Insights

Structuring insurance data for AI: a practical guide for MGAs

What an insurance book must look like before a model can learn from it: one row per risk, every answer a named field, the claim joined to the policy.

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

Every conversation about AI in insurance eventually arrives at the same place: the data. Not the volume of it, which most MGAs have plenty of, but its shape. A model learns from rows and columns with consistent meaning. An insurance book usually exists as documents, e-mails and a policy system that stores the outcome but not the inputs. Structuring insurance data for AI is the work of closing that gap, and the cheapest moment to do it is before the first policy is written.

What a model actually needs

Strip away the vocabulary and a risk or pricing model needs four things for every risk it learns from:

Features
The facts known at the moment of quoting: the answers to the application questions, the sum insured, the cover chosen, the duration, the channel. Each one in its own column, with a consistent name and type across every row.
Exposure
How much of the risk was carried: policy-years, trip-weeks, insured persons, or whatever unit claims scale with. Without it, a policy that ran three months looks like one that ran three years.
Target
What happened: whether a claim occurred, how many, how much was paid and reserved. This is the label the model learns to predict.
Lineage
The join between them. The claim must point to the policy, the policy to the quote, the quote to the answers. If any link is a free-text reference typed by hand, the row is unusable.

Everything else, including which algorithm to use, is secondary. A generalised linear model on a clean table beats a sophisticated model on a table that had to be reconstructed from PDFs.

Where insurance data usually breaks

  • The application is a form, not a schema. Answers live in a PDF, a scanned proposal or a notes field. The facts that priced the risk were read once by a human and never stored as data.
  • The premium is a number without a reason. The rate card changed three times; nobody can say which version priced a 2023 policy or which factor moved it.
  • The decision is implicit. The system knows a policy was bound; it does not know the risk was referred first, or why the underwriter accepted it anyway.
  • Claims live elsewhere. A claims system keyed to a policy number by hand, with cause and amounts in a different structure, often in a different currency convention.
  • Each product has its own spreadsheet. Ten products, ten layouts, no common vocabulary for the same fact.

How to structure it from day one

  1. 1Define every application question as a field with a name and a type. 'dog_age' as a number, 'breed_size' as an option from a fixed list, 'trip_start' as a date. Never 'describe the risk' as a text box if the answer will ever be used for pricing.
  2. 2Store the answers per field on the quote, not in the document generated from them. The document is an output; the fields are the record.
  3. 3Price from data, not from a spreadsheet beside the system. Base rates and factors as rate-table rows keyed on the same field names; the premium is then re-computable from the answers and the rate rows.
  4. 4Make the decision explicit. Accept, refer and decline as rule outcomes on the quote, and the underwriter's final status (quoted, declined with a reason, bound) as a separate fact, so a model can learn both what the rules said and what the humans did.
  5. 5Put the claim on the policy and the policy on the quote, by foreign key, in the same system. Record date of loss, cause, reserve, paid and settled as numbers.
  6. 6Keep exposure explicit: inception, expiry, endorsements and cancellations with effective dates, so earned exposure can be computed per row.
  7. 7Log every change with actor and time. A model trained on a book should be reproducible: which answers, which rates, which decision, when.
  8. 8Make the whole thing exportable as tables, not as reports. If the only way out is a PDF or a dashboard, the structure is lost at the door.

What this looks like in practice

One row per quote, the features as the columns your underwriters already ask for, the premium and the outcome beside them, and the claims joined on. A data scientist can start modelling the same afternoon. The alternative, a six-month project to extract fields from documents before any modelling begins, is where most insurance AI initiatives quietly end.

Questions

How much data does an MGA need before a model is useful?
Less than people assume, if it is structured. A frequency model on a few thousand policy-years with a few hundred claims already separates risk deciles meaningfully; the public motor datasets actuaries benchmark on are in that range. What kills a project is not small data but data that has to be reconstructed by hand.
Can we structure the data we already have?
Partly. Policy and claims amounts and dates usually survive; the application answers often do not, because they were never stored as fields. You can extract them from documents, but it is expensive and lossy. The better strategy is to start capturing properly now and treat history as a bonus.
Does structuring data for AI mean putting AI into the quote process?
No. Structuring the data is about what you record; it says nothing about how you price. On Soldek the engine stays deterministic (rate tables and rules); the structured data is exported and modelled outside the platform, and what the model learns comes back as an updated tariff.

See it running.

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