Short answer. In Ableneo’s portfolio, 35 of 49 AI projects run in production, several on platforms that have operated for more than two years next to a bank’s core systems. Keep the AI reading from a stable data layer rather than from the core directly, so a migration changes what sits behind that layer without changing what the model consumes. Treat the migration as an ICT change under DORA and check whether it counts as a substantial modification under EU AI Act Article 43, because that work has to happen before cutover, not after.
A core banking migration moves the system of record. The data your AI depends on changes address, format, or refresh cadence, sometimes all three on the same weekend. An AI system that queries the core directly breaks the moment the core changes. One that reads from a decoupled layer, a canonical data model, an event stream, or a feature store, keeps answering while the plumbing behind it is rebuilt.
Ableneo has delivered 49 AI projects for 26 clients across finance, energy, and telco, with 35 of them running in production. Two of the longest running are banking AI platforms. Česká spořitelna’s AI platform has operated for about two and a half years, and Erste Group’s George assistant has run for more than three years with a team of 21 people. Platforms that survive that long share one trait: the AI is not wired straight into the core, so the core can be rebuilt underneath it.
What to put in place before the cutover:
In a bank, a core migration is never only an engineering event. Under DORA, in force since January 2025, a change to systems that support critical functions runs through the ICT change management process set out in Article 9 and the detailed procedures in Commission Delegated Regulation (EU) 2024/1774. The migration and its effect on any AI system have to be recorded, tested, approved by the right level of management, and verifiable through change logs and deployment history. Supervisors expect that evidence, and an AI system that fails silently after a cutover is exactly the kind of uncontrolled change the rules exist to catch.
The EU AI Act adds a second test. If the AI is a high risk system, for example credit scoring under Annex III, feeding it from a migrated core can count as a substantial modification under Article 43(4), which retriggers a conformity assessment unless the change was predetermined and documented at the first assessment. High risk standalone obligations apply from 2 December 2027, so a 2027 core migration and the AI Act deadline can land in the same year. Mapping both before the cutover keeps one project from quietly reopening the other.
35 of Ableneo’s 49 AI projects run in production, several on platforms live for more than two years alongside a bank’s core.
The data contract breaks first. A migration rarely keeps every field name, type, and refresh rhythm identical. A model trained on the old schema starts receiving nulls, renamed columns, or values on a different scale, and its output drifts with no error message to flag it. The second thing to break is latency. A new core that answers in 400 milliseconds instead of 40 can push a real time decision past its service window. Catch both by validating the AI against the new core’s data during a dual run window, not on go live weekend when there is no time left to react.
Run both in parallel wherever the business case allows it. Pausing a production AI system removes a control that staff and customers now depend on, and restarting it cold after a migration hides whether the model still behaves. A parallel run, where the old and new cores both feed the decoupled data layer and the AI scores against each, shows you the delta before you commit. Pause only the write path, the point where the AI acts on the core, and keep the read and scoring path live so you can measure. Ableneo’s promise of a first agent in 100 days rests on this discipline: ship behind a layer you can test, not straight onto the system of record.
Often, yes. Under DORA it is an ICT change to a critical system, so the change management controls in Article 9 apply from the start of planning. Under the EU AI Act, a migration that changes the data or logic a high risk system relies on can be a substantial modification under Article 43(4) and force a fresh conformity assessment, unless you declared the possible change in the original technical documentation. The practical move is to document the migration’s effect on the AI during planning. You then either stay inside the predetermined changes already assessed, or you schedule the reassessment as part of the project rather than discovering it in an audit after go live.
Put a layer between the AI and the core, and never let the model reach past it. A canonical data model, an event stream that publishes core changes, and a feature store together mean the AI depends on stable contracts, not on the core’s internal tables. When the core changes, you remap one layer instead of every model. This is the difference between a migration that costs the AI team a weekend and one that costs them a quarter. Build the layer on the first AI project, not the third, because retrofitting it once several models are wired into the core is the expensive path.
One named owner for the whole migration, accountable from planning to the post cutover review. A pilot or platform borrowed from the core team for one weekend is the most common way AI breaks during a migration, because no one is watching the model while everyone watches the core. The owner holds the data contract, signs off the parallel run results, and owns the rollback decision. In a regulated institution this role also carries the DORA change record and the EU AI Act modification call, so it sits with someone who can speak to both the engineering and the compliance side rather than one of them.
Ableneo builds AI as a system that outlives the infrastructure under it. Across 49 projects for 26 clients, with 35 in production and a deployment that passed an ECB vendor security assessment, the pattern that keeps AI running through core change is the same every time: a stable data layer, a named owner, and the compliance work mapped before the cutover. You can see how this fits the wider path from pilot to production in our AI transformation FAQ. Intelligence is only an asset when it keeps working the morning after the core changes.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.