⌘ K
Partner with us
Insights
All insightsResourcesAboutTalk to usPartner with us

The Training Toggle Changed What Your Vendor Is

Explore how AI training settings affect vendor relationships, data privacy, security obligations, and enterprise AI governance.

By Editorial Team

8 min read

The Training Toggle Changed What Your Vendor Is
AI-TRAINING · AI-VENDOR-RISK

Several software suppliers spent this year switching on AI training over customer data, by default and by plan tier. A vendor that trains its own models on your content has stopped acting on your instructions. Privacy regulators across the Spanish- and Portuguese-speaking world just put that in writing, and added that anonymizing the data first does not change the answer.


This month the Red Iberoamericana de Protección de Datos, the network of data protection authorities covering Spain, Portugal and Latin America, published 29 pages of standards on personal data in AI systems. Much of it restates familiar principles. One provision does more. According to PPC Land's November 12 reading of the text, section 2.2.2 treats a vendor's use of a client's personal data for the vendor's own purposes, model training included, as a separate purpose that the client's processing mandate does not cover without express authorization and a legal basis of its own. Anonymizing or transforming the data first falls under the same provision.

The standards are guidance. They have no enforcement mechanism, and their weight depends on national regulators choosing to cite them. A CIO in Chicago could reasonably file them under foreign soft law.

That would be a mistake, because the document describes something that happened inside most enterprise software estates in 2026 without anyone on the buying side making a decision.

A setting made the decision

Start with what changed. From April 24, GitHub began using interaction data from its Copilot Free, Pro and Pro+ plans to train its models unless the individual user opts out. Copilot Business and Enterprise are excluded. From August 17, Atlassian began using customer data from Jira, Confluence and Jira Service Management to train its Rovo AI products, a change that reseller Seibert Group says touches roughly 300,000 customers. Presago's September summary of Atlassian's trust pages gives the defaults: metadata contribution is on for every plan, and only Enterprise customers can turn it off. Contribution of in-app content, meaning page bodies, issue descriptions and comments, is on by default for Free and Standard and off for Premium and Enterprise.

HubSpot went the other way under pressure. A July 1 terms update tied to a shared contact-enrichment dataset drew enough customer reaction that the company withdrew it within days and committed that future features of that kind would be "fully and transparently opt-in."

Look at how the protection is distributed. It is sold by tier. "We're on enterprise terms" is usually a true statement about the main contract. It is a much weaker statement about the Standard-tier Jira site a business unit opened three years ago, the instance that came with an acquisition, or the developers paying for Copilot on personal plans. Each of those is governed by a default nobody in procurement chose.

A default-on training setting is a contract amendment you accepted by not logging in.

Why the role matters more than the setting

Privacy law in Europe, the UK and much of Latin America sorts companies that handle personal data into two roles. A processor acts on the customer's instructions. A controller decides for itself why the data is being used. Your data processing agreement with a SaaS vendor is built on the first role.

When that vendor trains its own model on your content, it is pursuing its own commercial purpose. Seibert Group's analysis of the Atlassian change concludes that for this use the vendor is a controller in its own right, and that the existing agreement's language about improving products predates generative AI. The Ibero-American standards say the same thing in general terms: providers that reuse data for their own purposes take on the corresponding responsibilities, and contract terms about model development do not substitute for a legal basis.

Three consequences follow for the buyer. The agreement you negotiated does not govern that use. The notices you gave your own employees and customers probably do not mention it. And it is hard to reverse. The standards concede that a trained model may be unable to fully honor erasure and objection rights, and require the controller to justify, document and tell the person when that is the case. Opting out stops future contribution. None of the public material reviewed for this article says what happens to what was contributed before the switch was flipped.

Vendors answer with de-identification and aggregation. The standards do not accept that as a closing argument. Anonymization counts only where the absence of reasonable re-identification risk is demonstrated, the burden sits with whoever claims the rules do not apply, and a model itself is in scope if personal data can be extracted from it. European regulators took a similar line in December 2024, when the European Data Protection Board said models trained on personal data cannot automatically be treated as anonymous.

Figure 1

Vendor

What changed in 2026

Default

Who can switch it off

GitHub Copilot

From April 24, interaction data from individual plans trains GitHub's models

On for Free, Pro and Pro+. Business and Enterprise excluded

Each individual user, in account settings

Atlassian (Jira, Confluence, JSM)

From August 17, customer data contributes to training Rovo

Metadata: on for all plans. In-app content: on for Free and Standard, off for Premium and Enterprise

In-app content: any organization admin. Metadata: Enterprise only

HubSpot

July 1 terms change tied to a shared contact-enrichment dataset

Withdrawn within days after customer objections

Company committed to opt-in for future features of this kind

Model APIs (Anthropic, Google Cloud, Microsoft Foundry, OpenAI)

No change reported: commercial terms exclude training on customer content without permission or opt-in

Off

Not applicable, but an application vendor built on these models sets its own terms

Where AI training on customer data was switched on in 2026, and who holds the switch. Compiled from Presago (September 23, 2026), Seibert Group (April 23, 2026, updated October 28, 2026) and Deepak Gupta in Security Boulevard (November 11, 2026, provider terms checked October 30). Vendor settings change; verify each in your own admin console. Not a complete list.

The exposure is one layer up

The last row of that table is the surprising one. The companies that build the large models have, on paper, the strictest position. Security Boulevard's November 11 review of provider terms found each of the four excluding training on business customers' content unless the customer permits or opts in.

The risk sits in the application layer, where a vendor trains its own features on what customers type into its product. And that review makes a point procurement teams should take literally: the binding answer lives in the master agreement and the data processing addendum. A trust page binds no one.

Five things to settle before the next renewal

  1. Inventory. List every tenant and every tier of each SaaS product in use, including individually purchased seats, and record the training setting on each.

  2. Role. Ask each vendor in writing whether, for model training, it acts as your processor or as a controller.

  3. Notice. Treat a change to training terms the way you treat a new sub-processor: advance written notice and a right to object.

  4. History. Ask what was contributed before any opt-out, and whether it can be removed from training sets or models.

  5. Document. For every assurance, ask which signed document it comes from.

What this means for technology leaders

This is an asset-management problem before it is a legal one. Counsel cannot assess a setting nobody has listed, and in most companies the list of SaaS tenants by tier does not exist in one place.

Tier has become a data-governance control. The gap between a Standard and an Enterprise plan now includes who decides what happens to your content. That belongs in the consolidation math when the license review comes around, alongside seats and features.

Expect the regulators' reading to travel. The Council of Europe committee that oversees its data protection convention meets November 17 to 19 with draft guidelines on large language models on the agenda, though no adoption date has been committed. Any company with staff or customers in Spain, Portugal or Latin America is already inside the territory of the authorities that wrote this month's standards.

The vendors announced these changes, on blogs and in updated terms. Announcing is a lower bar than asking. The buyers who caught it were the ones who had assigned someone to read the settings page.


Sources and notes. Luís Rijo, PPC Land, "Client data is off limits for vendors' AI training by default, RIPD says" (November 12, 2026): summary of the Red Iberoamericana de Protección de Datos document "Estándares Iberoamericanos sobre el uso de datos personales en sistemas de inteligencia artificial" (29 pages, dated November 2026), including sections 2.2.2, 1.2.5 and 5.5, Chapter VI, its non-binding status, the EDPB Opinion 28/2024 reference and the Council of Europe Convention 108 committee dates. The RIPD text was not read directly; provisions are as reported. Presago, "Your SaaS Vendors Are Training AI on Your Data" (September 23, 2026): GitHub, Atlassian and HubSpot changes, defaults and dates. Katie Lai, Seibert Group (April 23, 2026, updated October 28, 2026): Atlassian customer count, data categories, controller analysis. Seibert Group is an Atlassian reseller. Deepak Gupta, Security Boulevard (November 11, 2026): model provider terms and where binding commitments sit. Figure 1 reflects those sources, not a review of each vendor's current terms. Journalism, not legal or procurement advice. Corrections welcome.

Tagged

#ai-training#ai-vendor-risk#data-privacy#enterprise-ai#ai-governance