Skip to main content
LLM Squared

Guidance and the AI contract

Read the guidance. Then inspect the AI terms.

SR 26-2 leaves out generative AI and AI agents. Review the feature you plan to use, the terms you signed and the records behind your decision.

Updated .

Sources checked October 4, 2026.

The April guidance has a defined scope.

The agencies issued model risk guidance on April 17, 2026. It leaves out generative AI and AI agents. The agencies point banks to their current risk and oversight practices for those tools. The guidance is nonbinding. It mainly applies to banks above $30 billion. For smaller banks, consider their model use, how complex it is and how much risk it poses. Interagency guidance, printed pages 2 and 3.

Can a vendor use your data, train on an input or keep an output? Can its tool act through a connection? Ask the owner. Read the contract and check which controls apply.

A resource framework is not a list of 230 mandates.

Treasury announced an AI Lexicon and Financial Services AI Risk Management Framework on February 19, 2026, developed through public and private sector work. Treasury resource announcement.

The industry-led framework sets out 230 control goals. Pick the ones that fit the way you use AI and the risks you face. Write down why each fits. The goals do not carry the force of a Treasury rule. Cyber Risk Institute framework.

A vendor feature belongs in the use inventory.

CSBS asks about AI from outside firms and AI built into vendor tools. It also asks what sensitive data passes through AI. Each state chooses how to use the guide. The review fits the use, and the guide adds no new legal or supervisory requirements. CSBS guide, printed pages 2 and 3.

At renewal, ask which AI features are included, enabled, optional or separately charged. Identify the underlying provider and the source supporting each answer. Preserve a vendor disclosure gap as an open item.

Prepare six contract questions.

  1. Entitlement: Which AI features does the agreed package include, and which may be enabled without another approval?
  2. Allowed inputs: Which data categories may the use receive, and through which integrations?
  3. Training and retention: What do the controlling terms say about input, output, logging, model training, deletion and subcontractors?
  4. Action authority: Can the feature only draft, or can it send, approve, transact or write? Which human review is required?
  5. Change and exit: What notice, export, termination and access-withdrawal records are available?
  6. Acceptance: Who reviews the evidence, what remains unresolved, and which conditions must be met before enabling or renewing?

Vendor AI Decision Review links findings to the records for one feature. You get a one-page recommendation.

Keep the calculation separate from the contract decision.

Vendor Compare uses the fees you enter and the same usage for each quote. It runs in your browser. It has no live prices and does not fill in fees, exchange rates or tax. Missing terms may prevent a full cost comparison. The total does not settle AI rights, data terms or renewal terms.

The institution owns the decision.

We read the terms and notices you send. You keep a file for the agreed scope. Your counsel and reviewers decide which duties apply and whether to accept the contract. The file gives no certification or approval to turn on a feature. Method boundaries and source handling.

Bring the renewal date and the AI feature.

Tell us what the tool does, who owns the work and when you need a decision. Leave out private bank and member data.

Book a call