Use case

Churn early warning in insurance: data, method and verification

UNID AI Redaktion ·

A churn early warning is a prediction that ranks existing insurance policies by how likely they are to be cancelled within a set period. It needs linked policy, claims, payment and contact data, a model that learns patterns from them, and a backtest that lets the insurer check for itself whether the ranking beats the one it already uses.

This article walks through the three building blocks, data, method and verification, so that a business team can judge what such a project involves.

What does a churn early warning do, and what doesn’t it do?

A churn early warning produces a ranked list. At the top are the policies with the highest risk of cancellation in the chosen window, for example the next three or six months. Account managers or the sales team can then focus their limited time on the cases where a conversation is most likely to make a difference.

What the prediction does not do: it does not decide which retention action is right, and it does not guarantee that a call will prevent a cancellation. Whether retention measures work is a separate question, and it has to be measured with a control group.

What data does a churn early warning need?

Most insurers already hold the data needed, but spread across several systems. What matters is that it can be linked through customer and policy numbers.

Data source Examples Why it helps
Policies Line of business, start date, term, renewal date, cover, number of policies per customer Tenure and the policy cycle show when cancellation is possible and likely
Premiums Premium level, adjustments, indexation, payment frequency Premium increases can prompt customers to compare offers
Claims Reported claims, handling time, rejections Claims experience can affect satisfaction
Payments Reminders, returned direct debits, arrears Payment problems can precede a cancellation
Contacts Complaints, enquiries, change of address, change of account manager Changes in the customer relationship
Distribution Channel, account manager or broker Different channels show different retention patterns

Two practical notes. First, special categories of personal data under Article 9 GDPR, such as health data, are generally not needed for a churn early warning and should be kept out. Second, fields that only come into existence because of a cancellation, such as a cancellation reason or the date a notice was received, must not be used as inputs. Otherwise the model “sees” the answer and looks better in testing than it will be in production. This is known as data leakage.

How is the prediction made?

The classic approach goes like this: data scientists condense the linked tables into one large table with a row per policy and hundreds of engineered features, such as “number of claims in the last twelve months”, and train a model on it. It works, but it takes a lot of time, and every new question, such as a different prediction window, starts the feature work again.

A newer approach works directly on the linked tables. The research field is called relational deep learning: the database is treated as a network of rows and their links, and the model learns the patterns itself instead of relying on hand-built features. Relational foundation models build on this work. One published example is KumoRFM by Kumo. According to its research paper, it was pre-trained on public and synthetic databases and can make predictions on a new database from historical examples without being trained for the specific task, with fine-tuning available for production use.

UNID AI Predict uses a relational foundation model of this kind. For the business team the main change is that the question can be stated precisely (“which policies will be cancelled in the next six months?”) and the prediction is made from the tables as they are.

One point matters for trust in the result: the numbers come from the data and the prediction model, not from a language model. Where a language model is used, it only puts the results into words, for instance as a summary for an account manager.

How can an insurer check that the prediction is right?

An insurer should not buy a prediction on the strength of a slide deck, but on the strength of its own test. We recommend a backtest with a hidden cut-off date, which we also call the “sealed envelope”:

  1. Set a cut-off date. The insurer picks a date in the past, for example twelve months ago.
  2. Hand over data up to that date. The model only receives data that was known on that day.
  3. Keep the outcomes sealed. The insurer keeps to itself which policies were actually cancelled afterwards.
  4. Deliver the ranking. The model ranks the policies by cancellation risk for the agreed window.
  5. Evaluate in-house. The insurer compares the ranking with the actual cancellations and with its existing model, for example: how many of the policies that were actually cancelled appear in the top five or ten per cent of the list?
  6. Decide. Only if the ranking is measurably better than the current method in your own test is the next step worth taking.

That is why we do not quote a percentage in advance. How good the prediction is, you measure in the backtest on your own data.

How is the ranking used day to day?

A prediction is only useful once it is part of a workflow. It pays to settle three things before you start.

Who receives the list? Usually in-house account managers, sales partners or a dedicated retention team. The list should contain only as many cases as those people can actually work through in the period.

What happens to a case? A call, an advisory appointment, a policy review, or a deliberate decision to do nothing. A person who knows the customer makes that call.

How is impact measured? To separate the effect of retention actions from the quality of the prediction, part of the high-risk cases should be left untouched as a control group. Only then can you see whether the conversations actually prevent cancellations.

The prediction is recalculated regularly, for example monthly, so that new claims, premium changes and contacts are taken into account.

Where do the data and the model run?

Policyholder data is sensitive. UNID AI Predict can therefore run on the insurer’s own infrastructure, with a local language model if required. Our article comparing on-premise, own server and cloud discusses the options, and our security and privacy page sets out the principles of data processing.

What does the EU AI Act say?

The AI Act, Regulation (EU) 2024/1689, lists in Annex III AI systems intended for risk assessment and pricing in relation to natural persons in life and health insurance as high-risk. A ranking that tells account managers whom to call first serves a different purpose. Whether it is nonetheless covered in a particular case depends on how the results are used, for instance whether they feed into pricing or underwriting decisions. The insurer has to make that assessment itself, ideally with its legal and compliance teams.

On timing: the Digital Omnibus on AI, published in the Official Journal in July 2026, moved the obligations for Annex III high-risk systems to 2 December 2027. The current text on EUR-Lex is what counts. The GDPR applies regardless, including the rules on automated individual decisions in Article 22.

What comes next?

The use-case page churn early warning in insurance shows the process in detail, and the UNID AI Predict product page describes the product behind it. Terms such as relational foundation model are explained in the glossary. To discuss a backtest on your data, contact us.

Sources

  1. Fey et al. (2025): KumoRFM – A Relational Foundation Model (research overview, NVIDIA documentation)
  2. KumoRFM-2: Scaling Foundation Models for Relational Learning (arXiv, 2026)
  3. Fey et al. (2023): Relational Deep Learning – Graph Representation Learning on Relational Databases (arXiv)
  4. Regulation (EU) 2024/1689 (Artificial Intelligence Act) – EUR-Lex
  5. Digital Omnibus on AI, Official Journal of the EU L, 24 July 2026 – EUR-Lex

FAQ

Frequently asked questions

How accurate is a churn early warning?

There is no general answer. Accuracy depends on the line of business, data quality and the prediction window. That is why we do not quote a number in advance; we measure it together with the insurer in a backtest with a hidden cut-off date.

Does a churn early warning fall under the high-risk rules of the EU AI Act?

Annex III of the AI Act lists AI systems used for risk assessment and pricing of natural persons in life and health insurance. A retention ranking serves a different purpose. Whether it is covered in a given case depends on how it is used, and the insurer has to make its own legal assessment.

Does the data have to leave the insurer?

Not necessarily. UNID AI Predict can run on the insurer's own infrastructure, with a local language model if required.

See it working on your own material.

In a 30-minute call we show the products live and work out which use case makes sense for you first. No obligation, in German or English, in Vienna or online.

UNID AI · Leonard-Bernstein-Straße 10 · 1220 Vienna · +43 660 476 30 89