How Do You Run an EU AI Act Readiness Assessment in a Bank?

7 min readAbleneo AI transformation team

Short answer. An EU AI Act readiness assessment answers one question in 4 to 6 weeks: which of your AI systems the Act treats as high-risk, and what each one is missing against Articles 9 to 15. It runs in five steps, inventory, classification, gap analysis, remediation plan, evidence pack. The hardest step is the first one, because most banks cannot produce a complete list of the AI and automated decision systems already running. Credit scoring and creditworthiness assessment are named high-risk uses under Annex III, so a bank that lends is almost certainly in scope.

1What This Means in Practice

A readiness assessment is a diagnostic, not a compliance project. Its job is to tell the board two things: where the bank stands today, and what the first three remediation decisions should be. It does not fix anything. It produces the map that makes the fix affordable, because the alternative is treating every model as high-risk and spending control budget where it is not required.

The work splits into five steps that run in sequence. Each one feeds the next, and skipping the first breaks the rest.

The single most common failure is an incomplete inventory. Cross-industry readiness research in 2026 found that the absence of a systematic AI system inventory is the largest single compliance gap, and that more than half of organizations still have no reliable list. A manual survey captures only what staff remember to report, which means shadow AI, the models and tools adopted outside IT governance, stays invisible until an audit finds it.

2Why This Matters for Regulated Industries

For a bank, the EU AI Act does not arrive on a clean desk. It lands on top of DORA, the GDPR, and existing model risk management rules, and the obligations overlap. The documentation and risk management a bank already maintains for DORA operational resilience and for internal model governance covers part of what Article 9 and Annex IV require. A readiness assessment that ignores this duplicates work and inflates the remediation bill.

The stakes are concrete. Annex III point 5(b) names AI used to evaluate creditworthiness or establish a credit score of natural persons as high-risk, alongside risk assessment and pricing in life and health insurance. A bank that runs credit scoring, transaction monitoring that feeds customer decisions, or AI-assisted onboarding is holding high-risk systems whether or not it has classified them as such. The common error runs the other way: banks underclassify, because a tool described as a “customer insight engine” turns out to drive a lending decision. The readiness assessment exists to catch exactly that gap between what a system is called and what it does.

A readiness assessment runs in five steps and produces a defensible inventory, classification, and gap map in 4 to 6 weeks.

3What Goes Into an AI System Inventory at a Bank?

Each entry needs enough detail to classify and govern the system, not just name it. At a minimum: the system owner and business unit, the vendor or whether it was built in-house, the data it consumes, the decision or output it produces, whether a human reviews that output before it takes effect, and the deployment status. The human-review field matters most for classification, because an output that influences a credit, insurance, or pricing decision without meaningful human review is what pulls an internal model into the high-risk tier.

The inventory has to reach beyond the model registry. Scoring logic embedded in a core banking system, a vendor fraud tool, a spreadsheet macro that ranks applicants, and a chatbot that answers in the bank’s name are all in scope. Automated discovery across the estate finds more than a survey does, because it does not depend on anyone volunteering a system they would rather not declare.

4How Do You Classify Which AI Systems Are High-Risk?

Classification runs each inventoried system through two gates. First, Article 5: is the use prohibited, for example social scoring or certain biometric categorization. Second, Annex III: does the use fall into a named high-risk area. For banks, the relevant entry is point 5(b), creditworthiness and credit scoring, with a carve-out for systems used only to detect financial fraud. Insurance risk assessment and pricing for life and health cover sits in the same point.

Two traps catch banks here. A system bought from a vendor still makes the bank a deployer with its own obligations under Article 26, so “the vendor handles compliance” is not a complete answer. And an internal risk model whose output shapes a lending or pricing decision without a human in the loop is caught even though no customer ever sees it. Classification is a documented judgment, recorded with its reasoning, because a supervisor will ask why a given system was placed where it was.

5What Does the Gap Analysis Actually Check?

For every system classified as high-risk, the gap analysis checks it against the provider and deployer obligations in the Act. The core set is a risk management system maintained across the lifecycle (Article 9), data governance for the training and input data (Article 10), technical documentation to the standard of Annex IV (Article 11), automatic logging (Article 12), transparency and instructions for use (Article 13), human oversight designed into operation (Article 14), and accuracy, robustness, and cybersecurity (Article 15). Post-market monitoring under Article 72 closes the loop once the system is live.

The output is a line per obligation per system: met, partial, or missing, with the evidence location. This is where the DORA and model risk overlap pays off, because a bank that already logs, monitors, and documents its internal models for prudential purposes can map that evidence straight onto several of these articles rather than building new controls.

6How Long Does an EU AI Act Readiness Assessment Take?

For a mid-size bank, 4 to 6 weeks of focused work produces a defensible inventory, classification, and gap map. The inventory takes the most calendar time, because it depends on reaching every business unit and every vendor relationship. Remediation is a separate program that runs for months after the assessment, and its length depends on how many high-risk systems carry real gaps. The assessment is the short, cheap step that tells the bank how long and how expensive the long step will be, which is why it comes first and why starting it now matters: the high-risk obligations for systems such as credit scoring are the longest-lead tranche of the Act, phasing in across the 2026 to 2027 window, and an Annex III remediation program runs 9 to 14 months end to end.

7The Ableneo Perspective

A readiness assessment is only useful if it ends in systems that reach production and stay compliant after they ship. Ableneo has delivered 34 production AI projects, with roughly 4 of 5 reaching production, and that record comes from treating governance, logging, and human oversight as build requirements rather than paperwork added at the end. The same discipline that gets a model live is the discipline the Act now requires: documented decisions, observable systems, and a named owner for each one. For the step that follows the assessment, see how a bank operationalizes the EU AI Act for credit scoring.

Key takeaways

Sources

Planning AI in a regulated business? Ableneo takes systems from classification to governed production.

Talk to Ableneo