Implementation Guide: Financial Services

    Shadow AI in bank branches and contact centers: bringing frontline AI use under policy

    Treat shadow AI in banking contact centers as an enforcement and evidence problem, not only a discovery or training problem. Map how frontline staff use AI, classify the data and interactions involved, define four policy outcomes (allow, redact, block, escalate for human approval), enforce them at runtime in the path of each interaction, and retain evidence of every decision for examiners and model risk management.

    • The policy version in force when each interaction occurred
    • The decision for each interaction: allow, redact, block or escalate
    • Data categories detected, and what was redacted or withheld
    • User or role, timestamp, and the AI tool or destination involved
    • For escalations, the approver and the approval outcome
    • Exportable records for internal audit, model risk management and examiners
    • Where is policy applied: network, endpoint, browser, or the interaction and tool call itself? Can it act before data leaves?
    • Can it make allow, redact, block and escalate decisions per interaction across AI sites, assistants and tool calls?
    • What is recorded for each decision, and can it be exported for audit and examiners?
    • Which specific NIST AI RMF and NYDFS provisions does the evidence map to?
    • How are policy changes versioned and tested, and does the system fail open or fail closed when a decision cannot be made?

    How shadow AI arises in branch and contact center work

    Branch staff and contact center agents handle call content, account details and product questions under time pressure. Public AI sites and browser-based assistants make it easy to summarize a call, draft a letter to a customer, or look up a policy answer. If the institution has not provided an approved alternative, or the approved option is slower, staff use whatever is reachable from their browser.

    The exposure is that customer account data, personally identifiable information and call content can leave the institution's control with no policy applied at the point of use. AI in call centers is a recognized banking use case, with federal agencies describing AI used to process and triage customer calls. That places frontline AI interactions inside governance scope. It does not tell you how often unsanctioned use occurs at your institution, and this guide does not assume a rate.

    Detection alone does not close the gap. The CFPB describes its own approach to unauthorized AI as monitoring IT activity, using the same method as shadow IT discovery. Discovery is a sound first stage because it shows that a tool was used. It does not show what policy applied to a specific interaction. Written acceptable-use policies and after-the-fact reviews share that limit: they describe policy as written, not policy as enforced.

    Classify interactions and data before writing rules

    Start with the interaction types your staff actually perform. Three are common in frontline work: call summarization, response drafting and knowledge lookup. Each carries different data, so each needs different outcomes.

    Summarization typically carries the most customer content, including account identifiers, verification details and the substance of the conversation. Drafting carries customer-specific facts in the prompt and produces text that may be sent to the customer. Lookup often carries less customer data, but the answer a staff member relies on may be wrong or out of date, so the concern shifts from disclosure to accuracy and reliance.

    Classify the data alongside the interaction: customer account data, personally identifiable information, call content, and internal product or policy material. The working artifact is a matrix of interaction type, data class and destination tool, with a policy outcome assigned to each cell. Rank the cells by your own risk assessment rather than a generic list.

    A staged path from unmanaged to policy-bound

    The sequence below is a recommended approach. Each stage produces something the next one needs, and the later stages are where most programs stop short.

    The four outcomes need clear definitions. Allow applies to approved tools and permitted data. Redact applies when the interaction is acceptable once identifiers are removed. Block applies when the data class or destination is prohibited. Escalate for human approval applies when the decision depends on context, such as a drafted customer communication on a disputed account, and a named reviewer must approve before the interaction proceeds.

    • 1. Map frontline AI use: Use discovery to identify which AI sites, assistants and tools staff reach, by role and workflow.
    • 2. Classify data and interactions: Build the matrix of interaction type, data class and destination.
    • 3. Define policy outcomes: Assign allow, redact, block or escalate to each combination, with an owner for each rule.
    • 4. Enforce at runtime: Apply the outcomes to individual interactions across AI sites, assistants and tool calls.
    • 5. Retain evidence: Record each decision in a form that internal audit, model risk management and examiners can review.

    Where enforcement has to sit in the execution path

    Our position is that policy must be applied before data leaves the institution, at the level of the individual interaction, with knowledge of the user, the data and the destination. That is an architectural judgment, and each placement involves tradeoffs. The descriptions below are general and vary by product.

    • Network controls: Broad coverage of destinations and a useful basis for discovery. They typically see traffic rather than the meaning of an interaction, so decisions tend to reduce to allow or block by destination.
    • Endpoint or browser controls: Closer to the user and the application context. Coverage depends on deployment across every device and browser staff use, and on keeping that deployment current.
    • Execution-path enforcement: Policy is evaluated on the interaction itself, including prompts, responses and tool calls. This supports redact and escalate decisions per interaction, and depends on the path being the one staff actually use.

    Four policy outcomes for each frontline AI interaction

    Allow

    Approved tool, permitted data, interaction proceeds and is recorded.

    Redact

    Customer identifiers are removed before content leaves institutional control.

    Block

    Prohibited data or destination, interaction stopped at the point of use.

    Escalate

    Held for human approval when the policy decision needs a reviewer.

    Test enforcement against your own frontline workflows

    Walk through call summarization, response drafting and knowledge lookup with an expert, and see what evidence runtime enforcement produces for each interaction.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

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