How to Standardize an AI Operating Model After a Bank Merger Under the 2026 MRM Guidance
Treat post-merger AI standardization as an operating-model and evidence problem. Consolidate one inventory, tier every inherited use case by materiality, assign single owners for policy and approval, then apply one policy set at runtime. Harmonized documents show intent. Only enforcement in the execution path, with the evidence it generates, shows what each system was allowed to do and did.
Criteria for evaluating runtime policy enforcement
- Applies one policy set across the heterogeneous models, agents, tools and coding assistants inherited from both institutions, without separate rule sets per estate.
- Acts in the execution path with explicit decisions (allow, redact, block, escalate for human approval) rather than flagging activity for later review.
- Records per-interaction evidence: policy version applied, decision taken and approver where escalated.
- Offers configurable retention, data residency and a deployment model that keeps evidence inside the bank's own environment, with a defined export path for examiners.
- Lets enforcement scope and strictness vary by use-case materiality, so the policy plane mirrors the combined risk tiers.
- Can respond to third-party model version changes and shared upstream providers, for example by triggering policy review or escalation.
What changed, and what it means for a combined bank
As of November 2026, the baseline for model risk management in U.S. banks is the revised interagency guidance: OCC Bulletin 2026-13, dated April 17, 2026, and Federal Reserve SR 26-2, with the FDIC participating. It rescinds prior model risk management guidance. It is risk-based, covers model development and use, validation and monitoring, governance and controls, and considerations for vendor and other third-party products, and is expected to be most relevant to banking organizations with more than $30 billion in total assets.
Two points matter for a merged institution. First, the principles apply to traditional statistical and quantitative models and non-generative, non-agentic AI. Generative and agentic AI are described as novel and rapidly evolving, so much of an inherited estate sits outside that scope and needs governance under the bank's wider risk framework. Second, the agencies state that practices appropriate for one banking organization may be inappropriate for another, and each organization is responsible for adopting practices that are effective for its own risks. No single design is prescribed. The combined bank has to choose its post-merger practices and be able to justify them.
On November 4, 2026, the OCC also issued Bulletin 2026-54, an examiner framework for bank use of what it calls super intelligence, including agentic systems. It states that it sets no supervisory expectations for banks, so it should not be cited as a requirement. It is still a useful signal of how examiners may look at the work. It tells them to tailor exams to the materiality of the use case, says risks belong under the existing risk management framework with no separate risk appetite, and describes Resilience (detect, contain, recover from and remediate failures) and Assurance (evidence that controls keep operating effectively).
A phased approach to standardization
The sequence below is a practical approach, not a regulatory prescription. It orders the work so that each phase gives the next one something stable to build on.
Start with the inventory because every later decision depends on it. The two institutions will rarely define a model, a use case or an AI tool the same way, so agree on one schema before merging records. Include third-party models with version identifiers, since a vendor changing a model can change the risk of a service even when the contract is unchanged. Both estates may also reach the same upstream provider through different routes, which the combined register should make visible.
- Phase 1: Consolidate the inventory: Record models, agents, tools, vendors and coding assistants in one schema. Capture owner, purpose, materiality, approval status, third-party dependencies, model version and the data involved, including prompts, outputs, embeddings and logs.
- Phase 2: Scope and tier: Separate use cases within the revised MRM guidance from generative and agentic use cases governed under general practices. Apply one materiality scale and document how each legacy tier maps to it.
- Phase 3: Harmonize policy and approval ownership: Name one accountable owner per policy and per approval decision. Responsibility for third-party AI stays with the board and management, so ownership should reach senior officers.
- Phase 4: Apply one policy set at runtime: Move from reviewing documents to enforcing the combined policy where interactions and tool calls happen, tiered by materiality.
- Phase 5: Standardize evidence: Define the same records for both estates: what was approved, which policy applied, what decision was taken, and what changed since.
What policy harmonization can and cannot settle
Inventory consolidation, risk tiering, approval ownership and validation procedures can all be harmonized through process. These are the parts of the operating model that a risk team can align with documents, committees and workflow changes, and they should be finished first.
They do not show how a live system behaves. A merged policy may say that sensitive data must not reach an external model, or that certain tool calls need human sign-off. If one legacy estate relies on developer discipline and periodic review while the other uses technical controls, the bank has one written policy and two behaviors. Review after the fact samples activity. It does not stop a prohibited action when it happens.
This is the distinction to put in front of examiners. Anything that depends on how a system acts in the moment is only reliable if policy is enforced in the execution path. Trussed is built for that layer. It enforces policy across models, agents, tools, MCP servers and coding assistants under one policy plane, with decisions to allow, redact, block, rate limit or escalate for human approval, applied to individual interactions and tool calls. It supports the inventory, validation and ownership work described above and does not replace it. Trussed is deepest in insurance, a regulated financial sector, and has no banking deployments to cite here.
Producing examiner-ready evidence across both estates
Examiners will ask what each system was permitted to do and what it did. If the two legacy institutions keep evidence in different formats, retention periods and systems, the combined bank answers with two partial stories.
The OCC's Assurance concept is a useful design target: evidence that controls continue to operate effectively, not only that they existed at approval. In practice that means documentation captured before deployment and after material changes, using the same fields for both estates. For each interaction in scope, it also means a record of the policy version applied, the decision taken and, where a human approved an action, who approved it.
Retention and export need a decision early. Set one retention period by tier, decide where evidence resides and who can read it, and test that a complete record for a single use case can be produced quickly for either legacy estate. Trussed generates audit-ready evidence as a by-product of enforcement, with configurable retention and data residency and self-managed deployment in the customer's cloud. It currently produces more than 30 GB of audit evidence per day in production use. Whatever the tooling, the test is the same: one question, one answer format, regardless of which institution built the system.
Five phases from two AI estates to one operating model
1. Inventory
One register of models, agents, vendors and coding assistants from both institutions.
2. Tiering
A single materiality scale and a documented mapping from legacy tiers.
3. Ownership
One set of policies, approvers and re-approval triggers.
4. Runtime
One policy set enforced in the execution path.
5. Evidence
Consistent records across both legacy estates.
Plan the runtime layer of your integration
Talk through how one policy set and one evidence format could apply across two legacy AI estates.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.