AI governance has become an infrastructure problem.

Generative AI, agents, third-party foundation models, retrieval systems and multi-model pipelines have turned the AI estate from a population of relatively stable models into an interconnected system of models, vendors, data, applications and autonomous actions. Governance functions are still frequently organized around a very different operating assumption: identify the model, assess it, approve it, validate it, document it, review it periodically.

That model was designed for a world of relatively stable statistical models. It is increasingly poorly matched to a world where AI systems change behavior, call external tools, consume continuously changing information and act on their own.

The Federal Reserve reported in April 2026 that roughly 18% of US firms had adopted AI by the end of 2025. A separate measure cited in the same work found 78% of the labor force working at firms that had adopted AI, and 54% at firms using large language models. The estimates differ because one is firm-weighted and the other employment-weighted — but they point the same way.

The financial system is seeing the same transition. The Financial Stability Board’s June 2026 consultation on sound practices documents institutions using AI across fraud detection, AML, customer service, risk management and operations. One large internationally active bank described an agentic system monitoring more than 80 million signals a day, contributing to developing or updating three quarters of its card fraud rules, and helping reduce fraud losses by over 20% in the first half of the 2026 financial year against the same period in 2025.

At that scale, governance cannot remain a periodic review exercise.

Traditional model governance asks whether a model performs acceptably. That becomes insufficient when the model is one component of a larger system. A generative application may combine retrieval, ranking, a foundation model, guardrails and human interaction. An agent adds something more consequential: authority to act. It can call tools, modify records, initiate workflows, or decide what to do next.

1. Is the model accurate?

   What traditional validation was built to answer.

2. Is the system’s behavior acceptable?

   Once retrieval, ranking, guardrails and a foundation model sit behind one interface.

3. Is the system authorized to take this action, under these circumstances, with this level of autonomy?

   Once the system can act. A control designed to review an output cannot govern a decision loop.

Institutions increasingly depend on models they did not build, cannot fully inspect and do not control. Have we performed sufficient due diligence? is no longer the hard question. These are:

• What happens when the vendor changes the model?
• What happens when its behavior changes without our application changing?
• Which internal systems depend on it?
• What happens if the provider becomes unavailable?
• How much of our AI estate depends on the same small group of providers?

The FSB and the BIS have both identified third-party dependency and concentration among major technology providers as material AI-related vulnerabilities — noting that providers of cloud infrastructure, hardware and foundation models are often highly concentrated, and that reliance on common models, datasets or infrastructure can produce correlated behavior across institutions. The governance object is therefore no longer the model. It is the dependency graph around the model.

The traditional operating model assumes the number of systems entering governance is manageable relative to the number of people reviewing them. That assumption breaks when adoption accelerates. When use-case volume grows faster than review capacity, governance does not simply become inefficient. It becomes a queue — and queues create predictable behavior.

1. Governance capacity is fixed
2. Review latency grows
3. Teams route around the queue
4. Visibility is lost
5. Controls degrade quietly

If the governed path takes weeks and the ungoverned path takes hours, teams under delivery pressure will eventually find another path. Not because they are irresponsible, but because they are responding rationally to the incentives the operating model created. Which is why the most serious governance failure may not be that governance approved the wrong system. It may be that governance never saw the system at all.

1. Inventory becomes infrastructure.

   Registration belongs in the path to production, not in a spreadsheet populated by asking teams what they have built. The inventory should know what the system does, who owns it, what data it uses, which models and vendors it depends on, what it can access, what versions are running, and what breaks if a dependency changes.

   The testCan you determine in minutes which systems depend on a particular AI provider, and what would be affected if that provider became unavailable tomorrow? If not, you have a register — not an operational inventory.

2. Classification becomes computation.

   Risk tier should be derived from system attributes — decisions influenced, data processed, people affected, autonomy, reachable systems, consequence of failure — rather than negotiated case by case. And when the system changes, the classification should change with it.

   The testThe governance committee sets the rules. It should not have to adjudicate every deployment. If your second line is a bottleneck, it is because it was made an operator rather than a rule-setter.

3. Monitoring becomes evidence.

   Point-in-time validation tells you the system was acceptable when someone reviewed it. Runtime evidence tells you whether it still is — guardrail activations, overrides, escalations, tool calls, version changes, behavioral drift, policy exceptions, material changes in operating context.

   The testThe objective is not to monitor everything equally. It is to generate continuous evidence where the consequence justifies it, automatically, as the system operates.

The next generation of AI governance will not be won by the institution with the longest policy. It will be won by the institution that can answer, continuously and with evidence: what AI do we have, what can it do, what does it depend on, what can go wrong, what controls are operating, and what changed?

That requires governance to move into the architecture of the enterprise. Registration connects to deployment. Classification connects to system attributes. Controls connect to runtime behavior. Evidence is generated automatically wherever possible, and human expertise is reserved for ambiguity, exceptions and consequential decisions rather than consumed by administrative throughput.

The objective is not to slow AI down. It is the opposite. The safest AI governance model is the one in which the governed path is the fastest path. That is the transition now underway: from governance as documentation to governance as infrastructure — and ultimately from governance as a gate to governance as the enterprise AI control plane.

Sources. Federal Reserve Board, FEDS Note, Monitoring AI Adoption in the U.S. Economy (3 April 2026), drawing on the Census Bureau BTOS and the Survey of Business Uncertainty; Financial Stability Board, Sound Practices for the Responsible Adoption of Artificial Intelligence, consultation report (June 2026); FSB, Monitoring Adoption of Artificial Intelligence and Related Vulnerabilities in the Financial Sector (October 2025); SR 26-2 and OCC 2026-13, Revised Guidance on Model Risk Management (17 April 2026). Full citations in the whitepaper.

Provided for general information; not legal advice. AI Risk Consulting is not a law firm. The FSB sound practices report is a consultation and may change before it is finalized.