Phasing shadow AI enforcement at a bank: observe, warn, then enforce
Run shadow AI enforcement as three policy modes applied to individual interactions and tool calls. In observe, record every interaction and the policy decision it would have received. In warn, notify, redact or escalate while tuning the policy. In enforce, block and rate limit what policy prohibits. Advance only when documented exit criteria are met, and keep a record of the policy version, action and outcome at every stage.
Enforcement is a progression of modes, not a cutover
Blocking everything on day one pushes business users and developers toward personal devices and unmonitored tools. Leaving use unmanaged leaves data exposure, model risk management gaps and examiner questions unanswered. A phased rollout resolves this by keeping one written policy and changing only the mode in which it is applied.
The unit of enforcement matters. Policy applies to an individual interaction or tool call: a prompt sent to a public AI site, a request from a coding assistant, a model call, or a call to a tool or MCP server. The possible actions are allow, redact, block, rate limit and escalate for human approval. Each phase uses a different subset of those actions.
The mode should be configurable by user group, tool and data class. A trading-technology team, a retail operations team and a general staff population will not reach the same exit criteria on the same date. Every mode change should itself be a recorded event with an approver.
There is precedent for graduated responses. As of November 2026, the OCC's supervisory framework (Bulletin 2026-54) says thresholds may trigger review, re-validation, reduced autonomy or suspension, and that a bank may define internal escalation and notification thresholds. The language is permissive and calibrated to the use case, but it fits a staged approach in which thresholds, not calendar dates, drive the next step.
Phase 1: Observe
The goal of the first phase is a baseline. Scope should be broad in coverage but passive in effect: AI sites, coding assistants, direct model calls and tool or MCP server traffic across the user populations you intend to govern. Existing Shadow IT discovery is a reasonable starting point. The CFPB's own AI compliance plan describes finding unauthorized AI through the same monitoring approach it uses for Shadow IT.
The only action in this phase is allow, with a record. For each interaction, capture who or what made the request, the destination, the data classes detected, and the decision the policy would have taken had it been enforced. That last field is what makes observe useful. It lets you measure the effect of enforcement before anyone is affected.
Exit criteria should be written down before the phase starts. Reasonable examples are an inventory of unsanctioned tools reviewed and categorized by risk, data classes mapped to policy rules, and a stable count of would-have-blocked events that the security team has sampled and judged accurate. Specific numeric thresholds are a bank decision, but they should be explicit and approved by the same people who will later approve exceptions.
Phase 2: Warn
The warn phase introduces user-visible actions while keeping work moving. Typical actions are a notice explaining the policy, redaction of sensitive data before a request leaves the bank, and escalation to a human approver for higher-risk requests such as a coding assistant sending proprietary source to an unapproved model. Blocking remains off for everything except data classes that are already prohibited outright.
This phase is where policy is calibrated. The OCC framework language notes that poorly calibrated thresholds can increase false positives, false negatives and other compliance risks. A warn period gives you real traffic to tune against, and it shows which rules generate noise and which gaps remain.
Track a small set of signals: unsanctioned tool usage by group, sensitive data exposure events, the rate at which users change behavior after a warning, approval turnaround for escalated requests, and exception request volume. A rising exception volume for one tool usually means a legitimate need that should be met with a sanctioned alternative, not a longer exception list.
Exit criteria can include warning-to-compliance rates holding at an agreed level, false-positive reviews below an agreed tolerance, a sanctioned alternative available for each major use case, and an approver process that responds within a time the business accepts.
Phase 3: Enforce, with exceptions that are recorded
In the enforce phase, policy decisions become binding. Prohibited destinations and data classes are blocked, high-volume or anomalous traffic is rate limited, and sensitive content is redacted by default. Escalation for human approval remains available for cases that policy designates as approvable.
Enforcement should stay scoped. Move one user group, tool category or data class at a time, and keep the ability to return a group to warn if a rule misfires. NIST AI RMF says that when an AI system presents unacceptable risk, deployment should cease in a safe manner until risks can be managed. That principle supports having a defined, tested block path, though it does not prescribe phasing.
Exception handling needs structure. Each exception should name the requester, the tool or data class, the business reason, the approver, the compensating control and an expiry date. Expired exceptions should revert automatically. Exceptions that are renewed repeatedly are a signal to either approve the tool formally through third-party risk and model risk processes or retire the need.
Ongoing success metrics shift from adoption of the policy to its effect: blocked attempts that recur, the volume of active exceptions, and sensitive data events that reach an external destination, which should trend toward zero.
Phase summary
Use this table as a template for the phase charter your steering group approves.
- Observe: Actions: allow and record. Primary metrics: tools discovered, would-have-blocked events. Evidence: interaction records with the policy decision that would have applied.
- Warn: Actions: warn, redact, escalate for human approval. Primary metrics: warning-to-compliance rate, false-positive rate, exception requests. Evidence: the notice shown, redaction applied, approver and decision.
- Enforce: Actions: block, rate limit, redact, escalate. Primary metrics: recurring blocked attempts, active exceptions, external data exposure events. Evidence: block records, exception approvals, mode change history.
What must sit in the execution path, and what evidence it produces
Review after the fact can tell you what happened. It cannot redact a prompt or stop a tool call. Redact, block, rate limit and human-approval actions have to run in the execution path, at the point where the request is made. That is the difference between a written AI policy and the behavior it describes. Discovery and reporting can stay out of band, but anything that changes the outcome of an interaction cannot.
For each enforcement point (AI sites, coding assistants, model calls, and tool or MCP server calls), ask which actions it can take inline and which it can only observe. Where a point can only observe, treat it as an observe-mode control and compensate elsewhere.
Every decision should produce a record containing a timestamp, user or agent identity, destination, data classes detected, policy version, current mode, action taken, outcome and any approver. Mode changes and exceptions need their own records. Together these show an examiner or model risk management reviewer which policy applied, what it did, and who changed it. Supervisory guidance for banks generally expects AI usage to be covered by existing model risk management and third-party risk processes, so these records should be reportable to those teams and retained for a period they set.
Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with allow, redact, block, rate limit and escalation for human approval applied to individual interactions and tool calls. Enforcement generates audit-ready evidence mapped to frameworks including NYDFS and NIST. Insurance, a regulated financial sector, is where Trussed is deepest. Whatever tooling you select, test it against the record fields above before enabling any block.
Three policy modes, one policy
Observe
Discover use, allow it, and record the policy decision each interaction would have received.
Warn
Notify users, redact sensitive data, escalate for human approval, and tune the policy.
Enforce
Block and rate limit prohibited use, with documented exceptions and evidence for every action.
Plan your observe-to-enforce rollout
Talk through scope, exit criteria and evidence requirements for each phase with a Trussed specialist.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.