Best Practices Guide for Bank CIOs

    Shadow AI Business Case for Banks: Costing the Sanctioned Alternative Under the 2026 Model Risk Guidance

    A defensible shadow AI business case for banks is a structured cost model, not a licence quote. It itemizes six categories: licensing and usage, deployment and integration, policy enforcement in the execution path, evidence and retention, model risk management effort, and change management. It sets them against the risk-adjusted exposure of unmanaged use, and it is framed against the model risk management guidance the OCC, Federal Reserve and FDIC issued on April 17, 2026.

    Evaluation criteria for enforcement and evidence

    • Enforcement happens in the execution path at runtime, not through review after the fact.
    • Coverage extends to the surfaces in use: models, agents, tools, MCP servers, AI sites and coding assistants.
    • Available actions include allow, redact, block, rate limit and escalation for human approval.
    • Evidence is generated by enforcement itself and shows what was permitted and what occurred.
    • Retention periods and data residency are configurable, and the deployment can run in the bank's own cloud.
    • Policy scales by use-case materiality, with named owners for approval, monitoring and review.

    What changed in the model risk baseline

    Many business cases still cite SR 11-7 as the model risk baseline. On April 17, 2026, the OCC, in coordination with the Federal Reserve Board and FDIC, issued updated model risk management guidance. It clarifies model risk management principles, sets out a risk-based approach, and rescinds prior model risk management guidance and other issuances. As of November 2026, the business case should cite the 2026 guidance and update any policy references that point to earlier issuances.

    The guidance covers three areas: model development and use, including testing; model validation and monitoring, including conceptual soundness and outcomes analysis; and governance and controls, including clear policies, roles and responsibilities. It also states that practices suitable for one bank may be inappropriate for another with a different risk profile or model purpose. That matters for costing, because effort should scale with materiality rather than apply uniformly.

    The agencies have also said they plan a request for information on model risk management that considers generative AI and agentic AI. Build the cost model so its assumptions can be revised when that request arrives.

    The six cost categories

    Proposals that count only licences understate the sanctioned path. The model below maps cost lines to the work a bank must actually perform. No benchmark figures are given here because the right numbers depend on user population, use-case mix and the bank's existing infrastructure. Each line should carry a one-time and a recurring component, with the recurring part owned by a named budget holder.

    Recurring cost deserves particular attention. Maintaining policy rules, reviewing exceptions and revalidating use cases do not end at go-live. Treating them as a one-time purchase is the most common way these programs become underfunded.

    • Licensing and usage: Seats, token or API consumption, and the growth in both as shadow users move to the sanctioned tool.
    • Deployment and integration: Identity, network, data access, coding assistant and browser configuration, and hosting in the bank's own environment.
    • Policy enforcement in the execution path: Building, testing and maintaining policy rules, plus handling of exceptions and human approval queues.
    • Evidence and retention: Storage, retention periods, data residency, access controls and the effort to produce evidence on request.
    • Model risk management and review effort: Inventory, validation, monitoring and governance work scaled by use-case materiality.
    • Change management: Training, communications, support, and retirement of unsanctioned tools and workflows.

    Why enforcement belongs in the cost model

    A sanctioned alternative only displaces shadow use if it is usable, and it only reduces risk if policy applies to what people actually do with it. Review after the fact detects problems once data has already left the bank or code has already entered a repository. Enforcement at runtime acts on the interaction itself: allow it, redact sensitive content, block it, or escalate it for human approval.

    This changes the cost structure. Runtime enforcement adds engineering and operating cost, including rule design, latency considerations and exception handling. It removes the need to rely on retrospective sampling as the primary control. Costing only the second approach hides the true workload, because manual review effort grows with usage.

    Failure handling is also a cost line. The OCC has published a framework document on bank use of Super Intelligence and agentic SI that includes a resilience area covering the ability to detect, contain, recover from and remediate system or agent failures. It also says audit may assess whether a human-process backstop is independently governed, tested and effective. Whatever the framework's final status for a given bank, the practical point holds: if escalation to a human is part of the design, the testing and governance of that step should be costed on its own line.

    Setting the cost against exposure from unmanaged use

    The comparison is not the sanctioned path against zero. It is the sanctioned path against the continued cost of unmanaged use. That side of the model should be built from the bank's own data: which tools are in use, by which groups, with what categories of data, and which of those uses touch significant business lines or sensitive information.

    A workable method is to list use patterns, estimate likelihood and impact for each in the bank's existing risk scale, and state assumptions explicitly. Public chatbots receiving customer or confidential content, coding assistants introducing code of unknown origin or licence, and browser-based AI sites outside any approved channel carry different exposure and should not be averaged together. Where the bank lacks measured data, label the estimate as management judgment. An examiner-facing document is more credible when it separates measured figures from assumptions.

    Avoid presenting the result as savings. The purpose is to show that the bank has identified a material exposure, selected a control approach in proportion to it, and funded that approach in full.

    What an examiner-facing business case must show

    The 2026 guidance expects clear policies, roles and responsibilities, so the document should name who approves each sanctioned use, who monitors it and who reviews exceptions. It should map effort lines to the three areas of the guidance: development and use, validation and monitoring, and governance and controls. It should show how materiality tiers drive different levels of validation and monitoring cost.

    It should also describe the evidence the bank can produce on request: the policy that applied, the decision taken on each interaction or tool call, who approved any escalation, and how long the record is retained and where it resides. Evidence volume has a real storage and retrieval cost, which is why retention and residency decisions belong in the model rather than in an appendix.

    Trussed provides policy enforcement in the execution path and generates audit-ready evidence as a result of that enforcement, with configurable retention and data residency and self-managed deployment in the customer cloud. It has run continuously in production in a regulated enterprise for more than 18 months, with 99.95% uptime on live production traffic. Its customers span insurance, pharma, healthcare and education. It does not claim banking deployments. Judge any option, including this one, against the criteria above.

    Cost model for the sanctioned AI path

    Cost side

    Six itemized categories, each with a recurring and a one-time component.

    Exposure side

    Risk-adjusted view of unmanaged public chatbots, coding assistants and browser-based AI sites.

    Control logic

    Policy enforced at runtime: allow, redact, block, escalate for human approval.

    Examiner view

    Evidence of what AI was permitted to do and what it did.

    Test your cost model against runtime enforcement and evidence

    Walk through the six cost categories and the enforcement and evidence criteria with a Trussed expert.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

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