Implementation Guide for Financial Services

    Shadow AI Employee Communication Plan for Bank CIOs

    An effective shadow AI employee communication plan treats messaging as one half of a control. Tell staff what counts as shadow AI, which data classes may never be submitted, which tools are approved and how to request others, what is monitored and enforced, and what the consequences are. Every statement should map to a control enforced in the execution path (allow, redact, block, escalate for human approval). Keep evidence that shows which policy version was in force when each interaction was handled.

    Rollout sequence

    1. 1

      1. Publish the acceptable use matrix

      One page listing tool, data class and role, with the permitted action for each.

    2. 2

      2. Confirm each row is configured

      Remove or reword any statement without a working control behind it.

    3. 3

      3. Brief leaders and managers

      Give them the matrix, consequence wording and the exception path before staff hear it.

    4. 4

      4. Announce to all staff, then train by role

      Use the intranet, manager briefings and role-specific sessions for developers, data teams, HR and legal.

    5. 5

      5. Open the request and reporting paths

      Document who decides on tools and exceptions, with each decision recorded.

    6. 6

      6. Review on a fixed cadence

      Use employee feedback, enforcement outcomes and new AI features in existing tools to update messages and controls together.

    Evidence that communication matched the controls

      What staff need to understand about shadow AI

      OWASP describes shadow AI as employees using unapproved online AI tools, unsafe browser plugins and third-party applications that add LLM features through updates or upgrades, circumventing standard software approval processes. It calls this the most pressing non-adversary LLM threat for many organizations. For a bank, that definition has a messaging consequence: a memo that says "do not use public chatbots" leaves out browser plugins, coding assistants and new AI features that appear inside tools already approved.

      A written acceptable-use policy and a staff memo state intent, but on their own they neither change behavior nor produce evidence. The plan therefore needs five core messages, each specific enough that an employee can act on it and a CIO can verify it.

      • Scope: Name the channels in scope: AI websites, personal accounts, browser plugins, coding assistants, and AI features added to approved tools.
      • Data that may never be submitted: List the data classes plainly, such as customer data, confidential supervisory information and proprietary code.
      • Approved tools and request path: Say which tools are sanctioned, who decides on new requests, and how long a decision should take.
      • What is monitored and enforced: Describe only controls that are configured today, including what happens to data a user submits.
      • Consequences and escalation: State outcomes consistently and give a route for reporting mistakes and requesting exceptions.

      Supervisory expectations that shape the message

      The model risk management guidance is hosted by the Federal Reserve as an attachment to SR 26-2, and the Fed, with the OCC and FDIC, has recently amended it, as of November 2026. Materials that cite earlier model risk guidance should be rechecked against the current text before they reach examiners. The guidance states that model use can present substantial risk and that effective policies are supported by clear policies and effective controls. It also refers to an ongoing monitoring plan. A staff communication that promises controls the bank does not enforce weakens that position.

      Third-party expectations point the same way. The Vice Chair for Supervision has asked how third-party risk-management expectations should apply to vendor-provided AI tools, and the OCC has said it expects robust due diligence, effective contract management and ongoing oversight of third parties. Approved-tool decisions should therefore be recorded as third-party decisions, not informal IT approvals.

      The NIST Generative AI Profile (an initial public draft) includes defining and communicating organizational access to generative AI and training employees on risks and reporting procedures. OWASP's checklist is non-regulatory. Treat both as reference practice, not bank obligations.

      Message content by staff group

      OWASP recommends permissible-use and security-awareness training for all employees, with more specialized content for human resources, legal, developers, data teams and security teams. Adding leadership and managers is a sensible extension. The core messages stay identical across groups. What changes is the depth and the examples.

      • All staff: Plain-language scope, prohibited data classes, approved tools, how to request a tool, and how to report an error without delay.
      • Managers: How to answer team questions, handle exception requests, and escalate suspected misuse consistently.
      • Developers: Rules for coding assistants, what code and secrets may not be pasted, and how AI-generated code is reviewed.
      • Data teams: Limits on uploading datasets, customer records or model inputs to external AI services, with approved alternatives.
      • HR and legal: Bank-specific consequence wording. The Federal Reserve Board's own internal policy names loss of or limits on IT resource access and disciplinary action up to separation, which is an example of plain statement, not a bank requirement.
      • Security and IT: Enforcement behavior, exception handling, reporting procedures and how policy changes reach the controls.

      Aligning messages with runtime enforcement

      The acceptable use matrix is the bridge between communication and control. Each row should correspond to one of four actions applied in the execution path. Allow covers an approved tool with a permitted data class. Redact covers an approved tool where specific data elements must not leave the bank. Block covers a prohibited data class or an unapproved destination. Escalate for human approval covers exceptions that need a named decision-maker. Staff messages can then describe real outcomes: this is permitted, this will be redacted, this will be stopped, this goes to a reviewer.

      Policy enforced at runtime, not reviewed after the fact, changes what a CIO can say. Statements about monitoring can be exact because the control acts on the interaction itself. Because third-party apps can add AI features through updates, the approved-tool inventory and enforcement scope need recurring review, and staff should be told that new AI features in existing tools are in scope.

      Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with allow, redact, block, rate limit and escalate for human approval, applied to individual interactions and tool calls. It supports self-managed deployment in the customer cloud with configurable retention and data residency. Insurance, a regulated financial sector, is where Trussed is deepest, and it has 18+ months of continuous production in a regulated enterprise.

      The tradeoff is friction. Tighter enforcement raises exception volume and user questions, so the messaging should explain what staff will see, including when redaction occurs, and the exception path needs capacity to respond.

      From message to evidence

      Say

      Scope, prohibited data, approved tools, consequences

      Enforce

      Allow, redact, block or escalate at runtime

      Record

      Action taken, tied to the policy version in force

      Review

      Update messages and controls under one change process

      Check your messages against what is enforced

      Talk through your acceptable use matrix and see how runtime enforcement and audit-ready evidence support what you tell staff.

      Talk to an Expert
      Next step

      How does your AI governance program compare?

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