Bank AI Gateway Sprawl: A Shared Runtime Policy Enforcement Model Under Current Model Risk Guidance
Bank AI gateway sprawl is best prevented by offering one shared runtime policy layer that enforces policy in the execution path, lets each business line vary policy within defined risk tiers, and generates evidence as a by-product of enforcement. Business lines adopt it when it is faster to use than building their own, so the central team must treat adoption as a product problem.
Separate routing from enforcement, then separate common controls from configuration
- 1
Common runtime layer
Enforcement mechanics, sensitive data detection and redaction, tool-call authorization, human-escalation routing, rate limiting, evidence format, retention and policy versioning.
- 2
Business-line configuration
Policy profiles by risk tier and use case, such as which data classes a model may see, which tools an agent may call, and which actions need human approval.
- 3
Approval of changes
A defined path in which business-line owners propose changes, and model risk management and security approve those above a set tier.
- 4
Coverage scope
Models, agents, tools, MCP servers and coding assistants pass through the same enforcement point, so no AI path sits outside the layer.
Evaluation criteria for a shared enforcement layer
Why each business line builds its own gateway
Gateway sprawl is rarely a failure of discipline. It is the predictable result of an operating model in which business lines carry delivery targets, while a central platform team cannot yet offer a policy layer that is quick to adopt. A retail team that needs a model endpoint this quarter will stand up a proxy, add an API key vault and a basic content filter, and move on. Markets, wealth and operations teams do the same, each with different assumptions.
The technical factors compound the organizational ones. Models, agents, tool and MCP server connections, and coding assistants each arrive with their own access paths. A team that solves only model access leaves tool calls and developer tooling outside any common control. Each new layer then carries its own redaction rules, rate limits, approval logic and log format.
The outcome is a set of gateways that look similar on an architecture diagram but behave differently on sensitive data, tool permissions and exception handling. That difference is the actual problem. It is an architecture and operating-model issue, not a tooling preference.
What fragmentation costs under current supervisory expectations
On April 17, 2026, the OCC, in coordination with the Federal Reserve Board and FDIC, issued updated model risk management guidance for OCC-supervised institutions. It rescinds OCC Bulletin 2011-12 and OCC Bulletin 2021-19. The guidance says practices should be risk-based, tailored, and commensurate with size, complexity and extent of model use, and it is expected to be most relevant to banking organizations with over $30 billion in total assets. It addresses factors that influence model risk, model development and use, validation and monitoring, and governance and controls. Control documentation that cites the earlier bulletins should be reviewed against the new issuance.
Separately, the OCC has published an examiner framework for bank use of super intelligence and agentic super intelligence, covering credit decisioning, payments, fraud detection, customer service, operations, risk management and market activities. As of November 2026, it says banks need not define a separate risk appetite for these systems. Risks are identified, measured, monitored and controlled within the existing risk management framework. It also says control intensity can increase with the potential consequences of an agent's failure or misuse, and it includes Resilience and Assurance components. Resilience concerns detecting, containing, recovering from and remediating failures. Assurance concerns evidence that controls continue to operate effectively.
A fragmented gateway estate makes each of these harder. Model risk management cannot apply one standard to controls that differ by business line. Control intensity cannot be tied to consequence if enforcement varies by who built the proxy. When an examiner asks what AI was allowed to do and what it did, the answer has to be assembled from several log formats with different retention periods, and the CIO has no single, consistent response.
Make the shared layer the easier path
Mandates alone push teams toward workarounds. Adoption holds when the shared layer removes work the business line would otherwise do. The central team should be measured on onboarding time and on how few policy changes require its direct involvement. Two caveats apply: a single enforcement point concentrates operational dependency, so resilience design matters, and tiering should be tested with business lines before it is fixed.
- Onboard in days: Provide a standard onboarding path with default policy profiles per risk tier so a team starts with an approved baseline.
- Delegate configuration: Let business-line owners change policies inside their tier without a central ticket, with approval required only above it.
- Cover all AI paths: Include agents, tool calls and coding assistants at launch so teams do not build side paths for uncovered use.
- Return evidence to owners: Give business lines their own enforcement evidence, so the shared layer also serves their audit and risk reporting.
- Design for resilience: Define behavior on failure, detection of faults and recovery, consistent with the Resilience emphasis in examiner expectations.
- Retire local gateways: Set dates for decommissioning existing proxies and tie new AI funding to use of the shared layer.
Where Trussed fits
Trussed is not a gateway. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, and it generates audit-ready evidence from that enforcement. Enforcement is at runtime, not reviewed after the fact. Deployment is self-managed in the customer cloud, with configurable retention and data residency.
In production, Trussed has run 100+ models under one policy plane and has operated continuously for 18+ months in a regulated enterprise. Insurance is where Trussed is deepest. Trussed does not claim bank or other financial-services deployments, so a bank should validate fit against its own supervisory requirements and the criteria above.
One policy plane, many business lines
Business lines
Retail, commercial, wealth, markets and operations each set policy variations within approved risk tiers.
Shared enforcement layer
Allow, redact, block, rate limit or escalate for human approval on each interaction and tool call.
Models, agents, tools, coding assistants
All AI traffic passes through the same enforcement point.
Evidence
Each decision, policy version and action is recorded for model risk management and examiners.
Assess your shared enforcement options
Review how in-path enforcement, per-business-line policy and audit-ready evidence would work in your own cloud.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.