AI is reshaping personalised services and customer journeys across ATMs, branches and digital channels. Therefore, banks need to be prepared to manage AI through existing operational and technology risk frameworks, with clear oversight, appropriate controls, human involvement and evidence that those controls are working.

Banks need visibility of where AI is used and what it affects

To start, banks require a complete inventory that records where AI touches core processes, what data it depends on and who is accountable for it. Banks must distinguish whether AI is used to inform, recommend or act because risk rises sharply when outputs influence decisions without traceability or human review. A centralised map lets banks prioritise controls for systems that influence lending, account actions or compliance workflows, and it provides the foundation that regulators and auditors will expect.

AI risk depends on the use case

Given that the EU AI Act uses a risk-based approach, banks should treat use cases differently. For example, AI systems used to assess creditworthiness or credit scoring for individuals may be classified as high risk and therefore require stringent controls. By contrast, fraud detection and transaction monitoring need different protections that focus on false positives, data integrity and escalation. For each use case, banks should define acceptable error rates, explainability requirements and testing regimes, including adversarial and data poisoning scenarios.They should also define access limits and remediation paths so business and risk teams can measure and mitigate impact consistently.

Traceability, oversight and accountability

As governance becomes more formal, robust AI oversight rests on four interlocking elements. First, end-to-end traceability of data, code, prompts, outputs and decisions. Second, defined human review and escalation points for high-impact outputs. Third, cybersecurity measures across models, APIs and pipelines. Fourth, named individuals accountable for risk management. Traceability must enable reconstruction of how a decision was reached, and human oversight must be operational with people empowered to review, override or remediate. Accountability must also be clearly documented so governance is auditable and enforceable.

Treating AI as a cyber-critical asset

An AI system is more than a model. It includes data, prompts, connectors, APIs, credentials, logs, pipelines and third-party components, and each layer increases attack surface and resilience needs. Data poisoning, exposed prompts, excessive API privileges or supplier dependencies can turn a model flaw into operational, regulatory and reputational harm, and inadequate logging can prevent reconstruction of incidents. Therefore, cybersecurity must be embedded in AI lifecycle management with vulnerability scanning, access controls, secure development, logging and incident playbooks to prevent, detect and respond to compromises that affect critical banking processes.

Aligning with DORA and third-party risk

Where an AI system is used within a financial institution, particularly where it supports an important or critical function, it may fall under both AI Act requirements and institution’s DORA risk-management framework. DORA addresses digital operational resilience and ICT risk while the AI Act focuses on trustworthy risk-based AI. Consequently, banks should treat critical AI Systems as ICT assets in the technology inventory and subject them to resilience testing, incident management and business continuity planning, including regulatory reporting where applicable. Supplier due diligence must also cover model governance, data provenance and security, and AI leads should work with enterprise risk, security and procurement to enforce model lifecycle controls through existing critical technology processes.

Proving AI control with auditable evidence

Regulatory compliance cannot be achieved by simply drafting policies. Crucially, this evidence should not be assembled only when an audit or regulatory request arises, it should be generated continuously as part of the system operation and lifecycle management. The real challenge will be in demonstrating that the bank remains in full control of the AI systems it deploys. This requires an up-to-date AI inventory, risk classifications and supporting documentation including decision traceability, test results, access and incident records, business continuity plans and third-party supplier assessments. The essential test is simple. If an AI system fails, produces an incorrect recommendation, behaves unexpectedly or is compromised, is the bank able to reconstruct exactly what happened, explain why it happened and respond quickly enough to correct it? If the answer is no, then the system is not truly under control.

Operationalising AI governance

2 August 2026 should be seen less as a deadline and more as a maturity checkpoint. The firms that gain advantage will be those that convert AI governance into repeatable measurable practices. That means integrating traceability, human oversight, cybersecurity and DORA-aligned resilience so AI is managed through the same risk frameworks as other critical technology.

Consequently, preparing for the EU AI Act and maintaining readiness will enable banks to preserve trust, limit operational and reputational exposure and recover rapidly from failures or false positives. Failure to prepare may increase regulatory exposure and business risk.

Néstor Santolaya, Cybersecurity Product Expert at Auriga