How Do You Get an Insurer's Data Ready for an AI Project?

7 min readAbleneo AI transformation team

Short answer. An insurer gets its data ready for AI by fixing data quality, lineage, and governance before the model, not after. The EU AI Act’s Article 10 requires the training, validation, and testing data behind a high-risk insurance system to be relevant, representative, and as free of errors as possible. That ordering is why roughly 4 of 5 Ableneo projects reach production: the data foundation is built first. The three moves that matter are a documented lineage from source system to model, a representativeness check against the real policyholder book, and a governance record an auditor can read.

1What This Means in Practice

Insurers hold decades of data in policy administration, claims, underwriting, and actuarial systems. That data was captured to record transactions, not to train models, so it rarely arrives in the shape an AI system needs. A claims or pricing model fails less often on the algorithm and more often on the data feeding it: missing fields, free-text notes with no structure, and codes that changed meaning across system migrations.

Getting data ready means three distinct jobs, not one. Clean it, so errors, duplicates, and gaps are resolved and recorded. Connect it, so there is a traceable path from the source record to the exact feature the model sees. Govern it, so there is a documented account of where each field came from and who is allowed to use it. An insurer that treats these as a single vague step underestimates all three.

Three common failures show what the work prevents:

2Why This Matters for Regulated Industries

For an insurer in the EU, data readiness is a legal duty, not a best practice. The EU AI Act classes AI used for risk assessment and pricing in life and health insurance as high-risk under Annex III. From 2 December 2027, those systems must meet Article 10, which sets explicit quality and governance duties on the training, validation, and testing data. General-purpose AI and the Article 50 transparency rules have been in force since 2 August 2026, so the compliance clock is already running while the data work is still being scoped.

EIOPA, the EU insurance regulator, reinforced this in its Opinion on AI governance and risk management of 6 August 2025. It folds AI into frameworks insurers already run under, Solvency II, the Insurance Distribution Directive, and DORA, and it applies a proportionality principle, so a small insurer is held to a lighter standard than a large one. The practical effect is that an insurer’s data governance for AI is read against rules it already knows, which makes the readiness work concrete rather than abstract.

Get the data right before the model: data quality, lineage, and governance are the main reason insurance AI pilots stall, not the algorithm.

3What Does the EU AI Act Require for Insurance Data?

Article 10 requires the datasets behind a high-risk system to be relevant, sufficiently representative, and, to the best extent possible, free of errors and complete for the intended purpose. It also requires documented data governance: the design choices, where the data came from, how it was prepared, the assumptions made, and an examination for possible bias. For a life or health insurer, that means a regulator can ask to see not just the model but the data lineage behind it. An insurer that cannot show where a training field originated, or how gaps were handled, has a compliance problem before it has a performance problem.

4What Does “Representative” Data Mean for an Insurer?

Representative data covers the people and risks the system will actually affect, not a convenient subset. For a pricing or underwriting model, the dataset has to reflect the full policyholder book across age, region, product, and risk class, so the model does not systematically misprice a group that was thin in the training data. The AI Act allows an insurer to process special-category data strictly to detect and correct bias, with safeguards, and to delete it once the bias is corrected. Representativeness is therefore both a performance requirement and a fairness requirement: a model trained on a skewed book produces skewed prices, and in a regulated market that is a supervisory finding, not a footnote.

5How Do You Get Data Out of Legacy Policy and Claims Systems?

Most insurers run core policy and claims platforms that are 15 to 20 years old and were never designed to feed a model. The data lives in formats and codes that shifted across migrations, with the same field meaning different things in different eras. Getting it out is an engineering job: build a traceable extraction path, map the historical codes to one current schema, and preserve the lineage so each model feature points back to its source record. The target is a repeatable pipeline, not a one-time export, because a model retrained next year needs the same data on the same terms. This is the last-mile work that separates a demo from a system that survives an audit.

6How Long Does Data Preparation Take Before an Insurer Can Start?

Data preparation is usually the longest phase of an insurance AI project, and planning for a few weeks when it needs a few months is the most common reason a pilot slips. The honest answer depends on the state of the source systems. An insurer with a modern data platform and clean lineage can be model-ready in weeks. One extracting from several legacy cores should budget months for cleaning, mapping, and governance. The way to shorten it is to scope narrowly: prepare the data for one well-defined use case first, then reuse that governed foundation for the next one, so the second project starts further along than the first.

7What Should an Insurer Do First?

Start with one use case that carries a clear business result, then work backward to the exact data it needs. Audit that data against the three Article 10 criteria, relevance, representativeness, and errors, and document the lineage as you go, so the governance record is a by-product of the work rather than a separate project later. Fix the data for that one case, ship it to production, and only then widen the scope. An insurer that tries to clean all its data before starting never starts. One that prepares data per use case builds a governed foundation that compounds across projects.

8The Ableneo Perspective

Ableneo treats data as one of the four layers of an AI system, alongside people, process, and technology, so the data foundation is built in the first sprint rather than retrofitted before launch. That discipline is the reason roughly 4 of 5 Ableneo projects reach production, and 94% of them run on large language models whose output depends entirely on the quality of the data behind them. Ableneo’s production work in regulated CEE industries, including the document-validation system at T-Mobile Czech Republic that cut processing from seven hours to thirty minutes at over 90% accuracy, shows what a governed data pipeline makes possible. For where this fits in an insurer’s wider programme, see Ableneo’s guide on where an insurer should start with AI under the EU AI Act.

Key takeaways

Sources

Planning AI in a regulated business? Ableneo takes systems from classification to governed production.

Talk to Ableneo