Somewhere in your pipeline a prospect’s security team is reading your terms of service. They aren’t reading all of it. They’re looking to find out if you train AI models on their data.

If your terms answer that clearly, the review moves on. If they’re silent, or if the answer hides inside a phrase like “to improve our services,” the deal slows down while someone asks you directly. And in 2026, silence reads as a yes.

This wasn’t a make-or-break clause just a few years ago. But it is now. Partly that’s because the market settled on a default. And partly it’s because a few well-known vendors spent the past year demonstrating exactly how not to handle it.

The market picked a default already

The large model providers resolved this question in their own paper. Anthropic’s commercial terms state that it doesn’t train on customer content from paid services. OpenAI’s enterprise and API terms take the same position. The companies with the most to gain from training data are promising not to train on paid customer content. As a result, that’s become the baseline every buyer expects.

That expectation puts a lot of existing startup contracts out of market. The catch-all license most SaaS templates carry (something like “we may use customer data to develop, improve, and enhance our products and services”) used to be benign boilerplate. Now it reads like permission to train and without more will derail your deal.

What happens when the promise changes?

In November 2025, a proposed class action was filed against Figma in federal court in California (Khan v. Figma, N.D. Cal.). The complaint alleges the company used customer design files to train its generative AI features after years of assuring users it wouldn’t. Enrollment, the complaint says, ran through default settings and background policy changes; nobody actually said yes. Figma denies training on customer content without permission, and the case is still in its early stages. But keep an eye on the legal theory. AI lawsuits used to be about copyright and fair use. This one is based on contract claim: what the vendor promised and what it actually did.

HubSpot ran the same play this July and reversed it in four days. On July 1, the company updated its terms to pool customer enrichment data across accounts and use it for AI training. Everyone was opted in by default. Opting out took three separate toggles. The backlash was immediate, and by July 5 it was over. HubSpot’s co-founder apologized on LinkedIn, its chief product officer published a post titled “We Got This Wrong. And We Are Fixing It,” and the company promised that any future data-sharing change will be opt-in from the start.

Atlassian is running the experiment right now. Starting August 17, 2026, data from Jira, Confluence, and its other cloud products feeds Atlassian’s AI models by default. Any admin on any plan can switch off the sharing of in-app content: the actual pages and ticket comments. The metadata layer is different. Below the Enterprise tier, it can’t be turned off at all. And read the carve-outs closely, because they tell you what Atlassian’s own lawyers think. HIPAA customers, government cloud, and certain financial services customers are excluded from collection automatically. The customers with active, nosy regulators got a pass and everyone else became the training set.

The FTC staked out its position on this pattern back in February 2024: quietly revising your privacy commitments to permit AI training may be unfair or deceptive under the FTC Act. Building AI isn’t what got these companies in trouble. Changing the data promise in the background and letting customers find out later did.

Five things your terms need to say

If you’re shipping an AI feature, your customer contracts need to settle five things.

First, state your training position in plain language. If you don’t train on customer data, say exactly that: no use of customer data to train, fine-tune, or evaluate any model, yours or a third party’s. If you want training rights, define them clearly and narrowly and make them opt-in. (And expect to fight over it anyway.)

Next, decide what “de-identified” means before your customer’s lawyer does. The standard buyer move in 2026 is to attack the aggregation loophole, like a no-training promise that excludes anonymized or aggregated data. If you intend to use de-identified data for anything, define the standard and say what it’s used for. If you don’t, extend the prohibition to cover it.

Then there’s the question of whose models these actually are. Most startups aren’t running their own models. Prompts flow to OpenAI, Anthropic, or Google, which makes those providers sub-processors of your customers’ data. If you handle personal data for European customers or other highly-regulated jurisdictions, GDPR’s processor rules mean those providers must be identified and authorized as sub-processors in your DPA. Even outside Europe, customers eventually find an undisclosed AI pipeline on their own. Being forthright here builds trust.

Don’t promise ownership you can’t deliver. Founders like to tell customers they own the outputs. But in March 2026, the Supreme Court declined to review Thaler v. Perlmutter, so the lower court’s ruling stands: purely machine-generated work has no copyright under U.S. law. You can assign customers whatever rights exist in AI outputs. But you can’t warrant that those rights include a registrable copyright, because for some outputs they won’t.

The last one is structural. Your customer terms can’t grant permissions your model license doesn’t give you. If your provider restricts certain uses, retains prompts for abuse monitoring, or won’t indemnify outputs, those limits flow through to your customers whether you write them down or not. Make sure that what you grant lines up with what your model provider grants you.

Decide before you ship

The hard part of all this is in making decisions and sticking to them. As you are building your AI features, consider and decide on your data positions. Write it into the terms and the DPA before launch. Make sure the product configurations match the promises you’re making in your terms. Then leave it alone, and if the position ever has to change, change it in the open, clearly and with explicit consent.