Best Practices Guide for Boards

    Shadow AI Board Questions for Banks: What to Ask the CIO and CISO Under the 2026 Model Risk Guidance

    Boards should ask the CIO and CISO whether shadow AI is observed in practice, whether policy is enforced at the moment AI is used, and what evidence would satisfy an examiner. Strong answers describe observed use, enforced controls and retained evidence. Weak answers rely on policy attestations, approved-tool inventories and after-the-fact review. As of November 2026, the April 2026 interagency model risk management guidance replaces OCC Bulletin 2011-12 as the baseline reference.

    The baseline has changed, and the questions should reflect it

    On April 17, 2026, the OCC, the Federal Reserve Board and the FDIC issued updated model risk management guidance. The OCC's release, Bulletin 2026-13, rescinds OCC Bulletin 2011-12 and OCC Bulletin 2021-19. The guidance is risk-based and is meant to be tailored to a bank's size, complexity and extent of model use. It is expected to be most relevant to banks with more than $30 billion in total assets, and may also apply to smaller banks with significant model risk exposure. As of November 2026, a board should not assume that its management reports still cite the rescinded bulletin.

    The first question for the CIO and CISO is therefore administrative but important: which model risk management guidance does the bank now apply, and how is AI use scoped within it? The follow-up is whether unsanctioned and embedded AI features are in scope or outside the inventory by default.

    The board's standing to ask is clear. OCC governance guidance expects directors to ask incisive questions, to require accurate and timely answers, and to understand the bank's IT risks and controls. It expects the CIO or CISO to brief the board periodically. OCC has also said examiners consider how new activities affect a bank's risk profile and how effective governance and risk management are. Shadow AI is a new activity that most banks do not yet measure.

    Visibility and data exposure

    Question 1: How do we know which AI tools employees actually use, including websites, coding assistants, browser extensions and AI features embedded in approved software? A credible answer names the observation method, its coverage across managed devices and networks, and the date of the last measurement. A weak answer points to the approved-tool list or to training completion rates.

    Question 2: What categories of customer, supervisory or confidential data have been sent to external AI services, and how was that determined? A credible answer separates confirmed events from estimates and states what the method cannot see. Discovery tools show that a service was reached, but often not what was submitted. Data loss prevention catches patterns it was configured to recognize and can miss free-form text. A weak answer says no incidents have been reported, which describes reporting, not exposure.

    Evidence to request: a recent report of observed AI destinations by business unit, and a documented list of coverage gaps such as unmanaged devices and personal accounts.

    Enforcement versus policy-only controls

    Question 3: When an employee submits sensitive data to an AI service that policy does not permit, what happens at that moment? The FTC has told boards to ask whether policies address security risks and whether actual practices address threats. That distinction applies directly here. A credible answer describes a control in the execution path that can allow, redact, block or escalate the interaction for human approval, with named owners for exceptions. A weak answer describes an acceptable-use policy, a periodic log review or a quarterly sampling exercise.

    Question 4: Which controls operate before data leaves, and which operate only afterward? After-the-fact review has value for investigation and trend analysis. It cannot undo a disclosure that has already occurred. Boards should expect management to say plainly which risks are prevented, which are detected and which are only documented.

    Question 5: How are exceptions handled? Ask how many interactions were blocked, redacted or escalated last quarter, who approved the escalations and how long approval took. A control that never produces any of these records is either unused or not measuring what it claims to measure.

    Evidence to request: enforcement rules in force, counts of enforced outcomes by type, and the exception and approval record.

    Code provenance, third parties and model risk alignment

    Question 6: What share of production code is produced with AI assistance, and how do we know the origin, licence and security context of that code? A credible answer shows that AI-generated code passes through the same review and scanning as other code and that provenance is recorded. A weak answer states that developers are instructed to review AI output.

    Question 7: How are third-party AI services covered by due diligence, contract management and ongoing oversight? OCC has said that AI risk management includes appropriate due diligence whether a tool was built by the bank or obtained from a third party, and that oversight should scale with the criticality of the service. Ask whether externally hosted services and embedded AI features that nobody formally procured are in the third-party inventory. Ask what happens to a service that is discovered in use but has no contract.

    Question 8: How does the AI use inventory connect to the model inventory? Under a risk-based approach, scope depends on the bank's profile, so management should be able to explain where each AI use sits, why it is in or out of model risk management scope, and who made that decision.

    Evidence to request: the AI use inventory reconciled to the model inventory and the third-party register, with discovered-but-unapproved services flagged.

    Evidence quality and where runtime control fits

    Question 9: If an examiner or internal auditor asked us to show that AI controls operate, what would we hand over today? OCC examiners scope their work around governance and risk management effectiveness, so evidence that controls operate carries more weight than evidence that policies exist. The insurance sector offers a parallel. NYDFS Insurance Circular Letter No. 7 (2024) tells insurers that the board should receive timely and relevant facts about AI use, and that senior management owns effective challenge, independent risk assessment, review of internal audit findings and prompt remediation. It is insurance guidance, not bank guidance, but it shows what board-level AI oversight looks like in a regulated financial sector.

    Strong evidence is produced by the control itself, at the time of the interaction, and is retained for a defined period. Evidence assembled afterward from scattered logs is harder to defend. Runtime policy enforcement in the execution path generates this kind of record as a by-product: what was requested, which policy applied and what the outcome was.

    Trussed enforces policy in the execution path across AI sites, coding assistants, models, agents and tool calls, and produces audit-ready evidence mapped to NYDFS and NIST requirements. Its deepest regulated experience is in insurance, and it has run continuously in a regulated enterprise for more than 18 months. These questions are useful whichever tooling the bank selects.

    Reading the answers: runtime control or after-the-fact review

    Use this summary to classify what management says. Several weak answers in a row suggest the board is receiving assurance about policy rather than evidence about control.

    • Visibility: Strong: measured AI use with stated coverage gaps. Weak: approved-tool inventory only.
    • Data exposure: Strong: confirmed events separated from estimates. Weak: no incidents reported.
    • Enforcement: Strong: allow, redact, block or escalate at the moment of use. Weak: policy and periodic review.
    • Code provenance: Strong: origin, licence and security context recorded. Weak: developers told to review output.
    • Third parties and MRM: Strong: discovered services reconciled to registers and inventories. Weak: separate lists that are never compared.
    • Evidence: Strong: records generated by the control and retained. Weak: reports assembled on request.

    Six oversight themes for the board

    Visibility

    AI use observed in practice, not only approved tools

    Data exposure

    What customer and supervisory data reached external services

    Enforcement

    Policy applied when AI is used, not reviewed afterward

    Code provenance

    Origin, licence and security context of AI-generated code

    Third parties and MRM

    AI use tied to due diligence and model inventories

    Evidence

    Artifacts an examiner or auditor could test

    Put the question bank to work

    Talk with an expert about how runtime policy enforcement and audit-ready evidence can support board oversight of AI use in a regulated enterprise.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

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