Compliance Guide

    Is an AI Control Plane Vendor a Bank Service Provider Under the Bank Service Company Act?

    It depends on the facts, and the vendor's own description does not settle it. Under FDIC rules as of November 2026, a bank service provider is a person that performs services subject to the Bank Service Company Act. Whether an AI runtime enforcement vendor does so turns on what it performs, whether it performs it for the bank, what data it can reach and how it is deployed. That is a legal determination for counsel and the primary regulator. The due diligence, contract and evidence work is the same whatever the answer.

    Mapping runtime functions to the services question

    1. 1

      Policy enforcement

      The layer evaluates each model call, tool call or agent action against rules the bank sets. Document who authors the rules, who operates the enforcement point and whether the vendor can change policy without the bank.

    2. 2

      Redaction

      Redaction requires reading the content of prompts or outputs. Record whether this happens inside the bank's environment or on vendor infrastructure, and whether any unredacted content is retained.

    3. 3

      Blocking

      A layer that can stop a request sits in the execution path. Define fail-open and fail-closed behavior and what a vendor outage does to the bank's AI-supported activity.

    4. 4

      Human escalation

      Escalation routes a decision to a bank reviewer. Note who defines approvers, who sees the escalated content and how decisions are recorded.

    5. 5

      Evidence generation

      Evidence records what the AI was allowed to do and what it did. Identify who holds the records, how long they persist and whether the bank can export them without vendor involvement.

    What to define regardless of classification

    • Document the vendor's actual functions, data flows and deployment model before asking counsel or the regulator for a view, and do not rely on vendor assertion.
    • Evaluate section 7 notification at contracting time against the 30-day period, and record the analysis and its outcome.
    • Write contract terms for incident notification, business continuity, audit and examination access, subcontractors and exit in enough detail for the bank to manage them. FDIC examiners have noted shortfalls in business continuity and incident response terms in technology service provider contracts.
    • Set limits on agent and tool activity: scope of authority, identity, permitted tool use, data use and escalation, with the ability to limit actions without ending the vendor relationship.
    • Retain evidence an examiner can review, including contracts, monitoring results, incidents, overrides, remediation and the underlying records, not only management summaries.
    • Track concentration, switching constraints and opaque provider chains as ongoing third-party risk indicators.

    What the statute and the rules establish

    The Bank Service Company Act is codified at 12 U.S.C. 1861-1867. Section 7 deals with notification of service relationships. As the FDIC quotes section 7(c)(2), an FDIC-supervised institution that has services performed by a third party must notify its agency of the existence of the service relationship within 30 days after the making of the service contract or the performance of the service, whichever occurs first. The FDIC states that, depending on the services provided, entering into a contract or having a service performed may trigger this requirement, including for technology-related services such as internet banking and data processing.

    The classification also matters beyond notification. The Congressional Research Service describes the Act as giving federal depository institution regulators authority to examine and regulate technology service providers that provide services to banks. FDIC rules on computer-security incidents define covered services as services performed by a person that are subject to the Bank Service Company Act. A bank service provider is a bank service company or other person that performs covered services. The definition therefore follows the scope of the Act, not the vendor's marketing category or contract language.

    The examples in the regulatory material, such as check and deposit sorting and posting, statement preparation, bookkeeping and accounting, describe core banking processing. Whether runtime policy enforcement on AI activity is comparable is a question of statutory interpretation. It should be answered by counsel reading the section 7(c) definition against a documented description of the vendor's functions. OCC testimony offers one general data point: a bank's use of third-party AI tools can involve services under the Act, which could, depending on the facts and circumstances, include banking services supported by AI.

    Deployment model and data access

    The regulatory material links deployment origin to visibility and control. An OCC 2026 supervisory framework says AI systems may be built in-house, supplied as third-party models or consumed as a fully external vendor service, and that each carries a different visibility and control profile. It also identifies data a bank may consider, including data provided to the system and new data created through prompts, outputs, embeddings, logs and feedback. None of this says that self-managed deployment removes or confirms status under the Act. It does show which facts to establish.

    In a vendor-hosted model, the vendor may store or process prompts, outputs and logs on its own infrastructure, and a service disruption is the vendor's operational event. In a self-managed model in the customer cloud, the enforcement software runs in the bank's environment, so data residency and retention are set by the bank. Counsel should then examine what remains with the vendor: telemetry, administrative access, update channels and support access. Trussed is one example of the self-managed pattern. It enforces policy in the execution path, with deployment in the customer cloud and configurable retention and data residency. Even so, the operational role and data access of any vendor must be assessed on its facts in each model.

    One further consequence follows from the incident rule. If a vendor is a bank service provider, it must notify the bank-designated contact as soon as possible after determining that an incident materially disrupted, or is reasonably likely to materially disrupt, covered services for four or more hours. Check whether the architecture can disrupt services the bank relies on.

    How the 2023 interagency guidance and model risk fit in

    The 2023 interagency guidance on third-party relationships applies to all supervised banking organizations. It covers any business arrangement with another entity, including outsourced services. It is supervisory guidance, so it does not have the force and effect of law and imposes no new requirements. Its expectations are risk-based: management should reflect the bank's risk, complexity and size and the nature of the relationship. It notes that third parties using new technologies may present elevated operational, compliance and strategic risks. The guidance applies whether or not a vendor is ultimately classified as a service provider under the Act, so the two analyses run in parallel.

    The OCC 2026 framework reinforces that accountability for managing risk generally remains with the bank when vendors are used. It names third-party and vendor risk as a control area, citing opaque provider chains, weak contractual rights, switching constraints and concentration in common models and infrastructure. It also says banks may define the scope of authority, identity, permitted tool use, data use and escalation limits for provider agents, and may monitor and potentially limit what those agents do. The framework describes these considerations as illustrative.

    For risk leaders, the practical link to model risk management is evidence. An enforcement layer is useful to an examiner conversation only if the bank can show what the AI was permitted to do and what it did, enforced at runtime, not reviewed after the fact. A bank should confirm with its model risk function how enforcement records fit its existing validation and monitoring practices.

    Questions to take to counsel and the primary regulator

    Bring a factual record, not a conclusion. Counsel can then apply the section 7(c) definition and advise on whether notification is required. Where the answer is unclear, the primary regulator is the appropriate party to consult. Useful questions include the following.

    Which specific functions does the vendor perform on model calls, tool calls and agent actions, and are any performed for the bank in a way comparable to services within the Act? In the chosen deployment model, what data, telemetry or administrative access does the vendor retain, and what does it store or process? Can enforcement decisions such as blocks, redactions, escalations and overrides be exported and retained in a form examiners can review? Has the bank documented its view, who approved it and when it will be revisited if the vendor's role changes?

    The last question matters because classification rests on facts that can change. A move from self-managed to vendor-hosted operation, or an expansion of vendor access to logs or content, may change the analysis. For context, Trussed has been in continuous production for more than 18 months in a regulated enterprise, with insurance as its deepest sector. That experience does not bear on how a bank is classified, which remains a determination for the bank's counsel and regulator.

    Factors that bear on classification

    Nature of the services

    What the vendor does on model calls, tool calls and agent actions.

    Performed for the bank

    Whether the functions are carried out on the bank's behalf.

    Data access

    Whether the vendor can see, store or process prompts, outputs and logs.

    Deployment model

    Self-managed in the customer cloud or vendor-hosted.

    Bring a documented architecture to the classification discussion

    Review how runtime enforcement and audit-ready evidence work in a self-managed deployment, then take the facts to counsel.

    Talk to an Expert