Shadow AI Ownership in Banks: Splitting Accountability Across the CIO, CISO, CRO, and CDO
No single executive should own shadow AI at a bank. Ownership works better as a split of five decisions: discovery, policy definition, runtime enforcement, exception handling, and evidence retention. The CISO typically operates discovery and enforcement, the CIO supplies the platform and retention base, the CDO defines data rules, and the CRO sets risk appetite and challenges the result as second line. Runtime enforcement and audit-ready evidence give each owner a control to operate and proof to show an examiner.
What to ask of any control that claims to cover shadow AI
- Which AI traffic does it see and enforce on: AI websites, coding assistants, agents, MCP servers, embedded features? What happens to usage that bypasses it?
- Can it apply per-agent identity and scope of authority, and route actions to human approval?
- What does it record for each AI interaction, and how is that evidence retained, protected from alteration, and mapped to NIST and other frameworks?
- Can policy decisions and exceptions be attributed to a named owner?
- Can reporting separate first-line operation from second-line oversight?
- Can it feed an AI inventory with ownership, provenance, autonomy and monitoring fields, including third-party and embedded AI?
What the current guidance changes
Shadow AI is AI deployed outside the governed program, typically because governance is too slow or unclear. In practice it includes unapproved online AI tools, unsafe browser plugins, and third-party applications that add LLM features through updates. At a bank it also includes coding assistants and model features embedded in tools that were approved for other reasons.
The model risk reference point has changed. The OCC, Federal Reserve Board and FDIC issued revised interagency model risk management guidance (OCC Bulletin 2026-13, dated April 17, 2026, with the Federal Reserve transmittal as SR 26-2). It clarifies principles, sets out a risk-based approach, and rescinds prior model risk management guidance. It discusses governance and controls, including clear policies and roles and responsibilities, and it states that practices suited to one bank may be inappropriate for another with a different risk profile or model use.
The earlier guidance, SR 11-7, split model risk roles into ownership, controls and compliance. It allowed banks to assign them in different ways, provided reporting lines and incentives were clear and conflicts of interest were addressed. That logic still helps when assigning shadow AI duties. Separately, the OCC issued a supervisory framework on November 4, 2026 (Bulletin 2026-54) to support examiners evaluating bank use of what it calls super intelligence. Related OCC material emphasizes ownership, inventory, and controls that prevent unauthorized, inappropriate or unsafe actions. As of November 2026, the direction for examiners is consistent: name the owner, keep an inventory, and show the controls.
A RACI-style split by decision
The guidance does not prescribe which executive holds which duty, so the mapping below is a working proposal. Adjust it to your reporting lines. The principle is to keep first-line operation of controls separate from second-line challenge, so that the group setting risk appetite is not the group grading its own controls.
- Discovery of AI use: CISO accountable and operating detection of AI sites, coding assistants and embedded features. CIO supplies network, endpoint and SaaS telemetry and the sanctioned platform list. CDO identifies which data domains are exposed. CRO reviews coverage and gaps as second line.
- Policy definition: CRO owns risk appetite and challenges policy. CDO defines classification and permitted data use per class. CISO writes security and leakage rules. CIO sets technology standards for sanctioned platforms.
- Runtime enforcement: CISO accountable for operating allow, redact and block decisions. CIO responsible for the network and platform paths that make enforcement possible. CDO supplies classification inputs. CRO tests effectiveness independently and does not operate the control.
- Exception handling: CRO approves risk acceptance. CDO approves data-use exceptions. CISO assesses technical exposure and sets compensating controls. CIO implements approved changes. Each exception needs a named approver and an expiry.
- Evidence retention: CIO responsible for storage, retention and data residency. CISO accountable for producing enforcement evidence. CRO consumes it for monitoring and examiner response. CDO uses it to show data handling by class.
Where runtime enforcement applies, and where it does not
Runtime policy enforcement acts on individual interactions and tool calls in the AI execution path. The available actions are allow, redact, block, rate limit, and escalate for human approval. For the CISO this is the operating control for data leakage. For the CDO it turns classification into a decision made at the moment data would leave. For the CIO it gives sanctioned platforms a consistent policy point. Escalation for human approval corresponds to the human-in-the-loop and human-on-the-loop arrangements the OCC names as controls, provided approvals route to a defined approver role and the decision is recorded.
Enforcement has a boundary. It can only act on traffic that passes through the enforcement point, so usage that bypasses sanctioned paths is not covered by it. That is why discovery sits beside enforcement as a separate duty, and why shadow AI findings should feed back into the inventory and the policy.
Some responsibilities are not enforcement problems at all. Risk appetite, accountability assignment, and the decision about which AI uses are acceptable remain governance decisions held by the CRO and senior management. Enforcement carries those decisions out; it cannot make them. Because shadow AI grows when governance is slow or burdensome, pair enforcement with a fast, defined path for approval requests, and tier controls by use case and data sensitivity in line with the risk-based approach in the revised guidance.
The evidence each owner must be able to produce
Examiners ask two questions: who is accountable, and what proof exists. Many banks cannot answer either one consistently because the answer lives in several functions. Build the evidence set around the same five decisions.
For discovery, that means an AI inventory in which each entry carries a named accountable owner and the controls that apply, covering third-party and embedded AI as well as internally built models. The OCC material describes inventories that link ownership, provenance, autonomy, monitoring and continuity. For policy, it means the version in force and who approved it. For enforcement, it means a record of what each policy decision was for each interaction or tool call. For exceptions, it means the approver, the rationale, and the expiry. For retention, it means evidence that is kept for a configured period and under a defined residency rule.
Separate reporting for first line and second line matters. The CISO should be able to show how the control operated, and the CRO should be able to show independent challenge of it. Evidence generated by enforcement is stronger than evidence reconstructed after the fact, because it records the decision when it was made. When evaluating any evidence source, test for completeness across the traffic in scope, retention that matches your obligations, protection from alteration, and mapping to the frameworks your examiners use, such as NIST.
Where Trussed fits
Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, using allow, redact, block, rate limit, and escalate for human approval. Enforcement is applied at runtime, not reviewed after the fact, and it generates continuous audit-ready evidence mapped to NIST and other frameworks including NYDFS, NAIC, HIPAA and the EU AI Act. It can be self-managed in the customer cloud with configurable retention and data residency.
Insurance is where Trussed is deepest among regulated financial sectors. Approved production proof includes more than 18 months of continuous production in a regulated enterprise, 99.95% uptime on live production traffic, and more than 100 models under one policy plane. Trussed does not assign ownership, set risk appetite, or discover usage that never reaches an enforcement point. Those remain decisions for your CIO, CISO, CRO and CDO.
Shadow AI ownership by decision
CIO
Sanctioned platforms, network and endpoint telemetry, evidence storage
CISO
Discovery, data leakage controls, runtime enforcement
CRO
Risk appetite, exception approval, second-line challenge
Chief Data Officer
Data classification and permitted data use
Map your shadow AI responsibilities to enforceable controls
Review how enforcement in the execution path and audit-ready evidence support each owner in your RACI.
Explore Agent and Tool-Call Enforcement