Shadow AI Program Roadmap for Banks: A Phased Plan Under Current Guidance
A shadow AI program roadmap for banks has four phases: discover and inventory unsanctioned AI use, set risk tiers and policy, enforce policy at runtime, then produce and sustain evidence that controls operated. Each phase needs an accountable owner, an exit criterion and a decision gate. As of November 2026, the OCC's examiner framework and the revised model risk management guidance both point to risk-based scaling and evidence that controls keep operating. The phase design here is a recommendation, not a regulatory requirement.
- Discovery of AI sites, coding assistants and unmanaged agent or tool connections, not only sanctioned applications.
- Per-interaction enforcement in the execution path: allow, redact, block, rate limit and escalate for human approval.
- Measured effect on latency and on legitimate use, with a defined exception path.
- Evidence generated by each enforcement decision, with agreed fields, retention and data residency.
- Framework mapping of that evidence, with a named owner who validates it.
- Fit with existing model risk, third-party risk, DLP and information security processes.
What current guidance means for phasing
Two developments frame the program as of November 2026. The OCC issued Bulletin 2026-54 on November 4, 2026, a supervisory framework to help examiners evaluate banks' use of what it calls super intelligence. Its scope includes credit decisioning, payments, fraud detection, customer service, operations, risk management and market activities, and it covers agentic systems. The OCC states that the framework sets no new supervisory expectations or enforceable standards, that banks are not expected to use it for self-assessment, and that examiners tailor their work to the materiality of the use case. It also includes an assurance element: evidence that system or agent controls continue to operate effectively.
Second, SR 26-2 reports that the agencies updated model risk management guidance to clarify principles and emphasize a risk-based approach tailored to a bank's model risk profile and the size and complexity of its operations. The original SR 11-7 attachment remains part of the reference set, so cite SR 26-2 alongside it rather than relying on SR 11-7 alone.
Neither document prescribes phases for shadow AI. Together they support three design choices: scale control depth to materiality, state policy clearly, and keep evidence that controls operated rather than only that a policy existed. The roadmap below follows those choices and should be presented to approvers as a proposal.
Phases 1 and 2: discover, then tier and set policy
Phase 1 objective: build an inventory of AI use the bank has not sanctioned. The CFPB's published AI compliance plan describes monitoring IT activity to find unauthorized AI using the same approach as Shadow IT discovery, so existing monitoring is a reasonable starting point. Typical inputs are web and proxy traffic, endpoint and browser telemetry, identity and SaaS records, and repository activity for coding assistants. Add unmanaged agent and tool connections, and ask business units to report uses they do not recognize as AI. The owner is usually the CISO, with technology and business unit leaders. Exit when the inventory records tools, user populations, data types touched and connection paths, with a dated description of the discovery method.
Phase 2 objective: turn the inventory into risk tiers and enforceable policy. Tier by use-case materiality, such as customer data exposure, confidential information, and influence on decisions. Define for each tier which tools are permitted, which data classes may be used and which actions need human approval. Write the policy so it can be tested: a rule that cannot be evaluated against an interaction cannot be enforced.
Map the program to existing processes at this stage. Model risk management is built around models. Third-party risk covers vendor onboarding. Data-loss prevention covers data movement. It is worth testing whether any of them decides, per employee prompt or tool call, what an AI service may receive or do. Treat that as a likely gap to confirm, not a settled finding.
Phase 3: enforce policy at runtime
Phase 3 objective: apply the Phase 2 policy where AI interactions occur, in the execution path, so policy is enforced at runtime, not reviewed after the fact. The available actions are allow, redact, block, rate limit and escalate for human approval. Enforcement at this level covers individual interactions and tool calls, which is where unsanctioned use creates exposure.
Roll out by tier. Start with the highest-materiality tier and run in observe mode first to measure how many interactions each rule would affect. Then move to enforcement with a defined exception path. Provide an approved alternative for each blocked use, or users will route around the control. Track latency, false positives and exception volume, since these determine whether legitimate adoption slows.
Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, and supports self-managed deployment in the customer cloud with configurable retention and data residency. Exit Phase 3 when each in-scope tier has enforced policy, an exception process and measured operating impact. The owner is typically the CISO, with the business owner accepting the effect on workflows.
Phase 4: produce evidence and sustain the program
Phase 4 objective: show that controls operated, continuously. Agree with internal audit and the model risk function what counts as evidence that controls continue to operate effectively, echoing the OCC's assurance concept. A workable record for each enforcement decision includes the policy version, the risk tier, the identity involved, the action taken, the time, and the approver for any escalation. Aggregated reporting should cover coverage by tier, blocked and redacted volumes, exception ageing and policy changes.
Sustaining the program means a review cadence. Re-run discovery to catch new tools, re-tier use cases as materiality changes, and retire exceptions that no longer apply. Evidence produced by enforcement is more reliable than evidence assembled for an examination, because it exists regardless of whether anyone asked.
Trussed generates continuous audit-ready evidence from enforcement, mapped to NYDFS and NIST among other frameworks. It has run for 18+ months in continuous production in a regulated enterprise, and insurance, a regulated financial sector, is where it is deepest. The bank's compliance and audit teams should validate any framework mapping against their own examination context.
Four phases, four decision gates
Phase 1: Discover and inventory
Find AI sites, coding assistants and unmanaged agent or tool connections.
Phase 2: Set policy and risk tiers
Tier use cases by materiality and define what each tier may do.
Phase 3: Enforce at runtime
Apply allow, redact, block, rate limit or escalate in the execution path.
Phase 4: Evidence and sustain
Show examiners that controls operated, and review them on a cadence.
Take the roadmap to your approvers
See how enforcement in the execution path produces audit-ready evidence, and how the phases map to your existing risk processes.
See How Evidence Is Produced