Implementation Guide | Financial Services

    Standardizing AI Agent Tool Access Across Business Lines in Banks

    Define one tool catalog with risk tiers, one baseline policy that sets minimum controls, and business-line overlays that can only tighten or scope that baseline by risk tier, data sensitivity and action type. Enforce the result at runtime with allow, redact, block, rate limit and human escalation outcomes, and retain a record of what each agent was permitted to do and what it did.

    Shared baseline with business-line overlays

    1. 1

      Baseline

      Catalog fields, tiers, minimum controls and evidence requirements, owned centrally.

    2. 2

      Overlay

      Business-line scoping by tier, data sensitivity and action type; tightens only.

    3. 3

      Exception

      Named owner, stated reason, expiry date, recorded as a versioned change.

    Criteria for evaluating a runtime enforcement layer

    • Does it decide in the execution path, before the tool call or MCP request executes, rather than reviewing afterward?
    • Can one baseline carry business-line overlays by risk tier, data sensitivity and action type, with versioned changes attributable to an owner?
    • Are allow, redact, block, rate limit and human escalation configurable per tool and per tier?
    • Does it record what each agent was permitted to do and what it did, in a form exportable to examiners and model risk reviewers?
    • Does it use existing identity and credential systems, support distinct agent identities and scoped grants, and run in the bank's own cloud?

    What a tool access standard has to cover

    When retail, commercial, wealth, markets and operations each deploy agents on their own, each also decides how those agents reach internal APIs, data stores and MCP servers. Permissions, approval rules and logging then diverge. The result is inconsistent entitlements, unclear ownership of tool-call decisions and uneven evidence. Treat this as a policy design and enforcement problem, not a documentation exercise. A document describing intended access does not show what an agent was actually allowed to do.

    The OCC supervisory framework for bank examiners, as of November 2026, offers a useful structure. It is organized around five elements: purpose (the actions an agent is permitted to execute), risk (the consequences if an output or action is not fit for purpose), control (what prevents unauthorized, inappropriate or unsafe actions), oversight (who monitors the agent and who is accountable) and resilience (the ability to detect, contain, recover from and remediate failures). It also states that control intensity could increase with the potential consequences of an agent's failure or misuse.

    The framework notes that model risk management principles in OCC Bulletin 2026-13 apply to non-generative and non-agentic models, and describes the framework as more applicable to generative and agentic systems. Architects should confirm with their own supervisors how agentic tool access is reviewed. A tool access standard that maps cleanly to purpose, risk, control, oversight and resilience gives both examiners and model risk reviewers a familiar frame.

    Build the tool catalog and classify risk

    Begin with an inventory of agents, MCP servers, APIs and non-human identities in each business line. Then define catalog fields once, centrally, and require every tool to carry them before any business-line overlay is written. Ownership and tier come first because every later policy decision depends on them.

    Classify by effect, not by label. A read/write flag alone may not establish risk, and tool annotations supplied by a server should be treated as untrusted unless the server itself is trusted. Assess whether an operation involves a sensitive read, a write, data disclosure or an irreversible action. Inspect tool descriptions, parameter names, types and return schemas before approval, then pin the reviewed definition with a cryptographic hash so that any change triggers re-review.

    • Owner: A named person or team accountable for the tool and its tier.
    • Data sensitivity: The most sensitive data the tool can read or return.
    • Action type: Sensitive read, write, disclosure or irreversible action.
    • Risk tier: Assigned from the effect of the action, not the tool label.
    • Pinned definition: A hash of the reviewed tool definition, with re-review on change.

    Mapping enforcement outcomes to risk tiers

    Runtime enforcement should be able to return one of five outcomes for each tool call: allow, redact, block, rate limit or escalate for human approval. The mapping below is one reasonable design for a bank, not a regulatory prescription. Adjust tier boundaries to your own risk assessment.

    For escalation, never auto-approve, especially when several MCP servers are in play. Show the approver the full call parameters, not only the tool name. Authority should also attenuate through delegation: a sub-agent or downstream service never holds more authority than its parent.

    • Tier 1: read-only, low-sensitivity data: Allow within the agent's scoped grant. Apply rate limits to contain runaway loops.
    • Tier 2: sensitive reads: Allow with redaction of sensitive fields not needed for the task. Rate limit. Block requests outside the overlay's data scope.
    • Tier 3: changes to records: Allow only for approved tools inside a just-in-time grant. Escalate for human approval on conditions the overlay defines.
    • Tier 4: irreversible actions, including moving funds: Escalate for human approval on every call with full parameters shown. Block everything else.
    • Unlisted tool or changed definition: Block until the catalog entry is reviewed and the definition is re-pinned.

    Evidence, and where runtime enforcement fits

    Examiners and model risk reviewers need to see two things for each agent: what it was permitted to do and what it did. For the first, retain the catalog version, the policy baseline and overlay versions in effect, the agent's identity and the grants it held. For the second, retain each tool call with its parameters, the enforcement outcome, any approver and the time. Evidence assembled after the fact from application logs depends on every team having logged consistently. Evidence produced by the decision point itself is uniform across business lines because the same layer made every decision before the action executed.

    This is why enforcement belongs in the execution path. Current agent identity guidance proposes identities distinct from human users and service accounts, zero standing privilege, just-in-time, time-bound grants, and continuous verification of each tool invocation against current context and risk. The policy layer should consume the identity and credential systems you already run rather than replace them. API management continues to govern API lifecycle and consumers. Model risk management reviews tier definitions, overlays and exceptions, and uses the evidence to test whether controls behaved as designed. Note that the NIST tenets are proposed design guidance, not regulatory requirements.

    Trussed enforces policy in the execution path on tool calls and MCP servers, with allow, redact, block, rate limit and escalate-for-human-approval outcomes. Enforcement produces audit-ready evidence as it runs, and deployment can be self-managed in the customer cloud with configurable retention and data residency. Its scope is individual interactions and tool calls, not end-to-end tracing of agentic workflows.

    One standard, scoped by business line

    Tool catalog

    Every tool and MCP server has an owner, data sensitivity, action type and risk tier.

    Policy baseline

    Minimum controls set once and applied to every business line.

    Business-line overlays

    Retail, commercial, wealth, markets and operations tighten or scope the baseline.

    Runtime enforcement and evidence

    Each tool call is decided before it executes and recorded for examiners.

    See how enforcement and evidence work in practice

    Walk through how policy on tool calls and MCP servers is enforced at runtime and how audit-ready evidence is produced for examiner and model risk review.

    See How Evidence Is Produced
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.