Short answer. Plan for 9 to 14 months of program work before a single credit-scoring model clears its conformity assessment. The EU AI Act classifies credit scoring as high-risk, and the Digital Omnibus moved the stand-alone deadline to 2 December 2027. Backward planning from that date means a bank starting in the fourth quarter of 2026 has close to the runway it needs. A bank that waits until 2027 runs out of time before the first model is ready.
Credit scoring sits in Annex III, point 5(b) of the EU AI Act: any system that evaluates a person’s creditworthiness or sets a credit score is high-risk, with a narrow carve-out for fraud detection. High-risk status is not a label a bank applies at the end. It is a set of obligations built into the model’s full lifecycle, from the data that trains it to the monitoring that watches it in production.
The work is wide, not deep in one place. A bank preparing one scoring model runs several tracks at once:
Ableneo’s delivery record shows why this timeline is measured in quarters, not weeks. In 2025, 4 of 5 projects Ableneo shipped reached production, across 34 production AI projects and 20 clients, and 94% of them used large language models. The distance between a model that works in a notebook and one that holds up under audit is exactly the lifecycle work the AI Act now makes mandatory.
For a bank in Central Europe, the credit-scoring deadline does not arrive alone. The European Central Bank asked significant banks to submit AI-related cyber action plans by 31 October 2026, so the governance muscle for AI oversight is already being tested this year. DORA has been in force since January 2025 and already requires a register of information and incident reporting for the systems a scoring model depends on.
The AI Act sits on top of that stack rather than beside it. The 2 December 2027 date set by the Digital Omnibus (Regulation (EU) 2026/1744) applies to stand-alone high-risk systems, which is where most credit-scoring engines fall. Systems embedded in a regulated product move separately, to 2 August 2028. A bank that already runs model risk management for its internal ratings has a head start, because the evidence a supervisor expects and the evidence a conformity assessment expects overlap heavily.
Plan for 9 to 14 months to get a first credit-scoring model through its conformity assessment.
No. The deadline is the date the model must already be compliant, not the date work begins. Work backward from 2 December 2027. A conformity assessment needs a complete technical file. That file needs finished data governance and risk documentation. Those need a model inventory and a classification decision before any of it. Each layer depends on the one beneath it, so they run partly in sequence, not all at once.
That dependency chain is why a realistic first pass takes 9 to 14 months for a bank that has not done this before. Starting in the fourth quarter of 2026 lands the first model with a few months to spare. Starting in mid-2027 does not. The deadline moved once already, from 2 August 2026, and a second delay is not something to plan a lending book around.
This single question changes who carries the heaviest work. A bank that builds its own scoring model, or materially changes a vendor’s model, is a provider and must run the Article 43 conformity assessment itself. A bank that uses a third-party model unchanged is a deployer, and its duty is to confirm the provider completed the assessment, then meet the lighter deployer obligations around human oversight and monitoring.
Many banks assume they are only deployers, then find their data science team retrained or re-weighted the vendor model, which makes the bank a provider for that model. Settle this per model early. A provider’s program runs months longer than a deployer’s, and the answer sets the size of everything that follows.
For credit scoring, the conformity assessment runs through internal control (Annex VI), not a notified body, so the bank assesses itself against the requirements in Articles 8 to 15. Before that assessment can pass, five things have to exist and hold together: a documented risk management system (Article 9), governed and bias-tested datasets (Article 10), the full technical file (Article 11 and Annex IV), human oversight a person can actually exercise (Article 14), and logging that supports traceability (Article 12).
Post-market monitoring under Article 72 then starts the day the model goes live. The order matters. You cannot document a control that does not yet operate, so the controls have to be built and running before the paperwork that evidences them is worth anything to an auditor.
The first quarter is about scope and ownership, not documentation. Build a complete inventory of every model that touches a credit or affordability decision, including internal risk models whose output feeds a lending decision. Classify each as high-risk or not, and record the reason. Name one accountable owner for the AI Act program, separate from the individual model owners.
Run a gap assessment against Articles 8 to 15 for the two or three models that carry the most lending volume. By day 90 a bank should know how many high-risk models it has, which are provider-built, and where the documentation gaps are. That inventory is what makes the remaining 6 to 11 months plannable instead of a scramble.
Data governance is the long pole. One Tier 2 bank’s data governance build took 10 calendar months on its own, covering data management processes, metadata, and attribute documentation. For a credit model the AI Act adds bias testing and representativeness checks on top of that. Legacy core systems are the second drag, because training data often sits in formats that were never built to be exported, documented, or audited.
The third is organizational. A scoring model usually crosses risk, data, compliance, and IT, and a program with no single owner stalls at every handoff. None of these is a pure technology problem. Each is a reason the clock should start in the fourth quarter of 2026, not the following year.
Ableneo builds AI systems in regulated Central European industries where the standard for done is a model running in production under audit, not a pilot in a sandbox. In 2025, 4 of 5 projects Ableneo delivered reached production across banking, insurance, and telecom, and the same lifecycle discipline that gets a model live is what an AI Act conformity assessment measures. A bank that treats the 2 December 2027 deadline as a 2027 problem loses the runway, because the work to find the gaps starts with a model inventory and a classification call. See Ableneo’s approach to AI transformation in regulated industries.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.