Your Contract Says You Can Train on Their Data. Your Buyer Just Read It.

Somewhere in your master services agreement is a sentence written before generative AI existed. It says something close to this: the vendor may use customer data to operate, maintain, and improve the services. Nobody negotiated it. It has been in the template for years, copied from an older template that was itself copied from something a law firm drafted in 2016.

That sentence reads differently now. "Improve the services" used to mean fixing bugs and building better dashboards. Today it can also mean fine-tuning a model that ships to the customer sitting across the table from your prospect. Buyer-side counsel worked this out sometime last year, and the paragraph everyone used to skip has become one of the slowest stretches of enterprise legal review.

The scale of the exposure is measurable. An analysis of contract terms by TermScout, drawing on research from Stanford Law School's CodeX center, found that 92% of AI contracts claim data usage rights beyond what is required to deliver the service. Across software contracts generally, that figure is 63%. Some AI agreements go further and permit use of customer data for competitive intelligence.

For Revenue Leaders, Sales Executives, Product Marketers, and GTM Teams Selling AI-Powered Products

Your data terms have quietly become part of the product buyers evaluate. Not the security posture, which most vendors already document. The specific question of what happens to a customer's inputs after the session ends, who else benefits from them, and for how long. Most revenue orgs have never read their own answer to that question. Their prospects have.

The language came from a world that no longer exists

Standard SaaS worked one way for fifteen years. Customer data went in, got stored and processed, and came back out. The vendor kept logs, ran aggregate analytics, and used the word "improve" loosely because nothing much turned on it.

An AI engagement breaks that model. The vendor may want to use customer data, or things derived from it, to train or fine-tune models that serve other accounts. Most contracts signed in the last three years draw no distinction between those two situations, which means the old permission sits there covering the new behavior by default.

Regulators noticed early. In February 2024 the FTC warned that a company collecting data under one set of privacy commitments cannot quietly revise those commitments to enable AI training, and that doing so may be unfair or deceptive under the FTC Act. The warning was aimed at terms-of-service updates, but the underlying principle travels: what you told customers when you collected the data constrains what you can do with it later.

The commercial risk goes past the obvious version of the story, where a document gets seen by someone it shouldn't. A model trained on a customer's proprietary workflows absorbs patterns. Pricing structures, deal shapes, escalation paths, the way a particular team writes its renewals. Those patterns can surface in a product sold to that customer's competitor without anyone doing anything malicious. Try explaining the difference to a CISO at 4pm on a Friday.

Most vendors answer this question by saying nothing

Cybernews scored 500 AI companies across 36 countries this year on security, data privacy, organizational transparency, and public perception. The finding that matters commercially is not who ranked first. It is that 63% of the companies assessed do not clearly disclose whether they train models on user data. Within that group, 42% said nothing at all and 21% used vague language.

Retention was worse. Around 65% do not clearly state how long they keep user data, and 56% referenced retention without ever committing to a timeline.

I find this genuinely strange, because the fix is cheap. A paragraph on a trust page and a defined term in the MSA costs a few hours of legal time. The reason most companies haven't written it is not cost. It is that writing it down forecloses options the product team has not decided about yet, and nobody wants to be the person who narrowed the roadmap in a contract clause.

That is a real tension, and I do not think it resolves cleanly. But leaving the question open no longer reads as flexibility. It reads as evasion, because "we may retain data to improve our services" is exactly the sentence a buyer would expect from a company that plans to train on their data and would rather not say so.

Buyers stopped taking your word for it

The verification behavior on the other side of the table changed faster than most sales orgs adjusted for.

TrustRadius surveyed 1,862 buyers and 444 vendors for its 2026 B2B Buying Disconnect report. Sixty-three percent of buyers used AI somewhere in their buying process, and 94% fact-check what it tells them. The share who verify always or very often climbed from 58% to 72% in one year. Meanwhile 47% said they trust online resources less than a year ago, up from 39%, and vendor marketing collateral ranked dead last among the sources buyers actually consult.

Gartner's May 2026 survey found the same instinct pointed at people: 69% of B2B buyers go to sales reps specifically to validate AI-generated insights. Buyers are not avoiding your team. They are using your team as a fact-checking service, which means every claim a rep makes about data handling gets tested against the contract, the trust page, and whatever a language model says about your company when asked directly.

Fifty-nine percent of the purchases in the TrustRadius data were AI tools or tools with AI features. So the verification instinct and the AI purchase are landing in the same deal, at the same time, in front of a buying committee that has been told by its own risk function to get specific.

Cisco's 2026 Data and Privacy Benchmark Study, based on responses from 5,200 professionals across 12 markets, shows what those risk functions have been doing with their budgets. Ninety percent of organizations expanded their privacy programs because of AI. Thirty-eight percent spent at least $5 million on privacy in the past year, up from 14% in 2024. That money buys people, and those people write questions.

The examples that made it concrete

Nothing moves a procurement policy like a story with a company name attached.

In November 2025, Figma was hit with a proposed class action in California federal court alleging it used customer designs and intellectual property to train generative AI tools without permission, including a claim that users were automatically opted in. Figma denied training on customer data without explicit authorization. The litigation will take years. The procurement policy took about a month.

Adobe went through a version of this in June 2024, when updated terms triggered a customer backlash over what rights the company had claimed to their content. Adobe clarified its position and committed not to train its AI systems on customer work, pointing instead to a licensed image collection behind Firefly, with revenue shared back to contributors.

Both cases produced the same artifact on the buyer side: a question in the security questionnaire that did not exist before, phrased in a way that a marketing page cannot satisfy. When the buyer asks for the training answer in writing, they are asking for a contractual commitment, not a link to your FAQ.

What saying "never" actually costs you

Here is where I part ways with most of the advice on this subject, which treats "we never train on customer data" as an obvious win.

It is a strong sales position. It is also a promise about your product's future that a lot of companies make without pricing it. Customer data is the fastest feedback loop most software businesses have. Aggregate benchmarks, model tuning against real usage, error correction from live corrections: these are how a product gets better than its competitor's in the places customers actually feel. Cisco's data shows 65% of organizations already struggle to access relevant, high-quality data. Cutting yourself off from your own is not a free move.

There is also a dependency problem running underneath all of this. A 2026 Zapier survey of 542 US leaders with active paid AI vendor contracts found 74% expected operational disruption, or described themselves as fully reliant, if a key AI vendor stopped service. Buyers are aware of how deep these tools go. That awareness raises their standards on every term that touches ownership and continuity, including data rights.

So the choice is not between saying yes and saying no. It is whether you design the answer or inherit one. Blanket promises made under sales pressure tend to be broad, permanent, and written by someone who will not be in the room when the model team asks for training data in eighteen months.

What a defensible position looks like

The vendors handling this well have stopped treating it as one question with one answer.

They separate customer content from operational metadata, and they say so explicitly. Prompts, uploads, and outputs sit in one bucket with tight commitments. Usage telemetry, error rates, and latency data sit in another, and the contract defines the line rather than leaving it to interpretation. Vague terms are where trust dies, because a buyer who cannot tell which bucket their data falls into assumes the worse one.

They define derived data. This is the term most agreements skip and most sophisticated buyers now catch. If a model learns a pattern from a customer's data, who owns the pattern? An answer that says "aggregated and anonymized insights" without explaining the aggregation method is not an answer, and buyers who have read a few of these will say so.

They commit to a retention number. Not "as long as necessary." Thirty days, or ninety, or the length of the term plus a defined tail, with a deletion path the customer can trigger. Given that 65% of vendors will not commit to a timeline, being one that does is cheap differentiation.

They handle the version problem. The model a buyer evaluates during procurement may be materially different from the one running in production six months later, which is one reason enterprise legal teams now ask for notification before material model changes. Committing to notice is easier than defending silence after the fact.

They build a value exchange for the cases where they do want training rights. Discounted pricing in return for consented use, opt-in programs with visible benefits, or design partner arrangements that name what the customer gets. Data rights you paid for in the contract are stable. Data rights you assumed through a 2019 boilerplate clause are a lawsuit waiting for a plaintiff's firm to notice you.

Cisco's respondents pointed at the same lever from the buyer side: 46% said clear communication about how data is collected and used is the single most effective way to build customer confidence, and 95% said privacy is essential to trust in AI-powered services. Consumer research from Usercentrics this year found 52% of people would pay more for AI transparency, at an average 7% premium, and 47% had already canceled, switched, or cut spend in the previous six months over how their data was being used. Consumer behavior is not enterprise behavior. But the direction of travel is hard to miss, and the people making enterprise decisions go home and cancel subscriptions like everybody else.

This is not the legal team's problem to solve alone

The mistake I see most often is routing this entirely to counsel. Legal writes a clause, the clause goes in the paper, and nobody in the revenue org can explain it in a live conversation.

Data terms are enablement content now. Your reps need a one-page position they can say out loud without checking with anyone: what you do with prompts, what you do with outputs, what you retain and for how long, what changes if the customer opts into anything, and what you will put in writing. If the answer requires escalation to legal every time, you have added days to deals that were already slow.

Product marketing owns the public version, and it belongs on the trust page next to the SOC 2 report, not buried in a terms document. Given that 63% of vendors publish nothing clear on this, publishing something specific gets you compared favorably by any buyer running a shortlist.

RevOps owns the deal data. Track how often data-rights questions appear, at what stage, and what they cost in cycle time. Most teams cannot answer that today, which is why the issue stays invisible until a large deal stalls and somebody goes looking for a pattern.

Three in four organizations report having an AI governance body, but only 12% describe it as mature, per Cisco. Immature governance on the buyer side is not good news for sellers. It means inconsistent questions, longer reviews, and committees that resolve ambiguity by asking for more paper.

The next ninety days

Read your own MSA first, specifically the improvement and derived-data language, and get an honest read on whether it permits model training today. A lot of executives are surprised by their own paper. Then decide what you actually want, with the product team in the room, because the sales answer and the roadmap answer have to be the same answer.

Write the position down in plain language and publish it. Give the reps a version they can deliver in a meeting. Put the definitions and the retention number in the contract, not just the trust page, since buyers who care will ask for the contractual version anyway.

Then watch what happens in legal review. If your data clause stops generating redlines, you did it right. If it keeps generating them, the specific objection tells you what the market now expects, which is more useful information than any analyst report on the subject.

Where this settles

The companies that get this right are not the ones with the most restrictive terms. They are the ones whose answer is the same in the sales deck, the trust page, the contract, and the product. Buyers are checking all four, and the gap between them is what creates the doubt.

The uncomfortable part is that many vendors will have to give up rights they have technically held for years, and some of those rights were genuinely useful. That is a real cost, not a rhetorical one. But those rights were only ever worth something while nobody was reading the clause, and that period ended. What replaces it is narrower, negotiated, and paid for, which is a worse deal for the vendor and a more durable one for the relationship.