How Much Unsanctioned AI Use Is Acceptable? Setting a Shadow AI Risk Appetite for Banks
A bank should tolerate unsanctioned AI use only in tiers defined by data sensitivity, tool category and business impact, and should state which tiers are tolerated, conditional, restricted or not tolerated. Each tier needs measurable indicators and escalation thresholds. The parts of the appetite that can be enforced on individual interactions and tool calls should be enforced at runtime, with evidence retained, because a written statement alone does not show the appetite is operating.
Evidence a reviewer is likely to ask for
- The approved appetite statement and tier definitions, with version history and approval dates
- Per-interaction policy decisions: tier, tool, data classification and the action taken (allow, redact, block, rate limit, escalate)
- Escalation records showing who approved or denied, when, and the expiry of any exception
- Indicator reports against thresholds, with the follow-up taken when a threshold was crossed
- Retention and data residency settings for the evidence itself, with their rationale
- Records of tolerance reviews and the changes made, tied to the model risk and third-party inventories
Why prohibition and silence both fail
Banks cannot fully prevent staff from reaching public chatbots, browser-based AI sites, coding assistants and unapproved agents or tools. A risk appetite statement that bans all of it is unenforceable in practice, and it pushes usage out of view. A statement that says nothing leaves the bank with no defined tolerance, no thresholds and nothing to measure against.
The more defensible position treats shadow AI as a tolerance question. Some unsanctioned use carries little risk and can be allowed with visibility. Some should be constrained by conditions. Some should never occur. NIST's AI Risk Management Framework takes this stance on tolerance generally: it does not prescribe a tolerance, it describes it as the organization's readiness to bear risk to achieve its objectives, and it expects organizations to define a reasonable tolerance where no established guidance exists. It also expects tolerances to change over time.
This does not require a new appetite category. In its November 2026 supervisory framework for examiners on what it calls Super Intelligence and agentic Super Intelligence, the OCC said such use is not a separate risk category and that banks are not required to define a separate appetite for it. The framework is non-binding, and examinations are tailored to a bank's size, complexity and risk profile. By extension, a shadow AI tolerance is best placed inside the operational risk, model risk, technology, third-party and compliance appetite statements the bank already maintains.
A four-tier tolerance model
The OCC's framework says risk classification can rest on potential consequences, transfer of agency and the degree of human oversight, rather than technical sophistication. Those factors translate well to unsanctioned use. The tiers below are an editorial framework for banks to adapt, not a regulatory requirement.
- Tier 1: Tolerated: Public or non-sensitive information in chat-style tools, no action taken on the user's behalf, a person reviews any output. Allow, with usage visible to risk.
- Tier 2: Conditional: Internal but non-confidential content, or coding assistants on non-production code. Allow with redaction of identifiers and secrets, rate limits and recorded user attestation.
- Tier 3: Restricted: Confidential business or supervisory information, or any tool that acts on a user's behalf, such as agents and tool calls. Blocked by default, with escalation for human approval and a time-bound exception.
- Tier 4: Not tolerated: Customer nonpublic personal data, credentials, or unapproved agents and MCP servers that can change production systems. Block, and treat every attempt as a reportable event.
Indicators and escalation thresholds
A tolerance is only meaningful if it can be measured. No regulator has set numeric thresholds for unsanctioned AI use, so the values are the bank's to choose and should be labelled as bank-set. Indicators worth tracking include the volume of unsanctioned interactions by tier, the share redacted or blocked, the number of newly discovered tools awaiting assessment, exceptions open past their expiry, repeat events by business unit, and time from event to triage.
Escalation thresholds should be written as triggers rather than aspirations. For example, any Tier 4 attempt that reaches a tool without being stopped escalates to a named senior risk owner. A sustained rise in Tier 3 requests from one function prompts a review of whether a sanctioned alternative is missing. A discovered tool that handles Tier 2 or higher data enters third-party due diligence or is blocked pending review.
Because tolerances are expected to shift, set a review cadence for the thresholds themselves and record each change with its rationale. A tolerance that has never been revisited is harder to defend than one that has been adjusted against observed behavior.
What policy can state and what must be enforced at runtime
Policy can define tiers, ownership, exception rules and consequences. It cannot stop an employee pasting restricted data into a browser tab at 4 p.m. on a Friday. The line to draw is between conditions that are about people and process, such as attestation and training, and conditions that apply to a single interaction or tool call, such as data type, destination tool and requested action.
The second group can be enforced in the execution path. NIST's generative AI profile says organizations can restrict AI applications that exceed stated risk tolerances. Applying that to individual interactions is an inference, but a practical one: each tier maps to an action. Tier 1 is allowed, Tier 2 is allowed with redaction and rate limiting, Tier 3 is escalated for human approval, and Tier 4 is blocked. Agent and tool-call actions deserve a distinct tier from plain chat, since examiner attention follows autonomy and transfer of agency.
Trussed enforces policy in the execution path across AI sites, coding assistants, agents, tools and MCP servers, with allow, redact, block, rate limit and escalate-for-approval actions applied to individual interactions and tool calls. This is enforced at runtime, not reviewed after the fact. It does not extend to transaction-level governance or end-to-end tracing of agentic workflows, so those parts of an appetite still rely on other controls.
Model risk management and third-party risk need a link to this. Tools discovered through enforcement should feed the inventory so that anything influencing decisions is brought into model risk processes or blocked, and tools without a contract are either denied or limited to the lowest tier. The OCC has described third-party expectations as due diligence, contract management and ongoing oversight; how those apply to tools used without a contract is a design decision the appetite should make explicit.
From appetite statement to evidence
Define tiers
Classify unsanctioned use by data, tool category and impact.
Set thresholds
Attach indicators and escalation triggers to each tier.
Enforce at runtime
Allow, redact, block, rate limit or escalate each interaction.
Retain evidence
Keep decisions, approvals and policy versions for review.
Test your appetite against the execution path
Talk through how tiered tolerances map to runtime enforcement of individual interactions and tool calls, and what audit-ready evidence results. Trussed is deepest in insurance, a regulated financial sector, and supports self-managed deployment in your cloud with configurable retention.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.