AI Budgeting for Banks: Planning Model Validation Costs Under the 2026 Model Risk Guidance

What changed: the planning baseline moved in 2026

Many bank planning documents still cite SR 11-7 and OCC 2011-12. On 2026-04-17 the Federal Reserve, the OCC and the FDIC issued Revised Guidance on Model Risk Management. SR 26-2 states that it supersedes and replaces SR 11-7 and SR 21-8, and the OCC bulletin states that it rescinds prior model risk management guidance. As of October 2026, planning documents should cite the revised guidance and treat the older annual-review language as legacy.

The agencies describe the revision as risk-based and tailored to a banking organization's model risk profile and the size and complexity of its operations. It covers model development and use, validation and monitoring (including conceptual soundness and outcomes analysis), governance and controls, and vendor and other third-party products. The OCC also notes that practices appropriate for one organization may be inappropriate for another.

For budgeting, the consequence is direct. The nature and rigor of validation align with a model's approach, use and materiality, and timing and frequency vary with purpose, methodology, the frequency and scope of changes, and data limitations. Validation cost is therefore a function of your inventory and tiering decisions, not a fixed percentage of AI spend.

Which cost lines belong where

Funding usually fragments because delivery teams, model risk management and control functions each plan separately. Naming the owner of each line in a single plan prevents the validation and evidence work from being left unfunded. The Federal Reserve Board's own AI compliance plan evaluates technology, governance, operations, incremental cloud compute and skill development together, which is a useful starting list to extend.

• AI delivery: Model development, licensing, infrastructure, incremental cloud compute and skills. Also pilot-to-production work such as code review and security risk assessment.
• Model risk management: Inventory maintenance and tiering, initial and change-driven validation, ongoing monitoring, vendor model due diligence, independent review and documentation.
• Runtime control and evidence: Policy enforcement while systems operate, and the production of records showing what each system was permitted to do and what it did. Often unowned today.

Start with the inventory and let the risk tier set capacity

A detailed AI inventory with risk ratings is the structure that links use cases to monitoring effort. FINRA describes this for the securities industry, and NIST AI RMF describes testing, evaluation, verification and validation as tasks performed throughout the AI lifecycle. The same logic applies to a bank plan: every use case, including generative and agentic systems, needs an entry, a tier and an owner before any capacity is estimated.

The supported capacity drivers are model purpose, methodology, frequency and scope of model changes, data limitations, and vendor or third-party status. In practice that means counting use cases by tier, separating vendor models from internally built ones, and estimating how often each system changes. Convert those counts into validation effort and monitoring hours using your own historical data. Generative and agentic systems often do not fit traditional tiering, so decide explicitly how your framework rates them and record the rationale.

Two further items are easy to omit. Where a current model is being replaced, budget for parallel running before retirement. Budget also for explainability documentation, such as written summaries of key input factors and output rationale. Drift is a further pressure: an FDIC request for information noted that AI that learns from new data can challenge validating, monitoring, tracking and documenting the approach, including for independent reviewers. Reserve capacity for that independent review rather than assuming it scales with development.

Where runtime enforcement and continuous evidence change the workload

Examiners and internal audit want to see what a production AI system was permitted to do and what it actually did. If validation frequency depends on how often a system changes, and drift complicates documentation, then reliable records of configuration and behavior are a plausible input to both validation and examination workload. Whether such records reduce, shift or simply formalize that work depends on your tiers and how the records are produced. Treat it as an assumption to test in the plan, not a saving to book.

There are two ways to produce this evidence. One reconstructs it after the fact from reviews and sampled records, which consumes validation and audit hours each cycle. The other enforces policy at runtime and generates evidence as a by-product of that enforcement. The second moves effort from periodic reconstruction to up-front policy definition and an ongoing control layer, and it needs its own budget line.

Trussed addresses this evidence and control layer. It enforces policy in the execution path across models, agents, tools and MCP servers, with actions such as allow, redact, block, rate limit and escalate for human approval, scoped to individual interactions and tool calls. Enforcement generates continuous audit-ready evidence mapped to frameworks including NIST. Insurance is the regulated financial sector where Trussed is deepest, and it can be self-managed in the customer cloud with configurable retention and data residency. It does not replace model validation or inventory management, which remain with model risk management.