ECB raises the bar on AI inventory and audit readiness for banks

At a glance

What this is: This is a Saidot analysis of how ECB supervision is pushing AI governance in banks from policy statements toward complete inventories, accountable ownership, and audit-ready control evidence.

Why it matters: It matters because IAM, risk, and audit teams now need governance structures that can trace each AI system from registration through review, especially where third-party models and embedded AI expand the control surface.

By the numbers:

• In November 2025, the ECB moved AI-related strategies, governance, and risk management from Priority 3 to Priority 2 in its 2026–2028 supervisory cycle.
• The ECB reported that AI use cases, particularly in credit scoring and fraud detection, grew significantly between 2023 and 2024.

Context

AI governance becomes operationally meaningful only when an organisation can identify every system in scope, assign accountability, and show evidence of ongoing review. In European banking, that means inventories, risk classification, and audit trails are now part of supervisory readiness rather than optional governance decoration.

The ECB's shift reflects a wider control problem across regulated enterprises: AI adoption is moving faster than governance structures, especially where third-party models, embedded tools, and generative AI blur ownership boundaries. For identity and access teams, the lesson is that governance must extend beyond policy to lifecycle control and traceable accountability.

Key questions

Q: How should banks govern AI in corporate lending without losing auditability?

A: Banks should place AI inside a controlled workflow rather than outside it. The AI layer can interpret data, flag anomalies, and draft recommendations, but the process layer must still define roles, approvals, and audit trails. That keeps lending decisions explainable, preserves accountability, and prevents automation from eroding regulatory control.

Q: Why do AI inventories fail when organisations track only models?

A: AI inventories fail when they track only models because the real governance object is the system, not a single algorithm. A use case may also include prompts, datasets, integrations, vendor products, and cloud services, and each can change risk, accountability, and control requirements. Without that context, oversight becomes incomplete and inspection evidence breaks down.

Q: What are the signs that AI governance evidence is too weak for an audit?

A: Weak evidence usually shows up when teams can only explain intent, not activity. If they cannot produce records showing what was allowed, blocked, redacted, or held, the control is not generating audit-grade proof. Another warning sign is when answering a simple time-bound question requires manual reconstruction instead of querying an existing trail.

Q: Who should own AI risk accountability in a regulated bank?

A: Board and executive leadership should own AI risk accountability, with clear operational responsibility assigned across data, security, and AI teams. Regulators are no longer looking for informal consensus or departmental policy statements. They want a traceable decision model, defined approvals, and evidence that the institution can produce on demand.

Technical breakdown

Why AI inventories fail when they stop at model registers

A model register records a model, but not the full AI system around it. In practice, an AI use case can include the model, datasets, prompts, integrations, cloud services, and vendor products that shape behaviour and risk. When banks inventory only statistical models, they miss the operational dependencies that supervisors care about. ECB scrutiny pushes organisations toward a system-level inventory because governance, classification, and review need to apply to the whole service, not just one component.

Practical implication: build the inventory around the AI system and its dependencies, not around isolated models.

How the Three Lines Model changes AI governance accountability

The Three Lines Model makes AI governance a shared control structure rather than a development-only task. The first line owns the system, its intended use, and the control evidence. The second line sets policy, risk appetite, and challenge, while the third line tests whether the inventory, classification, and review process actually hold up. This matters because AI risk often sits at the intersection of business ownership, technical implementation, and oversight. If any one line lacks capability, the governance chain breaks.

Practical implication: assign explicit first-, second-, and third-line responsibilities for each AI system and document the handoffs.

What audit-ready AI governance looks like in practice

Audit-ready governance means a reviewer can trace each AI system from registration through risk assessment, control assignment, review, and sign-off without relying on verbal explanation. That requires evidence that is current, linked, and consistent across the lifecycle. In a bank, the control failure is rarely absence of policy alone; it is fragmented documentation, stale ownership records, and disconnected spreadsheets that prevent independent challenge. The ECB's focus on on-site inspection readiness makes that documentation chain part of the control itself.

Practical implication: centralise evidence, lifecycle status, and review history so inspectors can follow one coherent governance trail.

NHI Mgmt Group analysis

AI governance debt is now a supervisory issue, not a documentation problem. The ECB's direction shows that policy-only programmes are no longer enough when AI adoption is expanding across business units and third-party tools. Banks need governance structures that prove ownership, classification, and review at system level. The practitioner conclusion is clear: treat AI governance like a live control environment, not a static policy library.

AI inventory completeness is the new control boundary. Once AI is embedded in vendor products, cloud services, and generative workflows, a model-only register cannot answer who owns the risk or where the dependencies sit. That creates a governance blind spot that internal audit and supervisors can exploit during review. The practitioner conclusion is to inventory AI as a service chain, not as a loose list of models.

Audit readiness depends on evidence integrity across the lifecycle. ECB-style inspection questions are difficult to answer when risk assessments, control assignments, and review outcomes live in separate documents. AI governance trail: the real control is the ability to reconstruct decision history from registration to re-evaluation. The practitioner conclusion is to design the governance process so evidence is produced as part of operations, not reconstructed after the fact.

The identity dimension of AI governance is becoming more explicit. As AI systems take on clearer operational roles, ownership, access boundaries, and delegated authority become governance issues, not just technical settings. That creates a direct bridge between AI governance and identity controls, especially where humans, models, agents, and vendors interact in the same workflow. The practitioner conclusion is to align AI oversight with IAM and accountability structures, not leave it isolated in model-risk teams.

Supervisory scrutiny is converging with operational resilience. The ECB placing AI governance beside operational resilience signals that resilience, traceability, and control evidence are becoming inseparable. This affects how banks design oversight for high-impact use cases such as credit scoring and fraud detection. The practitioner conclusion is to evaluate AI programmes using the same discipline applied to other regulated change, with explicit tolerance for review failure and control drift.

From our research library:

• 52% of respondents see AI security decision-making power shifting toward platform and infrastructure teams rather than the executive suite, according to the 2026 Infrastructure Identity Survey.
• Only 44% of organisations have implemented any policies to manage their AI agents, despite 92% agreeing that governing AI agents is critical to enterprise security, according to the 2026 Infrastructure Identity Survey.
• Read next: Ultimate Guide to NHIs — Lifecycle Processes for Managing NHIs

What this signals

AI governance debt: banks that rely on policy documents without a central inventory will struggle to prove control when supervisors ask for ownership, classification, and review history. That problem will intensify as embedded AI and third-party systems expand beyond the direct control of the model team.

A bank that cannot connect its AI use case to a named owner, a lifecycle stage, and a control record has already created an assurance gap. The practical response is to make governance evidence part of the operational workflow, not a separate after-the-fact exercise.

Only 5.7% of organisations have full visibility into their service accounts, according to the Ultimate Guide to NHIs. The same visibility gap now appears in AI governance when inventories cannot reconcile systems, owners, and delegated access across platforms.

For practitioners

• Centralise the AI system inventory Record every AI system, including embedded and third-party AI, with owner, lifecycle stage, risk class, intended use, and linked components.
• Map governance to the Three Lines Model Define first-, second-, and third-line responsibilities for registration, challenge, approval, monitoring, and independent audit of each AI use case.
• Attach evidence to each control decision Store risk assessments, control assignments, review outcomes, and sign-offs in one traceable record so the governance trail survives inspection.
• Reclassify AI systems on change, not only on launch Trigger re-review when model providers, data sources, prompts, or business use cases change, because governance drift is often introduced after deployment.
• Test internal audit capability before the supervisor does Validate that second- and third-line teams can challenge model assumptions, inspect control status, and reconstruct inventory history without system-owner help.

Key takeaways

• The ECB's supervisory stance shows that AI governance in banking is now judged on traceable control, not policy intent.
• Inventory completeness, ownership, and lifecycle evidence are the controls most likely to separate mature AI governance from paper compliance.
• Banks that cannot reconstruct AI decisions and sign-offs will struggle to satisfy on-site inspection and audit expectations.

Key terms

• AI System Inventory: A governed record of every AI system, its owner, purpose, data scope, and risk tier. In practice, it is the control register that lets security, compliance, and audit teams decide which obligations apply and which evidence must be retained.
• Three Lines Model: The Three Lines Model is a governance framework that separates operational ownership, risk oversight, and independent assurance. In AI governance auditing, security and IAM teams provide technical evidence, management and GRC own control mapping and remediation, and internal audit assesses whether the controls and the evidence are credible.
• Audit-Ready Compliance: Audit-ready compliance means maintaining evidence, controls, and decision records in a form that can be reviewed without last-minute reconstruction. For third-party risk, this includes clear due diligence, remediation tracking, and governance documentation that shows how vendor risks were assessed, approved, and monitored over time.
• Governance Trail: The connected record showing how a system was registered, assessed, controlled, reviewed, and approved across its lifecycle. In AI governance, the trail matters because supervisors assess whether the organisation can prove control continuity rather than simply describe its policy framework.

Deepen your knowledge

NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, IAM, identity lifecycle, and machine identity security. It helps practitioners connect governance evidence to the identity controls that underpin resilient access programmes.