What AI terms should we require from vendors?
When it comes up
When any team buys an AI product, when an existing vendor adds AI features to a product you already use, or at renewal.
Your options
Accept the vendor's standard AI terms
Fits when: Low-risk uses with public or internal data, and vendors whose standard terms already exclude training and offer indemnities.
Watch out: Standard terms change; check the version and your product tier.
Standard AI addendum for every AI purchase
Fits when: Most companies: one addendum covering data use, training, retention, output IP, indemnity, security and model changes.
Watch out: Large vendors may refuse addenda; be ready to map their terms against yours instead.
Risk-tiered negotiation
Fits when: High-volume buyers: light checks for low-risk uses, full negotiation for sensitive data or decisions about people.
Watch out: Needs a reliable triage step so high-risk deals don't slip into the light lane.
How to decide
- Triage the use (data sensitivity, decisions about people, regulatory exposure).
- Compare the vendor paper to your clause library: preferred, fallback, walk-away or missing for each clause.
- Weight by risk: data use and training, confidentiality, output IP and indemnity usually weigh most.
- Escalate by rule: any high-risk clause at walk-away or missing goes to the General Counsel.
- Record outcomes to improve the next negotiation.
Trade-offs
Pushing for every clause slows deals and can lose access to leading tools; accepting weak terms moves data and IP risk onto you. Risk-tiering keeps legal effort where it matters.
Real examples
- Microsoft's Copilot Copyright Commitment (September 2023) promised to defend paying commercial customers against copyright claims over Copilot output and pay adverse judgments, if they used the built-in guardrails and content filters. Source: Microsoft On the Issues — Microsoft announces new Copilot Copyright Commitment for customers (7 September 2023) ↗
- Google Cloud announced in October 2023 a two-part generative AI indemnity covering its use of training data and the generated output of certain services. Source: Google Cloud — Protecting customers with generative AI indemnification (12 October 2023) ↗
- The EU's public procurement Community of Practice published updated Model Contractual Clauses for AI (MCC-AI) in March 2025, aligned with the EU AI Act, as non-binding templates for public buyers. Source: European Commission Public Buyers Community — Model Contractual Clauses for AI procurement (MCC-AI, updated March 2025) ↗ · Covington Inside Privacy — EU's Community of Practice publishes updated AI model contractual clauses (April 2025) ↗
What to check first
- Does the contract exclude training on your inputs and outputs?
- What are the retention, deletion and human-access terms?
- Who owns outputs, and is there an IP indemnity that covers your actual use?
- What notice do you get of model changes, sub-processor changes and security incidents?
Learn and use
Sources
- European Commission Public Buyers Community — Model Contractual Clauses for AI procurement (MCC-AI, updated March 2025) ↗
- Microsoft On the Issues — Microsoft announces new Copilot Copyright Commitment for customers (7 September 2023) ↗
- Google Cloud — Protecting customers with generative AI indemnification (12 October 2023) ↗
- OpenAI — Enterprise privacy (business data is not used for training by default) ↗
- In re OpenAI copyright litigation (S.D.N.Y. 1:23-cv-11195) — Order of Magistrate Judge Ona T. Wang, 13 May 2025 (Justia) ↗
Last reviewed 2026-10-01.