For community banks and credit unions.
Prepare the records behind the eight AI questions.
Use these questions to get your records ready for the board or an exam.
Who the framework covers
CSBS released the guide on September 16, 2026 for state bank and nonbank exams. Each state decides if and how to use it.
It creates no new legal obligations. The scope of each exam depends on how the institution uses AI.
The OCC oversees national banks. Credit unions can use these questions to sort their records. See the NCUA sources below.
CSBS framework and Core Examiner Guide. Checked October 16, 2026.
The eight starting questions
- Where does AI support your institution?
- Which AI tools and uses have owners?
- Does AI face customers or influence decisions?
- Which AI capabilities come from vendors?
- Where is AI embedded in vendor products?
- Which generative AI tools are in use?
- How does your institution separate uses by risk?
- What sensitive information passes through AI?
The paraphrases follow Initial Scoping on guide page 3.
The 19 records examiners may request
The guide lists 19 items in four groups on page 4.
The records you need depend on how you use AI. If an item does not apply, note why.
Follow each link to see where that record fits in the example. Check your own files for the records you hold.
If they apply, find your own examples of AI outputs, complaint records and model tests.
A link tells you where to look. Check the record before you rely on it. Keep your decision separate.
| Record examiners may request | Where your answer lives |
|---|---|
| AI policies, standards or procedures | Policy register and policy gaps |
| Governance, committee or oversight materials | Accountable roles and review responsibilities |
| Management or board reporting on AI use and risks | Board briefing and conditions |
| Reports or dashboards monitoring AI use, performance, issues and risks | Action register and change history |
| Employee guidance or training on AI | Safe-use guidance and unresolved policy questions |
| Record examiners may request | Where your answer lives |
|---|---|
| Inventory of AI systems, tools, models or uses | AI use register |
| Business purposes, ownership and use | Use descriptions and accountable roles |
| Use descriptions, risk assessments, ratings or tiering | Proposed tier worksheet and reasons |
| Vendor products with known or possible embedded AI | Vendor disclosure and open questions |
| Change management records for AI systems, models, tools or uses | Retained baseline and change record |
| Record examiners may request | Where your answer lives |
|---|---|
| Policies, guidance or limits for generative AI and language models | Safe-use policy gaps and treatment |
| Approved or known generative AI tools in use | Use register and approval questions |
| Business purposes, data controls and output review | Use boundaries and evidence assessment |
| Record examiners may request | Where your answer lives |
|---|---|
| Vendor or third-party oversight of AI capabilities | Vendor sources and assigned questions |
| Model validation, testing, monitoring or control materials | Evidence assessment and missing test records |
| Consumer compliance, notices, complaints or remediation | Evidence gaps and actions for the institution |
| Recent AI-generated or AI-assisted consumer-facing outputs | Output samples linked to their versions and review |
| Terms on data use, retention, sharing or AI training | Contract sources and unanswered handling questions |
| Internal audit, compliance reviews, control testing, tracking logs or independent assessments | Scoped review observations and board recommendation |
CSBS Core Examiner Guide, page 4. Checked October 16, 2026.
How the optional tiers work
The optional worksheet asks four things. Who could be affected? Who checks the work? What if it fails? What data can it touch?
- Tier 1: Lower-impact internal use with reviewed outputs and limited sensitive data.
- Tier 2: Consumer-facing or decision-support use, or greater impact and less direct review.
- Tier 3: Significant consumer outcomes, sensitive data, operational dependence or limited human review.
The highest score usually sets the first tier. Write down your reason for any change. Seek governance approval if you lower the tier.
Tiers follow the optional CSBS worksheet. Your institution assigns the final tier.
Read the official optional worksheetChecked October 11, 2026.
Agent controls deserve their own evidence
Guide page 7 looks at agents that can act with little human input. It asks what they can do and where a person steps in. It also asks what gets logged, what can be undone and who can stop the agent.
Review agent authorityCredit union and federal applicability
OCC framework for national banks
The OCC issued its super intelligence framework on November 4, 2026. It guides examiners and sets no supervisory expectations for banks. Examiners may use it for community banks and tailor their review to the bank and use. Super intelligence is part of existing risk categories. Controls may increase if a failure could cause more harm. OCC Bulletin 2026-54. Checked November 7, 2026.
NCUA AI resources cover governance and vendor due diligence. Its 2026 priorities do not explicitly list AI.
SR 26-2 leaves out generative AI and agents. It is most relevant to banks the Fed regulates with more than $30 billion in assets. Checked November 7, 2026.
On September 11, 2026, the FDIC, Federal Reserve, NCUA and OCC requested comment on proposed third-party risk management guidance. Comments due November 16, 2026. Read the Federal Reserve release. OCC bulletin. Checked November 10, 2026.
Sources checked November 10, 2026.