Executive Summary
- Reuters reports that the Fed, OCC, and FDIC are asking banks how they control AI agents, with questions on authority, oversight, auditability, human involvement, and shutdown.
- SR 26-2 excludes generative and agentic AI from its scope. On November 4, the OCC issued nonbinding guidance for examiners assessing banks’ AI use. The supervisory questions Reuters reported on October 13 are a separate development. For contact centers, the useful preparation is evidence: what each agent may do, who approves what, and how an action is reconstructed or stopped.
- Governing execution in systems of record is one part of that evidence. Model risk management, data governance and vendor oversight sit alongside it.
Supervisors Are Asking About AI Agents Before Any Rulebook Exists
According to Reuters, the Fed, OCC, and FDIC are putting written and verbal questions to large, mid-size, and smaller banks. The questions cover “what limits are placed on systems, how their actions are monitored and approved, and whether banks can clearly track and audit AI agent decisions” (Reuters, October 13, 2026).
Supervisors are also asking about safeguards against agents “exceeding their authority,” whether “humans remain involved in key decisions,” and “how banks can shut down autonomous systems” (Reuters, October 13, 2026). The same report notes that regulators “are not yet being prescriptive about what lenders should be doing.”
Four layers sit behind those questions, and they carry different weights.
Reported Supervisory Interest
The Reuters questions show what supervisors are looking at. They are not a published standard.
Established Requirements and Guidance
SR 26-2, issued in April 2026, states that generative and agentic AI “are not within the scope of this guidance.” It adds that a bank’s “risk management and governance practices should guide the determination of appropriate governance and controls” for systems it does not cover (Federal Reserve, OCC, FDIC, 2026).
Existing obligations, including applicable GLBA requirements for protecting customer information, remain relevant when banks use AI agents.
Examination Guidance
On November 4, the OCC issued its first framework for how examiners assess banks’ AI use (OCC Bulletin 2026-54, “Super Intelligence: Supervisory Framework”). The OCC states that it “does not establish supervisory expectations for banks.”
Reuters reported that under it “a simple AI agent that is authorized to initiate transactions could be considered riskier,” while “higher-risk uses of AI may require greater restrictions on autonomy or more direct human oversight” (Reuters, November 4, 2026).
Broader Governance Guidance
The Financial Stability Board’s October 15 report sets out 12 sound practices. It states that they “do not establish an international standard” and that practices specific to agentic AI “remain nascent” (FSB, 2026).
None of this makes the five questions mandatory controls. It does make them predictable. A bank that deploys agents in its contact center should be able to show its answers. The Execution Layer Gap: 4 Signals Your Bank’s AI Agents Are Operating Ungoverned covers the warning signs.
Would Your Answers Hold Up Today?
See how CCIP enforces policy at runtime and records what happened.
The Five Questions, Translated for the Contact Center
Gartner predicts that agentic AI will autonomously resolve 80 percent of common customer service issues by 2029 (Gartner, March 2025). In banking, resolving an issue often means a card lock, a payment, or a profile change. Each of those actions gives supervisors something to ask about.
1. Authority: Can You Show What Each Agent Is Allowed to Do?
Picture an IVA that may lock a debit card. A caller then asks it to raise a daily limit. What stops the agent should be an explicit list of permitted actions, one per agent, with everything else out of bounds by default. A system prompt that says “only help with cards” is not that list. Permissions also need conditions. An agent allowed to read account data should reach it only after identity verification succeeds.
The FSB notes that financial institutions recognise “not every AI agent requires the same autonomy and access to achieve its goals,” and that the risks of an agent’s access to tools and systems “may compound multiplicatively (rather than additively)” as permissions expand (FSB, 2026).
A bank should be able to show who approved each permission and when it last changed. For more on why permissions provisioned in advance fall short for agents that act at runtime, see our post on least privilege for AI agents.
For SI and CCaaS partners: Document each agent’s permitted actions during solution design, with an approver and a change history, so the client has the list at go-live.
2. Oversight: Who Monitors the Agent, and Which Actions Need Approval?
Requiring a supervisor to approve every agent action defeats the purpose of the agent. The FSB names “the impracticality of real-time human monitoring of agent decisions as their use scales.” It describes financial institutions strictly separating an agent’s pre-defined goals, permissions and rules, including “rules (e.g. when to seek human approval),” from the data it can access to achieve those goals ” (FSB, 2026).
In practice, a bank should be able to show two things: routine actions such as balance lookups are monitored as a stream, with thresholds that flag anomalies, and consequential actions, such as a fee waiver above a set amount or an account closure, trigger an approval before they execute. The approval trigger should be written policy that any reviewer can read.
For SI and CCaaS partners: Agree on the approval matrix during scoping, so the client has a written policy from the first release.
3. Auditability: Can You Reconstruct One Transaction End to End?
When the CCaaS platform manages the conversation and the core holds the record, reconstructing why an agent moved money can mean joining those records by hand.
The FSB describes traceability as following “what triggered the action; what information and tools it accessed; what steps it took; and whether it acted within its permissions throughout this chain of steps.”
It also describes unique identifiers that let an institution see “which humans configured and oversaw” each agent (FSB, 2026).
Try This Test
Pick one agent-initiated transaction from last month. Trace it from trigger to outcome, including the permissions checked at each step. Time how long it takes.
| Download Case Study | How did one US bank make self-service transaction-ready? Read how it used CCIP. |
4. Human Involvement: Which Decisions Stay With People?
A bank may reserve decisions involving disputed transactions, hardship arrangements and vulnerable customers for human review, depending on the nature and risk of the interaction. The FSB’s Sound Practice 10 calls for human oversight proportionate to the materiality, risk, autonomy, complexity, and explainability of each AI use case (FSB, 2026).
A bank should be able to name those decisions and show the escalation path. The handoff matters as much as the rule. The human agent should receive the conversation, the action requested, and the reason the AI agent stopped, so the customer does not start over.
5. Shutdown: When Did You Last Rehearse Stopping an Agent?
Reuters reports that supervisors are asking “how banks can shut down autonomous systems” (Reuters, October 13, 2026).
The FSB’s final report also discusses implementing and testing risk-based kill-switch mechanisms that can pause an agent, limit or override its capabilities, or remove its access to systems and tools.
It treats fallback processes, recovery expectations and escalation routes as part of operational resilience planning (FSB, 2026). For a contact center, that question turns operational quickly.
Someone must hold the authority to stop an agent. Interactions in progress need a defined outcome. Calls and chats need a defined fallback, which may involve an alternative IVR path, other routing arrangements and sufficient live-agent capacity.
A shutdown plan that has never been tested is an assumption. A bank should be able to show when it last rehearsed one.
A Readiness Check Shows Where the Evidence Is Thin
The table lists what a bank should be able to demonstrate in each area. These are not regulatory requirements. They are the gaps that tend to surface when the questions arrive.
Governing Execution Is One Part of AI Governance
Gartner expects more than 40 percent of agentic AI projects to be canceled by the end of 2027, citing inadequate risk controls among the causes (Gartner, June, 2025). In a bank contact center, many of those controls apply at the moment an agent acts.
| Download Whitepaper | 4 in 10 Agentic AI Projects Won’t Survive 2027 |
NovelVox CCIP governs that moment. It sits between the AI agent, the CCaaS platform, and core systems such as Jack Henry, Fiserv, and FIS. CCIP enforces policy, approvals, and sequencing before an action reaches the system of record. Every tool an agent can call is tagged Low, Medium or High risk, and CCIP can require sign-off before high-risk actions such as account closure execute.
It records each action against the policy applied, in an audit trail built to your institution’s requirements. Each tool call is logged by timestamp, actor, status and response time, with records filterable by session or correlation ID. Model risk, data governance, vendor oversight, and board accountability sit outside that layer, and they remain the bank’s work.
Pick One Transaction. Can You Reconstruct It?
See how CCIP records every action against the policy behind it.
