Short answer. Czech banks have one hard high-risk date to plan for: 2 December 2027, when the EU AI Act obligations for credit scoring and creditworthiness systems apply after the Digital Omnibus delay. AI literacy and the prohibitions have applied since 2 February 2025, and transparency duties since 2 August 2026. The practical window is about 13 months, and the work is an operating model with named owners, not a document.
A Czech bank that uses AI for credit scoring, creditworthiness assessment or customer-facing chat has three clocks running. The first is already past: AI literacy (Article 4) and the prohibited practices (Article 5) apply since 2 February 2025. The second started on 2 August 2026, when transparency duties under Article 50 began to apply. The third is the one that drives budgets: high-risk obligations for Annex III systems, which include credit scoring of natural persons, now apply from 2 December 2027.
Those dates come from the Digital Omnibus, Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026. It moved the stand-alone high-risk date from 2 August 2026 and moved AI embedded in regulated products to 2 August 2028. Anyone still quoting the old August 2026 date is working from outdated material.
In a Czech bank, the first questions are concrete:
Five dates matter. 2 February 2025: literacy and prohibitions. 2 August 2026: Article 50 transparency, with a grace period to 2 December 2026 for watermarking of generative systems already on the market. 2 December 2027: Annex III high-risk systems, including creditworthiness evaluation and credit scoring (fraud detection is carved out). 2 August 2028: AI embedded in products under sectoral safety law, which rarely touches a bank.
Penalties scale with the breach. Prohibited practices carry fines up to EUR 35 million or 7% of worldwide turnover, and most other obligations up to EUR 15 million or 3%. Czech banks that are deployers of a vendor scoring model carry deployer duties under Article 26 even if they built nothing themselves.
The high-risk date for Czech bank credit scoring is 2 December 2027, set by Regulation (EU) 2026/1744.
Article 74(6) of the AI Act sends market surveillance of high-risk AI used by financial institutions to the authority that already supervises them. For a Czech bank, that points to the Czech National Bank (ČNB). The ČNB already expects careful AI use: its July 2023 supervisory benchmark derives AI expectations from existing banking law.
One point is not settled. The Czech implementing act, prepared by the Ministry of Industry and Trade, was still a draft in the sources checked for this article, and it names the Czech Telecommunication Office (ČTÚ) as central market surveillance authority with the ČNB as sectoral authority for finance. Treat the ČNB designation as expected, not yet confirmed in force, and track the bill in the Czech legislative portal.
They overlap on the same system, so a bank should run one control set. DORA (Regulation (EU) 2022/2554) has applied since 17 January 2025. It covers ICT risk, incident reporting and third-party risk, so every AI model in production and every vendor behind it already sits inside your DORA register and your incident process. A DORA audit asks who owns the system, who approves changes and how a failure is escalated. An AI Act serious incident should reach the same escalation path.
CRR3 (Regulation (EU) 2024/1623) has applied since 1 January 2025 and tightens the capital rules behind credit risk. Where a scoring model feeds capital or provisioning, model governance for prudential purposes and the AI Act high-risk file describe the same asset. One model inventory, one validation record and one monitoring log serve all three regimes. Banks that keep three separate files will pay for the same evidence three times.
Start with a single inventory of every AI system, tagged by Annex III relevance, owner and vendor. Classify each one, because the high-risk file only exists for systems that qualify. Then assign a named business owner and a named risk owner per system. Without those two names, no regulator conversation goes well.
Next, build the evidence once: data governance records, human oversight design, logging, accuracy monitoring, and for credit scoring deployers a fundamental rights impact assessment under Article 27. Run a dry audit in Q2 2027, six months before the deadline. Ableneo’s delivery data shows why this order works: about 4 of 5 of its 34 production AI projects in 2025 reached production.
Ask for production references in regulated banking, not slide decks. Ask which of their projects reached production and who signed off. Ask how their delivery produces Article 11 documentation, logging and oversight as a by-product of the build, and how they map those artefacts to DORA evidence.
Ask who stays accountable after go-live, and ask for fixed-scope outcomes in place of billable hours. A partner that cannot name a Czech or Slovak bank production deployment, or cannot explain the ČNB supervisory angle, is learning on your budget.
Czech banks answer to the ČNB, and many belong to EU banking groups, so group-level AI governance and the EU AI Act apply side by side. A local bank often has to meet the stricter of the two.
The Digital Omnibus bought 16 months for credit scoring. It did not change the work. Inventory, classification, ownership and evidence still have to be built, and the banks that start in 2027 will compete for the same validators and auditors.
Ableneo has shipped 34 production AI projects across Slovakia, the Czech Republic, Austria and the US, with about 80% reaching production, and works with banks and insurers in the region on exactly this operating-model question. Compliance holds up when governance is built into the delivery of the model, which is how Ableneo’s AI delivery practice approaches regulated work: named owners, monitored models and audit evidence that exists on launch day.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.