Enterprise Baseline AI Policy for Banks: Layering Business-Line Rules and Enforcing Them at Runtime
Treat the enterprise baseline as a floor that no business line can lower, and treat business-line rules as versioned overlays that may only tighten controls or add scoped ones. Define precedence so any deny or escalate decision prevails, route every relaxation through a time-bound exception, and evaluate each interaction and tool call against the effective policy at runtime. Record which layer and rule version produced each decision, so you can show an examiner what an interaction was permitted to do and what happened.
Ownership and the exception path
- 1
Request
The line of business names the use case, the control affected and the business rationale.
- 2
Compensating controls
The requester proposes controls that reduce the added risk, such as narrower scope or added human approval.
- 3
Approval
Enterprise risk approves or rejects, and the approver is recorded by name and role.
- 4
Expiry
Every exception carries an end date. No exception is open-ended.
- 5
Enforcement
The approved exception is encoded as a scoped, versioned rule with its own identifier.
- 6
Revocation and reporting
Exceptions can be revoked early, and counts and ageing are reported to enterprise risk.
Evaluation criteria for tooling that enforces layered policy
- Does it evaluate the baseline and the business-line overlay on each interaction and tool call, and can it reject an overlay rule that relaxes a baseline control?
- Is precedence explicit and visible to risk reviewers, including which rule won when layers disagreed?
- Can exceptions carry an approver and an expiry, and do they lapse automatically?
- Which actions are supported at runtime (allow, redact, block, rate limit, escalate for human approval), and across which models, agents, tools and MCP servers?
- Does each decision record the layer and rule version that applied?
- Can evidence be exported for examiners and internal audit, with configurable retention, and mapped to NIST and NYDFS?
Baseline as a floor, overlays as scoped additions
Most banks already have an enterprise AI policy. The difficulty starts when retail, commercial lending, wealth, markets and operations each need different rules for data classes, models, tools and approval thresholds. If those rules are written as separate documents, they drift from the baseline and from what AI systems actually do. Some quietly weaken the baseline. Others conflict with it and nobody decides which wins.
A layered model removes that ambiguity. The baseline holds the controls that apply everywhere and cannot be overridden. A business-line overlay can restrict a data class further, narrow the list of permitted models or tools, lower an approval threshold, or add a control the baseline lacks. It cannot remove or relax a baseline control. NIST's SP 800-53 overlay concept is a useful precedent: overlays sit on top of organization-wide policies and procedures and assume they already exist. The source concerns security controls, not bank business lines, so the analogy supports the structure and no more.
Scope each overlay explicitly by business line, data class, model, tool and approval threshold. Keep overlays versioned and stored separately from the baseline. Separation is what makes drift detectable, because any overlay rule can be compared mechanically against the baseline it sits on.
Precedence and conflict resolution
Write precedence down before writing the first overlay. A workable order is simple. The baseline sets the minimum. The overlay can only move a decision in a stricter direction. Where several rules apply to one interaction, the most restrictive outcome prevails: a deny or an escalation from any layer beats an allow from another. No published standard prescribes this order, so it is a design choice you should document and have approved.
Two practical rules prevent most conflicts. First, validate every overlay rule against the baseline at authoring time and reject any rule that would relax it. Second, inventory existing business-line rules now and compare each one with the baseline. Rules that weaken it should be removed or reclassified as exceptions, with the controls that come with an exception.
Tie overlay strictness to use-case risk. NIST AI RMF material says policies and resources should be prioritized by assessed risk level and potential impact, and OCC Bulletin 2026-54, an examiner-facing framework, tells examiners to tailor examinations to the materiality of the use case. A tiering scheme gives you a defensible reason for different approval thresholds across lines of business. Bulletin 2026-54 states that it does not establish supervisory expectations for banks, so use it to anticipate how examinations may be scoped, not as a requirement.
What to enforce at runtime, and what evidence results
A policy that exists only as a document cannot show what an AI system did. Enforcement has to happen on individual interactions and tool calls, using the effective policy: the baseline plus the applicable overlay plus any active exception. The action vocabulary below is design vocabulary rather than a standard taxonomy: allow, redact, block, rate limit, and escalate for human approval.
The categories that most often need business-line variation map to these actions directly. Data classification rules decide when to redact or block. Permitted model and tool lists decide when to allow or block. Approval thresholds decide when to escalate. Volume limits per business line or use case drive rate limiting. Each is decidable on a single request, which is why interaction-level enforcement is the right granularity.
Every decision should leave a record an examiner can follow. NIST's GenAI Profile draft calls for defining organizational access to generative AI and for compiling statistics on policy violations. Both depend on knowing which policy applied and what was decided. At minimum, retain the timestamp, the requesting identity and business line, the use-case tier, the model or tool involved, the data classes detected, the layers and rule versions evaluated, the exception identifier if one applied, the action taken, and the approver and outcome for any escalation. Retain these records under a defined schedule and map the fields to the framework requirements your reviewers use, so that internal audit can sample them without engineering help.
Where Trussed fits
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, at the level of individual interactions and tool calls. Decisions are enforced at runtime, not reviewed after the fact, and the actions are allow, redact, block, rate limit and escalate for human approval. That enforcement produces audit-ready evidence, mapped to NIST and NYDFS among other frameworks. Trussed runs as a self-managed deployment in the customer cloud, with configurable retention and data residency.
Trussed has run 100+ models under one policy plane, which is the property a layered model needs: one place where the baseline and overlays are evaluated rather than separate rule sets per tool. Insurance, a regulated financial sector, is where Trussed is deepest. The approach is framed here through examiner scrutiny and model risk management, and it does not imply bank deployments.
From layered policy to examiner evidence
Enterprise baseline
Non-negotiable floor owned by enterprise risk.
Business-line overlay
Tighter or additional scoped rules, versioned separately.
Approved exception
Time-bound, with approver, compensating controls and expiry.
Runtime decision
Allow, redact, block, rate limit or escalate per interaction and tool call.
Evidence record
Layer, rule version and action retained for review.
Test your layered policy against runtime enforcement
Walk through your baseline, an overlay and an exception with an expert, and see the evidence each decision produces.
Request a Demo