Investing in an AI compliance assistant looks like a capability decision, with accuracy benchmarks and citation quality and coverage of the regulatory library. However, it’s also important to understand what happens to the bank's information once it leaves the bank, as that decides whether staff can use the tool for the work they need it for.
This article covers the data classes a regulatory Q&A platform touches, the restriction on disclosing supervisory information to third parties, how the interagency third-party guidance applies to a conversational tool, the contract terms that decide the outcome, and where model risk guidance stops being useful.
The Confidential Supervisory Information Boundary
Confidential supervisory information is the category that decides how useful an AI compliance assistant can be in practice. The table below sorts the four data classes by who may see them and what the platform has to demonstrate.
| Data class | Example | Who may receive it | What the platform must prove |
|---|---|---|---|
| Published guidance | A CFPB rule, an OCC bulletin | Anyone | Nothing. Public record |
| Internal policy | The bank's BSA procedures | Staff and contracted parties | Confidentiality, tenancy isolation, no training use |
| Customer NPI | Loan files, account records | Staff and service providers under GLBA safeguards | Encryption, access control, documented service provider oversight |
| Supervisory information | A draft exam finding, an MRA | Officers, directors, employees, attorneys, auditors, and consultants under written contract | A written contract meeting 12 CFR 4.37(b)(2), or Comptroller permission |
For national banks, non-public OCC information is defined at 12 CFR 4.32(b) and covers records the agency creates or obtains in supervising, licensing, regulating and examining institutions, including reports of examination and supervisory correspondence.
Section 4.37(b)(2) permits a bank to disclose that information to people officially connected with it as officer, director, employee, attorney, auditor or independent auditor, where disclosure is necessary for business purposes.
The same provision extends to consultants under written contract, provided the consultant agrees in writing that it is aware of the restrictions on dissemination and will use the information solely for the services covered by the contract.
Anything outside those categories requires the prior written permission of the Comptroller under 4.36(d) and 4.37(b)(1). The Federal Reserve's parallel rule sits at 12 CFR 261.21 and the FDIC's at 12 CFR part 309.
A vendor whose master services agreement acknowledges the dissemination restrictions and limits use of the information to the contracted service is standing in the position the rule already contemplates.
The OCC proposed changes to these rules on 3 August 2026. This would establish confidential supervisory information as a distinct subcategory of non-public OCC information and add exceptions for:
- Business efficiency
- Government accountability
- Supervisory coordination,
Until it is finalised, the existing provisions govern.
Vendor Due Diligence for an AI Compliance Assistant
The Interagency Guidance on Third-Party Relationships issued by the Federal Reserve, FDIC and OCC on 6 June 2023 already supplies the structure. It runs on a lifecycle of planning, due diligence and third-party selection, contract negotiation, ongoing monitoring, and termination. The general third-party lifecycle is covered elsewhere.
- At planning, define which of the four data classes staff will be permitted to enter, and write that into the use policy before procurement.
- At due diligence, ask where inference runs and on whose infrastructure.
- At contracting, the terms below carry the weight.
- At ongoing monitoring, see the model behind the assistant and the list of parties processing data on the vendor's behalf.
- At termination, establish what happens to the query history and any uploaded documents, in what format they come back, and how deletion is evidenced.
Contract Terms That Decide the Outcome
Six terms stand out when it comes to contracts:
Training use
The agreement should state that bank inputs and outputs are not used to train the vendor's models or any third party's models.
Retention and deletion
Specify how long prompts, uploads and outputs persist, and what deletion on termination means technically, including backups.
Tenancy and encryption
Logical separation of the bank's data from other customers', encryption in transit and at rest, and key management arrangements the bank can describe to an examiner.
Subprocessor disclosure
A current list, advance notice of additions, and a defined objection route.
Audit rights and evidence
The right to receive a SOC 2 Type II report annually, plus penetration test summaries.
Incident notification
A defined timeline, a named contact, and content requirements, aligned with the bank's own incident response obligations.
Query Logs Are Compliance Records
A pattern of questions about flood insurance determination timing is a leading indicator worth reading, and query data has predictive value for examination scope. It is also a document that exists, that can be requested, and that may be discoverable.
Treat the query log as a compliance record with an owner, a retention schedule and access controls. Three capabilities follow from that:
- A bank should be able to export the full log in a usable format.
- It should be able to restrict who reads it
- It should be able to apply its existing records retention schedule
Platforms like Ask Kaia, built for regulated institutions, are self-hosted and trained exclusively on federal banking regulations. It does not share client data with commercial AI models and logs all agent executions so they can be audited.
These are the properties worth confirming in any vendor, in writing.
Where Model Risk Guidance Stops
On 17 April 2026 the OCC, Federal Reserve and FDIC issued revised guidance replacing OCC Bulletin 2011-12 and the corresponding SR 11-7. That revision takes a risk-based approach tailored to an institution's model risk profile.
The revised guidance expressly excludes generative and agentic AI models, describing them as novel and rapidly evolving, and the agencies said they intend to seek comment on bank use of AI separately. A bank therefore cannot present model validation as the governance answer for a generative Q&A tool.
What remains is the framework already described:
- Information security standards
- Third-party risk management
- Records governance
- The institution's own AI use policy
That is a narrower foundation than model risk management, and it puts more weight on the contract.
Frequently Asked Questions
What is confidential supervisory information?
Confidential supervisory information is information a banking agency creates or obtains through supervision, licensing, regulation and examination, including reports of examination, supervisory correspondence and ratings. The OCC defines non-public OCC information at 12 CFR 4.32(b). The Federal Reserve addresses the category at 12 CFR 261.21 and the FDIC at 12 CFR part 309.
Is ChatGPT safe for bank compliance questions?
For questions about published regulation that any member of the public could ask, the confidentiality risk is limited. For questions containing customer information, internal policy detail or supervisory correspondence, a general consumer service is the wrong venue, because the terms of service, retention behaviour and subprocessor arrangements are not negotiated with the bank.
Does model risk management guidance cover generative AI?
Not currently. The revised interagency model risk management guidance issued on 17 April 2026 explicitly excludes generative and agentic AI models on the grounds that they are novel and rapidly evolving, and the agencies indicated they would address bank use of AI through a separate request for comment. Governance for a generative assistant therefore rests on information security standards, third-party risk management and the institution's own policy.
Capability evaluation comes next, and it is a different exercise with different criteria. The companion explainer on evaluating an AI compliance solution covers accuracy, citation quality and coverage testing.
Learn how Ask Kaia can assist your organization’s compliance team in gaining clarity on regulatory changes.
Request Demo- Policy Drafting
- Compliance Automation
- Audit Trails
- Regulatory Intelligence