Standardizing AI across business lines in a bank: a runtime policy and evidence architecture
To standardize AI across business lines in a bank, centralize the policy taxonomy, risk tiering, data-handling rules, approval and escalation paths, and evidence requirements. Leave use-case selection and approved local exceptions to the business lines. Then enforce the central policy at runtime, in the execution path of every model, agent, tool and assistant, so that each decision produces evidence an examiner can review.
Runtime enforcement versus after-the-fact review
- 1
Central policy store
Taxonomy, tiers and rules, versioned and owned centrally.
- 2
Enforcement points
Standard hooks in the execution path of models, agents, tools and coding assistants.
- 3
Exception handling
Business-line overrides expressed in policy, with owner, approval and expiry.
- 4
Human-approval escalation
A routed path for actions the policy does not permit automatically.
- 5
Evidence store
Durable records of each decision, with retention and residency set by policy.
Evaluation criteria for a runtime enforcement and evidence layer
- One policy set applies across multiple models, agents, tools and coding assistants, and business-line exceptions are expressed and approved within it.
- Supported actions in the execution path include allow, redact, block, rate limit and human-approval escalation.
- Evidence records what was permitted and what happened for each decision, with defined fields and an export format examiners can use.
- Retention is configurable by tier and data residency can be enforced to meet the bank's requirements.
- The layer can be self-managed in the bank's own cloud.
- The vendor states its scope and limits plainly, including whether it covers model lineage, transaction-level governance or end-to-end tracing of agentic workflows.
What standardization means in a multi-business-line bank
Retail, commercial, wealth, capital markets and operations usually adopt AI on their own schedules. Each ends up with its own models, vendors, approval processes, data-handling rules and evidence practices. The result is that the bank cannot easily answer a basic examiner question: what was this AI system permitted to do, and what did it do?
Standardization does not mean one model, one vendor or one tool. It means one consistent answer to the control questions. Which uses are in which risk tier? What data may reach which model? Who approves what, and when does a human have to be involved? What record must exist afterward? Treating this as a tooling consolidation exercise tends to fail, because business lines keep adopting new tools faster than any consolidation can absorb them. Treating it as a policy and evidence problem scales better.
Industry commentary on the interagency model risk management guidance describes it as already applying to models including AI and machine learning, with development, implementation and use controls, validation and governance. The same commentary argues that oversight should be proportionate to the risk of the business application. That is a sound basis for the operating model: extend the existing model risk management framework instead of building a parallel one. Runtime enforcement supplements validation and governance. It does not replace them.
What to standardize centrally and what to leave to business lines
The split below is a recommendation, not a regulatory requirement. The test is whether inconsistency would prevent the bank from showing one coherent control environment. If it would, the element belongs at the center. If the element depends on local business knowledge, it stays local, but within the central policy.
- Policy taxonomy: Central: shared categories of permitted, restricted and prohibited behavior. Business line: maps its use cases to those categories.
- Risk tiering: Central: tier definitions and criteria, aligned to the model risk management framework. Business line: proposes the tier for each use case.
- Data-handling rules: Central: what data classes may reach which models and tools, and where. Business line: identifies the data its use case actually touches.
- Approval and escalation paths: Central: who approves each tier and when human approval is required. Business line: names the accountable owners.
- Evidence requirements: Central: minimum record content and retention by tier. Business line: adds local fields where its use case needs them.
- Use-case selection and exceptions: Central: the exception process and its expiry rules. Business line: chooses use cases and requests exceptions, which are explicit, approved, time-bound and recorded.
A phased rollout sequence
The sequence below is an editorial proposal. It moves from visibility to policy to enforcement, so that enforcement is switched on first where risk is highest and the policy has already been tested. Plan for integration effort with existing bank systems and suppliers; compatibility across multiparty software is a recognized difficulty.
- Phase 1: Inventory: Catalog AI use by business line: models, vendors, approval paths, data-handling rules and evidence practices. This is the fragmentation baseline.
- Phase 2: Policy and governance: Define the taxonomy, risk tiers, data rules, escalation paths and evidence requirements, mapped to the existing model risk management framework. Stand up a cross-business-line review body.
- Phase 3: Observe: Run policy in observation mode to see what would have been blocked or escalated, and tune rules and exceptions with business-line owners.
- Phase 4: Enforce on higher tiers: Turn on enforcement for the highest-risk tiers first, with human-approval escalation, then extend to lower tiers.
- Phase 5: Evidence review: Test whether the evidence answers a realistic examiner request, and adjust retention and fields.
Regulatory context and where Trussed fits
The EU AI Act requires a risk management system for high-risk AI systems and effective human oversight by design. Its transparency concept includes traceability and explainability, which implies the system must record enough to support oversight. The European Commission proposed delaying the high-risk application dates in November 2025, to 2 December 2027 for stand-alone systems and 2 August 2028 for systems embedded in products. As of November 2026, confirm the final text before fixing EU timelines.
Trussed is one way to apply a single policy plane across many models and tools. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants: allow, redact, block, rate limit, or escalate for human approval. Its scope is individual interactions and tool calls. Enforcement produces continuous, audit-ready evidence mapped to frameworks including the EU AI Act and NIST. Deployment can be self-managed in the customer cloud, with configurable retention and data residency. Trussed has run 100+ models under one policy plane, and has 18+ months of continuous production in a regulated enterprise.
It does not provide model lineage, dataset provenance, transaction-level governance or end-to-end tracing of agentic workflows. Its customers are in insurance, pharma, healthcare and education, and insurance is where it is deepest. It makes no claim of bank deployments. Banks should evaluate it against the criteria above, as they would any layer.
One policy set, many business lines
Central policy
Taxonomy, risk tiers, data-handling rules, escalation paths and evidence requirements.
Runtime enforcement
Allow, redact, block, rate limit or escalate for human approval in the execution path.
Business-line exceptions
Explicit, approved, time-bound and recorded.
Audit-ready evidence
A record of what each AI system was permitted to do and what it did.
Plan a consistent AI operating model
See how policy enforcement in the execution path produces audit-ready evidence that risk teams can review.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.