Rail Governance

What Bank Examiners Expect to See in a Vendor SOC Report

A clean SOC opinion is only a start for AI vendors.

· 9 min read
Cover illustration for “What Bank Examiners Expect to See in a Vendor SOC Report”
SOC 1 vs SOC 2 Decisions · November 15, 2026 · 9 min read · 2,069 words

A vendor SOC report gives bank examiners evidence about defined controls within the broader assessment of AI risk. For AI systems that can execute transactions, a clean opinion is only the starting point. The bank must also show that the report’s scope, its vendor agreements, and its own oversight reach the decisions and actions for which the bank remains accountable.

Why examiners are scrutinizing AI vendor SOC reports more closely now

Examiners already review third-party SOC reports, but the guidance for those reviews changed in 2026. The April revised interagency model risk guidance, identified as OCC Bulletin 2026-13, SR 26-2, and FIL-15-2026, excluded generative and agentic AI as “novel and rapidly evolving.” That left a visible governance gap as deployment accelerated.

OCC Bulletin 2026-54, “Super Intelligence: Supervisory Framework,” addressed that gap on November 4, 2026. It is an examiner resource. The framework treats AI as amplifying or modifying existing operational, compliance, and credit risks. It also identifies third-party risks that conventional service-level and information-security reviews do not fully address. The bank’s accountability remains unchanged regardless of who builds or hosts the system.

What a SOC report actually covers, and the structural limits that apply before AI enters the picture

A SOC 2 Type II report assesses the system and criteria the vendor chose to describe over a defined audit period. It is not continuous or comprehensive certification. The Trust Services Criteria contain one mandatory category, Security, also called Common Criteria, and four optional categories, Availability, Processing Integrity, Confidentiality, and Privacy. None contains AI-specific language, and assessment depends on the vendor’s description of its system.

Those boundaries are part of the report’s design. A vendor can define scope narrowly enough to exclude inference pipelines, foundation model integrations, or retrieval layers without that exclusion appearing as a deficiency in the opinion letter. A clean opinion therefore concerns the described controls within the defined scope. It does not establish that those controls answer every question about an agent executing banking transactions. The examiner must first determine what was actually assessed.

Six areas where AI-specific risk goes untested in a standard SOC report

Answer accuracy is the first gap to examine. Processing Integrity asks whether all the data is processed correctly and on time, as the vendor described. It does not establish that a generative model’s answer is factually correct. The bank should ask the vendor whether output accuracy has been assessed separately from correct processing.

A system can give wrong answers that sound authoritative while processing data as designed. For an agent executing transactions, that misleading answer can become an unauthorized action. What evidence shows how often the system produces such responses?

Model governance and provenance are the third gap. The framework’s Control pillar asks who is responsible for behavior, how model changes are documented, and what validation has occurred. A SOC report from an earlier period does not show how a later foundation model swap, training-data change, or inference adjustment was governed. Has the vendor documented and disclosed those changes?

Retrieval-time permissions are a separate control question. Application-level logical access controls do not necessarily show whether the AI checks a user’s permissions each time it retrieves information. Access established at index time may not preserve the bank’s role, department, and document boundaries at query time. The examiner needs evidence that retrieval respects those boundaries.

The bank bears the compliance risk when outputs show bias. No Trust Services Criterion specifically tests for disparate impact in model outputs. Fair lending and consumer protection obligations still attach to the bank when a vendor’s AI produces credit-related outputs. A clean SOC report cannot substitute for evidence addressing that risk.

Queries sent through a hyperscaler’s foundation model API may place customer data in an environment the vendor does not fully control. Even when the vendor chooses Confidentiality or Privacy coverage, that environment may fall outside it. Does the report actually include the inference layer and its data handling?

How subservice organization coverage determines whether a vendor's SOC report reaches the AI components that actually matter

A clean opinion on the vendor’s systems may say nothing about its foundation model provider, cloud inference layer, or data pipeline vendor. Under the carve-out method, subservice organizations are excluded from the report’s scope. Under the inclusive method, their controls are covered within the same report. That choice determines whether critical AI dependencies appear in the assessment.

The framework’s third-party risk section flags opaque provider chains and weak contractual rights. Its companion, OCC Bulletin 2023-17, supplies lifecycle guidance on subcontractor due diligence and audit rights. Together, they leave coverage of a carved-out foundation model provider unresolved for a Tier 1 or Tier 2 use case.

The chain may extend beyond one subcontractor. An AI vendor can depend on a foundation model provider and additional data or compute subcontractors, each with separate governance gaps. Examiners will seek documentation addressing training data practices, model change governance, and incident notification across those dependencies. Can the bank obtain adequate evidence, and do its audit rights reach the organizations needed to assess it?

The six OCC pillars as an examiner checklist for agentic AI vendor relationships

Purpose asks what the system is permitted to do. The bank’s evidence must define what the agent is designed to do and not do, which data and systems it can reach, and when a person must intervene or the tool must pause. An outcome that looks reasonable does not excuse an action outside that purpose. Does the vendor file define, in testable terms, what the agent may initiate, approve, and decline?

Risk links the strength of controls to what could happen in each use case. The framework’s illustrative four-tier scheme places autonomous payment agents and real-time credit decisioning at Tier 1, critical. It places AI-assisted underwriting and agent-initiated customer communications at Tier 2, high, and document summarization at Tier 3, moderate. Sophistication alone does not determine the required scrutiny. Has the bank assigned each use case a tier, supported by the vendor’s documentation?

Control asks what prevents unauthorized or unsafe action. Its scope includes human involvement, model risk management, data governance, privacy, compliance, and cybersecurity. The framework treats agent credentials and permissions as comparable to privileged user access, so the agent’s identity is itself an examination subject. Does the system enforce transaction limits, role-based permissions, and delegated authority constraints at that identity level?

Oversight requires both monitoring and an accountable owner. A named bank employee needs to know how the vendor’s system behaves to oversee it properly. The examiner therefore needs evidence of the bank’s governance assignment and the visibility available to that owner. Does the vendor provide real-time or near-real-time monitoring that supports the bank’s responsibility for the use case?

Resilience extends beyond keeping a service available. The framework describes an agent that “could remain available while taking inappropriate or harmful actions” and names circuit breakers, rate limits, least privilege, and kill switches as safeguards. A Wolters Kluwer survey published June 10, 2026 found that kill-switch protocols or regulatory reporting of AI failures were areas where most U.S. banks said they were least prepared. Can the bank immediately suspend or revoke agent access without vendor involvement, and has it tested that capability?

Under Assurance, the framework calls for testing that seeks unknown failure modes, including adversarial testing, red teaming, hallucination rate measurement, and expert review. Documentation must also let a third party reconstruct a material AI decision without developer support. Does the vendor produce sufficient human-readable audit trails, and has the bank conducted or obtained independent testing beyond the vendor’s attestation?

Four vendor contract terms OCC examiners will specifically look for in an AI vendor relationship

Permitted scope must be specific enough to test. A general description of the agent’s role does not establish what it may initiate, access, or decline inside bank systems. Examiners will look for contractual limits that correspond to the bank’s approved purpose for the tool.

AI incident definitions must cover more than ordinary outages. A model can remain available while producing wrong outputs, an agent can exceed its transaction scope, or an authorization chain can fail silently. The agreed definitions and notice periods show whether the vendor must report those failures as well as downtime and data breaches.

Liability allocation must state who is responsible for errors and harm to customers. Examiners will scrutinize whether the agreement clearly assigns responsibility for AI-related failures. A contract that defaults all liability to the bank for outputs produced by the vendor’s model raises a risk-management finding.

Advance notice must cover vendor-added AI features. New agentic capabilities in an existing product can create an unreviewed high-consequence use case without a new purchase. The notice provision therefore supports more than contract administration. It gives the bank information needed to keep its AI inventory current.

What the bank's AI inventory must show before the examiner reviews a single vendor SOC report

The inventory establishes the context for each vendor review. Before interpreting a SOC report, the examiner needs to know whether the bank has recorded the system, captured vendor-enabled features, assigned each use case a tier, and named an accountable owner. Without that map, report coverage cannot be matched to the bank’s actual use. A technically thorough vendor assessment may still concern an incomplete picture of deployment.

Feature additions can make that picture stale silently. A vendor that adds an agentic payment capability to an existing fraud-detection product can move a Tier 3 use case to Tier 1 without an internal purchase triggering review. Advance notice provides the opportunity to reassess it, but the inventory must then be updated.

Data visibility is part of this preparation too. Rapid enterprise AI rollouts can leave compliance unable to describe what data each system touches. That condition makes the missing inventory entry a data governance finding. The practical prerequisite is an inventory detailed enough to connect the system, its use, its risk tier, and its owner to the controls under review.

Why agentic AI executing real transactions raises the examiner's questions to a different order of magnitude

Autonomous payments act directly, are hard to reverse, and have a large financial impact. That combination explains their Tier 1 treatment, regardless of the model’s sophistication. A gap in permissions or oversight has different consequences when an agent moves funds than when a tool drafts text. The framework’s privileged-user treatment means the bank must audit the agent’s credentials and permissions as it would those of a privileged human user.

The SOC scope question consequently becomes more exacting. Does the report cover payment authorization logic, how permissions are enforced at each transaction step, and how audit trails are generated? If a subservice organization performs any of those functions, the examiner needs to know whether its controls are included or carved out. A general description of application security cannot show where that coverage starts and ends.

Authorization traceability is the decisive test for each payment. The record must explain why the payment was authorized, which policy governed it, whose delegated authority applied, and where authorization occurred in the transaction chain. A compliance officer or regulator must be able to reconstruct that decision without developer support. Without that human-readable record, the bank cannot demonstrate the accountability the Assurance pillar calls for.

What well-structured AI vendor governance gives an examiner before they ask for it

A prepared bank presents records that connect to one another. Its inventory entry identifies the vendor system, tier, owner, and an update date after the latest vendor feature change. Its signed agreement sets out the permitted scope, incident definitions and notice periods, liability allocation, and requirements for advance notice. Those records establish what the bank approved before the examiner evaluates how it is controlled.

The SOC report then has a defined evidentiary role. Its scope covers the payment authorization logic, retrieval permissions, and audit trail generation used by the bank, and its subservice coverage includes the foundation model or inference layer where used. Separate testing records show what adversarial testing, hallucination measurement, or expert review examined and found. The bank also retains human-readable records for every agent-initiated action and a tested kill-switch process documenting capability, authority, triggers, and test results.

Vendors designed for institutional governance support this posture through normal operation. Full audit trails, configurable controls, human overrides, and SOC 2 certification each supply a different part of the evidence, so none can take the place of the others. Documentation produced as the system operates is more dependable than a file assembled under examination pressure.

Sources

  1. OCC Bulletin 2026-54: Examiner Framework for Agentic AI in Banking

More in SOC 1 vs SOC 2 Decisions