Short answer. Most bank AI projects stall on data, not models. Fewer than 25% of banks rate their data platforms as ready for AI. You get data ready by scoping it to one use case, not the whole bank: profile the 3 to 5 sources that use case touches, fix accuracy and completeness in those sources, put governance and lineage around them, and confirm you can refresh the data on the cadence the model needs. A focused readiness check takes 2 to 3 weeks, not the multi-year data lake program most vendors sell.
Data readiness is a decision about scope, not a platform you buy. A bank that tries to make all of its data AI-ready before starting any project waits years and ships nothing. A bank that scopes readiness to a single use case starts in weeks. The difference is where you draw the boundary.
Start from the use case and work backward to the data it needs. A fraud model needs clean transaction records and confirmed fraud labels. A customer service assistant needs current product terms and accurate account data. A credit decision model needs application data, bureau data, and outcome history. Each of these touches 3 to 5 systems, not 300. You make those systems ready, and you leave the rest for the next use case.
In practice the readiness work falls into four checks:
Ableneo has shipped 34 production AI projects across banking, insurance, and other regulated sectors, and roughly 4 of 5 reach production. The projects that stall almost always stall on one of these four checks, discovered late. The projects that ship run the check first, before a line of model code is written.
For banks the data question is also a compliance question. The EU AI Act, in force since August 2024 with high-risk obligations applying from August 2026, sets data governance rules for high-risk AI systems in Article 10. Training, validation, and testing data must be relevant, sufficiently representative, and, to the best extent possible, free of errors and complete for the intended purpose. The Act also requires you to document data collection, its origin, and the preparation steps such as labelling, cleaning, and aggregation, and to examine the data for bias that could lead to discrimination. Credit scoring and insurance pricing sit squarely inside this scope.
Banks already carry a data quality standard that maps closely to this. BCBS 239, the Basel Committee’s principles for risk data aggregation, requires accuracy, completeness, timeliness, and adaptability, backed by lineage and controls. A bank that treats AI data readiness as an extension of its BCBS 239 discipline, rather than a separate program, reuses controls it already runs and gives auditors evidence they already recognise. The infrastructure that satisfies a data audit is close to the infrastructure that makes AI defensible.
Most bank AI projects stall on data, not models, and fewer than 25% of banks rate their data as AI-ready.
AI-ready data is data that is fit for a specific model’s purpose, governed, and repeatable. Fit for purpose means it covers the real distribution of cases the model will see, with labels the model can learn from. Governed means every field has a known origin, an owner, and a lineage record. Repeatable means the pipeline that delivered the training data can deliver the same shape of data in production, on schedule, without a person exporting a spreadsheet. Data that is clean once but cannot be refreshed is a demo, not a ready pipeline.
A focused data readiness assessment for one use case takes 2 to 3 weeks when the right people have access to the source systems. The output is concrete: a profile of each source, a list of the accuracy and completeness gaps, an estimate of the effort to close them, and a go or no-go on the use case. This is deliberately short. The point is to find the blockers before the project commits budget, not to boil the ocean. A readiness check that runs for six months has become the project, and it is the wrong project.
No. You do not need a new data lake, a lakehouse, or a full platform migration to start an AI use case. Those programs are useful, and many banks are right to run them, but making them a precondition is how AI stalls for years. Scope the data to the use case, connect to the systems of record you already have, and build the minimum governed pipeline that use case needs. A platform earns its place once several use cases share the same data, not before the first one has proven value. Sequencing the platform after the first production use case, rather than before it, is the difference between AI in production this year and a roadmap slide.
For high-risk systems the Act turns data preparation from an engineering task into a documented, auditable one. Under Article 10 you must record where data came from, how it was prepared, and what you did to check it for bias. That means the bank needs a data sheet for each high-risk model: sources, preparation steps, coverage of protected groups, known limitations, and the checks run before deployment. Most banks already produce similar evidence for model risk management. The change is that AI data documentation is now a legal obligation with penalties attached, not a good practice, so it belongs in the readiness work from day one rather than bolted on before an audit.
Fix the data the first use case depends on, and fix it in the source, not in a copy. Patching data inside the AI pipeline hides the problem and breaks the next project that uses the same source. Prioritise the fields the model weights most, confirm you can reconcile them to a source of record, and put a quality control and a refresh schedule around them. Assign an owner for each source before the model goes live, because data that no one owns degrades, and a model on degrading data drifts. Everything else waits for the use case that needs it.
Ableneo builds the data, governance, and process foundations that let AI reach production, not just a pilot. Across 34 production projects in 2025, the pattern holds: the readiness check comes first, scoped to the use case, and the pipeline is built to run again, not once. That discipline is why roughly 4 of 5 Ableneo projects reach production in a market where most AI pilots do not. Our work with banks and insurers in Central and Eastern Europe pairs this data readiness practice with the governance evidence regulators expect. See how we approach this in AI transformation.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.