What will enterprise customers ask before buying our AI product?

The first enterprise deal for an AI product often stalls in procurement rather than on price. Security, privacy and legal teams now send AI-specific questions alongside the usual security questionnaire, and many have an AI addendum they want attached to your contract. The questions are predictable, so you can prepare.

What paperwork should we expect?

• An AI questionnaire, asking what models you use, what data they see, whether you train on customer data, how you test output and what oversight exists.
• A security review, often asking for a SOC 2 report or similar evidence, penetration test summaries and your incident response process.
• A data processing agreement, the customer’s or yours, with a sub-processor list.
• An AI addendum, adding AI-specific commitments to your SaaS agreement or to the customer’s master agreement.

Answers in questionnaires can end up incorporated into the contract, or be treated as representations. Answer accurately and don’t describe controls you plan to build as if they exist. The Competition Act addresses misleading claims to the public, and a misstatement to a specific customer can support a contract or misrepresentation claim.

Will they ask us not to train on their data?

Almost always. The usual ask is that you won’t use customer data, inputs or output to train or improve models, yours or anyone else’s. Before agreeing, check two things:

1. Your own practice. If you use customer data to improve the product, say what you do, define any aggregated or de-identified data narrowly, and make the contract, privacy policy and product settings match.
2. Your providers’ terms. Your commitment depends on your model providers’ commitments. See what OpenAI, Anthropic and Google’s API terms mean for your product.

Be careful with promises to delete data on a fixed timeline if a provider retains inputs for abuse monitoring.

Which commitments do they usually ask for?

Ask	What it means	How to approach it
No training	No model training on customer data	Agree if true, and align with providers’ terms
Sub-processor notice	Advance notice and a right to object before new providers	Give notice for new model providers, with a realistic objection process
Data location	Processing in Canada or a named region	Only commit to what your hosting and model providers support in writing
Output IP indemnity	You defend them if output infringes third-party IP	Limit to what your provider indemnifies, with exclusions for customer inputs and modifications
Liability cap carve-outs	Uncapped liability for data breaches or confidentiality	Offer a separate, higher cap rather than unlimited liability
Model change notice	Notice before you change models or materially change behaviour	Commit to notice of material changes, not to freezing a model version
Human oversight	Human review for consequential decisions	Describe the controls your product offers and who is responsible for using them
Audit rights	Right to audit your AI and security controls	Offer third-party reports and questionnaires instead of on-site audits

For more on caps, see limitation of liability and indemnity clauses.

Why do they care about accuracy and oversight?

Because they carry the risk with their own customers and regulators. A business is generally responsible for what its AI tools tell people, as Moffatt v. Air Canada showed. If your product helps make decisions about individuals, Quebec’s private sector act requires the business to give notice of decisions based exclusively on automated processing (section 12.1). Customers will want your contract to say who gives that notice and what information your product can provide to explain a decision.

Your terms should say output can be wrong, that the customer is responsible for reviewing it and for how it uses it, and what your product does to support review.

Are some customers subject to extra rules?

Yes. Regulated buyers pass their own obligations down to vendors:

• Federally regulated financial institutions will apply OSFI’s Guideline E-23 on model risk management, effective May 1, 2027, and ask for documentation on model development, validation and monitoring.
• Federal departments apply the Treasury Board Directive on Automated Decision-Making and may ask for information for their Algorithmic Impact Assessment.
• Health customers may need you to meet health privacy law, such as Ontario’s PHIPA. See PHIPA for health apps.
• Customers with EU operations may ask about the EU AI Act. See our AI Law Tracker.

How do we prepare before the first big deal?

• Write a standard AI and security questionnaire response and keep it current.
• Publish or prepare a DPA and a sub-processor list that names your model providers.
• Have your own AI product terms and an AI addendum template, so you can offer your paper first.
• Decide your positions in advance on training, data location, output indemnity, liability caps and model changes.
• Collect your providers’ current terms, DPAs and security reports.
• Keep a record of how you test output quality and safety.

What to do next

We prepare SaaS agreements with DPAs and AI terms, and review customer paper through contract review, at fixed fees on our fees page. For the wider picture, see what legal work an AI startup needs and legal help for AI startups.

This guide is general information about Ontario and Canadian law, not legal advice for your situation. To talk through an enterprise deal, book a free 15-minute consult.