Implementation Guide for Financial Services

    Shadow AI to Sanctioned AI Gateway Migration in Banks

    Banks move employees from shadow AI to a sanctioned gateway by treating it as a sequenced migration: discover current use, tier use cases by data sensitivity, build a sanctioned path that is as easy to use as the unsanctioned one, enforce allow, redact, block and escalate decisions at runtime, and retain evidence of each decision for examiners and model risk management reviewers. Blocking alone tends to push use onto unmanaged devices and networks.

    Stage 5: retain evidence for examiners and model risk reviewers

    • Timestamp, user identity and the tool or model addressed
    • Use case and the tier it was assigned
    • Data classifications detected in the request
    • Policy version applied and the action taken: allow, redact, block, rate limit or escalate
    • For redactions, what category of data was removed; for escalations, the approver and outcome
    • Retention period and an export path for examiner requests

    Shadow AI is a symptom of friction

    Employees reach for unsanctioned AI sites, browser tools, coding assistants and personal accounts because they solve a problem quickly. When the sanctioned option is slower, narrower or harder to reach, the unsanctioned one wins. The OWASP LLM AI Cybersecurity and Governance Checklist describes shadow AI as the most pressing non-adversary LLM threat for many organizations, and other practitioners have raised it as a concern in comments to NIST.

    For a bank, the practical consequence is that a blanket block does not remove demand. It relocates it to personal devices and networks where the institution has no visibility and no ability to apply policy to data leaving the firm. The migration therefore has two goals that must advance together: reduce the reasons for bypass, and enforce policy on whatever traffic remains in scope.

    That makes this a change-management and policy problem as much as a technical one. Tool selection comes late in the sequence, after the institution knows what employees are doing and what data is involved.

    Stages 1 and 2: discover current use and tier it by data sensitivity

    Discovery starts from the vantage points the institution already has. Network and proxy telemetry shows which AI destinations are being contacted. Browser controls show which extensions and web tools are in use on managed endpoints. CASB and DLP signals show where sensitive content is moving, and SaaS audits show which applications have been adopted by business units. No single vantage point describes all use, so treat each as partial and reconcile them into one inventory. Add direct conversations with business units, because employees will usually explain what they are trying to accomplish.

    For each discovered use, record the tool, the team, the task and the type of data involved. Then assign a tier based on data sensitivity. A workable scheme has a small number of tiers, for example public or general content, internal content, confidential or customer data, and regulated or restricted data. Tiering drives every later decision: which tools are sanctioned for which tier, which runtime actions apply, and who approves exceptions.

    This inventory also serves a second purpose. The Federal Reserve publishes Supervisory Guidance on Model Risk Management as an attachment to SR 26-2 (as of November 2026), and describes it as setting out sound principles for effective model risk management. An AI use inventory that records tier, owner and decision history gives model risk management reviewers a defined starting point. NIST's Generative AI Profile likewise calls for documenting uses that exceed organizational risk tolerances, which the tier assignment makes straightforward.

    Stage 3: build a sanctioned path employees will choose

    The sanctioned path has to remove the reasons people avoided it. Design it from the discovery data rather than from a vendor feature list. The usual friction points are model choice, response speed, sign-in effort and coverage of the tools people already use, particularly coding assistants and browser-based AI sites.

    A few design decisions matter most. Offer a range of models suited to the approved tiers, so employees are not forced into one model for every task. Use existing single sign-on so access takes seconds, not a separate request process. Measure latency added by enforcement and keep it low enough that it is not the reason someone returns to a personal account. Support coding assistants and tool-using agents explicitly, because developers will leave any path that does not cover their editor.

    Policy should be published alongside the path in plain terms: what is allowed, what is redacted, what is blocked and how to request an exception. Where the path is a network-layer interception point, confirm that it covers AI sites, coding assistants and tools, not only model API traffic. The Federal Reserve Board's own stated approach to its internal AI use is to apply appropriate safeguards while removing unnecessary barriers to adoption. That posture describes the Board's use rather than a bank requirement, but it is a useful parallel for a CISO weighing control against usability.

    Stage 4: enforce allow, redact, block and escalate at runtime

    Enforcement belongs in the execution path, so that a decision is made on each interaction before data leaves, not reviewed afterward. The mapping from data tier to action is a policy decision each bank must make. The table below is a starting design that should be adjusted to the institution's own classification scheme and risk appetite.

    Each action should be tied to the tier of the use case and the classification of the data detected in the request. Decisions about who can approve an escalation, and how quickly, deserve as much attention as the detection logic, because a slow approval queue recreates the friction that drove employees away.

    • Allow: Lower-tier use with no restricted data detected. The interaction proceeds and the decision is recorded.
    • Redact: Prompts containing identifiable or confidential fields that the task does not need. The fields are removed or masked before the request leaves.
    • Block: Restricted data sent to an unapproved destination or tool. The request is stopped and the employee is told why.
    • Rate limit: Unusual volume, such as bulk extraction or repeated attempts after a block. Throttling limits exposure while the activity is reviewed.
    • Escalate for human approval: Higher-tier requests or tool calls with material consequences. A named approver decides before the action proceeds.

    Track adoption and risk signals during the migration

    Define signals before rollout so progress is measured rather than assumed. The following are practitioner suggestions, not regulatory metrics. On the adoption side, watch active users on the sanctioned path as a share of the population discovered earlier, the volume of traffic moving from unsanctioned destinations to the sanctioned one, coverage by business unit, and the number of exception requests, which shows where the path is still too narrow. On the risk side, watch continued contact with unsanctioned AI destinations, the rate of redactions and blocks by tier, escalation turnaround time, and activity from unmanaged devices where it can be observed.

    A rising block count is not automatically a failure. Early on it often shows enforcement working while employees learn the rules. A block rate that stays high for one team usually points to a workflow the sanctioned path does not yet serve, which should feed back into stage 3.

    Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with actions that include allow, redact, block, rate limit and escalate for human approval. It produces audit-ready evidence from that enforcement, mapped to frameworks including NYDFS and NIST. Trussed is deepest in insurance, a regulated financial sector, and does not claim bank deployments. The staging above applies whichever enforcement layer a bank selects.

    Five-stage migration path

    Discover

    Find which AI sites, browser tools and coding assistants employees already use.

    Tier

    Group use cases by the sensitivity of the data involved.

    Sanction

    Offer a path that matches how employees actually work.

    Enforce

    Apply allow, redact, block and escalate decisions in the execution path.

    Evidence

    Retain a record of every decision for examiner and model risk review.

    Plan the migration with the evidence in mind

    See how policy enforcement in the execution path produces audit-ready evidence for each decision, and how it applies to AI sites, coding assistants and tool calls.

    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.