Is your data ready for AI: four tests for banks and insurers

Banks and insurers continue to improve the completeness, reconciliation and traceability of their data. Even where these controls are established for reporting, it may still be unclear which source an AI should trust when two systems disagree, or how old a figure can be before it should no longer inform a decision.

This article sets out four tests that Chief Data Officers (CDOs), Chief Operating Officers (COOs) and heads of business functions in banking, insurance, wealth, payments and capital markets should think about to check if their firm can support AI in live processes.

The challenge facing organisations

Many firms are finding it difficult to answer this question with confidence. The measures they have (data quality scores, lineage, governance) were built for reporting and for people, and they remain necessary. The four tests set out below assess what AI needs in addition: a governing source, a limit on age, meaning made explicit and a record of what was used. They can be applied one workflow at a time, without a new data programme.

Your data was prepared for people

Banks and insurers have spent more than a decade improving data for regulatory reporting and analytics. The goal of that work was to produce data that is complete, reconciled and traceable for a report. It did not necessarily capture what experienced staff add when they use it.

A relationship manager knows the customer record misses the directors' other companies, so real exposure is larger than the figure on screen. A claims handler knows the newest document in the folder does not govern, because an earlier endorsement changed the cover.

An AI system reading the same records will not automatically have this context unless it is explicitly provided or built into the workflow. Without it, the AI may not be able to determine whether the value it retrieves is current, whether another system holds a better one or whether it is sufficient to support a decision.

The data has not deteriorated. Its user has changed and the new user does not bring its own contextual knowledge.

Readiness depends on what the AI is asked to do

The bar rises with the AI's role: reading, analysing, recommending, preparing an action or executing it.

Take payments. Identifying a failed transaction can be done on yesterday's data. Deciding whether to retry it needs the live status, the retry rules for that payment type and assurance that a second instruction will not duplicate the first.

So "Is our data ready for AI?" has no useful answer at enterprise level. "Is the data behind SME credit renewals ready for AI to recommend a decision?" can be answered with evidence from a single real case.

Four tests of AI-ready data

These are the tests we apply to the data behind a workflow. A firm can pass every data quality control it has and still fail all four.

What does this mean for the CDO and the COO

The CDO typically leads data standards and governance, while the COO or relevant business owner owns the process, where unwritten knowledge and checking costs often sit. Neither can close the gap alone. The source that governs a decision must be agreed by the relevant business and data owners, then translated into a rule the AI-enabled workflow can use.

Regulation points the same way. Under the Senior Managers and Certification Regime, the use of AI does not remove senior management accountability. The Consumer Duty requires firms to monitor and evidence good outcomes for retail customers. For banks, AI applications that meet the definition of a model also need to be considered within the PRA's SS1/23 model risk management framework.

Five questions for one workflow

1. For each data item the decision relies on, which source governs?
2. How old can it be before the AI must not use it?
3. When two sources disagree, what happens?
4. What do experienced staff know about this data that is recorded nowhere?
5. Can we reconstruct one decision from last month: the data used, the AI output and what the person changed?

What’s next in practice?

Start with one workflow where AI is already live or close to implementation. Take a recently completed case and trace the decision back through the data, source by source, applying the four tests to identify where the workflow is ready and where gaps remain.

This provides a practical way to identify the changes that matter most, whether that is clearer data ownership, better-defined business rules, stronger controls or changes to the underlying process.

Forvis Mazars can help organisations apply this assessment to a priority workflow, bringing together business, data, technology and risk perspectives to identify practical actions and prioritise them according to their impact on efficiency, control and AI readiness.

FAQs