Implementation Guide | Financial Services

    When KYC and Customer Service Agents Share a Customer Record: Access Boundaries Step by Step

    Treat the shared record as a data-classification and tool-access problem. Assign every field a sensitivity tier, define which tools and operations each agent role may call, set an allow, redact, block or escalate action per role and tier, enforce those rules at each tool call and interaction in the execution path, and keep a decision record for every enforcement event. Datastore roles alone cannot describe what an agent retrieves, passes to a model or returns.

    Step 4: Enforce in the execution path, not after the fact

    1. 1

      Retrieval

      Evaluate role, tier and token audience before the record or document is fetched.

    2. 2

      Model input

      Redact or block restricted fields before they enter a prompt.

    3. 3

      Tool write

      Permit only allowlisted fields. Block or escalate everything else.

    4. 4

      Returned output

      Check the response against the role's tier limits before it reaches the customer.

    5. 5

      Derived copies

      Carry tier labels onto embeddings and caches and apply the same checks.

    Evaluation questions for any enforcement approach

    • Does it evaluate each tool call and MCP request per agent role, and block or redact before data reaches the model or the response?
    • Does policy operate at field level on the shared record?
    • Do tier labels persist on embeddings, caches and other derived copies?
    • Does each decision record include identity, role, tool, arguments, policy version, action and approver?
    • Can retention and data residency be configured, and can it run in your own cloud?
    • Are escalation approvals tied to specific actions and arguments?

    Why datastore permissions do not describe agent access

    In financial services, a KYC agent and a customer service agent often work against the same customer record. The KYC agent needs identity documents, screening results, risk ratings and sensitive investigative notes. The service agent needs a narrow subset to resolve an inquiry. When both reach the record through shared tools, connectors or MCP servers, the datastore sees one connector identity, not two different agent purposes.

    Datastore role permissions govern who can read or write a row or column. They do not say what an agent may retrieve for a given task, what it may pass to a model, or what it may return to a customer. Agent guidance from NIST and OWASP focuses on a different layer: per-invocation scope, token audience and retrieval-time filtering. That points to three separate control points: retrieval, model input and returned output. Each needs its own decision.

    The risk is concrete. An over-scoped MCP server token creates aggregation risk, because one broad credential lets a service agent reach KYC fields regardless of how the datastore roles were designed. OWASP recommends narrow scopes, such as read-only, and validation of the audience of access tokens. The remaining steps assume this framing: the problem is classification and tool access, not model behavior.

    Steps 1 and 2: Tier the fields, then define roles and tool grants

    Start by inventorying every field and attached document in the shared record and assigning a sensitivity tier. A workable scheme has five tiers: service-safe fields such as case status and contact preferences; sensitive identifiers such as Social Security, passport and taxpayer identification numbers; identity verification artifacts; screening results and risk ratings; and internal investigative notes. This is an architectural convention, not a regulatory taxonomy. The SEC's Regulation S-P final rule offers one reference for identifiers, listing Social Security numbers, driver's license or ID numbers, passport numbers, taxpayer identification numbers and biometric records as examples of sensitive customer information.

    Record the rationale for each tier assignment in your risk assessment. The FTC Safeguards Rule (16 CFR 314.3) requires a written information security program based on a risk assessment that identifies reasonably foreseeable internal and external risks to the confidentiality and integrity of customer information, and that assesses the sufficiency of existing safeguards. Tier decisions belong in that record.

    Next, define agent roles: KYC, customer service and any others. For each, list the exact tools, MCP servers and operations it may call, and default-deny everything unlisted. Issue role-specific, narrowly scoped, short-lived credentials rather than one shared service credential. NIST's design tenets describe least privilege constrained by separation of duties, zero standing privilege and time-bounded, audience-specific grants. Apply the same logic to delegation: a sub-agent or downstream service should never hold more authority than the agent that called it.

    Step 3: Set an action for each role and tier

    Read and write policy should be defined separately. The matrix below is a starting point to adjust with the risk owner, not a prescribed standard. Escalation thresholds in particular should be set with the people who own the risk, since NIST's tenets call for organizational risk thresholds that trigger human review of agentic actions.

    • Service-safe fields: Customer service: allow read, and write only to a named field allowlist. KYC: allow.
    • Sensitive identifiers: Customer service: redact at field level, for example showing a partial value only if the workflow needs it. KYC: allow for the active case.
    • Identity verification artifacts: Customer service: block. KYC: allow read, escalate for human approval on any write.
    • Screening results and risk ratings: Customer service: block, including from model input. KYC: allow read for the active case.
    • Investigative notes: Customer service: block. KYC: allow read, escalate writes and any export outside the case.
    • Writes outside the allowlist: Customer service: block, or escalate when a legitimate case exists. Rate limit repeated attempts.

    Step 5: Evidence for examiners and model risk reviewers

    An examiner or model risk reviewer will want to see what each agent was permitted to do and what it did. That requires two sets of artifacts. The first is the configuration: the field classification with rationale, role definitions, tool grants, the action matrix and approved changes to each. Version these. Tier reassignments and new tool grants should be approved and dated, so a reviewer can reconstruct the policy in force at the time of any interaction.

    The second is the decision record. For every enforcement event, retain the agent identity and role, the tool and its arguments, the policy version, the fields affected, the action taken, the approver if the request was escalated, and a timestamp. Bind escalation requests to the specific action and arguments, so an approval cannot be reused for a different call. Present this set as a proposal for your model risk reviewers to validate against your own framework, not as a claim about what any guidance mandates.

    Set retention and data residency according to your institution's own schedule. The five-year retention in the Customer Identification Program rules applies to specified CIP records, such as the methods and results of verification measures, not automatically to agent decision records. Trussed generates audit-ready evidence from enforcement itself, with configurable retention and data residency.

    Decision records also help after an incident. The interagency guidance on unauthorized access expects response programs that identify the affected systems and types of customer information. A record of which fields an agent accessed makes that scoping faster. Third-party tools and MCP servers that touch the record belong in your access reviews, since Regulation S-P extends protections to customer information handled on a covered institution's behalf, as of November 2026.

    Five steps from shared record to examiner-ready boundaries

    1. Tier the fields

    Assign a sensitivity tier to every field and attached document.

    2. Define roles and tools

    List the tools, MCP servers and operations each agent role may call. Deny the rest.

    3. Set actions

    Choose allow, redact, block or escalate for each role and tier.

    4. Enforce at runtime

    Evaluate each tool call and interaction before data moves.

    5. Retain evidence

    Record who or what acted, under which policy version, with what result.

    Review your shared-record boundaries with an expert

    Walk through your field tiers, agent roles and tool-call enforcement points, and see how audit-ready evidence is produced for each decision.

    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.