Implementation Guide | Financial Services

    KYC AI Agent Runtime Controls for Banks Under the 2026 OCC Model Risk Guidance

    KYC AI agent runtime controls are policies enforced at each interaction and tool call an onboarding agent makes. For each control point, the bank decides whether to allow, redact, block, rate limit or escalate for human approval, and records the policy version, action, tool, approver and timestamp as evidence. Pre-deployment validation shows how a model was assessed. Runtime enforcement shows what the agent was permitted to do and what it did.

    Map the KYC agent to its interactions and tool calls

    1. 1

      1. Document intake

      The agent reads passports, licences and proof of address. Allow approved document types, redact identifiers and other PII before prompts are built, and rate limit bulk uploads.

    2. 2

      2. Sanctions and PEP screening

      Allow calls only from an agent with a verified identity. Block unauthenticated or self-declared identities and rate limit repeated screening of the same subject.

    3. 3

      3. Adverse media lookup

      Allow approved sources only. Redact customer identifiers that are not needed in outbound queries and block unapproved external tools and MCP servers.

    4. 4

      4. CDD and EDD drafting

      Allow the draft. Redact PII in outputs that leave the case file and escalate any recommended outcome, such as approve or escalate to EDD, for human approval.

    5. 5

      5. Core system reads

      Allow reads scoped to the customer the agent is working on. Block queries that reach other customers' records.

    6. 6

      6. Core system writes

      Block by default. Escalate changes to risk rating, account status or case disposition for human approval before they execute.

    Criteria for evaluating runtime enforcement

    • Coverage: does enforcement cover model calls, agent actions, tool calls and MCP servers, and which are enforced inline rather than only recorded?
    • Actions: can each tool call take allow, redact, block, rate limit or escalate for human approval, and what happens if the policy service is unavailable?
    • Identity: is each agent's identity cryptographically verified and bound to every request in the evidence record?
    • Evidence: does each action produce the policy version, action, tool, approver identity and timestamp, and can it be exported for examiners and internal audit?
    • Deployment: can it run in the bank's own cloud with configurable data residency and retention?
    • Mapping: can evidence be mapped to the frameworks the bank is examined against?

    What changed in model risk guidance, and what has not

    On April 17, 2026, the OCC issued Bulletin 2026-13, revised model risk management guidance covering model development and use, validation and monitoring, governance and controls, and vendor and third-party products. The accompanying news release says the update serves in part to rescind prior model risk management guidance and other issuances. As of October 2026, internal policies that still cite earlier guidance by name should be reconciled with the current OCC text by the bank's model risk function.

    The revised guidance says practices should be risk-based, tailored, and commensurate with a bank's size, complexity and extent of model use, and that what works for one institution may not suit another with a different risk profile or model purpose. It also describes generative and agentic AI as novel and rapidly evolving. The OCC, Federal Reserve Board and FDIC have said they plan a request for information on bank use of AI, including agentic AI, so expectations may continue to develop.

    None of this changes a practical gap. Validation and monitoring assess a model before and after deployment. They do not limit which tools an agent can call, which data it can read, or which actions it can take in the middle of an onboarding case. Using an AI agent also does not remove AML or sanctions obligations. Compliance owners still define the screening, monitoring, reporting and retention rules. Runtime controls give those rules a point of enforcement.

    Choose the policy action by exposure and impact

    Two questions drive the choice at each control point: how much sensitive data is exposed, and how hard is the action to reverse. High data exposure with low decision impact, such as document intake, usually calls for redaction. High decision impact, such as a write that changes a customer's risk rating, calls for human approval. Actions with no legitimate use in the workflow, such as a call to an unapproved MCP server, should be blocked rather than reviewed later.

    Rate limits are a quieter control. They bound the damage from loops, runaway retries and bulk extraction without stopping normal casework.

    Identity belongs in the same design. Each agent should present a cryptographically verifiable identity, and self-declared headers such as an agent ID or role without proof should be rejected before the request reaches screening or internal systems. Binding that identity to each request means the record shows which specific agent performed a check and for which customer.

    Decide in advance what happens when a policy check cannot complete. A fail-closed default, where the action does not proceed, suits screening and writes. The tradeoff is operational: a policy service outage can pause onboarding, so the bank should set that expectation with business owners. The NYDFS industry letter on AI cybersecurity risk supports the broader principle of layered, overlapping controls so that if one fails, others limit the impact. For NYDFS Covered Entities, it also notes the MFA requirement in force as of November 2025 and expects third-party provider policies to address access controls, encryption and due diligence.

    Evidence that answers examiner questions

    An examiner or internal auditor will ask two things: what was this agent allowed to do, and what did it do? Answering both requires records created when the action happens, not reconstructed from application logs afterward.

    For each enforced action, the record should show the policy version in force, the action taken (allowed, redacted, blocked, rate limited or escalated), the tool or MCP server called, the agent identity, the customer case context, the timestamp, and, where a human approved or rejected the action, the approver's identity and decision time. Those fields tie the permission to the behavior.

    The difference from after-the-fact log review is evidentiary. Logs collected for debugging capture what a system emitted, often without the policy that applied or proof that a control actually ran. Evidence generated by enforcement is a by-product of the decision itself: the control fired, so the record exists. Retention and data residency settings matter here too, because evidence containing case details must be kept under the bank's own rules and for as long as those rules require.

    The OCC guidance emphasizes clear policies, roles and responsibilities, and ongoing validation and monitoring. Runtime evidence supports those expectations by giving monitoring teams a record of control activity, not only model outputs.

    How runtime enforcement fits with model risk and compliance

    Runtime enforcement does not replace model validation, compliance review or the bank's AML program. Model risk management still assesses conceptual soundness and outcomes. Compliance still owns the screening and reporting rules. Runtime controls take the output of those functions, such as which tools are approved, which fields are sensitive and which actions need a second pair of eyes, and enforce them in the execution path.

    The scope should be stated plainly. Enforcement of this kind operates on individual interactions and tool calls. It does not by itself provide governance of a whole onboarding case across agents and tasks, or an end-to-end trace of an agentic workflow. Banks that need those views should plan for them separately and avoid describing interaction-level evidence as more than it is.

    Trussed is one way to enforce policy at runtime. It applies allow, redact, block, rate limit and escalate-for-human-approval actions across models, agents, tools and MCP servers, produces continuous audit-ready evidence mapped to frameworks including NYDFS and NIST, and can be self-managed in the customer's cloud with configurable retention and data residency. Insurance, a regulated financial sector, is where Trussed is deepest. The platform has run 18+ months in continuous production in a regulated enterprise at 99.95% uptime on live traffic.

    Runtime control points in an AI-assisted KYC workflow

    Document intake

    Redact PII before it reaches the model.

    Screening and adverse media

    Verify agent identity, rate limit queries, block unapproved tools.

    CDD and EDD drafting

    Allow drafts, escalate recommendations for human approval.

    Core system writes

    Block by default, approve high-impact changes.

    See how enforcement produces examiner-ready evidence

    Walk through a KYC agent's tool calls, the policy applied at each one, and the evidence record each action generates.

    See How Evidence Is Produced