AI Vendor Contracts: Key Clauses
AI vendor contracts need terms that standard SaaS templates miss: limits on training with customer data, output rights, IP and privacy indemnities, realistic accuracy commitments, notice of model changes, audit rights, liability caps fit for AI risk and a workable exit. Canadian privacy law keeps the customer accountable for personal information it shares.

Key takeaways
- Define customer data broadly and control its use for model training.
- Output assignments are only as good as the vendor's underlying rights.
- IP indemnities often carry conditions that can limit coverage.
- Accuracy disclaimers should be balanced by defined performance commitments.
- Model changes need notice, testing time and exit rights.
- PIPEDA keeps customers accountable for personal information sent to vendors.
Why AI vendor contracts are different
Buying an AI product is not quite like buying conventional software. The vendor's model may change without notice, outputs are probabilistic rather than deterministic, the vendor may want to use customer data to improve its models, and the legal status of training data and outputs is still developing. Standard software-as-a-service templates often fail to address these points, and vendor paper tends to allocate the new risks to the customer. Negotiating AI vendor contracts therefore requires a specific checklist, adapted to the risk of the use case and the bargaining position of the parties.

This insight reviews the clauses that matter most when licensing AI models, platforms and AI-enabled services in Canada, from the customer's perspective, with notes on what vendors typically resist. It is general information, not legal advice.
Use of customer data for training
The most important commercial question in many AI vendor contracts is whether the vendor may use customer inputs, outputs or other data to train or improve its models. Many enterprise offerings now commit not to train on customer data by default, while consumer and lower-tier offerings often reserve that right. The contract should define customer data broadly, including prompts, uploaded files, outputs, embeddings, feedback and metadata, and state whether and how each may be used. If any use for improvement is permitted, it should be limited to aggregated or de-identified data, with clear restrictions on sharing with other customers or third parties.
This is not only a commercial issue. Under the Personal Information Protection and Electronic Documents Act, an organization remains responsible for personal information it transfers to a third party for processing and must use contractual or other means to provide a comparable level of protection. If a vendor uses personal information for its own purposes, the arrangement may become a disclosure that requires consent. The privacy regulators' principles for generative AI emphasize legal authority, appropriate purposes and safeguards for personal information used in AI systems.
Ownership of outputs
Customers generally want to own, or at least have unrestricted rights to use, the outputs generated for them. Many vendors now assign whatever rights they have in outputs to the customer, but that assignment is only as good as the underlying rights, and whether AI-generated material is protected by copyright in Canada is unsettled. The contract should confirm that the customer may use outputs for any lawful purpose, that the vendor will not claim rights in them, and how outputs that are similar across customers are treated. For custom models, fine-tuned weights, prompts and evaluation data created for the customer, the contract should specify ownership and what happens to them on termination.
IP indemnities
Intellectual property claims are a central risk for generative AI: claims that training data infringed copyright, or that outputs reproduce protected works or trademarks. Several major providers now offer indemnities for third-party IP claims arising from outputs, typically subject to conditions such as using the service as documented, not disabling safety filters, not intentionally prompting for infringing content, and not modifying outputs in ways that cause the infringement. Customers should read these conditions closely, confirm that the indemnity covers both training data and outputs, check that it is not swallowed by the liability cap, and confirm the vendor's control-of-defence and mitigation rights.
Privacy, security and data protection terms
AI vendor contracts should include a data processing addendum that reflects Canadian law, covering purpose limitation, security safeguards, data location, sub-processors, breach notification and cooperation, assistance with access and correction requests, and return or deletion at termination. Where Quebec residents' information is involved, transfers outside Quebec require a privacy impact assessment and a written agreement. Customers should seek an indemnity, or at least a separate higher liability cap, for breaches of data protection and security obligations. See our insight on service provider agreements.
Key clauses in AI vendor contracts at a glance
| Clause | Customer objective | Typical vendor position |
|---|---|---|
| Training on customer data | No training without opt-in | Enterprise tiers often agree; others resist |
| Output rights | Customer owns or may freely use outputs | Assignment of vendor's rights, no warranty of protectability |
| IP indemnity | Covers training data and outputs | Offered with conditions and exclusions |
| Accuracy | Performance commitments for defined uses | Broad disclaimers of accuracy |
| Model changes | Advance notice and right to test or exit | Right to update models at will |
| Audit and transparency | Documentation, testing results and audit rights | Third-party reports only |
| Liability cap | Super-cap for data, IP and confidentiality | Fees paid in prior 12 months |
| Exit | Data return, deletion and transition support | Limited assistance |
Accuracy disclaimers and performance commitments
Vendors almost always disclaim the accuracy, completeness and fitness of AI outputs, often stating that outputs may be incorrect and must be reviewed by the customer. Some disclaimer is reasonable given how the technology works. But customers buying AI for defined business purposes should seek meaningful commitments: performance against agreed benchmarks or acceptance tests, service levels for availability and latency, warranties that the service will perform materially in accordance with its documentation, and remedies if performance degrades. For high-impact uses, the contract should support the customer's own testing for accuracy and bias, which may be needed to meet human rights and regulatory expectations.
Define the use case in the contract. An AI vendor that knows its system will be used to triage customer complaints or summarize medical records is in a weaker position to rely on a blanket accuracy disclaimer than one selling a general-purpose tool with no stated purpose.
Model change notices
AI services change frequently: new model versions, changes to safety filters, deprecation of older models and changes in sub-processors or hosting locations. Any of these can affect accuracy, behaviour, cost and compliance. AI vendor contracts should require advance notice of material changes, a period during which the customer can test the new version, the ability to remain on a prior version for a reasonable time where feasible, and a right to terminate without penalty if a change materially degrades the service or creates compliance issues. Deprecation timelines should be stated.
Audit, transparency and documentation
Customers need information to govern their AI use: documentation of the model's intended uses and limitations, testing and evaluation results, data sources at an appropriate level of generality, security certifications and incident history. The federal voluntary code of conduct for advanced generative AI contemplates developers publishing information on capabilities and limitations and providing guidance to downstream managers. Contracts can convert those expectations into obligations. Audit rights for AI are often limited to third-party reports and questionnaires; for high-risk deployments, customers should seek rights to receive testing results, to conduct or commission assessments, and to obtain information needed to respond to regulators or complainants.
Liability caps and exclusions
Standard caps, often tied to fees paid over the prior twelve months, may be inadequate for AI failures that cause data breaches, IP claims, regulatory penalties or harm to customers. Common negotiated positions include uncapped or super-capped liability for breach of confidentiality, data protection obligations and IP indemnities, and carve-outs from exclusions of indirect damages for data loss and regulatory fines where legally permitted. The right answer depends on the deal size and risk, but the allocation should be deliberate.
Regulatory and incident cooperation
If an AI system causes harm or attracts regulatory scrutiny, the customer will need the vendor's help to investigate, explain outcomes and remediate. The contract should require prompt notice of incidents affecting the customer, cooperation with investigations and regulator inquiries, and preservation of relevant logs. Customers in regulated sectors should check that the contract satisfies their regulator's outsourcing and third-party risk expectations.
Upstream model providers and sub-processors
Many AI products are built on foundation models supplied by a small number of upstream providers, and the vendor in front of the customer may simply pass prompts to one of them. The customer's protections then depend on the terms between the vendor and its upstream provider, which the customer rarely sees. Contracts should require the vendor to identify the models and sub-processors used for the service, to flow down obligations on data use, confidentiality, security and location, to remain responsible for their performance, and to give notice before switching to a different model provider. Customers should also ask whether upstream providers retain prompts for abuse monitoring, for how long and where, because that retention may matter for privacy and confidentiality analysis.
Confidentiality of prompts and outputs
Prompts often contain the customer's most sensitive information: strategy documents, source code, client files and personal information. The confidentiality clause should expressly cover prompts, uploaded content, outputs and any derived data such as embeddings, and should not be limited to information marked confidential. Human review of prompts by vendor staff, for example for safety or quality purposes, should be disclosed, limited and subject to confidentiality obligations. Where the vendor offers zero-retention or private deployment options, the contract should state which option applies.
Agentic and integrated AI features
AI features that take actions, such as sending messages, updating records, executing code or making purchases, raise additional contracting issues. The customer needs clarity on what permissions the AI component has, how actions are logged, how errors are reversed and who bears the loss when the system acts outside its instructions. Contracts should require configurable limits on actions, human approval for defined categories of transactions and detailed audit logs. See our insight on agentic commerce for related questions about AI agents that buy and pay.
Exit and portability
Switching AI vendors can be harder than it looks, especially where prompts, workflows, fine-tuned models and embeddings are tied to one platform. Exit provisions should cover return of customer data in usable formats, deletion with certification (including from backups and sub-processors, subject to defined exceptions), transition assistance, continued access for a wind-down period and treatment of customer-specific models. Planning for exit at signing preserves leverage.
Negotiation strategy for AI vendor contracts
Customers have more leverage on AI vendor contracts than they often assume, particularly in enterprise deals and with vendors seeking reference customers. Prioritize the clauses that match the risk of the use case: data use and confidentiality for almost every deal, IP indemnities for content generation, accuracy and audit for decision-support systems, and model change and exit for business-critical deployments. Where the vendor will not move on its terms, consider compensating controls, such as limiting the data shared, adding human review or restricting the use case. See our insight on AI governance programs.
How TRW & Co helps
We negotiate AI vendor contracts and model licences for customers and AI providers, including data use and training terms, output rights, IP and privacy indemnities, performance commitments, model change provisions, audit rights and exit. See our technology transactions and outsourcing and artificial intelligence pages or contact us.
Frequently asked questions
Can an AI vendor train on our data?
Only if the contract allows it. Many enterprise offerings commit not to train on customer data by default. Define customer data broadly, including prompts, files, outputs and feedback, and restrict any permitted use.
Do we own AI outputs?
Contracts commonly assign the vendor's rights in outputs to the customer, but whether AI-generated material is protected by copyright in Canada is unsettled. Secure the right to use outputs for any lawful purpose.
Are vendor IP indemnities reliable?
They can be valuable but usually include conditions, such as using the service as documented and keeping safety filters enabled. Check coverage of training data and outputs and how the indemnity interacts with the liability cap.
What if the vendor changes the model?
Seek advance notice of material changes, time to test, the option to stay on prior versions where feasible, and a right to terminate if the change degrades the service.
What privacy terms are needed?
A data processing addendum covering purpose limitation, safeguards, data location, sub-processors, breach notice, individual rights assistance and deletion, consistent with PIPEDA's accountability principle and Quebec transfer rules.
Sources
General information only, not legal advice. References to statutes, courts and authorities describe published law and practice at the date shown.


