Implementation Guide for Risk Leaders

    Supervising AI Features Embedded in Wealth Management CRM and Financial Planning Software

    Treat vendor-embedded AI as a third-party supervision problem. Inventory every feature a vendor release enables, tier each by the client data it touches and the actions it can take, and define permitted use. Then place each control where it can be enforced: at runtime for traffic the firm controls, and through due diligence, contract terms and configuration for everything else. Keep records that show the policy applied, the action taken, the data redacted and the approver. FINRA states that its rules apply to embedded features in existing third-party products.

    Step four: decide what evidence should exist

    • Feature and version or release identifier, linked to its register entry.
    • Policy applied to the interaction, with the policy version.
    • Action taken: allowed, redacted, blocked or escalated.
    • Data redacted, recorded by category rather than by content.
    • Approver identity and timestamp where human approval was required.
    • Retention and access controls for the evidence itself.

    Why embedded AI is a third-party supervision problem

    Meeting summarization, client communication drafting, next-best-action suggestions and planning recommendations often arrive inside a CRM or planning platform through a vendor release. The firm did not select a model or run a deployment project, yet the feature may touch client data and shape client-facing output.

    FINRA Regulatory Notice 24-09 states that its rules are technology neutral and apply whether a firm develops Gen AI tools or uses a third party's technology, including through embedded features in existing third-party products. FINRA's 2026 Annual Regulatory Oversight Report says GenAI use can implicate supervision, communications, recordkeeping and fair dealing, and cites FINRA Rule 3110. It also describes a supervision, governance or model risk management framework with clear policies, procedures and documentation, and data management that addresses quality, integrity, retention and security.

    The practical consequence is that a feature enabled by a vendor is the firm's to supervise. Because the firm is not building the model, the main questions are about visibility, boundaries and evidence rather than model construction. The OCC's 2026 supervisory framework for bank use of advanced AI points the same way. It flags diffuse ownership as a limit on oversight and notes that point-in-time validation may not address non-determinism or continuous change. Its stated scope is banks, but vendor release cycles raise the same concern for any firm whose features change without a firm-led deployment. This reflects published material as of November 2026.

    Step one: inventory and tier every feature

    Start with a register of every AI feature present in the CRM, planning and adjacent tools, including those switched on by default. Pull from vendor release notes, administrator settings and vendor notices, and assign each feature a named owner. Diffuse ownership is a recurring weakness, so a feature without an accountable person should not stay enabled.

    Tiering then needs two inputs per feature: what client data it can read, and what it can do with the result. A summarizer that produces internal notes for an advisor to read is a different risk from a drafting tool that produces client-facing text, and both differ from a feature that proposes planning outputs a client may rely on. A reasonable starting point is to map each feature to the rule areas FINRA links to GenAI: supervision, communications, recordkeeping and fair dealing. The tier is a firm decision, not a sourced standard, so document the reasoning.

    • Record the feature, vendor, version or release that enabled it, enable date and approver.
    • Name an accountable owner for each feature.
    • List the client data categories the feature can access, and whether data leaves the platform or is retained.
    • Record the output type: internal note, client-facing draft, recommendation or plan output.
    • Record whether the feature can take an action, such as sending, writing to a client record or triggering a workflow.
    • Set re-review triggers tied to vendor releases, so a changed feature returns to the register.

    Step two: define permitted data use and actions

    For each tier, write down what the feature may receive, what it may produce and what requires a human. Permitted data use should name the client data categories allowed into the feature and those that must be removed first. Permitted actions should state whether output can reach a client without review and which actions need approval.

    NIST AI RMF MAP 3.5 calls for human oversight processes to be defined, assessed and documented in line with organizational policy, and the framework lists mechanisms such as override, incident response and change management. These can anchor your escalation and approval points. FINRA's 2020 report on AI in the securities industry gives a precedent for limits on AI-triggered actions, describing thresholds where models trigger actions autonomously. Translating that into allow, redact, block or approve decisions for a CRM feature is a design choice, but a defensible one.

    If any AI is used inside the supervisory system itself, for example to review electronic correspondence, Notice 24-09 says the firm should have policies covering technology governance, including model risk management, data privacy and integrity, and the reliability and accuracy of the model. Include those uses in the register.

    Step three: decide where each control can be enforced

    Not every control can run at the firm's boundary. A runtime control can only act on interactions and tool calls that pass through a path the firm operates. Anything the vendor does internally, such as training data use, retention or its own logging, cannot be verified from that path and belongs in due diligence and contract terms. The split below is a design guide, and the right placement depends on how each vendor feature is built and configured.

    • Redacting client data before it reaches an AI feature: Runtime where the request passes through a firm-controlled path. Otherwise configuration settings and contract limits on data use.
    • Blocking prohibited outputs or tool calls: Runtime for interactions in the path. Otherwise vendor feature toggles and acceptance of the vendor's own controls.
    • Human approval before client-facing use: Runtime escalation where the path carries the action. Otherwise supervisory procedures applied inside the platform workflow.
    • Vendor training use, retention and internal logging: Due diligence and contract terms. Not verifiable at runtime by the firm.
    • Notification of model or feature changes: Contract terms and release monitoring, with re-review triggers in the register.
    • Understanding how outputs are derived: Due diligence. FINRA's 2020 report describes supervising AI as understanding how applications function and how outputs are derived.

    What a runtime enforcement layer can and cannot cover

    Trussed enforces policy in the execution path across models, agents, tools and MCP servers. It can allow, redact, block, rate limit or escalate for human approval, and enforcement is applied at runtime, not reviewed after the fact. Each enforcement decision produces audit-ready evidence, and deployment is self-managed in the customer's cloud with configurable retention and data residency. Trussed's customers are in insurance, pharma, healthcare and education, and it does not claim financial-services deployments.

    Its scope is individual interactions and tool calls that pass through the path. It does not observe what happens inside a vendor's platform outside that path, and it does not replace due diligence or contract terms for vendor-side behavior. It does not provide model provenance, transaction-level governance or end-to-end tracing of agentic workflows. For a risk leader, the value is narrow and concrete: where the firm can place a control in the path, the policy decision and its evidence exist for each interaction, which supports the examiner conversation about permitted and actual behavior.

    Supervision sequence for embedded vendor AI

    Inventory

    Every AI feature enabled by a vendor release, with an owner and an approver.

    Tier

    Classify by client data accessed and actions the feature can take.

    Place controls

    Runtime enforcement where possible; due diligence and contract terms elsewhere.

    Retain evidence

    Records of policy applied, action taken, data redacted and approver.

    Map your embedded AI to controls and evidence

    See how policy enforcement in the execution path produces audit-ready evidence for individual interactions and tool calls, and where due diligence and contract terms must carry the rest.

    See How Evidence Is Produced