Is your AI governed ? Get your maturity score and roadmap. Request Access to the AI Governance Advisor →

How to Tier AI Agent Use Cases Like an Examiner

Table of Contents

Summarize and analyze this article with
ChatGPT

Chat GPT

ChatGPT

Perplexity

 
ChatGPT

Grok

 
ChatGPT

Google AI

ChatGPT

Claude

 

The question

How should a bank, insurer, or health system sort its AI agents by risk so that the reasoning still holds up when a supervisor asks why one agent gets more controls than another? Until this month there was no federal reference point for that question. The April 2026 interagency model risk guidance, SR 26-2 and OCC Bulletin 2026-13, placed generative and agentic AI outside its scope. On 4 November 2026 the OCC issued Bulletin 2026-54 with a supervisory framework on bank use of advanced AI, written for its examiners, that describes how they may think about classifying generative and agentic AI use cases. It does not create rules for banks, but it shows how an examiner is likely to reason, which makes it the most useful design reference available.
This guide turns that framework into a tiering method you can run and defend. It is general information and not legal advice.

1. Start from the inventory entry

A tier needs something to attach to. The framework describes an inventory that may capture the use case, its owner, its provider, its permissions, its deployment status, and its monitoring and failover arrangements. Keep one row per use case, not one row per model. The same model can sit behind a low-risk summarizer and a high-risk payment agent, and the two need different tiers.

2. Score the seven factors

The framework’s illustrative methodology lists seven factors. Score each as low, medium, or high, with a short note naming the evidence. “Reversibility: high, because every action queues for a person to release it” is a usable note. “Reversibility: medium” is not.
Factor What it asks Evidence that supports the score
Decision consequence Does the system inform, decide, or act? Workflow showing where the output goes and who acts on it
Reversibility Can a wrong action be undone, and how fast? Queue, hold, or rollback mechanism; time to reverse
Customer or financial impact Who is harmed, and how much, if it is wrong? Population affected; transaction or exposure limits
Autonomy level How much does it do without a person? Approval map; which actions run without review
Data sensitivity What data can it reach? Data classifications in scope; access grants
Explainability Can a third party reconstruct why it acted? Decision logs, retrieved sources, tool-call records
Provenance Built in house, fine-tuned, or consumed as a service? Model and vendor lineage; contract terms

3. Combine the scores into a tier

The framework describes a composite tier and gives four illustrative tiers. Illustrative is the important word: a bank defines its own tiers and cutoffs. Many teams find a short rule works better than a weighted formula, for example: any high score on consequence or autonomy sets a floor of Tier 2, and a fully autonomous, customer-facing action that moves money is Tier 1. That last rule follows the framework’s own illustration, which places autonomous payment agents in the critical tier.
Illustrative tier Profile Example use cases
Tier 1, Critical Direct action, high impact, autonomous or partly autonomous, little reversibility Autonomous payment agents, real-time credit decisioning
Tier 2, High Direct action or heavily weighted decision support, human on the loop AI-assisted underwriting, agent-initiated customer messages, AML alert disposition
Tier 3, Moderate Decision support with a mandatory human in the loop, reversible Credit-memo summaries, agent-assisted research
Tier 4, Low No customer or financial data, no direct action Internal productivity tools, general code assistance

4. Attach controls to the tier

The framework says control intensity may increase with the potential consequences of an agent’s failure or misuse. Write that down as a table. The layout below is PiTech’s suggested design, not a requirement from any regulator. The control types come from the framework’s own discussion: decision traceability detailed enough for a third party to reconstruct a material decision, agent credentials treated with the same care as privileged human accounts, and resilience measures that include transaction limits, narrowly scoped permissions, and emergency shutdown capability.
Control Tier 1 Tier 2 Tier 3 and 4
Decision traceability Full: inputs, outputs, sources, tool calls, approvals, overrides Full for customer-facing actions Summary logging
Agent credentials Own identity, privileged-account controls, tight scope Own identity, least privilege Own identity
Suspension Tested kill switch, limits, rate controls Tested suspension path Documented disable step
Human review Real review with time and authority; no bulk approval Sampled review, escalation Periodic review
Re-scoring Every configuration change Every material change Annual or on change

5. Name the triggers for a fresh score

A tier is a statement about a configuration. Changes to permissions, tools, data sources, the underlying model, or the vendor’s terms should trigger a fresh score. Put the triggers in the same document as the tiers, so a reviewer can see why a tier was last set and what would move it.

What usually goes wrong

  • Tiering the technology instead of the use. A team scores a model once, labels it moderate, and reuses the label across every deployment. The framework points the other way: classification may rest on potential consequences, the transfer of agency, and the degree of human oversight rather than technical sophistication. A simple rules-based agent that can move money can carry more risk than a powerful model that only drafts internal documents.
  • Counting review that is not real. The framework warns that reviewers need the expertise, authority, and time to perform the task, and that high volumes or bulk approval can reduce review to a formality. If a person approves hundreds of agent actions in a batch, the tier should reflect the agent’s autonomy, not the existence of an approval screen.
  • A tier nobody can explain. A score with no written reasoning is the first thing to fall apart under questioning. One paragraph per use case, naming the deciding factors, is enough.

How this relates to the CSBS worksheet

State-chartered institutions may also meet the CSBS AI Use Case Risk Tiering Worksheet, published in September 2026, which uses four factors and three tiers with the highest factor setting the tier. The two approaches are compatible. An institution supervised by both can keep one inventory and record both scores against each use case, rather than running two tiering programs. The seven OCC factors are a superset of the CSBS four, so a seven-factor score usually yields the CSBS tier directly.

Where PiTech fits

PiTech Solutions helps regulated enterprises turn frameworks like this into inventories, tier definitions, and evidence records that teams can maintain: one row per use case, seven-factor scoring with evidence notes, written tier reasoning, a tier-to-control table, and re-scoring triggers tied to configuration changes. It connects tiering to the AI inventory and agent governance beneath it and to the BSA/AML and data-engineering work that sits under many high-tier banking agents. Delivery runs under CMMI Level 3 and ISO 27001, 9001, and 42001 certifications. See the banking practice, AI, GenAI and ML, and Data Solutions. PiTech Solutions Inc. is headquartered in Durham, North Carolina (UEI GNLRY5LNNVH6, CAGE 530K4) and is distinct from similarly named companies.

The bottom line

The OCC framework gives examiners a way to classify AI by consequence, and gives banks a reference design worth adopting before anyone asks. Start from the inventory, score seven factors with evidence, combine them with a short rule, attach controls by tier, and name the triggers that force a fresh score. Then test it: take ten use cases and check whether two reviewers reach the same tier from the same notes.

Frequently Asked Questions (FAQs)

What is OCC Bulletin 2026-54?

OCC Bulletin 2026-54, issued on 4 November 2026, introduced the OCC’s supervisory framework on bank use of advanced AI, including generative and agentic systems. It is written for OCC examiners and describes how they may assess the way banks use, govern, and control AI, including an illustrative method for classifying AI use cases by risk, four illustrative risk tiers, and discussion of controls such as decision traceability, credential management, and emergency shutdown. The OCC states that the framework does not establish supervisory expectations for banks. Its value to banks is as a window into examiner reasoning: it shows the questions an examiner is likely to ask and the logic they may apply, which makes it a strong reference design for an internal tiering program.
No. The OCC says the framework does not establish supervisory expectations for banks and is not meant to be used by banks as an assessment of their own AI risk management. It is an examiner resource. That said, a framework that describes how examiners may classify and evaluate AI use cases is a strong signal of the questions banks will face, and aligning an internal tiering method to its logic means the bank’s records will make sense to the examiner reading them. The practical posture is to treat it as a reference design, not a rule: adopt the parts that improve your program, document where your approach differs and why, and avoid describing it internally as a mandate, since it is not one.
The framework’s illustrative methodology, as summarized in published research notes, lists decision consequence (whether the system informs, decides, or acts), reversibility, customer or financial impact, autonomy level, data sensitivity, explainability, and provenance (whether the model is built in house, fine-tuned from a third-party base model, or consumed as an external service). Each can be scored low, medium, or high. The useful discipline is to attach a short evidence note to each score, naming the mechanism or record that justifies it, because a bare score falls apart under questioning while a score with its evidence can be reviewed and reproduced. Confirm the exact factor wording against the primary OCC document before quoting it in board materials.
The framework describes four illustrative tiers. Tier 1, critical, covers direct action with high impact and full or partial autonomy, such as autonomous payment agents and real-time credit decisioning. Tier 2, high, covers direct action or heavily weighted decision support with a human on the loop, such as AI-assisted underwriting or AML alert disposition. Tier 3, moderate, covers decision support with a mandatory human in the loop and reversible outcomes, such as credit-memo summaries. Tier 4, low, covers uses with no customer or financial data and no direct action, such as internal productivity tools. The tiers are illustrative, so a bank defines its own definitions and cutoffs, but these profiles give a sensible starting structure.
Because risk comes from what the system does in a particular deployment, not from the model itself. The same model can power a low-risk internal summarizer and a high-risk agent that initiates payments, and those two deployments carry very different consequences if they go wrong. The OCC framework makes this point directly, indicating that classification may rest on potential consequences, the transfer of agency, and the degree of human oversight rather than technical sophistication. A team that scores a model once and reuses the label everywhere ends up under-controlling its riskiest deployments. One inventory row per use case, each scored on its own consequences, is what keeps controls matched to actual exposure.
With a short, written rule rather than an opaque formula. A weighted score can produce results nobody can explain, while a rule such as “any high score on consequence or autonomy sets a floor of Tier 2, and a fully autonomous customer-facing action that moves money is Tier 1” is easy to apply consistently and easy to defend. The rule should be documented alongside the tiers, and any exception, a use case placed in a tier other than the rule suggests, should carry a written rationale and a named approver. Whatever method you choose, test it by having two reviewers tier the same ten use cases independently; if they disagree often, the rule or the factor definitions need tightening.
The heaviest set your program defines. Drawing on the control types the OCC framework discusses, a Tier 1 agent would typically carry full decision traceability (inputs, outputs, retrieved sources, tool calls, approvals, and overrides recorded so a third party could reconstruct a material decision), its own identity with credentials managed like a privileged human account and tightly scoped permissions, a tested emergency shutdown capability with transaction limits and rate controls, genuine human review where review is part of the design, and re-scoring on every configuration change. The exact layout is the bank’s to define; the tier-to-control table in this guide is a suggested design, not a regulatory requirement.
When the reviewer has the expertise, authority, and time to perform the task, which is the standard the OCC framework describes. Review that happens in high volumes or through bulk approval can become a formality, and the framework warns about exactly that. If one person approves hundreds of agent actions in a batch, the review is not meaningfully reducing the agent’s autonomy, and the tier should reflect the agent’s real autonomy rather than the existence of an approval screen. To count review as a control, record who reviews, what they see, how long they take, and how often they reject or change the agent’s proposed action; a review that never rejects anything is a warning sign.
They are compatible and can share one inventory. The CSBS AI Use Case Risk Tiering Worksheet, published by state supervisors in September 2026, rates four factors (consumer impact, human oversight, harm potential, and data sensitivity) into three tiers, with the highest factor generally setting the tier. The OCC framework’s seven factors cover the same ground and add explainability, reversibility, and provenance, with four illustrative tiers. An institution that answers to both federal and state supervisors can record both scores against each use case in a single inventory rather than running two programs. Because the OCC factors are effectively a superset, a well-documented seven-factor score usually yields the CSBS tier directly.
Yes. PiTech Solutions helps regulated enterprises turn frameworks like the OCC’s into working programs: an inventory with one row per use case, seven-factor scoring with evidence notes, written tier reasoning, a tier-to-control table, and re-scoring triggers tied to configuration changes. It connects tiering to the AI inventory, agent governance, and data-engineering work beneath it, including the BSA/AML systems that often sit under high-tier banking agents. Delivery runs under CMMI Level 3 and ISO 27001, 9001, and 42001 certifications. A practical first engagement is a ten-use-case pilot, testing whether two reviewers reach the same tier from the same notes, before scaling the method across the estate.