Bank AI Operating Model Implementation: A 180-Day Plan Under the 2026 Supervisory Framework
A bank can consolidate fragmented AI under one operating model in four phases: inventory and risk tiering (days 1 to 30), shared policy and ownership (days 31 to 60), runtime enforcement in the execution path (days 61 to 135), and continuous evidence (days 136 to 180). As of November 2026, the revised model risk management guidance and the OCC's super intelligence framework do not set enforceable requirements, so the bank's own policy defines the standard it must be able to show.
Criteria for evaluating runtime control and evidence infrastructure
- Enforcement in the path: can it allow, redact, block, rate limit or escalate for human approval on model calls and agent tool calls, with each decision recorded against a named policy version?
- Linked evidence: can records be tied to inventory entries so permitted actions and actual actions are shown together?
- Deployment and residency: can it run in the bank's own cloud, with configurable retention and data residency for the evidence?
- Framework mapping: can evidence be mapped to NIST AI RMF functions and to the bank's model risk documentation?
- Vendor change detection: how does it surface vendor-side AI features or model changes that should trigger reassessment?
- Stated limits: does the provider say plainly what it does not cover, such as transaction-level governance or end-to-end workflow tracing?
What changed in 2026 and what one operating model means
Plans built on the earlier model risk management guidance need updating. On April 17, 2026, the agencies issued revised model risk management guidance (OCC Bulletin 2026-13, Federal Reserve SR 26-2). It describes effective practice in model development and use, validation and monitoring, and governance and controls, including clear policies and roles and responsibilities. The OCC states that it sets no enforceable standards or prescriptive requirements, and that non-compliance will not result in supervisory criticism. The Federal Reserve's Vice Chair for Supervision has said the revised guidance now applies narrowly to traditional models, and that banks rely on existing risk-management frameworks to guide AI use.
On November 4, 2026, the OCC issued a supervisory framework for examiners evaluating bank use of super intelligence and agentic super intelligence (Bulletin 2026-54). It covers use in credit decisioning, payments, fraud detection, customer service, operations, risk management and market activities. The OCC says it does not establish supervisory expectations. Examiners are to tailor exams to the materiality of the use case and look for considerations that supplement existing model risk, governance, technology, compliance and operational-risk practice.
For a risk leader, the consequence is practical. No external document hands the bank its AI standard, so the operating model is the bank's own policy, and the bank must decide and document how AI sits relative to model risk management. NIST AI RMF is a voluntary mapping reference. Its GOVERN function applies across the whole lifecycle, while MAP, MEASURE and MANAGE apply per system. One operating model means one inventory, one tiered policy set, one ownership structure and one body of evidence, applied to vendor tools, coding assistants and agents as well as internally built models.
Days 1 to 60: inventory, tiering, policy and ownership
Phase 1 (days 1 to 30) produces the inventory. The OCC framework describes inventory records that link boundaries to ownership, provenance, autonomy, monitoring and continuity arrangements, and that identify authorized users and use cases. Use those attributes as the record schema. Include business-unit pilots, coding assistants, agents and the AI features that vendors add to products already in use. Under the OCC framework, accountability for vendor-provided AI generally remains with the bank, and a bank may require notice when vendors change models. Then tier each use case by materiality, using autonomy and consequence as the main inputs. Exit criteria: every known use case has a named owner and tier, and a defined process exists to discover new ones. The deliverable examiners would expect to see is the inventory itself, with its change history.
Phase 2 (days 31 to 60) turns tiers into policy. Define a pre-deployment assessment and a reassessment trigger for material changes, including vendor model or feature changes. For consequential or autonomous use cases, require documentation of architecture, permissions, known limitations, expected operating conditions and whether the proposed controls address the identified risks. Write policy so each rule can be expressed as an action: allow, redact, block, rate limit or escalate for human approval.
Ownership needs an explicit decision, because the guidance does not prescribe one. A workable default is that the first line owns each use case and operates its controls, second-line risk owns policy, tiering and effective challenge, and internal audit independently tests the evidence. Technology owns the enforcement layer. Record the allocation and how AI relates to the existing model risk framework. Exit criteria: approved policy set, signed ownership matrix, and a documented position on AI within model risk management.
Days 61 to 135: enforcement in the execution path
Policy that is reviewed after the fact leaves a gap between what was permitted and what happened. Phase 3 closes that gap by enforcing policy at runtime, in the path of each model call and tool call. Start with the highest-tier use cases, then extend to lower tiers.
The following mapping of control points to evidence is a practical inference, not supervisory text. For model calls, enforce data-handling rules on prompts and responses and record the decision against the policy version. For agent tool calls and MCP servers, enforce the permissions documented for the use case, and escalate sensitive actions for human approval. For coding assistants, enforce which repositories and data they may reach, and capture the origin, licence and security context of generated code. For third-party AI sites, decide which are allowed, which are blocked and which require redaction.
Each control point should produce a record with the same minimum fields: the inventory record, the policy version, the action taken, the identity involved and a timestamp. Enforcement applies to individual interactions and tool calls. Linking those events into a business transaction or a full agentic workflow is a separate problem, and a bank should not assume its enforcement layer solves it.
Tradeoffs appear here. Strict blocking in early weeks can disrupt pilots, so many banks run new rules in a monitoring mode on lower tiers first. Escalation for human approval adds latency and needs reviewers who are staffed and trained. Exit criteria: all top-tier use cases route through enforcement, exceptions are tracked with owners, and policy changes can be made without redeploying applications.
Days 136 to 180: continuous, examiner-ready evidence
Phase 4 makes the evidence usable. An examiner's core question is what the AI was permitted to do and what it did. Answering it requires the permission record and the activity record to be linked, so evidence should reference the inventory entry that holds owner, autonomy level, monitoring arrangement and documented permissions. Because examiners tailor review to materiality, evidence depth should scale by tier: full permission documentation and decision records for consequential or autonomous use cases, lighter records for the rest.
The NIST GenAI Profile draft includes the action of comparing generative AI outputs against pre-defined risk tolerances and reviewing them against those standards. Apply the same idea to the bank's own tolerances by showing the tolerance, the rule that enforced it and the outcome. Map the evidence to the NIST AI RMF functions and to the bank's own model risk documentation so second line and internal audit work from one record set instead of rebuilding documentation for each deployment.
Run a mock examination before day 180. Select a high-tier use case and a vendor-embedded feature, and have someone outside the first line reconstruct permissions and activity for a chosen period using only the evidence layer. Gaps found here are cheaper to fix than gaps found in an examination. Exit criteria: the mock request is answered from the evidence layer without manual assembly, retention periods are set, and a quarterly reassessment cadence is in place.
Milestones and exit criteria by phase
Use these gates to decide whether to advance. A phase that misses its exit criteria should extend rather than let later phases build on an incomplete base.
- Days 1 to 30: inventory and tiering: Deliverable: inventory with ownership, provenance, autonomy, monitoring and continuity fields, including vendor-embedded AI. Exit: every use case has an owner and tier, and a discovery process exists.
- Days 31 to 60: policy and ownership: Deliverable: tiered policy set, assessment and reassessment triggers, ownership matrix across three lines. Exit: approval by risk leadership and a documented position on AI within model risk management.
- Days 61 to 135: runtime enforcement: Deliverable: enforcement for model calls, tool calls, coding assistants and third-party AI sites, starting with top tiers. Exit: top-tier use cases enforced, exceptions tracked, decisions recorded against policy versions.
- Days 136 to 180: evidence: Deliverable: evidence tied to inventory records and mapped to NIST AI RMF and internal model risk documentation. Exit: mock examination answered from the evidence layer, retention set, reassessment cadence running.
180 days, four phases
Days 1 to 30
Inventory and risk tiering
Days 31 to 60
Shared policy and ownership across the lines of defense
Days 61 to 135
Enforcement for interactions and tool calls
Days 136 to 180
Continuous, examiner-ready evidence
Test your plan against an examiner's question
Walk through your inventory, policy and evidence design with a Trussed expert, and see how runtime enforcement produces audit-ready evidence for individual interactions and tool calls.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.