Best Practices Guide

    Central Funding vs Chargeback for a Shared AI Control Plane in Banks

    Neither pure model is sufficient alone. Central funding best protects uniform enforcement and consistent evidence, chargeback best preserves cost accountability, and a hybrid is usually the easiest to defend: fund baseline coverage centrally and allocate only incremental usage. Whichever model you choose, mandatory coverage should be set by risk policy, not by who pays. This comparison is analytical, built on the revised interagency model risk guidance as of November 2026.

    Questions to raise before procurement

    • Which function owns the budget, and who approves exceptions or opt-outs from enforcement?
    • Under this model, can any business unit send AI interactions or tool calls around the shared layer, and how would we detect it?
    • What evidence must every line of business produce, who retains it, and does funding affect its completeness?
    • Which allocation driver best avoids discouraging high-risk use cases?
    • Are early-stage units charged, subsidized, or exempt from cost, and are they still covered by enforcement?
    • When will we revisit the model after the revised model risk guidance and the third-party proposal are applied in examinations?

    Funding is a risk-control decision, not only a finance one

    A shared layer that enforces policy on AI interactions and tool calls across business lines only works if every business line is behind it. How it is paid for therefore shapes who is covered, who can step out, and whether evidence looks the same everywhere. Those are risk outcomes that model risk management teams and examiners will ask about.

    The supervisory baseline has also moved. On April 17, 2026, OCC Bulletin 2026-13 announced updated interagency model risk management guidance from the OCC, the Federal Reserve Board, and the FDIC. It clarifies model risk management principles, sets out a risk-based approach, and rescinds prior model risk management guidance and other issuances. Funding cases written against earlier guidance should be rechecked against the revised text.

    The revised guidance covers governance and controls, model validation and monitoring, and considerations for vendor and third-party products. It describes effective governance as including clear policies and roles and responsibilities, and it notes that practices appropriate for one banking organization may be inappropriate for another with a different risk profile or model purpose. Read plainly, coverage should follow risk, and the funding model should not decide scope.

    Comparing the three models against risk criteria

    The assessments below are analytical judgments, not findings from a survey of bank practice. They assume the shared layer sits in the execution path, so any interaction that bypasses it has neither enforcement nor evidence.

    • Central funding: Uniformity of enforcement: high, because coverage does not depend on a business-unit budget. Evidence completeness: high, if one schema is mandated. Business-unit incentives: weak, since consumption has no visible cost. Budget governance: concentrated in one owner, with a risk of weak prioritization.
    • Usage-based chargeback: Uniformity of enforcement: variable, because units may delay onboarding, opt out, or route around the layer to avoid cost. Evidence completeness: uneven where usage is partial. Business-unit incentives: strong cost awareness. Budget governance: distributed, with disputes over allocation.
    • Hybrid: Uniformity of enforcement: high for the funded baseline. Evidence completeness: high where the baseline includes evidence generation. Business-unit incentives: moderate, tied to incremental usage only. Budget governance: shared, which needs explicit approval roles.

    Where funding choices create coverage gaps

    Chargeback carries the most visible gap risk. If a unit pays per interaction, the cheapest path is to send fewer interactions through the layer, or none. Opt-outs, unsanctioned tools, and shadow AI usage all follow from that incentive. Each leaves interactions without enforcement and without evidence, and the unit least likely to onboard may be the one with the highest-risk use cases.

    Central funding has the opposite weakness. When consumption is free to the consumer, prioritization and accountability can fade. Low-value use cases may crowd the platform and the central owner may struggle to justify growth. That is a governance problem, but it is less likely to produce a coverage gap an examiner finds.

    Staged adoption adds a wrinkle. Commenters to the Federal Reserve have described production rollouts that test on small representative slices, such as 1% of business and then 5%, and expand as confidence grows. A funding model should state whether early-stage units are charged, subsidized, or exempt from charges. They should never be exempt from enforcement.

    The design principle is to separate who pays from who must be covered. Mandatory coverage can be set by risk policy regardless of the funding mechanism. Likewise, one evidence schema and retention standard owned centrally keeps evidence uniform even when costs are allocated unevenly.

    Choosing an allocation driver and fitting it to your organization

    If any portion of cost is allocated, the driver matters. Per-interaction volume tracks consumption closely but penalizes high-traffic units and rewards throttling. Counting models or agents is simple, but it can discourage registering new agents and encourage unregistered ones. Business-unit count is predictable and ignores actual use, so small units can feel overcharged. Risk tier is the most consistent with a risk-based approach, since commenters describe validation prioritized by complexity and the risk of failure or misuse. Not every driver has been shown in use for chargeback, so test each one against a single question: does it discourage onboarding of high-risk or low-volume use cases?

    As a rule of thumb, central funding fits institutions still establishing baseline coverage, where uniformity matters more than cost signals. Chargeback fits only where mandatory coverage and evidence standards are already enforced independently, so cost allocation cannot change who is covered. Hybrid fits most institutions between those poles: the central budget covers the enforcement baseline and evidence generation, and business units fund incremental volume or specialized policy needs.

    Third-party expectations are also in motion. Interagency third-party guidance was issued as OCC Bulletin 2023-17 in June 2023, and the FDIC published a proposed interagency third-party risk management guidance on September 11, 2026, which is pending as of November 2026. NYDFS also announced guidance on AI-related cybersecurity risk in October 2024. If the shared layer comes from an outside provider, plan to revisit the funding model once the proposal is finalized.

    Best practices for structuring the funding model

    These practices keep the funding decision from changing the risk-control scope.

    • Set coverage by risk, not budget: Define mandatory coverage by model and use-case risk tier, and apply it to every business line regardless of who funds it.
    • Fund the baseline centrally: Put enforcement of the baseline policy set and evidence generation in the central budget so they cannot be cut at unit level.
    • Own one evidence standard: Define a single evidence schema and retention period centrally, and state who retains evidence and for how long.
    • Control exceptions: Name who can approve an opt-out or exception, log every one, and review them on a schedule.
    • Detect bypass: Document how shadow AI usage and unmonitored tool calls are identified and escalated. This needs internal design.
    • Approve jointly: Have model risk management, technology, and finance approve the model together, with a planned review after guidance changes.

    Funding model test: does coverage stay uniform?

    Central funding

    Strong on mandatory coverage; weaker on cost accountability.

    Chargeback

    Strong on consumption alignment; exposed to opt-outs and uneven evidence.

    Hybrid

    Baseline coverage funded centrally; incremental usage allocated.

    Deciding test

    Can any interaction or tool call bypass enforcement or evidence?

    Test your funding model against enforcement and evidence

    Trussed enforces policy in the execution path, with allow, redact, and block decisions on individual interactions and tool calls, and produces audit-ready evidence from that enforcement. Talk to an expert about what uniform coverage and evidence require in a regulated enterprise.

    Talk to an Expert