AI Governance Works Better Inside Existing Business Controls

The Office of the Comptroller of the Currency issued a new AI supervisory framework on November 4 with a useful message for operating businesses: AI should not be managed as a detached category of risk. It should be governed through the business controls already responsible for the process, decision and consequence the system can affect.

That does not mean AI needs no special controls. It means the controls belong inside operations, finance, quality, security, compliance, commerce and vendor management—not in a policy document that sits beside the work.

What the OCC actually changed

The OCC’s framework is an examiner resource for evaluating AI use by national banks, federal savings associations and federal branches and agencies. The agency says the framework does not create new supervisory expectations, enforceable standards or prescriptive requirements.

Its central design principle is still consequential. Examiners may use AI-specific considerations to supplement existing model-risk, governance, technology, compliance and operational-risk practices, but should not treat AI as a separate risk category or require a separate AI risk appetite.

Sources: the OCC’s November 4 announcement, Bulletin 2026-54 and framework summary, Reuters’ independent reporting, and the ABA Banking Journal’s industry summary.

Documented facts, interpretation and boundaries

  • Documented: the OCC instructs examiners to tailor work to the materiality of the use case and the bank’s size, complexity and risk profile.
  • Documented: the framework says control intensity can increase with the consequences of failure or misuse.
  • Documented: risk classification may consider consequences, transfer of agency and human oversight rather than the technical sophistication of the model.
  • Not a new rule: the OCC expressly says the examiner guide is not legally binding and does not create new regulatory expectations.
  • Bridge Road inference: although the framework applies to OCC-supervised banks, the structure is useful for manufacturers, distributors and commerce operators because their existing process owners already manage financial, operational, safety, quality and customer risks.
  • Unresolved: the framework will be updated as technology and supervisory understanding mature. It does not prescribe one universal control set or prove that a particular AI system is safe.

Classify the business consequence, not the model label

Reuters highlighted an important example from the framework: a simple agent allowed to initiate transactions may be riskier than a technically sophisticated model used only for internal analysis. That distinction should shape every AI intake process.

Do not start with “Which model are we buying?” Start with:

  • Which business process changes?
  • What data can the system read or alter?
  • Can it recommend, decide, communicate, transact or control equipment?
  • Who or what receives the output?
  • What is the credible consequence of error, misuse or unavailability?
  • How much agency transfers from a person to software?
  • Can the action be reversed, reconciled or performed manually?

This puts AI governance in the same language the operating team already uses.

Give every use case a business control home

AI use case Primary control home Existing controls to extend
Product descriptions and campaigns Marketing and commerce Brand review, claim substantiation, pricing and publishing approval
Inventory, allocation or replenishment Operations and supply chain Planning tolerances, exception ownership, reconciliation and override rules
Invoice, refund or payment action Finance and commerce Segregation of duties, limits, beneficiary validation and ledger reconciliation
Quality inspection or disposition Quality and manufacturing Acceptance criteria, calibration, hold authority and traceability
System configuration or code change Technology and security Change control, testing, approval, deployment evidence and rollback
Customer or employee decision Process owner, legal and compliance Eligibility rules, notice, appeal, privacy and record retention

An AI committee can set standards and review exceptions. It should not replace the accountable owner of the underlying process.

Use a consequence-and-autonomy control ladder

Control intensity should rise when the potential consequence or delegated autonomy rises:

  1. Assist: AI drafts or summarizes; a person reviews before use.
  2. Recommend: AI proposes a decision; a named owner approves and records the rationale.
  3. Execute within limits: AI acts only inside deterministic thresholds, narrow permissions and monitored workflows.
  4. Execute material actions: dual approval, independent validation, stronger logging and immediate interruption are required.
  5. Control safety- or mission-critical work: validated engineering constraints, fail-safe behavior, formal change control, incident drills and tested manual operation are mandatory.

This ladder avoids both extremes: treating every chatbot like a production-control system or treating every agent like an ordinary software feature.

Add AI-specific controls to the existing process

Once the use case has a business home, perform a control-gap review. Common additions include:

  • Least privilege: a dedicated identity with only the data, tools and actions required for that workflow.
  • Deterministic validation: schemas, totals, business rules, safety limits and acceptance tests that the model cannot override.
  • Human approval: named approvers for consequential, exceptional or irreversible actions.
  • Auditability: the user, agent, model, prompt or instruction version, data sources, tool calls, approvals and outcomes.
  • Monitoring: quality, drift, exceptions, cost, latency and business-outcome signals—not merely application uptime.
  • Recovery: rollback, compensating action and a manual fallback tested before production.

The objective is not a parallel AI bureaucracy. It is a stronger version of the control system the business already depends on.

Keep ownership with the operating process

A use-case register should name the business owner, technical owner, data owner, approver, exception owner and incident commander. It should also record the process, systems, model or vendor, permissions, autonomy level, consequences, controls, evidence, fallback and next review date.

For operations, commerce and business systems, the accountable owner should be able to pause the workflow without waiting for the AI vendor. That person also owns unresolved exceptions and the decision to return to normal operation after an incident.

Review vendors through the risk they inherit

Vendor and model review should follow the business consequence. A copywriting assistant and an agent with ERP write access should not receive the same diligence merely because both use a large language model.

For higher-impact use cases, verify data handling, subprocessors, retention, model-change notice, identity and permission design, log export, testing evidence, incident notification, service continuity, rollback support and termination assistance. Map contract commitments to the existing vendor-risk and continuity programs.

Re-review when the vendor changes the model, tool set, data access, action authority or hosting arrangement. A new connector or write permission is a material control change even if the product name stays the same.

Trigger formal change review when agency changes

AI risk can change without a new model. Moving from draft to publish, read to write, one system to several systems, recommendation to execution or human approval to automatic approval changes the agency transferred to software.

Require a new review when any of those boundaries change. Re-run deterministic tests, update permissions, confirm monitoring, rehearse rollback and manual fallback, and obtain approval from the process owner. For industrial automation, include safety, quality and engineering owners before production authority expands.

If an incident occurs, use the established incident-response process, supplemented with model, prompt, agent and tool-call evidence.

The practical takeaway

The OCC’s framework does not tell every business how to govern AI, and it does not apply directly to most manufacturers, distributors or merchants. Its organizing principle is still sound: govern the consequence where the consequence lives.

Register the use case. Map it to the existing business risk. Name the process owner. Classify consequence, autonomy and human oversight. Extend the current controls with AI-specific permissions, validation, evidence, monitoring and recovery. Escalate the control intensity as the action becomes more consequential or autonomous.

Bridge Road helps operating businesses turn AI governance into an implementation-ready control map across commerce, systems, operations and automation. Contact Bridge Road to define the use-case register, ownership model, control gaps and production-readiness workflow.