Central AI Platform Team vs Embedded AI Leads: Choosing a Bank AI Operating Model
Neither design wins on its own. A central AI platform team gives the most consistent policy and the simplest evidence trail, but it sits further from business context and can become a queue. Embedded AI leads move faster and know the business, but they tend to produce uneven controls, duplicated tooling and fragmented evidence. Many banks end up with a hub-and-spoke hybrid: policy and evidence standards set centrally, delivery owned by the lines of business. Whichever design you choose, runtime enforcement and evidence format should not vary by line.
Requirements to set regardless of the design you choose
- One inventory across all lines of business with owners, risk ratings and escalation paths.
- Policy definition separated from line-of-business ownership, with a named approver for policy changes and exceptions.
- Runtime enforcement that covers agents and tool calls, not only model outputs.
- Evidence available per use case showing what the AI was permitted to do and what it did, produced continuously rather than at approval time.
- An independent validation and challenge owner in the second line, with a stated scope for agents.
- Deployment model, data residency and retention set up front, with responsibilities mapped to NIST AI RMF govern, map, measure and manage.
Org design is a control and evidence question
Debates about central versus embedded AI teams usually start with staffing and reporting lines. For a risk leader, a better starting point is where three things happen: policy definition, runtime enforcement and evidence production. An org design that is efficient but leaves any of the three unowned will be hard to defend in an examination.
Supervisory materials point in this direction. The Federal Reserve's model risk management guidance states that effective policies define risk management expectations, and it frames supervisory review around whether those policies are actually implemented. The OCC's supervisory framework for bank use of what it terms Super Intelligence (SI), set out in Bulletin 2026-54 dated November 4, 2026, is written for examiners. It covers normal-course activities such as credit decisioning, payments, fraud detection, customer service, operations, risk management and market activities, and it explicitly includes agentic SI. Its Assurance element is defined as evidence that SI system or agent controls continue to operate effectively.
Two implications follow. A written standard is not enough, because implementation has to be demonstrable. And evidence has to be ongoing and available per use case, not assembled once at approval time. Both are properties of the operating model, not just of the tooling.
Central, embedded and hybrid models against six criteria
The table summarizes typical tendencies. Individual banks vary, and each weakness can be mitigated with deliberate design.
- Policy consistency: Central: one team sets and applies one standard. Embedded: each lead interprets policy for their line, so drift is likely. Hybrid: central standard with a defined path for local exceptions.
- Speed of delivery: Central: intake queues and distance from business context can slow delivery. Embedded: fastest locally. Hybrid: lines deliver quickly inside pre-approved standards, but only if the hub does not become a gate for everything.
- Model risk management alignment: Central: a single inventory and intake simplifies validation and challenge. Embedded: inventories and documentation vary, which burdens validators. Hybrid: shared inventory and risk tiering, with lines supplying the use-case detail.
- Three-lines accountability: Central: clear ownership, but the team may both build and set the rules. Embedded: strong business accountability, but leads may write policy for their own work. Hybrid: hub defines, lines operate, second line tests and monitors.
- Evidence production: Central: one repository and one format. Embedded: evidence scattered across lines and tools. Hybrid: workable only if all lines write to a common evidence schema.
- Staffing: Central: scarce expertise concentrated, with bottleneck risk. Embedded: skills duplicated across lines and uneven in depth. Hybrid: small expert core plus capable leads in each line, at the cost of coordination.
Mapping each design to the three lines of defense
In both designs, the AI team is first line: it builds, deploys and operates AI for the business. The question is who sits in the second line and how independent that challenge is. The Institute of Internal Auditors, in comments to NIST, urged positioning AI risk within existing governance and enterprise risk management programs rather than creating siloed AI-only governance and risk functions. It also described controls that review, test and monitor policies and procedures as second-line management.
A central platform team makes the second-line relationship simple, since there is one counterpart to challenge. The risk is role overlap: if the platform team writes the policy and also builds to it, independent challenge has to be explicit. With embedded leads, accountability to line-of-business heads is natural, but second-line oversight must reach many teams with different practices, and reporting lines can compromise independence.
Validation ownership deserves its own decision. SR 11-7 includes a dedicated section on the components of an effective validation framework. Decide who performs independent validation of AI use cases, where it reports, and how it applies to agents and tool use, not only to models. NIST AI RMF adds a useful check: it treats governance as a cross-cutting function across map, measure and manage, so policy definition cannot live only in a delivery step.
Where a hub-and-spoke model fits
A hub-and-spoke model separates what should be uniform from what should be local. One NIST submission, a third-party view rather than a regulator position, suggests that depending on adoption scale a formal approval process may be governed by a central body with cross-disciplinary expertise, that central monitoring and escalation is in many cases essential, and that a Center of Excellence can serve as a knowledge-sharing platform. The same submission describes an AI system inventory recording business owners, developers and risk ratings.
In practice, the hub owns policy standards, the inventory, approval of higher-risk use cases, monitoring and escalation, and evidence requirements. The spokes own use-case selection, business context, day-to-day operation and input to risk ratings. The model fails when the hub is nominal, with standards no one enforces, or when it approves everything and recreates the central bottleneck.
If you stand up a central approval body, define its escalation path and the criteria for evolving or retiring it as adoption matures. The same submission notes that such a structure may later be disbanded or replaced.
One enforcement layer that supports either design
Org design determines who writes the rules. It does not, by itself, ensure the rules are applied when a model answers, an agent acts or a tool is called. A single policy enforcement layer in the execution path addresses that gap and lets policy definition stay separate from line-of-business ownership. The hub or second line defines policy, lines of business operate within it, and every interaction passes through the same point of enforcement regardless of which team built the use case.
Trussed provides this kind of layer. It enforces policy in the execution path across models, agents, tools and MCP servers, with outcomes of allow, redact, block, rate limit or escalate for human approval. Scope is individual interactions and tool calls. Enforcement happens at runtime, not review after the fact, and it generates continuous audit-ready evidence mapped to NIST and other frameworks. It is self-managed in the customer cloud, with configurable retention and data residency. It runs 100+ models under one policy plane. Insurance, a regulated financial sector, is where Trussed is deepest.
A runtime layer does not choose your org design, replace independent validation or substitute for an inventory. It makes the chosen design enforceable and evidenced.
Three questions for any AI org design
Define
Who writes AI policy, and who can change it for a given line of business?
Enforce
Where is policy applied to models, agents and tool calls while they run?
Evidence
Who can show an examiner what the AI was permitted to do and what it did?
Test your org design against examiner questions
Walk through policy definition, runtime enforcement and evidence production for your own lines of business with a Trussed expert.
Talk to an Expert