted. Whether it was referred, declined and reconsidered, or priced off-tariff is in an e-mail thread.
The claim
In a separate claims system, keyed to the policy by a number typed by hand, with cause as free text and amounts split across fields that changed meaning after an upgrade.
The changes
Endorsements as new documents rather than dated changes to fields; cancellations as a status with no effective date.

What can still be recovered

  • Premiums, inception and expiry dates, and claim amounts usually survive and are reliable. That is enough for a loss-ratio by product and period, not for a risk model.
  • Application answers can be extracted from documents with modern tooling, at a cost per document and an error rate per field. Worth it for the few fields that matter most, rarely for all of them.
  • Rating factors can sometimes be reconstructed by re-rating historic policies against archived rate cards, if the cards were kept and the inputs can be recovered.
  • Decisions and overrides are usually gone.

The decision that actually matters

The question is not how to rescue the legacy book but how to stop producing another one. Every policy written today on a system that stores answers as fields, rates as data and claims on the policy is a row a model can use; every policy written on a system that does not is another document to extract later. An MGA that switches captures a clean book from the day of the switch and treats the legacy history as a partially recoverable bonus.

Migrating what you have

When moving a book, migrate the structured facts that survive (clients, policies with dates and premiums, open claims with amounts) and accept that the application answers for historic policies will be sparse. Record them where they can be recovered cheaply; leave the fields empty where they cannot, rather than filling them with guesses that will poison the model later. The renewal of each policy is the moment the answers get captured properly.

See it running.

Book a demo and we show you the complete platform live. Still stuck? Write to hello@soldek.io.