How to Standardize AI Controls Across Bank Subsidiaries and Non-US Entities
Standardize AI controls by splitting policy into a versioned group baseline and explicit entity and jurisdiction overlays, then enforce both in the execution path for each interaction and tool call. Record every decision in one evidence schema. Keep data residency, retention and regional regulatory mappings local. Standardize the policy model, the enforcement outcomes and the evidence format.
The layered policy model
- 1
1. Group baseline
A single, centrally versioned set of controls covering permitted use, data handling, tool access and human approval thresholds. Every entity inherits it.
- 2
2. Entity and jurisdiction overlays
Local restrictions, parameters and approval requirements expressed as explicit policy. An overlay is selected by the originating entity and jurisdiction of each interaction.
- 3
3. Runtime enforcement
Each model call, tool call or MCP server action is evaluated against the baseline plus the applicable overlay before it completes. Outcomes are allow, redact, block, rate limit or escalate for human approval.
- 4
4. Uniform evidence
Every decision produces a record in one schema, so records from different entities can be compared directly.
Criteria for evaluating runtime enforcement tooling
- Does enforcement happen before a model call, tool call or MCP server action completes, or only afterward?
- Can it apply one group baseline together with entity and jurisdiction overlays, and show which policy version and overlay applied to each decision?
- Which outcomes are supported (allow, redact, block, rate limit, human-approval escalation), and can they be set per entity, model, agent or tool?
- Can it be deployed in the bank's own cloud or region, with retention and evidence storage location configurable per entity?
- Is evidence emitted in a uniform schema across entities, exportable, and mappable to frameworks such as NIST AI RMF and the EU AI Act?
- Can overlays be restricted so they cannot weaken the baseline, with every deviation visible to group risk and internal audit?
Why divergence appears and what examiners look for
When each subsidiary interprets group policy on its own, local tooling, approval processes and evidence formats drift apart. The group policy document says one thing, entity practice says another, and runtime behavior is often a third. An examiner asking what an AI system was permitted to do, and what it did, then gets a different answer format in every entity.
The supervisory material points toward demonstrable operation, not document completeness. As of November 2026, the OCC's supervisory framework for bank use of 'Super Intelligence' (OCC Bulletin 2026-54, issued 4 November 2026) is written to support examiners. It states that it creates no new supervisory expectations or enforceable standards. It is still a useful indicator of how an examination may probe. Its 'Control' domain concerns controls that may prevent unauthorized, inappropriate or unsafe actions by an AI system or agent. Its 'Assurance' domain concerns evidence that those controls continue to operate effectively.
The Federal Reserve's SR 26-2 guidance on model risk management says each banking organization is responsible for adopting practices appropriate to its own risks. It also refers to internal audit's role in assessing whether policies are implemented accordingly. For a multi-entity bank, that means group policy needs demonstrable implementation in every entity. The OCC also says AI is not a separate risk category and needs no separate risk appetite, so these controls belong inside the existing risk and model risk frameworks, not a parallel structure.
What to standardize and what must stay local
Standardize the policy language, the set of enforcement outcomes, the risk classification method, the approval workflow design and the evidence schema. The OCC's classification factors are a practical starting point: potential consequences, transfer of agency and degree of human oversight. Applying them identically in every entity means a high-consequence action carries the same minimum control everywhere, and human-approval escalation is reserved for those cases.
Keep local the items that depend on where data sits and which regulator is involved: data residency, evidence storage location, retention periods and the mapping of local regulatory requirements to controls. The residency and retention rules for each entity need confirmation from legal counsel, because they vary by jurisdiction. As a design judgment, the policy decision point and the evidence store for a jurisdiction may need to run in that jurisdiction, with only schema-conformant summaries or exports moving to group.
Extraterritorial reach also matters. The EU AI Act is intended to apply to providers and deployers established in a third country where the output is intended to be used in the Union. A non-EU entity whose AI output is used in the EU may therefore need the EU-relevant overlay. The Commission's Digital Omnibus proposal of 19 November 2025 would set fixed dates of 2 December 2027 for stand-alone high-risk systems and 2 August 2028 for those embedded in products. Avoid hard-coding AI Act dates into overlays without confirming current law.
For sequencing, inventory AI use cases per entity, classify them, map each to baseline controls and overlays, and document the legal and regulatory requirements that drive each overlay. The NIST AI RMF Govern functions, such as Govern 1.1 on understanding, managing and documenting legal and regulatory requirements, give a shared structure for that mapping.
Evidence that can be compared across entities
Examiners and internal audit need to compare control behavior, not read nine different report formats. The AI Act's record-keeping and traceability expectations for high-risk systems, and the OCC's assurance emphasis on controls that continue to operate, both point to evidence produced as part of operation and retained for review.
A workable minimum record per decision includes the entity, the jurisdiction, the policy baseline version, the overlay that applied, the decision outcome, a timestamp, the model, agent or tool involved, and for escalations, who approved or rejected the action. These attributes are a design recommendation, not a prescribed regulatory schema. Their value is that every entity emits the same fields with the same meaning.
Two practical points follow. First, evidence generated by the enforcement point records what the control actually decided, while evidence reconstructed from application logs records what a system happened to capture. The first is easier to defend. Second, storage and retention should be configurable per entity so that records stay in the required region for the required period, while the schema stays identical. Significant decisions and periodic risk management reviews should also be documented as recurring processes, not one-time approvals.
Where Trussed fits
Trussed is built for regulated enterprises that must show a regulator or auditor what their AI was allowed to do and what it did. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with outcomes of allow, redact, block, rate limit and escalate for human approval. The scope is individual interactions and tool calls. Policy is enforced at runtime, not reviewed after the fact.
That enforcement generates continuous audit-ready evidence, mapped to frameworks including NIST and the EU AI Act. Trussed deploys self-managed in the customer's cloud, with configurable retention and data residency, which supports keeping evidence local while the policy model stays common.
On the production record, Trussed has run continuously for more than 18 months in a regulated enterprise, at 99.95% uptime on live production traffic, with 100+ models under one policy plane. Its customers are in insurance, pharma, healthcare and education, and insurance is where it is deepest. Its certifications and attestations include SOC 2 Type II, ISO 27001, HIPAA, GDPR, FERPA and NIST AI RMF.
A layered policy model for multi-entity AI controls
Group baseline
One versioned control set applied to every entity.
Entity and jurisdiction overlays
Explicit, reviewable local restrictions and parameters.
Runtime enforcement
Allow, redact, block, rate limit or escalate for human approval.
Uniform evidence
Comparable decision records across every entity.
Review your layered policy model with an expert
Walk through how a group baseline, local overlays, runtime enforcement and uniform evidence would apply to your entities.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.