AI governance and the three disciplines it is mistaken for

Every regulated firm already runs three functions that look, from a distance, like AI governance. Each of them will volunteer to absorb it. Understanding precisely where each one stops is the first practical decision you make, because the gaps between them are where obligations go unowned.

AI ethics stops at the point of decision

An ethics programme produces principles: fairness, transparency, human agency, accountability. These are worth having and they are not obligations. Nobody is fined for breaching a principle, no auditor asks to see one implemented, and a principle cannot tell you whether a specific system needs a conformity assessment before 2 December 2027.

Governance starts where ethics ends. It converts a position into a decision right, a named owner, an artefact and a date. "We believe in human oversight" is ethics. "The claims fraud model is overseen by the named senior claims handler, who has authority to override any output, whose interventions are logged, and whose throughput is reviewed quarterly against the time available per decision" is governance. The second sentence is examinable; the first is not.

The failure mode is a firm that has published principles and believes it has a programme. Ask for the register, the classification reasoning and the evidence a mitigation was implemented. If those do not exist, what exists is a statement of intent.

Model risk management stops at models you validate

Banks have run model risk management for years, and the discipline is strong. The US interagency guidance issued on 17 April 2026 as SR 26-2, OCC 2026-13 and FDIC FIL-15-2026 superseded SR 11-7 and carried the core across intact: effective challenge, validation covering conceptual soundness, ongoing monitoring and outcomes analysis, a model inventory, and documentation sufficient for a knowledgeable third party to understand the model. It added proportionate tiering by materiality and an expectation that firms look at aggregate model risk across the estate rather than model by model.

That is a serious foundation and you should build on it rather than beside it. But its scope is models used for business decisions of material consequence, and its questions are about whether the model is fit for its purpose. It does not ask whether the people subject to the output were told, whether a visible label was applied to generated content, whether workers' representatives were informed before workplace use, or whether the system's intended purpose puts it in a regulated category at all. Those are not model risk questions. They are rights questions, and model validation has never been built to answer them.

Two structural mismatches follow. First, a summarisation feature switched on by a vendor update carries AI Act duties and would never reach a model risk committee, because nobody validates it as a model. Second, the model risk inventory is organised around models, while the obligations attach to systems and their uses. The same model in two uses can sit in two different tiers.

Data governance stops at the data

Data governance owns lineage, quality, classification and retention, and AI is entirely dependent on it. The EU AI Act's data governance requirements read like a data management standard: design choices, collection and origin, the original purpose where personal data is involved, preparation operations including annotation, labelling, cleaning, enrichment and aggregation, the assumptions the data encodes, its availability, quantity and suitability, examination for bias, and the measures taken to detect, prevent and mitigate it. Datasets must be relevant, sufficiently representative, to the best extent possible free of errors and complete, with appropriate statistical properties, and must reflect the geographical, contextual, behavioural or functional setting in which the system will be used.

Every one of those is a data governance capability. None of them tells you who may switch the system off, how an affected customer complains, or what happens fifteen days after a serious incident.

What is left over, and why it needs an owner

Strip out ethics, model risk and data governance and a residue remains that belongs to nobody by default: knowing what AI the firm actually runs including what it did not choose to adopt; deciding and recording which regime tier each system falls into; establishing whether the firm is a provider or a deployer for each one; disclosure to the people affected; oversight designed so it can actually be exercised; incident classification against statutory deadlines; and retention of the record after the system is switched off.

AI governance is the function that owns that residue and coordinates the three neighbours around it. It is a thin function done well, not a thick one duplicating work others already do. The most common structural error is the opposite: a new AI governance team rebuilding model validation from scratch while nobody notices that eleven AI features arrived by vendor update in the last quarter.

Worked example: two committees, one model

A retail bank builds a model scoring applicants for an unsecured personal loan. Model risk management validates it: conceptual soundness reviewed, outcomes analysis scheduled, documentation written to the standard a knowledgeable third party could follow, and the model tiered as material. Validation passes. The model goes live.

Nothing in that process asked whether creditworthiness and credit scoring appear in Annex III of the EU AI Act, which they do. Nothing asked whether the bank owes a fundamental rights impact assessment, which deployers of Annex III creditworthiness systems do owe. Nothing established who tells a declined applicant that an Annex III decision was made about them. Nothing set the clock for a serious incident report.

The validation report is excellent and answers a different question. The gap is not a failure of model risk management. It is the predictable consequence of assuming that a function scoped to model performance will produce rights compliance as a by-product.

As of 24 September 2026.