Best Practices Guide

    What a Bank CIO Should Present to the Board on Enterprise AI Standardization

    For board purposes, enterprise AI standardization means a shared inventory of AI systems, a consistent policy baseline, policy enforced at runtime, and durable evidence of what each system was allowed to do. Present it around three questions: what AI is in use, what it is allowed to do, and what can be proven afterward. Tool consolidation is optional. Common control and evidence are not.

    Questions directors should ask

    • Can management produce one inventory of models, agents, tools and coding assistants, and how is completeness checked?
    • Which policy applies to each material system, and who owns it?
    • Is policy enforced while the system runs, or reviewed after the fact?
    • What was a given AI system permitted to do, and what evidence shows it?
    • How do the records map to our model risk management framework and to SR 26-2, where it applies?
    • What does the baseline not yet cover, and what is the plan?

    What standardization means for a board

    Standardization is often presented as consolidating onto fewer AI vendors. That is an architecture decision, and it is not what directors are asking about. When a board asks whether AI use is consistent, controlled and defensible to examiners, it is asking whether the same rules and the same records apply wherever AI is used.

    Business units that adopt models, agents, tools and coding assistants independently usually end up with separate policies, separate approval paths and separate records. The CIO then cannot say which systems exist, which policy applies to each, or what evidence is available. Tool diversity is manageable. Fragmented control and evidence is the problem to report on.

    A board-level definition therefore has four parts. Each can be demonstrated, and each maps to a question a director can ask.

    • Shared AI inventory: A single record of every model, agent, tool and coding assistant in use, with an owner and a materiality tier.
    • Consistent policy baseline: One set of rules covering data use, access and permitted actions, with documented tiering for exceptions.
    • Runtime enforcement: Policy applied while the AI operates, so permitted behavior is a property of the system and not a statement of intent.
    • Durable evidence: Records of what was permitted, blocked or escalated, retained in a form an examiner can review.

    Supervisory context as of November 2026

    The Federal Reserve has issued SR 26-2, revised guidance on model risk management. It is expected to be most relevant to banking organizations regulated by the Federal Reserve with over $30 billion in total assets, so smaller institutions should confirm how it applies to them. The guidance emphasizes sound principles while recognizing that practices vary with a bank's risk profile and model usage, and it ties risk management expectations to clear policies and effective controls. That is the frame for the board report: policy and control, demonstrably applied.

    On 4 November 2026 the OCC issued Bulletin 2026-54, a supervisory framework to help examiners evaluate banks' use of what it calls "super intelligence" in normal-course business. The OCC states that it sets no new regulatory expectations or enforceable standards, that banks are not expected to use it to assess their own use, and that examiners should tailor review to the materiality of the use case. It also says this use is not a separate risk category and does not require a separate risk appetite.

    Three practical points follow. Do not present the OCC framework to the board as a new compliance standard. Report within existing risk categories. Tier inventory and reporting depth by materiality, as examiners are told to do.

    Industry comment letters from 2021 describe model risk management as covering development, implementation and use controls, sound validation, and governance. They are industry views, not requirements, but they support starting from the existing framework. They also point to three challenge areas that board materials for material use cases should be ready to address: explainability, data usage and dynamic updating.

    Metrics and artifacts to present

    Board reporting differs from operational reporting. Operations needs per-event detail. The board needs coverage, trend, exceptions and open gaps, with the detail available underneath on request. Six artifacts make a workable pack. The specific metrics are a recommendation to adapt to your risk committee's format, not a published standard.

    • Inventory coverage: Number of AI systems by type and materiality tier, and the share confirmed as inventoried versus discovered outside approval paths.
    • Policy mapping: For each material system, the policy that applies and the date it was last reviewed.
    • Enforcement outcomes: Counts and trends of allow, redact, block and escalate decisions, by tier and by system type.
    • Escalations and exceptions: Items sent for human approval, time to resolution, and approved deviations from the baseline.
    • Evidence coverage: Share of material systems producing retained evidence, with retention period and storage location.
    • Open gaps: Named areas the baseline does not yet cover, with owners and target dates.

    Judging whether the evidence would satisfy an examiner

    The central distinction is when policy meets the AI system. Review after the fact depends on logs and recollection assembled once a question is asked. Enforcement in the execution path decides each interaction or tool call against policy as it happens, and the decision itself becomes the record. The second approach gives a board a stronger statement: not that policy existed, but that it was applied and what resulted.

    Several technical criteria help test any standardized runtime policy layer.

    First, policy actions. It should support the outcomes the bank needs, such as allow, redact, block and escalation for human approval, across coding assistants and agent tool access.

    Second, identity and permission. NIST's work on software and AI agent identity and authorization is still a concept paper, but its premise is useful: you cannot apply appropriate controls to agents unless you can state what each agent or tool is permitted to reach. Coding assistants and enterprise copilots also connect to different data, such as codebases versus email and files, so the baseline should cover both.

    Third, evidence properties. Ask what is retained, for how long, where it is stored, and whether retention and data residency are configurable. Ask whether it can run self-managed in the bank's own cloud. NIST's draft generative AI profile also points to provenance of outputs as a theme in what can be reconstructed later.

    Fourth, mapping. Records should be traceable to the bank's model risk management framework, so reviewers do not need a separate translation exercise.

    What standardization does not yet cover, and where Trussed fits

    Be explicit with the board about limits. A common policy and evidence baseline does not by itself provide model lineage or dataset provenance. It does not provide end-to-end tracing of an agentic workflow across tasks and agents. Enforcement at runtime operates on individual interactions and tool calls. Validation, explainability and ongoing performance monitoring remain part of the model risk management framework. Listing these as open gaps in the board pack is more credible than implying full coverage.

    Trussed is one way to enforce policy at runtime and generate audit-ready evidence across many models. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with actions that include allow, redact, block, rate limit and escalate for human approval. Enforcement produces continuous audit-ready evidence, mapped to HIPAA, NAIC, NYDFS, EU AI Act and NIST. It can be deployed self-managed in the customer's cloud, with configurable retention and data residency.

    Production proof includes 18+ months of continuous operation in a regulated enterprise, 99.95% uptime on live production traffic, 100+ models under one policy plane and 30+ GB of audit evidence generated per day. Insurance is the regulated financial sector where Trussed is deepest. Banks should assess it against their own examiner and model risk requirements, using the criteria above.

    The three board questions

    What AI is in use

    One inventory of models, agents, tools, MCP servers and coding assistants, tiered by materiality.

    What is it allowed to do

    One policy baseline, applied to each system and enforced in the execution path.

    What can be proven afterward

    Durable, audit-ready evidence generated by enforcement, not reconstructed later.

    Build the board baseline on control and evidence

    See how policy enforced in the execution path produces audit-ready evidence across models, agents, tools and coding assistants.

    See How Evidence Is Produced