An AI inventory can show that a bank knows which tools exist. It does not, by itself, show an examiner which use cases merit a closer look, what evidence supports their controls, or who owns the unresolved gaps.
That is the practical opportunity in the CSBS AI Supervisory Framework, released September 16, 2026. CSBS describes it as a discretionary resource for state examiners assessing AI at state-chartered banks and state-licensed nonbank financial institutions. It is designed to help identify AI use, assess associated risk, and decide when a deeper review may be appropriate; it is not a declaration that every institution must use one fixed scoring model.
This guide turns that examiner-facing question into a reusable risk-tier evidence pack. The goal is not to have an AI system declare a use case “compliant.” The goal is to create one inspectable record per use case: what it does, what it can affect, what documents exist, who must answer for it, and why the record may require deeper review.
The search intent: exam readiness, not a generic AI checklist
Search results for AI risk tiering are crowded with generic governance templates, consulting frameworks, and broad lifecycle advice. The gap is a bank-ready method that connects a use-case tier to a concrete document request and an explicit review depth. That connection matters because an examiner cannot assess a vague “AI program” in the abstract; the useful unit is a bounded use case and its evidence.
Use this approach for a customer-service copilot, fraud-triage model, credit-support workflow, document classifier, employee assistant, or vendor tool. Do not merge unrelated systems into one card. A low-impact drafting assistant and a system that shapes customer eligibility may both use the same model vendor, but they need different questions, evidence, and escalation paths.
The approach complements an AI banking integration cutover decision card, which tests one planned production change. This pack begins earlier: it creates an examination-ready inventory and a defensible way to decide which AI use cases need a deeper file.
Start with a bounded use-case record
Before assigning a tier, write a record that a business owner could defend without referring to marketing language. “AI for operations” is not a use case. “A vendor-hosted model summarizes call notes for an employee, who manually reviews them before a CRM entry is proposed” is.
| Field | Evidence to request | Why it matters in review |
|---|---|---|
| Use-case ID and business purpose | Inventory record, owner attestation, process map | Links the system to a defined financial workflow |
| Affected people and outcomes | User journey, decision map, customer-impact analysis | Separates internal assistance from consequential decisions |
| Data and system boundaries | Data-flow diagram, vendor list, retention and access settings | Makes sensitive inputs and third-party dependencies visible |
| AI role and human authority | Prompt/configuration version, action matrix, approval rule | Shows whether AI drafts, recommends, routes, or can cause a change |
| Monitoring and change path | Test results, monitoring plan, incident route, change tickets | Proves the record is maintained after launch |
| Evidence owner and review date | Named first-line owner, independent reviewer, next review date | Keeps missing proof from becoming an orphaned task |
Use source links or stable internal record IDs rather than pasting customer data, credentials, or full sensitive documents into a prompt. The same source-minimization discipline appears in this financial research agent data-control contract: an AI tool may organize approved evidence, but access and retention boundaries remain the institution’s decision.
Tier the consequence, then scale the evidence request
A useful tier is not a legal classification or a substitute for a regulator’s judgment. It is an internal routing decision: how much evidence and review should this specific use case receive before its record can be called complete.
The NIST AI RMF and its Playbook provide voluntary risk-management guidance through the Govern, Map, Measure, and Manage functions. Use them as a structure for questions, not as a box-ticking claim. CSBS itself notes that each state agency determines how, if at all, it incorporates the framework into supervisory programs.
A practical three-tier crosswalk
| Review tier | Observable trigger | Minimum evidence pack | Deeper-review trigger |
|---|---|---|---|
| Tier 1 — assistive | Internal drafting or retrieval; a trained employee checks all output before use; no direct customer, account, or financial action | Use-case record, approved data/access boundary, human-review statement, vendor/model identifier | Sensitive data without a documented boundary, or output reused beyond the stated purpose |
| Tier 2 — operational | AI sorts, prioritizes, classifies, or proposes work at volume; an error can delay or misroute a financial process | Tier 1 items plus test samples, quality/exception metrics, workflow owner, change and incident path | Material backlog, persistent exceptions, unclear override authority, or a vendor/system change |
| Tier 3 — consequential | AI materially informs or triggers a customer, account, payment, credit, compliance, fraud, or other high-impact outcome | Tier 2 items plus decision/action map, authority limits, validation evidence, impact assessment, escalation/appeal path, independent challenge where applicable | An uncontrolled write/action path, missing validation, unresolved impact, no accountable owner, or a material incident |
These are intentionally internal labels. Do not say that a use case is safe because it reached Tier 1, or that Tier 3 alone proves regulatory compliance. The value is operational: the record tells a reviewer what to ask for next and marks the cases that cannot advance with missing evidence.
For example, a model that proposes an analyst’s research outline is not the same as a model that changes a payment instruction. The latter must remain on the recommendation side of the authority boundary until a named human or approved control authorizes the action. That separation aligns with the principle in the agentic payments controls guide: a useful automated suggestion is never automatic permission to move money.
Build the evidence pack as a document-request index
An exam-ready file should make a missing artifact as visible as a present one. Treat the pack as an index, not a document dump. Each line should include a source link, artifact version or date, owner, reviewer, and state.
| Evidence domain | Example artifact | Owner | Acceptable state |
|---|---|---|---|
| Governance | Approved policy, committee minute, accountable executive | Risk or business owner | Linked, dated, and in scope |
| Use-case scope | Process map and decision/action boundary | Product or operations owner | AI role and prohibited actions are explicit |
| Data and vendor | Data-flow, access record, vendor due diligence | Data or third-party-risk owner | Inputs, retention, and dependencies are known |
| Testing and validation | Test plan, sample results, exception outcomes | Model, technology, or control owner | Relevant test evidence is traceable to the version used |
| Monitoring and incidents | Thresholds, issue log, escalation route | Operations owner | Trigger, recipient, and response owner are named |
| Change management | Configuration/model/prompt change record | Technology owner | Prior version and approval are recoverable |
| Consumer or financial impact | Impact assessment, override or appeal route when applicable | Compliance or business owner | Impact is assessed; unresolved items are not hidden |
The index prevents an easy failure mode: a team attaches a policy and calls the program controlled, while the record has no proof for the exact use case or version under review. For systems that affect finance operations, pair it with a reconciliation exception handoff so evidence gaps and production exceptions have owners and closure approval.
Copyable prompt: examiner-prep gap finder
Use the prompt only with approved, minimized internal records. It must organize evidence, flag absence, and route questions. It must not create records, make a supervisory determination, or infer that a missing control exists.
You are an AI use-case examination-preparation evidence reviewer for a financial institution.
Use only the supplied records. Do not state or imply that the institution is
compliant, approved, safe, or ready for an examination. Do not make customer,
credit, payment, fraud, or compliance decisions. Do not create, change, or
delete records.
For exactly one AI use case, return valid YAML with these fields:
use_case_id:
business_purpose:
ai_role: draft | retrieve | classify | prioritize | recommend | act
affected_people_or_processes:
data_and_system_boundaries:
vendor_model_and_configuration_refs:
human_authority_and_prohibited_actions:
proposed_review_tier: TIER_1 | TIER_2 | TIER_3 | UNDETERMINED
tier_rationale:
evidence_index:
- domain:
artifact_ref:
version_or_date:
accountable_owner:
state: PRESENT | MISSING | STALE | CONFLICTING
deeper_review_triggers:
unanswered_questions:
required_human_reviewers:
status: EVIDENCE_GAP | READY_FOR_HUMAN_REVIEW | BLOCKED
Set EVIDENCE_GAP when a required artifact, owner, version, data boundary, or
authority limit is absent or stale. Set BLOCKED when records show an
uncontrolled action path, conflicting ownership, or a material uncertainty
that prevents tiering. Leave unknown fields blank and list them under
unanswered_questions. Never fill gaps with assumptions.The output is a routing artifact, not a risk score with legal force. A person with the right authority must confirm the tier, determine review depth, and decide whether a gap blocks use, release, or examination readiness.
Run negative tests before relying on the pack
Do not test only the happy path where every PDF has a link. Feed the workflow sparse and conflicting records to make sure a plausible-looking summary cannot hide a control failure.
| Test case | Expected output |
|---|---|
| A chat assistant has a named vendor but no data-retention setting | EVIDENCE_GAP; data boundary and owner are requested |
| A fraud-routing tool can close an alert without a documented human authority | BLOCKED; uncontrolled action path is explicit |
| An internal classifier has sample tests and a human reviewer but the model version changed | EVIDENCE_GAP; the current-version test evidence is missing |
| A credit-support workflow has a documented recommendation boundary and appeal route, but no impact assessment | EVIDENCE_GAP; the pack cannot be called complete |
| An employee drafting tool has complete scope, access, human review, and no direct action path | READY_FOR_HUMAN_REVIEW; the human still confirms the proposed tier |
These tests are different from adversarial security tests. Use an AI red-team test harness for banking agents to probe misuse and policy bypass. Use this evidence pack to expose missing ownership, documents, or decision boundaries before an examiner—or an incident—finds them.
Measure evidence completeness, not “AI readiness”
Avoid a single percentage that makes a partially documented portfolio look safe. Track the specific conditions that make an examination conversation easier to conduct:
- In-scope use cases with a named business owner and review date.
- Use cases whose proposed tier has a recorded rationale and human confirmation.
- Tier 2 and Tier 3 use cases with a complete evidence index by domain.
- Open evidence gaps by age, accountable owner, and whether they trigger deeper review.
- Use cases with changed model, prompt, data, or vendor configurations awaiting evidence refresh.
- Exceptions where an action boundary, escalation route, or incident owner is absent.
Set targets through the institution’s risk appetite and policies. Do not reward teams for lowering a tier to improve a dashboard. The right outcome is an honest record of where review should be deeper.
Keep the conclusion narrow
CSBS’s framework provides a timely, practical reason to turn an AI inventory into use-case-level evidence. A risk-tier evidence pack gives examiners and internal reviewers the same starting point: a clear use case, its consequence boundary, documents, owners, and unresolved questions.
It cannot guarantee an examination result, regulatory acceptance, or safe model behavior. It is not legal, compliance, investment, or financial advice. Use it to prepare records and surface gaps, then have qualified risk, compliance, legal, technology, and business owners make the decisions the artifact deliberately leaves to them.
Sources
About Enis
AI Engineer specializing in Machine Learning and LLMs. Combining Computer Engineering and Economics to build data-driven financial tools.