Four Frameworks, One Exam: Getting AI-Ready Before the Examiner Calls
Executive Summary
Four AI documents now shape how a US financial institution should prepare its AI program. The FSB, CSBS and OCC documents each say they create no binding requirements, and the CRI framework is a sector-developed control library, not a regulation. Two of them show what examiners may ask.
- FSB (15 October 2026): 12 sound practices for responsible AI adoption. The global view of good governance.
- CRI FS AI RMF (version 1.0): 230 control objectives. The detailed control library.
- CSBS (approved 13 August 2026, released 16 September): an examiner guide for state-chartered banks and state-licensed nonbanks. It starts with one question: do you use AI, and can you prove you know where?
- OCC (4 November 2026): a framework for OCC examiners, built around 6 pillars and 16 components, aimed mostly at generative and agentic AI.
Read together, they agree on six things: an inventory, risk tiers, named owners, real human oversight, vendor visibility, and the ability to stop a system.
They also disagree on one detail that can cost you time. The OCC calls Tier 1 the highest risk. The CSBS worksheet calls Tier 1 the lowest. Pick neutral labels before you build anything.
This post shows how the four fit together, maps all 12 FSB practices to the examiner documents, lists the controls and evidence they point to, and ends with eight readiness questions and a 90-day plan.
A note on names
The OCC framework is not titled an AI document. OCC Bulletin 2026-54 and its attached framework use the term Super Intelligence (SI), which Executive Order 14434 (91 Fed. Reg. 63129, 2 October 2026) defines as replacing “Artificial Intelligence” in federal usage. The OCC’s own footnote says SI “refers to the technologies and systems previously encompassed by the term ‘Artificial Intelligence’” under 15 U.S.C. 9401(3). Nothing about the framework’s substance changes because of the new label, so this post uses “AI” throughout for readability and for comparison with the other three documents. The References section keeps the OCC’s own title, so expect to see “Super Intelligence” there even though the body text says “AI.”
Why this matters now
The CSBS and OCC frameworks came out seven weeks apart, on 16 September and 4 November.
Both are written for examiners. Neither creates new legal obligations. The CSBS guide says so directly. The OCC bulletin says the framework does not set supervisory expectations for banks, and that banks are not expected to use it to assess their own AI use.
That does not make them optional. They show the questions examiners may ask and the evidence they may want to see. The CSBS announcement says institutions can use its framework to assess their own AI programs and prepare for exams. The OCC takes a different position, so treat the OCC material as a view into examiner thinking, not a self-assessment standard.
Know which supervisor you answer to. The CSBS framework covers state-chartered banks and state-licensed nonbanks, and each state agency decides how much to use. The OCC framework is written for examiners of OCC-supervised banks.
Who does what
| If you are | Start with | Note |
|---|---|---|
| A bank supervised by the OCC | OCC framework for the examiner’s view; FSB and CRI to build the program | The OCC says banks are not expected to use its framework to assess themselves |
| A state-chartered bank or state-licensed nonbank | CSBS guide, work program and worksheet; nonbank supplements if you are a nonbank | Each state agency decides how much to adopt, so ask yours |
| Answerable to more than one supervisor | Both examiner documents, with neutral risk labels | The tier numbering runs in opposite directions |
| Any institution | FSB for principles, CRI for controls | Neither is an examination tool |
What each document gives you
| Document | What it is | What it gives you |
|---|---|---|
| FSB sound practices | 12 practices, non-prescriptive, not a binding international standard | The principles. Governance in practices 1 to 4, lifecycle in practices 5 to 12 |
| CRI FS AI RMF | Risk and control matrix: 4 functions, 19 categories, 72 subcategories, 230 control objectives | The control library you map your own standard to |
| CSBS framework | Core guide, work program, tiering worksheet, nonbank supplements | The examiner’s starting questions and a one-page tiering form |
| OCC framework | 6 pillars, 16 components, illustrative 4-tier method | The examiner’s view of generative and agentic AI |
CSBS says its framework builds on the NIST AI RMF and the CRI framework, and its examiner work program cites CRI control IDs by number next to its own scoping and governance questions.
Use the FSB for your principles, the CRI for your controls, and the CSBS and OCC documents as your exam rehearsal.
Where all four agree
Six themes appear in every one.
| Theme | FSB | CRI | CSBS | OCC |
|---|---|---|---|---|
| Inventory of AI use | Practice 3 | GV-1.6.1 | AI-1, IS-2 | Component 2.1 |
| Risk tiers based on consequence | Practice 5 | MP-5.1 | AI-4 and the worksheet | Component 2.2, section 3 |
| Named owners and governance | Practice 2 | GV-2.1, GV-2.3.3 | GO-1, GO-2 | Component 2.4 |
| Meaningful human oversight | Practice 10 | MP-3.5, GV-3.2 | GE-8, worksheet factor 2 | Component 2.3 |
| Vendor and embedded AI | Practice 12 | GV-1.6.3, GV-6.1 | IS-5, AI-6, GO-7 | Component 2.9 |
| Monitoring and the ability to stop | Practices 9, 10 | MS-2.4, MG-4.1 | GO-6, GE-7, GE-8 | Components 2.10, 2.12 |
Board reporting is a seventh shared theme: FSB practice 1, CRI GV-1.5.3 and GV-2.3.4, CSBS GO-4, and OCC component 2.16.
The CRI IDs for inventory, owners and vendors are the ones the CSBS work program cites directly. The other CRI IDs in this post are paired by subcategory title rather than a stated citation, so check them against your own mapping before you rely on them.
Three points worth repeating
- Inventory first. The CSBS work program says examiners look for “a current and usable understanding” of your AI footprint, not just general awareness that AI exists somewhere in the organization. The FSB report notes that banks with more limited AI adoption to date have been observed to rely on spreadsheets for their AI inventory. A spreadsheet is a fair start if it is complete and kept current.
- Oversight must be real. The OCC says approval should be “more than a nominal click.” High volumes or bulk approval could reduce review quality and contribute to alert fatigue. The FSB report suggests tracking when humans override AI recommendations and studying the patterns. If reviewers approve everything, your own data will show it.
- Know how to stop it. The OCC warns that an agentic system “could remain available while taking inappropriate or harmful actions,” requiring the bank to detect, contain, recover and remediate harm. The CSBS work program asks the same question in plain words: can you restrict or halt the system?
All 12 FSB practices, and where examiners look
The FSB report lists 12 sound practices. This table pairs each with the OCC component and CSBS item on the same topic. The pairing is by topic. Only the CSBS links to CRI are stated in the documents themselves.
| FSB practice | What it asks | OCC | CSBS |
|---|---|---|---|
| 1 Strategic direction and oversight | Board and senior management align AI with strategy and risk appetite | 2.4, 2.16 | GO-1, GO-4 |
| 2 Governance and accountability | Clear roles and a governance framework | 2.4 | GO-1, GO-2 |
| 3 AI risks in the risk framework, with documentation | Risk framework that covers AI, with records of AI use | 2.1, 2.14 | AI-1, AI-2, GO-5 |
| 4 Organizational adaptability | Adjust oversight and skills as AI evolves | 2.16 | GE-7, GO-6 |
| 5 Materiality and risk assessment | Assess each use case at the start and afterward | 2.2 | AI-4, IS-7 |
| 6 Selection process | Weigh needs, materiality and risk when choosing a model | No direct match | GO-3 (closest) |
| 7 Data governance | Data that is fit for training, testing and use | 2.6 | IS-8, GE-4 |
| 8 Explainability and transparency | Know the limits and add controls where needed | 2.5 | GE-5 |
| 9 Performance management | Test and monitor in proportion to risk | 2.12, 2.13 | GO-6 |
| 10 Human oversight | Oversight sized to risk, autonomy and complexity | 2.3 | GE-5, GE-8 |
| 11 Cyber and ICT risk management | Include AI scenarios in tests and exercises | 2.11 | GE-6 |
| 12 Third-party AI risk | Performance, transparency, data, concentration, continuity | 2.9 | IS-4, IS-5, AI-6 |
Practice 6 is the one gap. Neither examiner document has a dedicated item on model selection, so a documented selection step is yours to design.
Where they differ
The differences are small in number but easy to trip over.
| Point | OCC | CSBS worksheet |
|---|---|---|
| Number of tiers | 4 | 3 |
| Tier 1 means | Critical (highest risk) | Low risk |
| Highest tier | Tier 1 | Tier 3 |
| Scoring | 7 factors combine into a composite tier | 4 factors; the highest-rated factor sets the preliminary tier |
| Status of the method | Illustrative | A tool institutions can use; not official guidance from a state agency |
Take a fully autonomous, customer-facing payment agent. The OCC’s own illustrative table puts an autonomous payment agent at Tier 1, Critical. On the CSBS worksheet, a fully automated system (no human review before the output affects the customer) rates high risk on the human-oversight factor, and the highest-rated factor sets the preliminary tier. That makes the same agent a Tier 3 on the CSBS scale.
If both labels appear in the same report, someone will read them wrong.
Now take the opposite case, also illustrative: a tool that drafts summaries of published market research, where an analyst reviews every summary. On the CSBS worksheet it is internal use, every output reviewed, minor impact from errors, and non-personal data. All four factors rate low, so it is Tier 1. On the OCC’s composite scoring it would land toward the low end too, since it carries no direct action, full reversibility and no customer or financial data exposure, which the OCC’s own example places at Tier 4. Same tool, opposite tier numbers. Both mean “low risk.”
The fix is simple. Use four neutral labels in your own program, such as Low, Moderate, High and Critical. Then show how each maps to the OCC and CSBS tiers.
The CSBS worksheet adds three habits worth copying:
- Complete one worksheet per use case, and keep it as exam support.
- Update it when the purpose grows, the AI gets more decision-making authority, or a vendor issue arises.
- Require governance approval for any downward adjustment of a tier.
Illustration: the AI nobody listed
This is an illustrative scenario, built from the language in the documents.
A bank answers “no” to the first examiner question: does the institution use AI in any products, services or operations? The CSBS work program says a negative or unclear response should not end the discussion if other information suggests AI use, and may warrant confirmation against the institution’s vendor inventory, software inventory, approved tools, and recent product or platform changes.
In this scenario, the bank’s loan servicing vendor added an AI summary feature in a routine software update. Nobody in the bank was told. Now look at what each document says:
- FSB: in the case of embedded AI, financial institutions may be unaware that their third-party tools or services incorporate or rely on AI. Third-party providers might also change the underlying models powering their AI systems without adequate notice, leaving institutions unaware of changes that could affect performance or risk outcomes.
- CRI: GV-1.6.3 asks that purchased and shadow AI be identified, documented and added to the inventory.
- CSBS: questions IS-5 and AI-6 ask whether the bank has tried to find embedded AI, and what vendors have told it.
- OCC: component 2.9 says existing vendors may add AI features, functionality, or change models connected to products already in use. The bank may require notice and transparency of new functionality, and may reassess the relationship or retest the model before or when the changes are made.
Four documents point to the same gap and the same fix. Ask vendors in writing, add AI-change notice to contracts, and put every answer in the inventory.
Controls and evidence the examiners point to
Neither examiner document is a checklist. The OCC words its points as things a bank “may” do, and the CSBS guide says examiner judgment remains important. Together they describe a practical control set and the evidence that supports it.
OCC: 16 components
The OCC does not list required evidence. The evidence column reflects the records its component text points to.
| Component | Control to have | Evidence to keep |
|---|---|---|
| 2.1 Purpose and permitted use | Written purpose per use case: allowed actions, limits, users, data and systems | Purpose statements; inventory entries with owner, provider and permissions; change records |
| 2.2 Risk classification | Consequence-based tier before deployment and after material change | Tier assessments; reclassification records; view of shared providers |
| 2.3 Human oversight | Reviewers with the expertise, authority and time to reject, pause or halt | Reviewer assignments; override logs; workload analysis per reviewer |
| 2.4 Governance and accountability | Board-set limits on autonomy and actions; named executive and use-case owners | Decision-rights matrix; approval records; board and management reports |
| 2.5 Model risk extensions | Validation adapted for drift, non-determinism and limited explainability | Repeated-run and sampling results; consistency thresholds; reference sets |
| 2.6 Data governance | Controls over data quality, lineage, access and retention, including prompts and outputs | Lineage records; approved-source lists; drift monitoring |
| 2.7 Privacy | Privacy risks mapped to each use case and revisited on change | Privacy map; retention periods for prompts and logs; vendor terms on training with bank data |
| 2.8 Compliance | Limits on automated statements to customers; verification in higher-impact cases | Output reviews; hallucination test results; correction and remediation procedures |
| 2.9 Third-party and vendor risk | Defined authority for provider agents; AI incident and change-notice terms; liability terms | Contracts; vendor change notices; retest records after vendor changes |
| 2.10 Operational resilience | Detect and contain an agent that runs but acts wrongly; tested fallback; customer remediation plan | Tests of circuit breakers, kill switches and fallback; post-incident reviews |
| 2.11 Cybersecurity | Defenses against prompt injection, poisoning, extraction and excess agent permissions | Red-team results; agent credential and permission inventory; identity logs |
| 2.12 Performance monitoring | Production metrics and thresholds that trigger review, reduced autonomy or suspension | Dashboards for accuracy, hallucination, override and complaint rates; drift reports |
| 2.13 Testing | Testing sized to consequence, including adversarial, red team and hallucination tests | Test plans and results; red-team reports; samples of live outputs |
| 2.14 Documentation and auditability | Records that let an outsider reconstruct a material decision | Logs of inputs, outputs, tool calls, approvals and overrides; version control for model, prompt and data |
| 2.15 Internal audit and challenge | Independent audit with access to underlying evidence | Audit plan; findings with independent closure evidence |
| 2.16 Reporting and transparency | Risk-based reporting with set incident thresholds | Reports on material use cases, changes and incidents, reconciled to the inventory |
CSBS: what examiners ask for
The CSBS document request list and scoping procedures name much of this evidence directly. This table groups it by procedure area.
| Procedure area | Control to have | Evidence to keep |
|---|---|---|
| Scoping (IS-1 to IS-8) | A checked answer to each of the eight questions, confirmed against vendor and software inventories and recent platform changes | The signed-off answers and the lists used to check them |
| Governance and oversight (GO-1 to GO-7) | Named oversight; policies and committees; risk-based approval; reporting to management and the board; testing and monitoring after changes; coverage of internal, external and embedded AI | AI policies, standards or procedures; committee materials; management and board reporting; monitoring dashboards; employee guidance or training |
| AI inventory and use cases (AI-1 to AI-6) | An inventory that records purpose, owner and type; each use classed as internal, customer-facing, decision-support or decision-making; risk tiering; a search for embedded vendor AI | The inventory; purpose and ownership records; risk assessments, ratings or tiers; vendor products with known or possible embedded AI; change management records |
| Generative AI and emerging use (GE-1 to GE-8) | Rules on generative tools; data entry controls; output review; named risks such as hallucination and prompt injection; for agents, permitted actions, human checkpoints, logging, reversibility and a way to halt | Policies or restrictions; approved and known tool lists; business purposes, data controls and output review records |
| Existing resources (as applicable) | Vendor, model risk, consumer protection and operational risk reviews that cover AI | Vendor oversight files; validation, testing and monitoring results; notice, complaint and remediation records; samples of AI-generated customer outputs; contract terms on data use and training; audit and control testing reports |
For nonbank reviews, the CSBS supplements add AI questions to three existing reviews. Vendor oversight asks whether the institution treats vendor-provided AI “as a managed dependency rather than a black box.” Model risk asks whether the institution assesses vendor model performance itself instead of relying on vendor claims. Consumer protection asks whether the institution can “identify and communicate specific and accurate reasons” for adverse outcomes where AI informs them.
The CSBS tiering worksheet’s suggested controls
The worksheet adds suggested controls by tier. They are cumulative: a Tier 3 use case is expected to carry the Tier 1 and Tier 2 controls too.
| CSBS tier | Suggested controls |
|---|---|
| Tier 1 (all use cases) | Documented inventory with owner and purpose; written acceptable-use policy; vendor contracts that address AI risk and audit rights; staff trained on limitations; data quality controls; periodic management reporting; documented risk level |
| Tier 2 (adds) | Model risk or similar review; explainability documentation for consumer-facing outputs; escalation path for adverse outputs; third-party model oversight; pre-deployment testing; tested adverse action notice process; periodic performance review; complaints process that captures AI issues |
| Tier 3 (adds) | Ongoing monitoring of accuracy, fairness and consumer outcomes; independent model validation; AI-specific incident response; senior management review of the full inventory; consumer disclosures on automated decisions; documented pre-deployment fairness review; reporting to senior management, the board or an oversight body |
Build the evidence once. A complete inventory, a tier worksheet for every use case, and reliable logs answer most rows in both sets of tables.
The CRI publishes more than the matrix. Its site lists an adoption stage questionnaire, a guidebook, and a control objective reference guide with examples of controls and evidence. The framework was built through a public-private collaboration that included more than 100 financial institutions. This post did not review the guidebook or the reference guide in full, so use them to fill any evidence gap above.
Eight questions to test your readiness
These follow the CSBS scoping questions, which examiners use at the start of a review.
| # | Question | Evidence to have ready |
|---|---|---|
| 1 | Do we use AI anywhere, including internal support and compliance work? | A signed-off answer, checked against vendor and software lists |
| 2 | Do we have a maintained list of AI systems and use cases? | The inventory, with owner and purpose for each entry |
| 3 | Does AI touch customers or influence decisions? | Use case descriptions that show where AI informs, recommends or decides |
| 4 | Which AI do we get from vendors or platforms? | Vendor list marked for AI, with contract terms |
| 5 | Have we looked for AI embedded in vendor products? | Vendor questionnaires and written responses |
| 6 | Do we use generative AI, LLMs or agents, officially or not? | Approved tool list, policy, data controls, output review steps |
| 7 | Do we tier use cases by risk, and does the tier drive review? | Completed tier worksheets with rationale |
| 8 | Does customer or confidential data enter AI tools? | Data classifications, access controls, vendor data-use terms |
Answer each one with a document, not a statement. If the answer takes a meeting to assemble, that question is a gap.
What to do in the next 90 days
Days 1 to 30: find everything
- Build or refresh the AI inventory, with an owner and purpose for each entry.
- Send written AI questions to your vendors.
- Adopt four neutral risk labels and map them to the OCC and CSBS tiers.
Days 31 to 60: tier and assign
- Complete a tier worksheet for each use case, starting with anything customer-facing or generative.
- Name an accountable owner for each use case. For agents, name one individual, as the FSB report suggests.
- Put AI-change notice and audit rights into vendor contract templates.
Days 61 to 90: prove it works
- Test one stop procedure for your highest-tier agent or automated decision.
- Start tracking human overrides on one high-tier use case.
- Produce a first management report built from the inventory and tiers.
- Run a mock scoping interview using the eight questions above.
What to watch
Three limits are worth stating plainly.
- These documents are new. The CSBS framework is version 1.0 and the OCC framework says it will be updated regularly, since SI and agentic SI are novel technologies experiencing rapid evolution. The FSB report calls practices specific to agentic AI and multi-agent systems “nascent,” so expect that area to change.
- CSBS adoption varies. Each state agency decides how much to apply.
- Coverage is partial. The CRI matrix was checked only for the objectives cited here. The NIST and Treasury documents behind the CSBS material were not reviewed for this post.
Conclusion
You do not need four programs. You need one program that can answer four sets of questions.
Start with the inventory, because every framework starts there. Tier by consequence, because every framework asks you to. Name owners, make human review real, know your vendors, and keep the ability to stop a system.
Then clean up the one trap: use neutral risk labels so that “Tier 1” never means two opposite things in your own reports.
Organizations that do this can answer an examiner’s first question with a document, not a guess.
References
- Financial Stability Board. Sound Practices for Responsible Adoption of Artificial Intelligence (AI): Final report, 15 October 2026.
- Cyber Risk Institute. Financial Services AI Risk Management Framework (FS AI RMF), Risk and Control Matrix and Guidebook, version 1.0.
- Conference of State Bank Supervisors. The CSBS Artificial Intelligence Supervisory Framework (Core Examiner Guide v1.0, Examiner Work Program v1.0, AI Use Case Risk Tiering Worksheet v1.0, Nonbank Supplements v1.0).
- Conference of State Bank Supervisors. AI Supervisory Framework: Core Examiner Guide v1.0, approved 13 August 2026.
- Office of the Comptroller of the Currency. OCC Issues Supervisory Framework on Bank Use of Super Intelligence, News Release 2026-97, 4 November 2026.
- Office of the Comptroller of the Currency. OCC Bulletin 2026-54: Super Intelligence: Supervisory Framework.
- Executive Order No. 14434, 91 Fed. Reg. 63129 (2 October 2026), defining “Super Intelligence” as replacing “Artificial Intelligence” in federal usage.