8 minutes

The OCC's New AI Framework: 10 Implications for Companies Deploying Agents

Harold Byun

Chief Executive Officer

The OCC’s New AI Framework: 10 Implications for Companies Deploying Agents

What the newly released banking framework signals about agent authority, enterprise accountability, and control during execution.

The newly released Supervisory Framework for Bank Use of Super Intelligence (SI) from the Office of the Comptroller of the Currency (OCC) (link) offers a useful lens for the enterprise agent market: evaluate systems in light of their authority and the consequences of their actions.

The scope matters. This is an examiner resource for banking supervision. The OCC explicitly says it establishes no new regulatory expectations or enforceable standards, and banks are not expected to use it to assess their AI risk management.

The framework establishes six main pillars as supervisory considerations for the use and operation of SI agents and models.



1. Purpose: Defines the actions the SI system or agent is permitted to execute. (Defined Purpose and Permitted Use) 

2. Risk: Assesses the consequences if the SI system or agent's output or action is not fit for purpose. (Risk Classification and Criticality Assessment) 

3. Control: Identifies controls that may prevent unauthorized, inappropriate, or unsafe actions of the SI system or agent. (Human-in-the-Loop and Human-on-the-Loop Controls; Model Risk Management; Data Governance; Privacy; Compliance; Cybersecurity) 

4. Oversight: Establishes who monitors the SI system or agent and who is accountable for its use. (Governance & Accountability; Human-in-the-Loop and Human-on-the-Loop Controls; Performance Monitoring; Reporting & Transparency) 

5. Resilience: Addresses the bank's ability to detect, contain, recover from, and remediate SI system or agent failures while continuing critical business processes. (Operational Resilience; Cybersecurity) 

6. Assurance: Provides evidence that SI system or agent controls continue to operate effectively. (Testing; Documentation & Auditability; Third-Party & Vendor Risk; Internal Audit & Independent Challenge) 

The broader implications below are our interpretation and recommendations. They concern a question every company deploying agents will need to answer: what gives the organization a sound basis for allowing software to act on its behalf and how should they look to operationalize these deployments?

1. Component assessment becomes a mandatory part of deployment governance

Skills, tools, and MCP servers are the foundational connectors for how an agent operates. Their capabilities and dependencies require thorough assessment and examination before they become part of a business workflow.

Assessment should include a behavioral and intent-based analysis to inform whether the component should be approved for use in an agentic deployment. Or if scope should be narrowed or if other remediation is required. Component assessment should also include confidence and risk scoring to enable mapping to a broader risk framework.

2. Adding a tool can change the deployment’s risk. Understand what it can do.

A workflow does not need a new model to become more consequential. Connecting a transaction tool, enabling CLI wrappers, or granting write access can expand what it can do immediately.

Tools and skills can drift. With MCP, tools can be dynamically updated. Treat changes in tools as part of the lifecycle. Teams should understand the new capability and decide whether existing boundaries remain appropriate before enabling it.

3. Permissions and authorization are a critical part of the AI risk conversation

Two agents using the same model can create very different business exposure. One might summarize customer correspondence; another might modify an account.

Our recommendation is to include entitlements and data access as part of the agent execution layer. This includes credentials, connected resources, tools and actions alongside model evaluations in deployment reviews. The relevant unit is the agent as an entity and its authorization.

4. Agent controls and guardrails are foundational requirements

The ongoing concerns around rogue AI are rooted in a poor understanding of how the agentic execution layer actually works. The actions being taken when accessing or manipulating data or persistently attempting a course of action go far beyond prompt and response and tools.

Unless organizations understand that agents execute throughout their own runtime stack, escapes and destructive actions will continue to occur. Control needs to be implemented from agent-to-model to tool or skill to MCP and also govern process execution and the runtime. 

5. Human-in-the-loop (HIL) needs to be tied to both risk and action

Naming a reviewer does not establish whether that person can supervise the deployment effectively. The volume of activity, clarity of the evidence, and time available to intervene all matter. HIL is not just pushing a keyboard button to authorize as reviewer or approver.

Design approval and escalation around specific decisions and incorporate the risk of the action into the approval process. Give reviewers sufficient context and a practical mechanism for restricting execution in real-time.

6. Oversight requires complete and comprehensible observability 

Most solutions today still do not have any visibility into what an agent does underneath the covers. Further, with the volume of events driven by agentic workloads, much of the monitoring becomes nonsensical. 

Monitoring and reviewers need to understand the nature of an action beyond just which tools ran. They need to understand which attempts were made, what was denied and what executed in the hidden substrate below a tool or a skill. 

7. Baseline agent behavior and understand consistency and deviation

As agents perform through development, staging and production, a behavioral curve emerges that shows how they perform through reasoning and inference to the execution and reconciliation phases. This behavior can be baselined and further monitored and should not deviate over time. 

Thresholds for consistency and acceptable deviation should be established as part of the operating guidelines for any organization. 

8. Implement an agentic policy control plane for actions

To date, agent environments largely run with constraints invoked at launch. Given the non-deterministic nature of AI and agent behavior, enterprises need policy controls that are more dynamic in nature vs. re-instantiating agent environments every time a new risk or deviation is identified. 

A policy control plane will enable enterprises to manage fleets of agents without having to understand how the environment was launched and whether it needs to be modified. Without one, organizations will be left with an agentic configuration item nightmare that they cannot keep track of.

9. Automate controls for responsiveness and resilience

Service availability does not establish that an agent is performing the right operations. An agent could persist in attempts to repeatedly change the wrong records, for example.

Because agents operate at such a high rate of velocity, traditional approaches to implement controls and counter adverse actions become challenging. Fail-safe automated controls need to be implemented in order to ensure appropriate actions are the ones permitted.

Define incident conditions around behavior and business impact. Then test identification, containment, suspension, and recovery so the team knows how it will respond while an active deployment is still running.

10. Existing software vendors can introduce new agent authority

An established supplier relationship does not answer every question about an agent or MCP server added to its product. A feature update can introduce new actions, data access, or dependencies. Many of these capabilities are hidden beyond a third-party SaaS wall.

Procurement and technical owners should ask for greater transparency and perform the same assessments of agentic tooling that they conduct for internally developed solutions.

What this means for enterprise adoption

These implications connect to the operating foundation we are building at BlueRock: assess and understand agentic components, and monitor and control agentic via a foundational AI infrastructure. An agentic-aware infrastructure that isolates execution, observes all activity, and enforces policy across your deployments. Bringing those capabilities together helps teams evaluate how an agent should run and whether its deployment is ready to expand.

The next useful step is to examine one consequential workflow. Identify its authority, the boundaries around it, and the evidence available when it runs. That exercise gives business and technical leaders a concrete basis for deciding what gaps they have and what to improve before handing the keys over to more agents.

Explore the BlueRock platform to see how these capabilities support agent deployment.

FAQ

What is the OCC's supervisory framework for bank use of AI?

It is an examiner resource for banking supervision. The OCC says it establishes no new regulatory expectations or enforceable standards, and banks are not expected to use it to assess their own AI risk management. It organizes supervisory considerations into six pillars: purpose, risk, control, oversight, resilience, and assurance.

What are the six pillars in the OCC's AI framework?

Purpose defines the actions an AI system or agent is permitted to execute. Risk assesses the consequences if its output or action is not fit for purpose. Control, oversight, resilience, and assurance cover preventing unauthorized actions, who monitors and is accountable, the ability to detect, contain, and recover from failures, and the evidence that controls keep operating.

Why does adding a tool change an AI agent's risk?

A workflow does not need a new model to become more consequential. Connecting a transaction tool, enabling CLI wrappers, or granting write access can expand what an agent can do immediately. With MCP, tools can also be updated dynamically, so teams should treat tool changes as part of the lifecycle and decide whether existing boundaries still fit before enabling the new capability.

How should companies assess MCP servers and skills before deploying agents?

Skills, tools, and MCP servers are the connectors an agent operates through, so their capabilities and dependencies should be assessed before they join a business workflow. That assessment should include behavioral and intent-based analysis to decide whether to approve a component, narrow its scope, or require remediation, plus confidence and risk scoring that maps to a broader risk framework. BlueRock's Agentic Trust Registry assesses MCP servers, agents, skills, and tools for ownership, provenance, capabilities, dependencies, evidence, and assessed risk.

What should human-in-the-loop look like for AI agents?

Naming a reviewer does not establish that the person can supervise a deployment effectively. The volume of activity, the clarity of the evidence, and the time available to intervene all matter. Approval and escalation should be designed around specific decisions, incorporate the risk of the action, and give reviewers context and a practical way to restrict execution in real time.