Implementation Guide: Financial Services

    Shadow AI remediation budget for banks: what to fund in year one

    Fund shadow AI remediation in dependency order: discovery and inventory first, then acceptable-use policy, then runtime enforcement on interactions and tool calls, with evidence retention designed in from the start. Add human escalation and staff enablement as enforcement matures, and hold part of the budget in reserve for regulatory change. Detection-only tooling and policy documents alone leave unsanctioned use uncontrolled and leave little evidence for examiners.

    Test whether a proposed control produces examiner-ready evidence

    • Does it enforce policy in the execution path, or only detect and report after the fact?
    • Does it record what was allowed, blocked, redacted and escalated, as underlying records rather than summary dashboards only?
    • Can the bank export those records for examiners and internal audit?
    • Are retention periods and data residency configurable, including in a self-managed deployment?
    • Are human approvals and overrides captured, attributed to a person and reviewable by an independent party?
    • Can the bank show that enforcement and escalation were tested and continue to operate?

    What the year-one budget has to accomplish

    Shadow AI in banking means employees, developers and business units using generative AI sites, coding assistants, browser tools and agent or MCP-connected tools outside approved model risk management and third-party risk processes. The budget question is not which product to buy. It is which control outcomes the bank must be able to demonstrate by the end of the year: what is in use, what is permitted, what was actually stopped or constrained, and what record proves it.

    The supervisory context points toward evidence rather than narrative. The federal banking agencies issued revised model risk management guidance on April 17, 2026. It covers model development and use, validation and monitoring, and governance and controls, including clear policies, roles and responsibilities. It also says practices appropriately vary with a bank's risk profile and model usage, so the scale of funding should be justified by the bank's own exposure. The agencies have also said they plan a request for information on model risk management that considers bank use of AI, including generative and agentic AI.

    An OCC supervisory framework for bank use of advanced AI points the same way. It refers to testing detection, containment, suspension, escalation and remediation arrangements, and contemplates audit reviewing incidents, overrides and underlying evidence rather than only management summaries. Read together, these suggest that a budget weighted toward policy documents and after-the-fact reviews will be hard to defend.

    Sequencing spend across four quarters

    Each category depends on the one before it. Policy written without an inventory misses real usage, and enforcement without policy has nothing to enforce. Evidence design cannot wait until the end, because retention and export requirements shape how enforcement is deployed.

    • Quarter 1: discovery, inventory and evidence requirements: Fund discovery of AI sites, coding assistants, browser tools and agent or MCP tool connections. Record each in an inventory tied to owners and data exposure. Agree with internal audit and model risk management which evidence artifacts will be required later.
    • Quarter 2: acceptable-use policy and first enforcement: Convert inventory findings into rules by use case and data class. Begin enforcement on the highest-risk channels, ideally observing first and then enforcing. Set retention periods and data residency requirements before volume grows.
    • Quarter 3: tool-call enforcement and escalation: Extend enforcement from interactions to tool calls and agent connections. Stand up human approval workflows for actions that policy routes to a person. Test that containment and escalation actually work.
    • Quarter 4: evidence review, enablement and reserve: Run an internal audit style review of exported records. Fund staff enablement so users understand what is blocked and why. Keep a reserve to absorb outcomes from the planned AI request for information and the final third-party risk guidance.

    Where runtime enforcement fits, and where it does not

    Runtime enforcement is the category most often underfunded, because detection tools are easier to buy and policy documents are easier to produce. Detection reports that an unsanctioned tool was used. Enforcement in the execution path decides, at the moment of an interaction or tool call, whether it is allowed, redacted, blocked, rate limited or escalated for human approval. The difference matters to an examiner: one shows awareness of exposure, the other shows the exposure was constrained.

    Trussed operates in this layer. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, at the level of individual interactions and tool calls, and generates audit-ready evidence from that enforcement. It can be deployed self-managed in the customer cloud, with configurable retention and data residency. Trussed has 18+ months of continuous production in a regulated enterprise and is deepest in insurance. It does not claim financial-services deployments.

    Enforcement is not a complete program. It does not replace the inventory, the policy decisions, model risk management review, third-party due diligence or staff training. It also does not provide transaction-level governance, model or dataset provenance, or end-to-end tracing of agentic workflows. Budget it as one layer that depends on the layers before it and feeds the evidence layer after it.

    Assign an owner to every budget line

    Model risk management guidance stresses clear policies and defined roles and responsibilities, so each funding category should have a named accountable owner before money is released. A workable split is as follows. Security architecture owns discovery and runtime enforcement. Risk, compliance and model risk management own the acceptable-use policy and decide which uses require model validation. Third-party risk owns vendor due diligence for any AI tool or provider found in the inventory. Internal audit and compliance own the evidence requirements and the year-end review. Business line heads own adoption and staff enablement.

    Escalation needs particular care. Where a human process is the backstop, audit may assess whether it is independently governed, tested and effective. Approvers should therefore be separate from the business users they are approving, and every decision should be recorded.

    Two items warrant flexibility rather than fixed allocation. The proposed third-party risk management guidance from the OCC, Board, FDIC and NCUA is open for comment and not final as of November 2026, so vendor-related spend should allow for the final text. The planned AI request for information may add expectations for generative and agentic AI. Keeping a defined reserve is more defensible than committing the full budget in the first quarter and reopening the plan later.

    Year-one funding sequence

    Discover

    Inventory the generative AI sites, coding assistants, browser tools and agent or MCP-connected tools in use.

    Define

    Set acceptable-use rules that name what is allowed, redacted, blocked, rate limited or escalated.

    Enforce

    Apply policy in the execution path at runtime, not after the fact.

    Evidence

    Retain records that show examiners what was allowed, blocked or redacted.

    Review your year-one plan against the evidence test

    See how runtime enforcement in the execution path produces audit-ready records of what AI use was allowed, blocked or redacted.

    See How Evidence Is Produced