Building a Paved Road for AI in Banking Under the 2026 Model Risk Guidance
A paved road for AI in banking is a standard, pre-approved route for using models, agents, tools and coding assistants that is also the easiest route to use. It combines one entry point, approved models and tools, default policy enforcement, self-service onboarding and automatic evidence. Enforcement happens where interactions and tool calls occur, so teams have little reason to go around it.
Where enforcement belongs: the execution path
- 1
Allow
The request matches policy and proceeds, with the decision recorded.
- 2
Redact
Sensitive content is removed or masked before it reaches the model or tool.
- 3
Block
The request violates policy and does not execute.
- 4
Rate limit
Volume or frequency is capped for a user, agent or tool, which limits the effect of misuse or runaway behavior.
- 5
Escalate for human approval
A defined action pauses until a named role approves or rejects it.
Evaluating a runtime enforcement layer
- Does enforcement sit in the execution path, so decisions change outcomes at runtime and are not reviewed afterward?
- Can one policy set cover models, agents, tools, MCP servers and coding assistants?
- Can policies be tiered by use-case consequence, with escalation roles defined explicitly?
- Does each decision produce an evidence record that model risk management and internal audit can query?
- Can deployment, retention and data residency be configured to the bank's requirements?
- Is onboarding self-service, so the enforced path stays the lowest-friction option?
Why mandates alone do not hold
Shadow AI in banks tends to follow a familiar pattern. An engineering or business team needs a model, a coding assistant or an agent framework. The approved route is slower, harder to use or less capable than what is freely available, so the team uses the other one. A policy document that prohibits this does not change the incentive. It only moves the activity out of view.
The consequence is inconsistent controls and gaps in the evidence that model risk management, internal audit and examiners expect to see. Reviewing usage after the fact does not close those gaps, because the interaction has already happened and the record is incomplete.
The more useful framing is platform engineering. Treat the approved path as a product with internal customers, and treat control as a property of that product. If the paved road is faster to start, more capable and already compliant, the control objective and the adoption objective point the same way.
The five components of a paved road for AI
A paved road is defined by what it standardizes. Each component below removes a decision or a delay from the team using it, and adds a control the bank can rely on.
The components work as a set. An entry point without an approved catalog is a bottleneck. A catalog without enforcement is a list. Enforcement without evidence cannot be shown to anyone outside the platform team.
- Standard entry point: A single route through which models, agents, tools and coding assistants are reached, so every interaction passes through the same policy layer and identity context.
- Pre-approved models and tools: A maintained catalog with each entry's permitted use, data classes and tier, so teams start from approved options instead of requesting exceptions.
- Default policy enforcement: Baseline policies applied automatically to every interaction and tool call, with stricter policies attached to higher tiers rather than requested by the team.
- Self-service onboarding: Teams register a use case, declare its tier and receive access through templates and automation, with human review reserved for higher-consequence cases.
- Evidence generation: Each policy decision is recorded as it happens, so model risk management and audit receive evidence without teams assembling it.
Adoption incentives that make teams choose the approved path
Adoption is a product problem. The approved path should win on time to first use. If a team can reach an approved model or coding assistant in hours through self-service, while the alternative needs a procurement exercise and a security review, the choice is easy. Default policies matter here: a team that inherits sensible enforcement does not need to design controls before it can start.
Tiered approval is the second lever. Low-consequence uses, such as internal drafting or code assistance on non-sensitive repositories, should pass with default policy and no queue. Higher-consequence uses, such as agents that can take actions in customer-facing or financial systems, warrant added review. Tiering keeps scrutiny proportionate and stops review capacity from becoming the reason teams leave the path.
Capability parity is the third. The catalog needs to include the models and tools teams actually want, and it needs a short route for adding new ones. A paved road that is safe but outdated will be bypassed regardless of how well it is communicated.
Finally, the road should be measurable. Time to onboard, share of AI traffic passing through the entry point and the volume of exceptions all show whether the path is being chosen or avoided.
What the 2026 guidance changes for model risk and examiner evidence
As of November 2026, the baseline for model risk management has changed. In April 2026 the OCC, the Federal Reserve Board and the FDIC issued revised interagency model risk management guidance (OCC Bulletin 2026-13). It clarifies model risk management principles, takes a risk-based approach and rescinds prior model risk management guidance and other issuances. Internal policies that still cite the rescinded guidance as current should be checked and updated with counsel.
The revised guidance states that it does not set enforceable standards or prescriptive requirements, and that non-compliance will not result in supervisory criticism. It also notes that practices suitable for one banking organization may be inappropriate for another with a different risk profile or model purpose. For platform teams, this means the paved road should document how each tier maps to the bank's own risk assessment, not to a fixed external checklist.
The guidance describes effective practice in three areas: development and use, including testing; validation and monitoring; and governance and controls, including clear policies, roles and responsibilities. Runtime evidence supports the second and third directly. Per-interaction records of which policy applied and what decision was made give monitoring and validation teams material to work from, and they show that roles such as approvers were actually exercised.
Separately, an OCC supervisory framework on bank use of advanced AI agents states that control intensity could increase with the potential consequences of an agent's failure or misuse. Higher-consequence cases may warrant tighter limits on autonomy, stronger controls and more direct human oversight. It also says human oversight is most effective when the bank defines the nature of the human involvement. That maps onto tiered policies and explicit escalation rules: define which actions need approval, and by whom, before the agent is deployed.
An evidence record for each decision would reasonably capture the request context, the policy and tier applied, the decision, the approver where escalated, and a timestamp. These fields are a design proposal for the bank to validate with its own risk and audit functions.
Components of a paved road for AI
Standard entry point
One route for models, agents, tools and coding assistants.
Approved catalog
Pre-approved models and tools with known permissions.
Default enforcement
Allow, redact, block, rate limit or escalate, applied automatically.
Self-service onboarding
Teams join the path without a ticket queue.
Evidence generation
A record of each decision, produced as a byproduct of use.
Design the approved path with enforcement built in
Talk through how runtime policy enforcement and audit-ready evidence could fit your paved road for AI.
Talk to an Expert