Standard SaaS vendor contracts were not written for AI. The data processing addendum you've been using since 2018 doesn't cover the scenarios that matter most when the vendor's product is an AI system making consequential decisions about your customers, employees, or operations.
Here are the clauses that most AI vendor contracts still don't include — and what to ask for in negotiations.
The Gap That Standard DPAs Don't Fill
A data processing addendum (DPA) under GDPR covers: how the processor handles personal data on behalf of the controller, sub-processor relationships, data breach notification, data subject request assistance, and return or deletion of data.
It doesn't cover:
- What happens when the AI model is updated and its behavior changes
- Who is responsible when an AI output causes a documented harm
- What training data the AI system was built on
- Whether your data is being used to train the vendor's models
- What the vendor's human oversight and incident escalation procedures are
- How the AI system performs across demographic groups (bias and fairness)
- What security controls apply specifically to the AI components
These aren't hypothetical edge cases. They're operational risks that have materialized in real deployments — and the organizations with leverage in those situations are the ones that negotiated protections before signature, not after.
Clause 1: Model Update Notification and Stability
The problem: AI vendor contracts almost universally include broad rights to update the product. For traditional SaaS, this is fine — a UI change or a performance improvement doesn't change whether your business process works correctly. For AI systems, a model update can materially change outputs, introduce new failure modes, or break integrations that depended on consistent behavior.
What to negotiate:
Provider shall provide no less than [30/60] days' advance written notice prior to deploying any Material Model Update affecting the AI System. A "Material Model Update" includes any change to the underlying model that results in: (a) a material change in output characteristics; (b) changes to supported use cases or safety guardrails; or (c) changes to the data categories used for inference. During the notice period, Customer shall have the right to evaluate the updated model in a test environment prior to production deployment.
Minimum acceptable: Advance notice + access to changelogs that document what changed and what testing was performed.
Clause 2: Training Data Representation
The problem: AI systems trained on biased, incomplete, or harmful data produce biased, incomplete, or harmful outputs. Most AI vendor contracts say nothing about the training data underlying the model.
What to negotiate:
Provider represents and warrants that: (a) the training data used to develop the AI System was collected and processed in compliance with applicable data protection laws; (b) Provider has rights to use the training data for the purposes for which it was used; (c) reasonable steps were taken to identify and mitigate bias in the training data prior to model training; and (d) Provider will make available upon request a summary of the training data methodology, including data sources, curation processes, and bias mitigation measures undertaken.
Minimum acceptable: A model card or equivalent documentation you can review before deployment.
Clause 3: No-Training-on-Customer-Data Commitment
The problem: Many AI vendors — particularly those offering free or low-cost tiers — use customer inputs to improve their models. This creates data governance, IP, and confidentiality risks that are almost never disclosed clearly at the point of contract.
What to negotiate:
Provider shall not use Customer Data (including without limitation Customer inputs, outputs, feedback, and derived data) to train, fine-tune, or improve any AI model, algorithm, or system, including models made available to other customers, without Customer's express written consent. This restriction shall survive termination of this Agreement.
Why this matters for EU AI Act compliance: Deployers of high-risk AI systems must document data governance across the full supply chain, including what the vendor does with data. If the vendor is using your data to train models deployed to other customers, that's a data governance disclosure you likely need to make.
Clause 4: Accuracy and Performance Standards
The problem: AI vendor contracts routinely include uptime SLAs but say nothing about accuracy or performance of AI outputs. A system that's available 99.9% of the time but returns wrong answers 30% of the time meets the SLA and doesn't meet your business need.
What to negotiate:
Provider warrants that the AI System shall perform in accordance with the Performance Specifications set forth in Exhibit [X], which shall include: (a) baseline accuracy metrics by use case; (b) fairness metrics by applicable demographic categories; (c) testing methodology; and (d) performance monitoring procedures. Provider shall notify Customer within [5 business days] if performance falls below the specified thresholds, along with a remediation plan.
Minimum acceptable: A documented performance benchmark you can measure against — even if you can't negotiate a warranty, getting the baseline into the contract gives you evidence for disputes.
Clause 5: Explainability and Audit Rights
The problem: For AI systems making consequential decisions, you may need to explain those decisions to affected individuals, regulators, or internal audit. If the vendor can't or won't provide explanations for outputs, you can't fulfill your own obligations.
What to negotiate:
Provider shall make available to Customer, upon request, sufficient documentation and technical mechanisms to enable Customer to: (a) provide meaningful explanations of individual AI System outputs to affected individuals upon request; (b) conduct audits of AI System performance, including bias and fairness assessments; and (c) respond to regulatory inquiries concerning AI System operation. Provider shall cooperate with Customer's reasonable audit requests, subject to appropriate confidentiality protections.
EU AI Act relevance: For high-risk AI systems, deployers have specific obligations to enable human oversight and to provide affected individuals with information about automated decision-making. Vendor audit rights are how you fulfill those obligations when you don't own the underlying system.
Clause 6: Incident Notification for AI-Specific Events
The problem: Standard data breach notification clauses cover security incidents involving personal data. They don't cover AI-specific incidents: outputs that cause documented harm, discovery of systematic bias, safety filter failures, or adversarial attack success.
What to negotiate:
Provider shall notify Customer within [72 hours] of becoming aware of any AI System Incident, defined as: (a) AI outputs that have caused or are likely to cause material harm to individuals; (b) evidence of systematic bias or discriminatory outputs affecting protected categories; (c) safety control failures or adversarial attacks that compromised AI System integrity; or (d) third-party security researchers or regulators identifying material vulnerabilities in the AI System. Notification shall include a description of the incident, affected populations, and remediation steps.
Why this matters: EU AI Act Article 73 requires providers of high-risk AI systems to notify market surveillance authorities of serious incidents. Your vendor contract needs to ensure you receive notification fast enough to fulfill your own regulatory notification obligations.
Clause 7: Liability Allocation for AI Outputs
The problem: Standard limitation of liability clauses cap vendor liability at amounts that are often absurdly low relative to the potential harm from a wrong AI output — particularly in high-stakes contexts like credit, healthcare, employment, or legal decision support.
What to negotiate: This is the hardest clause to move on, but the negotiating position is:
Notwithstanding any limitation of liability clause, Provider's liability for: (a) AI outputs that cause harm due to Provider's failure to comply with representations and warranties; (b) data privacy breaches arising from training data misuse; or (c) failure to notify Customer of a known AI System defect, shall not be limited to less than [12 months of fees / a specified floor amount].
Minimum acceptable: A mutual understanding — documented in writing, even if not in the main agreement — of how liability would be apportioned if an AI output causes a documented harm. Getting this documented now is better than litigating the ambiguity later.
A Note on Leverage
Not every organization has the leverage to negotiate all of these clauses with every vendor. Large enterprise customers buying significant volumes have more leverage. SMBs buying standard plans typically don't.
The practical approach: prioritize Clauses 3 (no training on customer data) and 6 (AI incident notification) as non-negotiables — these protect against the risks with the highest potential consequence. Use the remaining clauses as a due diligence framework even if you can't get them in the contract — the vendor's responses tell you something about their governance maturity.
If a vendor refuses to represent that they don't train on customer data, that's information. If they can't produce any documentation of training data methodology, that's information. If they've never thought about AI-specific incident notification, that's information.
Contract negotiation is also due diligence.
Further reading:
Top comments (0)