"Our software vendor just switched on AI features in a platform we have used for six years. Did anything in our contract change?" That question comes up in nearly every vendor review we run with business owners right now, and the honest answer is usually: more than you think, and not in your favor.
The AI vendor contract clauses that matter most are the ones that decide what the vendor may do with your data, who pays when an AI output causes harm, and what you are told when the model changes. If your agreement is silent on those points, the vendor's standard terms fill the gap, and those terms were written to protect the vendor.
Here is the part most leaders miss. The biggest AI contract risk in a typical small or mid-sized business is not the shiny new AI tool going through procurement. It is the incumbent vendor, already holding your client files, financial records, or patient data, that became an AI company partway through your contract term. Nobody ran a fresh review, because nobody thought anything had been purchased.
What changed when your existing vendors added AI?
When a vendor adds generative or predictive features to a product you already use, three things typically shift at once: your data can flow into new processing (sometimes including a third-party model provider), the product begins producing outputs your staff may act on, and the vendor's online terms are updated to cover all of it. Many SaaS agreements incorporate online terms by reference, which means the paper you signed can change without anyone at your company signing again.
Zoom is the clearest public illustration. In March 2023 the company updated its terms of service with language that, as ISACA's review of the controversial terms update describes, drew heavy criticism by mid-August 2023 for appearing to grant broad rights to use customer content for AI. Zoom responded publicly, and its explanation of how its terms apply to AI features now states that it does not use customer audio, video, chat, screen sharing, attachments, or similar communications content to train Zoom or third-party AI models.
The lesson is not that Zoom misbehaved. The lesson is that a terms update with real consequences for confidentiality went largely unnoticed for months, and it took public pressure to clarify it. Most of your vendors will never get that level of scrutiny. If your firm owes clients confidentiality, or holds data covered by HIPAA, GLBA, or a client's security addendum, you need the protection written into your agreement, not inferred from a blog post.
Why "the AI did it" no longer works as a defense
Many leaders assume that if a vendor's AI produces a harmful result, responsibility sits with the vendor. Courts and legislatures are moving the other way. California's AB 316, signed on October 13, 2025, added Civil Code Section 1714.46, which says that a defendant who developed, modified, or used artificial intelligence alleged to have caused harm may not assert that the AI autonomously caused the harm. Other defenses, including causation and the comparative fault of other parties, remain available.
Read that from a business owner's chair. If your company used an AI feature and someone was harmed, "the system decided on its own" is off the table in California. What remains is an argument about who else was at fault, and that argument is won or lost on documentation and on what your contract says about the vendor's responsibilities.
Colorado is moving in a similar direction. Its replacement AI statute, SB 26-189 on automated decision-making technology, signed May 14, 2026, sets up how fault is allocated between developers and deployers in discrimination claims and requires developers to give deployers documentation on intended uses, known limitations, and human review, with core obligations starting January 1, 2027. In its analysis of the new Colorado law, DLA Piper notes that liability is apportioned by relative fault and that companies cannot avoid violations through contractual indemnity clauses. An indemnity clause can still decide who ultimately pays. It cannot make your obligations disappear.
Which AI vendor contract clauses matter most?
The clauses that matter most are data use and training restrictions, notice of AI and model changes, liability and indemnity for AI outputs, performance and testing rights, flow-down to the vendor's own AI providers, and exit terms covering data the AI has already processed. Each one closes a gap that standard SaaS paper typically leaves open.
| Clause | What it should say | Why it matters to the business |
|---|---|---|
| Data use and no-training | Your data, prompts, and outputs are not used to train or improve models without separate written consent | Protects client confidentiality commitments you have already made |
| AI change notice | Advance written notice before new AI features, model swaps, or new AI subprocessors touch your data | Stops the "silent upgrade" problem at renewal and mid-term |
| Output liability and indemnity | Vendor responsibility for harms caused by defects in its AI features, with a cap that reflects the data at stake | Determines who pays when the autonomy defense is unavailable |
| Performance and testing rights | A stated baseline for accuracy or reliability and your right to test against it | Replaces "as is" disclaimers with something measurable |
| Flow-down to AI providers | The vendor's model provider is bound by the same data terms the vendor accepted | Your protections fail if they stop at the first vendor |
| Exit and deletion | Return and certified deletion of data, including derived data, on termination | Prevents your records living on inside a system you left |
Two of these deserve a deeper treatment than one table row allows. Testing and audit rights determine whether you can prove compliance when a regulator or client asks, which our AI Lab covers in the AI vendor contract clause that decides whether you can prove compliance. Exit terms determine whether you can leave without losing a workflow, covered in planning your exit before an AI vendor goes down.
What is a no-training clause, and does my business need one?
A no-training clause states that the vendor will not use your data, inputs, or outputs to train or improve its models, or anyone else's, without your separate written consent. If your firm holds client, patient, or financial data under any confidentiality obligation, you need one. Confirm it lives in the signed agreement or an order form that overrides online terms, not only in a privacy FAQ the vendor can edit.
The federal government already wrote a template you can borrow
You do not have to invent this list from scratch. In April 2025, the White House Office of Management and Budget issued memorandum M-25-22 on acquiring AI in government. It directs agencies to include contract terms that permanently prohibit vendors from using non-public agency data and outputs to train publicly or commercially available AI, to address data and model portability to reduce lock-in, and to preserve the agency's right to evaluate system performance regularly. It also tells agencies to consider requiring advance notice before new AI features are added to contracted services.
Federal agencies have buying power you do not. Still, the memo is useful in a negotiation because it shows these are not exotic asks. When a vendor tells you no customer gets a no-training commitment or change notice, you can point out that the U.S. government now expects both as a matter of policy. Some vendors will move; the ones that refuse have told you something about how they view your data.
What a silent contract costs you
The cost of a weak AI contract rarely shows up on the day it is signed. It shows up when a client sends a security questionnaire asking whether their data trains any AI model and you cannot answer with confidence. It shows up at cyber insurance renewal, where carriers increasingly ask about third-party and AI exposure, a pattern we covered in what underwriters actually require in 2026 and in our look at how shadow AI and insurance exclusions are colliding on SMBs. And it shows up after an incident, when the question of who pays is decided by language nobody negotiated.
This is a governance problem more than a legal one. Contracts get reviewed when someone is buying. AI features arrive when nobody is buying. That is why we argue that AI adoption is a governance problem, not a technology one: the controls have to be triggered by change, not by purchase orders. Before your next renewal cycle, a practical sequence looks like this:
- List every vendor that holds sensitive data, then mark which ones have added AI features in the last two years.
- Pull the signed agreement and any online terms it incorporates, and note where each of the six clauses above is silent or vague.
- Ask each vendor in writing whether your data reaches a third-party model provider, and which one.
- Prioritize renegotiation by data sensitivity, starting with the vendors holding regulated or client-confidential records.
- Calendar a review trigger for any future AI feature announcement, not just for renewal dates.
That work also strengthens the broader vendor program described in protecting your business from vendor vulnerabilities, and it depends on knowing where your sensitive data actually lives, the foundation of any data governance program that supports growth.
How Securafy helps
We are not your attorneys, and we will tell you when a clause needs counsel. What we bring is the technical half of the review that legal teams often cannot do alone: identifying which of your platforms have quietly enabled AI features, tracing where your data goes once those features run, and confirming whether a vendor's model provider is in the data path at all. That gives your counsel a precise list of issues to negotiate instead of a generic checklist.
From there, we help build the operating rhythm that keeps the review current: a vendor inventory tagged by data sensitivity and AI exposure, a change-trigger process when vendors announce new features, and evidence your auditors and insurers can follow. The goal is simple. When someone asks how you vetted the AI touching your clients' data, you have an answer on file.
Where To Go From Here
Your contracts were written for software that stored and processed data. Many of those same vendors now generate outputs your people act on, and the terms have not caught up unless you made them catch up.
If you want a structured look at where AI already sits in your vendor stack and which agreements need attention first, our AI Discovery Program maps your AI exposure across the tools you already use.
If you would rather talk through your specific environment first, book a strategy call with Securafy and we will walk through your current vendors, the clauses that matter most, and where to start.
By Randy Hall
Join The Conversation
Have a question or perspective on this topic? Add it below.