AI data readiness in banking: proof beats good intentions
Everyone agrees AI needs a solid data foundation. Almost nobody can prove theirs is solid. Here is what evidence a bank’s data readiness should produce – an article written by IDS Consulting.
There is a sentence circulating in every data leadership feed right now, and it is correct: before you build agentic or generative AI, fix the data foundation — ownership, definitions, quality, lineage. Build on sand and the models will not simply fail. They will scale the mess, confidently, at machine speed.
Nobody serious disagrees. Deloitte’s 2026 banking and capital markets outlook puts it in almost the same words, describing AI implementation as “often throttled by brittle and fragmented data foundations.” In Deloitte’s 2024 data and analytics survey, more than 90% of bank data users said the data they needed was often unavailable or too slow to retrieve, and 81% named data quality as a top challenge. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027.
We made this argument ourselves in Why AI and Your Data Aren’t Best Friends Yet. So the advice is right. It is also, on its own, not actionable. Every bank we work with already agrees that the foundation matters. What almost none of them can do is answer the next question.
The question nobody is answering
How do you prove the foundation is fixed?
Not assert it. Not assume it because a governance programme ran for eighteen months and produced a catalogue. Prove it — to an auditor, to a regulator, to a Chief Risk Officer who was not in any of the project meetings.
“We have improved data quality” is an opinion. An opinion is not an acceptance criterion. And a model that acts on data does not read your governance policy. It reads the table.
What a defensible answer looks like
In our work inside European bank data departments, a defensible answer produces three kinds of evidence, in this order.
1. The definition is agreed before anything moves
Not documented afterwards. Agreed — by the business owner of the metric, in writing, before a single transformation is built. Most AI incidents we see are not quality failures at all. They are two correct numbers produced from two different, equally defensible definitions, with no one authorised to say which one the model should use.
2. Equivalence is demonstrated, not claimed
When a figure is produced in two places — the legacy on-premises estate and the new cloud platform, the warehouse and the agent’s answer — the two must be reconciled and the difference explained to the cent, or explicitly accepted with a reason. This is the single most useful discipline we carry into an engagement: build the reconciler before the converter. A migration is finished when equivalence can be shown to someone who was not on the project. Not when the pipeline runs green.
3. Lineage is traceable end to end
From source field to reported figure, with the transformation logic visible. This is why we built ATLAS, our source-to-target mapping and lineage platform. Lineage is not documentation. It is the only thing that lets you answer “where did this number come from?” in minutes rather than weeks.
Why the migration window decides this
Most European banks are currently running the hardest version of this problem: moving an on-premises data warehouse to cloud while regulatory reporting continues, uninterrupted, every single period. For months, two estates hold the same metric. If no one reconciles them, the organisation quietly acquires two versions of the truth — and then points an AI system at whichever one it can reach.
That is how the mess scales. Not through bad intentions. Through an unclosed gap during a migration window nobody treated as a governance event.
The regulator is about to ask the same question
Under the EU AI Act, credit scoring and creditworthiness assessment sit in the high-risk category, and Article 10 requires training, validation and testing data sets to be governed, relevant, sufficiently representative and — to the best extent possible — free of errors and complete. Those obligations were deferred by the digital omnibus and now apply from 2 December 2027. That is not a reprieve. It is a delivery window, and it closes in the middle of most banks’ current migration programmes.
The evidence a regulator will want is the same evidence a Head of Data should want anyway: agreed definitions, demonstrated equivalence, traceable lineage.
More details here
