Many enterprise software agreements in force today were negotiated before anyone at the table was thinking about generative AI. They described a defined service, a defined data flow and a familiar set of risks. Then the vendor shipped an update. A drafting assistant appeared in the document editor. The help-desk platform began summarizing tickets. The CRM (“Customer Relationship Management”) started proposing, and, in some configurations, sending follow-up emails on its own.
Notably, none of these changes is likely to have required a new signature. More often, vendors rolled them out as “feature enhancements” under the existing subscription, governed by online “AI terms” incorporated by hyperlink that the customer’s lawyers last reviewed at signing, if at all.
Customers generally want these features. The difficulty is that a contract built for a passive software tool now governs a system that takes in the customer’s data, generates new content from it and, increasingly, acts on the customer’s behalf. That gap is why the AI Addendum has become one of the most heavily negotiated documents in technology transactions on both sides of the border.
This article considers why the legacy MSA (“Master Service Agreement” or “Master Subscription Agreement”) may no longer be sufficient, what a well-built AI Addendum should contain, where the hardest negotiations take place, why the IP indemnity a vendor advertises may be considerably narrower than it first appears, and how agentic features change the analysis.
I.The existing MSA may no longer be sufficient
The first question is often whether the vendor was entitled to make the change at all. SaaS (“Software as a Service”) agreements usually reserve to the vendor a right to modify the service and to revise policies hosted at a URL. Those reservations were drafted for incremental product updates. Applied to AI features whose capabilities can expand materially within months, and to online terms that change just as often, they could allow the vendor to reshape the deal without ever returning to the negotiating table.
Regulators are also paying attention. In February 2024, the US Federal Trade Commission warned that it may be unfair or deceptive for a company to adopt more permissive data practices, such as using data for AI training, and to inform customers only through a “surreptitious, retroactive amendment” to its terms of service or privacy policy. The warning was a welcoming news, especially after incidents such as Zoom’s August 2023 retreat, within days of a public backlash, from terms that appeared to permit training on customer content.
For most enterprise customers, the result is the same: online terms can shift the bargain, the new features send data to places the original data processing terms never contemplated, and the existing indemnity says nothing useful about AI-generated content.
II.The anatomy of an AI Addendum
The AI Addendum has emerged as the principal means of closing these gaps. It supplements the MSA and the data processing agreement, governs the AI features specifically and should expressly take priority over the vendor’s online terms. Its core provisions typically include the following:
| Provision | What it does | Where the negotiation happens |
|---|---|---|
| Definitions | Defines “AI Features,” “Inputs” (prompts, instructions, uploaded files), “Outputs” and any third-party model providers | Whether metadata, usage logs and user feedback count as Customer Data |
| Activation and change control | Keeps features off until an administrator enables them; requires advance notice of new AI features or material changes | Default activation and whether customers can opt out. |
| Data use and training | Limits what the vendor and its model providers may do with Inputs and Outputs | The scope of “service improvement” and “de-identified data” carve-outs |
| Retention and logging | Sets how long prompts and outputs are stored and why | Abuse-monitoring windows; legal holds |
| Output ownership and confidentiality | Allocates rights in Outputs and treats them as the customer’s confidential information | The vendor’s right to reuse similar outputs for others |
| Accuracy and human review | Vendor disclaimers about accuracy; customer’s duty to review | How far “AI may be inaccurate” disclaimers reach |
| IP indemnity | Vendor defends third-party IP claims arising from Outputs | The exclusions and whether the indemnity is capped |
| Agent authority | Limits the actions the AI may take without human approval | Who bears the loss when an agent exceeds those limits |
| Regulatory cooperation and exit | Information for impact assessments and audits; right to disable features or terminate without penalty | Cost and scope of the vendor’s assistance |
Several of these provisions deserve closer attention, because they are where the legacy MSA is most likely to fall short once new AI features arrive.
i.Drafting the no-training covenant
The core promise is easy to state: the vendor will not use Customer Data to train, retrain, fine-tune or otherwise improve any AI model. The real work lies in the definitions. A well-drafted clause captures:
prompts, instructions and uploaded files;
Outputs, including drafts the user discards, in any format, including text, image or video;
user feedback, such as ratings and corrections, which vendors often treat as their own;
metadata and usage data (who asked what, when and how often), which can be commercially revealing even without the content; and
derived data, such as the searchable indexes vendors build so that the AI can draw on a customer’s documents.
Customers can expect three carve-outs in response: “aggregated and de-identified data,” “service improvement” and “security and abuse monitoring.” The first two can swallow the covenant. A common compromise permits de-identified usage statistics (query volumes, error rates) but not content, and confines “service improvement” to work that does not train any model shared with other customers. A model trained only for the customer, walled off from others and deleted on termination, can be expressly permitted.
ii.Flow-down to the model provider
SaaS vendors often do not build the foundation models behind their AI features. Instead, they license them from one of a handful of major developers, such as OpenAI, Anthropic or Google, and fine-tune them on a narrower set of data so that they perform better at particular tasks. A no-training covenant that binds only the SaaS vendor therefore leaves a significant exposure unaddressed: unless the upstream provider is itself contractually restrained, it may process, retain, review or learn from the same data.
The addendum should close that gap. At a minimum, it should require the vendor to identify each model provider that receives or processes customer data and disclose its role, to flow down restrictions at least as protective as those the vendor owes the customer, and to give advance notice before adding or replacing a provider. Where the vendor cannot secure equivalent terms upstream, the customer should know that before enabling the feature, not after.
iii.Retention and legal holds
Vendors often promise to delete prompts quickly, or not to store them at all (“zero data retention”). Litigation can override that promise. In the New York Times copyright case, a May 2025 order required OpenAI to preserve consumer and standard API output logs, but not those of ChatGPT Enterprise and eligible API endpoints operating under a Zero Data Retention amendment. Negotiated retention terms therefore do real work, and the addendum should also require prompt notice of any legal hold affecting customer data.
iv.Privacy considerations
Privacy law adds a further layer. In Canada, PIPEDA’s accountability principle (Schedule 1, clause 4.1.3) leaves an organization responsible for personal information it transfers to a third party for processing and requires a comparable level of protection through contractual or other means. Québec goes further, requiring a written agreement with service providers (s. 18.3) and a privacy impact assessment and any agreed mitigation before personal information is communicated outside the province (s. 17). Bill C-36, the proposed Protecting Privacy and Consumer Data Act, tabled on June 15, 2026, would draw a statutory line between de-identified and anonymized data and introduce fines of up to greater of C$25 million and 5% of prior-year gross global revenue. The Bill is currently at the second reading stage at the House of Commons.
In the United States, the California Consumer Privacy Act regulations permit a service provider to use personal information to build or improve the quality of the services it provides to the business, but not to perform services on behalf of another person (Cal. Code Regs. tit. 11, § 7050(a)(3)). The regulation’s own example allows an email marketing provider to analyze one client’s campaign results and offer the improved service to all of its clients, but not to use that client’s list to send emails for another business. Training a shared model to improve the product generally may therefore be permitted. The risk arises where one customer’s personal information is used to generate outputs for, or otherwise serve, other customers. The difficulty is that AI models can memorize. A vendor may intend only “general improvement,” but if the model retains and reproduces individuals’ details, the use could move into the prohibited category.
v.Who owns the Outputs?
Most vendors now assign Outputs to the customer, but the assignment may transfer only limited rights, and other users may receive similar outputs.. In Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), the D.C. Circuit held in March 2025 that US copyright requires a human author, and the Supreme Court declined review on March 2, 2026. Canadian law remains unsettled: the federal government’s 2025 consultation report identifies AI authorship and ownership as policy questions, and the Federal Court has yet to decide CIPPIC’s ("Canadian Internet Policy and Public Interest Clinic") challenge to a registration naming an AI tool as a co-author.
The practical protection is therefore contractual: treat Outputs as the customer’s confidential information and bar the vendor from reusing them. Customers should obtain an assignment of the vendor’s rights in Outputs, to the extent those rights exist, together with the permissions needed to use any embedded vendor material. Vendors, for their part, should avoid any blanket promise that all generated material will be unique, copyrightable or free of third-party rights.
In both the US and Canadian markets, the drafting lesson is the same: allocate whatever rights exist, address permitted uses and treat infringement protection as a separate issue. An assignment between vendor and customer cannot eliminate a third party’s rights.
III.Indemnities and liability caps
Infringement protection is where the indemnity negotiation begins. Commercial AI vendors will often indemnify customers against defined third-party IP infringement claims. What they generally will not accept is an open-ended obligation to cover outputs that are inaccurate, unsafe, misleading or otherwise harmful. That line shapes the entire negotiation: the indemnity addresses whether the customer’s use or distribution of the output triggers a covered third-party infringement claim, not whether the output is any good.
Scope is the first thing to check. An indemnity limited to infringement by “the services” may not reach the customer’s own publication or distribution of what the services generate. Customers should ask for express coverage of Outputs, confirm which IP rights and territories are covered, and review how defence, settlement and remedies are handled.
OpenAI’s published Service Terms illustrate the limits. For specified API and Enterprise (including ChatGPT Enterprise, Edu and Healthcare) output coverage, the exclusions include ignoring relevant filters or safety features, modifying or combining outputs, lacking rights to inputs, certain trademark claims and content from third-party offerings. Beta services are excluded from indemnification. Coverage depends on the product and the governing agreement, and these provisions should not be generalized to every subscription. Microsoft’s Customer Copyright Commitment offers another operational lesson. Its documentation requires applicable mitigations for Azure OpenAI and specified configurable services, which, depending on the use case, include filters, system instructions and evaluation records; customers tendering claims must demonstrate compliance.
More broadly, counsel for the customer should carefully assess the allocation of AI-related risk under the existing MSA, recognizing that indemnification obligations, liability caps, exclusions and consequential-damages waivers may operate differently depending on the nature of the claim. There is no uniform liability cap applicable to AI-related losses. OpenAI’s published Services Agreement, for example, expressly excludes its defined Service-Specific Terms Indemnity from the general liability cap, although this does not render all AI-related losses uncapped or indemnifiable. In negotiating appropriate contractual protections, counsel for the customer should consider the following:
Negotiate appropriate liability caps and carve-outs. Seek uncapped liability, or a separate and higher liability cap, for specified risks, including third-party IP infringement claims arising from AI-generated outputs, prohibited model training, breaches of confidentiality and material security failures. Ensure that the allocation of liability reflects the nature of the risks, the parties’ respective control over them and any applicable insurance coverage.
Clarify the scope of indemnification and recoverable damages. Recognize that a conventional IP indemnity generally addresses specified third-party infringement claims and may not extend to losses arising from inaccurate AI outputs, unauthorized transactions or the unauthorized incorporation of confidential information into AI models. Expressly address the recoverability of investigation, remediation, data restoration and replacement costs, rather than leaving their classification as direct or consequential damages open to dispute.
Scrutinize how liability caps are calculated. Avoid tying liability exclusively to the fees attributable to an AI feature, particularly where that feature is bundled into an existing subscription at no additional charge. Instead, consider a cap calculated by reference to the fees paid for the affected platform, a negotiated minimum amount, or the greater of the two.
Ensure consistency across liability provisions. Confirm whether indemnification obligations and defence costs fall within or outside the applicable liability caps, whether defence costs erode those caps, and whether the relevant categories of loss are recoverable notwithstanding any exclusion of consequential, incidental or indirect damages.
Preserve protections relating to mandatory safeguards. Require applicable output filters and other contractual safeguards to be enabled by default, with the vendor bearing the burden of establishing that the customer affirmatively disabled them. Any mitigation requirements should be identified at signing, with subsequent changes subject to reasonable notice and an adequate implementation period.
Narrow the exclusions from indemnification. Limit exclusions based on the customer’s knowledge of potential infringement to circumstances involving actual knowledge, rather than constructive or imputed knowledge. Similarly, ensure that exclusions relating to customer modifications apply only where the infringement claim would not have arisen but for the relevant modification.
Address experimental AI functionality. Expressly determine whether beta, preview or experimental AI features fall within the scope of the vendor’s indemnification and liability commitments. Where such features are excluded, require that they remain disabled by default unless the customer knowingly elects to enable them.
IV.When the AI acts: agentic features
These allocation questions become sharper still with the newest features, which do not merely draft; they act. “Agentic” AI can send emails, update records, approve invoices, interact with clients or place orders within limits the customer sets. That raises two important questions: whether the customer is bound by the agent’s actions, and whether the vendor will indemnify the customer for the agent’s mistakes.
On the first question, statutes on both sides of the border favour enforceability. Section 14 of the Uniform Electronic Transactions Act, adopted in nearly every US state, and s. 20 of Ontario’s Electronic Commerce Act both provide that contracts can be formed by electronic agents without human review; E-SIGN adds a U.S. Federal counterpart (15 U.S.C. § 7001(h)). The error relief in s. 21 of the Ontario’s Electronic Commerce Act protects an individual dealing with someone else’s agent, not a business whose own agent erred. The customer will therefore usually be bound by what its agent does.
The second question remains unsettled, and neither Canada nor the United States has yet provided a clear answer. Because liability can run to both the customer and the vendor, both parties have reason to allocate it clearly. The addendum should address:
authority limits the customer configures, such as spending thresholds, permitted tools, approved counterparties and actions that require human sign-off;
a vendor warranty that the agent will operate within those configured limits;
complete logs of every action the agent takes, available to the customer request; Note that this discussion must be reconciled with the discussion on zero-retention and data minimization.
an immediate shut-off right, without penalty; and
loss allocation that distinguishes between an agent exceeding its limits because of a vendor defect and one doing so because of customer misconfiguration.
V.Closing thoughts
Vendors are not acting in bad faith when they embed AI into existing products; for many, it is now the core of the product roadmap. But a contract negotiated for a tool that stored a company’s data should not govern a tool that reads, writes and acts on its behalf.
That mismatch runs through the entire agreement. Modification rights and hyperlinked online terms allow the bargain to shift without a new signature. Data provisions written for storage say nothing about model training, flow-down to upstream model providers, retention under a legal hold or the use of one customer’s personal information to serve others. Output assignments may convey less than they promise, the advertised IP indemnity is often narrower than it first appears, and a liability cap tied to a feature offered at no additional charge may leave little to recover. Agentic features add a further layer: because the customer will usually be bound by what its agent does, authority limits, logging, shut-off rights and loss allocation must be settled before the agent is switched on, not after it errs.
A well-built AI Addendum addresses each of these points in a single instrument that takes priority over the vendor’s online terms. For counsel on either side of the border, the time to close that gap is before the next feature release, not after the first claim.