Technical Guide | Financial Services

    Shadow AI in Bank Treasury: Spreadsheet Copilots, Add-Ins and Model Risk Management Under the 2026 Guidance

    Shadow AI enters bank treasury and finance when analysts use spreadsheet copilots or third-party add-ins that send workbook contents to external models outside sanctioned channels. The result is data exposure and AI-assisted outputs that feed calculations without appearing in the model inventory or leaving examiner-ready records. As of November 2026, the interagency model risk management guidance issued on April 17, 2026 is risk-based and non-enforceable, so the bank's own governance practices carry the weight. The practical response is to discover the tools, define policy by data class and downstream criticality, and enforce allow, redact, block or escalate decisions at runtime, with evidence produced by each decision.

    Sequencing discovery, policy definition and enforcement

    1. 1

      1. Discover

      Identify which copilots and add-ins finance users have, sanctioned or not, and which workbooks and data classes they touch. Compare the result with the model inventory.

    2. 2

      2. Classify

      Decide whether outputs that feed financial calculations are in scope as models or as tools under the bank's own governance practices. Tier by the criticality of the downstream calculation.

    3. 3

      3. Align onboarding

      Route approved tools through existing third-party risk management: due diligence, contract management and ongoing oversight scaled to criticality. Update any policy that still cites rescinded bulletins.

    4. 4

      4. Define policy

      Write rules by data class and tier: which ranges may be sent, which require redaction, which are blocked, and which need human approval.

    5. 5

      5. Enforce in a narrow scope

      Start with the highest-tier workbooks and data classes, observe decisions, and tune before widening scope.

    6. 6

      6. Monitor vendor change

      Set materiality thresholds and escalation paths for output shifts when inputs are unchanged, and record model or version identity where it is available.

    How shadow AI enters treasury and finance spreadsheets

    Treasury, liquidity and finance analysts work in spreadsheets because the work is iterative and the tools are familiar. A spreadsheet copilot or AI add-in adds a new step to that workflow. The user selects a range or a whole workbook, writes a prompt, and the tool sends that content to an external model. The model returns text, a formula, a summary or a set of values that the analyst pastes back into the sheet.

    That step is shadow AI when it happens outside sanctioned channels. No one approved the tool for the data class involved, the model inventory does not list it, and the interaction leaves no record the bank controls. The user is not acting in bad faith. The add-in is simply easier to reach than the approved process.

    Two features make spreadsheets harder than other AI entry points. First, the unit of exposure is a range of cells, so a single prompt can carry liquidity positions, forecast assumptions and counterparty details together. Second, the output returns into the same file that feeds downstream calculations, so an AI-generated value can become an input to a figure someone else relies on, with no marker showing where it came from.

    What changed in model risk management expectations

    Many internal policies and articles still frame this problem around SR 11-7. The current position is different. On April 17, 2026, the OCC, in coordination with the Federal Reserve Board and the FDIC, issued updated model risk management guidance that clarifies principles, sets out a risk-based approach, and rescinds prior model risk management guidance. The OCC is rescinding Bulletin 2011-12 and Bulletin 2021-19. A Federal Reserve speech in May 2026 describes the Fed as having amended its model risk management guidance together with the OCC and FDIC, so teams should confirm how their own policy references map to the current text.

    The guidance says practices should be risk-based, tailored and commensurate with an institution's size, complexity and extent of model use. It covers model development and use, validation and monitoring, and governance and controls. It does not set enforceable standards, and non-compliance will not result in supervisory criticism. It is expected to be most relevant to banking organizations with more than $30 billion in total assets, though smaller institutions with significant model risk exposure may also find it relevant.

    Two points matter for spreadsheet AI. The guidance says a banking organization's own risk management and governance practices should determine controls for tools, processes or systems it does not cover. Separately, the OCC has published a supervisory framework for examiners evaluating banks' use of AI in normal-course activities, including risk management and operations. Examiner attention therefore rests on whether the bank can show its own governance applied to these tools, not on a checklist.

    The data exposure and model risk management gaps

    The first gap is scope. The bank must decide whether AI outputs that feed financial calculations are models, or other tools governed by its own practices. Either answer requires that the tool be known. A copilot nobody inventoried cannot be classified, tiered or validated.

    The second gap is data handling. Liquidity positions, forecasts and counterparty details are exactly what a prompt can carry. Confidential supervisory information, where analysts hold it, creates the highest-consequence version of this problem. Privacy and information security obligations apply to AI use, and a bank cannot demonstrate compliance for transfers it cannot see.

    The third gap is vendor change. A public comment to the Federal Reserve in 2026 cited research indicating that vendor-driven model updates can shift financial risk assessments with no change to the bank's inputs, prompts or processes. That is a commenter's claim rather than a regulator finding, but the mechanism is easy to reason about. If the bank does not receive the vendor's code or data, it can only control and evidence its own side of the interaction: what went in, what came out and what decision was applied.

    The fourth gap is provenance of use. Risk leaders need to say which AI-generated outputs fed which calculations. Without a record at the moment of interaction, that answer is reconstructed from memory and file history.

    Review after the fact versus enforcement at runtime

    Detection and periodic review have a role in finding unknown tools. They do not change what happened at the moment of use. The table compares the two approaches on the questions a risk leader will be asked.

    • Timing of the decision: After the fact: data has already left the bank when a reviewer sees it. Runtime: policy is applied to the individual interaction before the request proceeds.
    • Possible outcomes: After the fact: alert, ticket or retrospective remediation. Runtime: allow, redact, block or escalate for human approval.
    • Evidence of permitted behavior: After the fact: reconstructed activity records, often without the policy that applied. Runtime: each decision is recorded with the policy version that produced it.
    • Handling of sensitive ranges: After the fact: exposure is documented. Runtime: counterparty fields or position data can be redacted or the request blocked.
    • Tiering by criticality: After the fact: severity is assigned during review. Runtime: stricter rules can apply when the workbook feeds a critical calculation.
    • Tradeoff: After the fact: low friction and incomplete control. Runtime: requires routing interactions through a control point and tuning policy to avoid blocking legitimate work.

    Audit-ready evidence and where runtime control fits

    Whatever mechanism a bank chooses, the evidence standard is practical. For each interaction, the bank should be able to show which user and tool were involved, what data was sent, what policy version was evaluated, what decision was taken, what was returned, and whether a human approved an escalation. Where the vendor model identity is exposed, it belongs in the record too. These records let a risk leader answer an examiner on what the tool was permitted to do, what it received, and which outputs reached a calculation.

    Trussed is one control option for this layer. It enforces policy in the execution path of individual interactions and tool calls, with allow, redact, block, rate limit and escalate for human approval decisions enforced at runtime, not reviewed after the fact. The audit-ready evidence is generated by that enforcement and mapped to frameworks including NIST and NYDFS. It can be self-managed in the customer cloud, with configurable retention and data residency. Whether a particular add-in can be routed through that path depends on how it connects, so that should be established during discovery.

    Trussed has run continuously in production in a regulated enterprise for more than 18 months and operates 100+ models under one policy plane. Insurance, a regulated financial sector, is where it is deepest. It does not claim bank deployments, and the control logic in this guide applies regardless of which tooling a bank selects.

    Path of a spreadsheet AI interaction

    Workbook

    Cash positions, forecasts and counterparty data sit in cells and ranges.

    Copilot or add-in

    A prompt and selected cell contents are assembled for an external model.

    Runtime decision

    Policy evaluates the interaction: allow, redact, block or escalate for human approval.

    Output into calculation

    Returned text, formulas or values may feed a liquidity or forecast figure.

    Evidence record

    Inputs, outputs, decision and policy version are retained for review.

    Evaluate runtime control for spreadsheet AI

    Review how policy enforcement and audit-ready evidence work for individual AI interactions, and how they map to your model risk management practices.

    Request a Demo
    Next step

    How does your AI governance program compare?

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