Mandatory AI Controls Bank-Wide: What to Fix and What Business Lines Can Tailor Under the 2026 MRM Guidance
Mandate bank-wide the controls whose failure causes harm regardless of use case: data redaction rules, tool-call restrictions, human approval for high-impact actions and evidence retention. Let business lines tailor parameters that depend on risk tier and purpose, such as thresholds, approved tools and escalation routing, inside locked limits with recorded approvals. Enforce both in the execution path, so that evidence is produced by enforcement itself rather than assembled after the fact.
Evaluating a runtime enforcement approach
- Can one policy baseline be defined once and inherited by every business line?
- Can overrides be scoped to a business line, use case or tool, with locked parameters that cannot be changed locally?
- Is every override change approved, versioned and recorded with approver, rationale and expiry?
- Are allow, redact, block, rate limit and escalate decisions made in the execution path, not reviewed afterward?
- Is a record produced for each interaction and tool call showing the policy version and decision?
- Can evidence be exported for examiners and internal audit, mapped to NIST and NYDFS-related requirements, with defined retention and data residency?
What the 2026 guidance changes for this decision
On 17 April 2026, the OCC, the Federal Reserve Board and the FDIC issued updated interagency model risk management guidance. It clarifies MRM principles, sets out a risk-based approach, and rescinds prior MRM guidance and other issuances. It covers model risk and MRM overview, model development and use, validation and monitoring, governance and controls, and vendor and other third-party products. Programs built around earlier guidance should be reviewed against the current text.
The guidance states that practices appropriate for one banking organization may be inappropriate for another with a different risk profile or model purpose. The OCC describes MRM practices as risk-based, tailored and commensurate with size, complexity and extent of model use. The guidance is expected to be most useful for models supporting significant business lines and most relevant to organizations with over $30 billion in total assets. The agencies also plan to issue a request for information on model risk management that considers bank use of AI, including generative and agentic AI.
The practical implication is that supervisors expect tailoring, but the guidance does not hand a bank a list of mandatory AI controls. The split described below is an editorial framework for making that decision explicit and defensible. It is not a list prescribed by any regulator.
For institutions with EU exposure, the EU AI Act requires a documented and maintained risk management system for high-risk AI systems, and requires that such systems be designed so people can effectively oversee them during use. Whether a given use case is high-risk needs a separate legal determination. The Commission's proposed Digital Omnibus on AI would set 2 December 2027 for stand-alone high-risk systems and 2 August 2028 for those embedded in products.
Criteria for classifying a control as baseline or tailorable
Start from the factors the guidance itself uses: risk profile, model purpose, materiality, complexity and extent of use. One illustration from an industry comment letter to the Federal Reserve, not a regulator's position, is that customer-facing models such as consumer lending may be held to a higher external explainability standard than internal operational models such as document processing. Risk tier legitimately changes what a control must demonstrate.
Three questions help sort each control. First, does the control only work if it is uniform? A redaction rule that differs by team creates gaps that cannot be explained to an examiner. Second, does the correct setting depend on the use case? A review threshold for a marketing summarizer should not match one for a credit decision aid. Third, would variation change the bank's exposure if set wrongly? If so, the control belongs in the baseline, or the tailorable range must be bounded tightly.
A control is baseline when its purpose is to define what AI may never do or must always leave behind. It is tailorable when its purpose is to calibrate to context. Many controls have both parts: the existence of a rule is fixed, and the parameter inside it is tailorable.
Fixed baseline versus tailorable parameter, by control category
The table shows one reasonable split for common control categories. Each bank should adjust it to its own risk tiers.
- Data handling and redaction: Baseline: redaction rules for restricted data classes and the requirement that they apply on every interaction. Tailorable: additional data classes a business line handles, within the bank's classification scheme.
- Tool and MCP access: Baseline: restrictions on which tool calls are never permitted and the default-deny posture. Tailorable: the approved tool list for a use case, added through an approval step.
- Human approval for high-impact actions: Baseline: the requirement that high-impact actions escalate for human approval. Tailorable: escalation routing, approver groups and the definition of impact for a business line, within a tier.
- Rate limiting: Baseline: the presence of limits to contain runaway or abusive usage. Tailorable: limit values by application, team or workload.
- Evidence retention: Baseline: what is recorded for each decision and the minimum retention period. Tailorable: longer retention or residency settings where a business line's obligations require them.
- Code provenance for coding assistants: Baseline: origin, licence and security context captured for generated code. Tailorable: which repositories or teams require stricter review before merge.
Limits on tailoring: who changes what, with what approval
Tailoring weakens the baseline when overrides are informal, unbounded or invisible. Four rules keep it contained.
First, make the baseline owned centrally, typically by model risk and security leadership, with business-line input. Industry practice described in a comment letter to the Federal Reserve is to coordinate controls across business lines, including MRM and cybersecurity teams. Second, scope every override to a named business line, use case or tool, so it cannot silently widen. Third, treat direction of change differently: a business line can tighten a control on its own authority, while loosening a parameter needs approval from the baseline owner. Fourth, record every change with the requester, approver, rationale, effective date and expiry, so exceptions are reviewed rather than forgotten.
Place these under the governance and controls heading of the MRM program. Third-party products fall within the guidance's scope, so approved-tool lists should cover vendor AI tools and coding assistants as well as internal models.
Parameters that change over time also belong under version control. Bank practices described in industry comments include drift thresholds that trigger review and rollback to previously approved states. A policy layer that versions thresholds and approved settings makes it possible to show which configuration was in force on a given date.
Enforcing the split in the execution path
A written baseline and a tailoring procedure only matter if the decision happens where the AI acts. Review after the fact shows what occurred; it does not stop a prohibited tool call or an unredacted data exposure. Enforcement in the execution path evaluates each interaction and tool call against the effective policy, which is the enterprise baseline plus any override scoped to that context, and returns a decision: allow, redact, block, rate limit or escalate for human approval.
This architecture also produces the evidence. Each decision can record the policy version, whether the baseline or an override applied, the action taken and, for escalations, who approved. That gives internal audit and examiners a record of what the AI was allowed to do and what it did, for both fixed and tailored controls, and it supports the use-time monitoring that the AI lifecycle requires beyond pre-deployment validation.
Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, at the level of individual interactions and tool calls. Its audit-ready evidence is generated by enforcement and mapped to NIST, NYDFS, NAIC, HIPAA and the EU AI Act. It runs self-managed in the customer cloud with configurable retention and data residency. It has operated for 18+ months in continuous production in a regulated enterprise, with 100+ models under one policy plane. Insurance is the regulated financial sector where Trussed is deepest.
One policy baseline, scoped overrides, per-call evidence
Enterprise baseline
Fixed controls every business line inherits.
Scoped overrides
Tailored parameters limited to a business line, use case or tool.
Runtime enforcement
Allow, redact, block, rate limit or escalate on each interaction and tool call.
Audit-ready evidence
A record of the policy version and decision behind each action.
Test your baseline-and-override design against runtime enforcement
Walk through your control categories, tailoring limits and evidence needs with a Trussed expert.
Talk to an Expert