Paying for AI won’t protect your customers’ data: what matters is the account

Can your company put customer data in AI tools? The answer depends far less on the tool (ChatGPT, Claude, Gemini, Grok or Copilot) than on the type of account it is used with. A personal account, even a paid one, runs on consumer terms. By contrast, a properly contracted business account turns the provider into a processor, with a contract and safeguards. In this article we review what each provider’s terms say today. We also explain what a company should do before letting its team work with AI.

The right question: which account is the AI used with?

Almost every provider offers two separate worlds:

• Personal plans (free or paid). They run on consumer terms. The provider does not sign a processing agreement with your company and, in several cases, may use conversations to train its models.
• Business plans (Business, Team, Enterprise, API, or the version included in Google Workspace or Microsoft 365). The provider processes data on the company’s behalf under a data processing agreement (DPA). It also commits not to train on that data.

The difference is legal, not technical. The GDPR requires the controller to use «only processors providing sufficient guarantees» (Article 28(1)). With a personal account there is no such contract. So there is no processor to choose: there is a third party receiving your customers’ data under its own rules.

Customer data in AI: what each provider says

We reviewed each provider’s official terms and help centres in October 2026. They change often, so check them again before signing up.

OpenAI (ChatGPT)

• Free, Go, Plus and Pro (personal accounts). OpenAI states that on these plans data sharing for training is on by default. Users can turn it off under Settings > Data controls > «Improve the model for everyone». Temporary chats are not used for training, although OpenAI may keep them for up to 30 days. Moreover, if a user rates a reply with a thumbs up or down, OpenAI may use the whole conversation for training, even after opting out.
• Business, Enterprise, Edu and the API. By default, OpenAI does not use this content to train its models. Its DPA incorporates the EU standard contractual clauses. For customers in the European Economic Area, the contracting party is OpenAI Ireland.
• Data residency in Europe. OpenAI offers it for Enterprise, Edu and the API, and only for new workspaces or projects. ChatGPT Business is not on that list.

Anthropic (Claude)

• Free, Pro and Max. Since 28 August 2025, each user chooses whether their conversations are used for training. If they agree, Anthropic keeps the data for up to five years; if not, for 30 days.
• Team, Enterprise, Education and the API (also through Amazon Bedrock or Google Cloud Vertex AI). These fall under the commercial terms. Anthropic does not train on those inputs and outputs, and the DPA is incorporated automatically with the standard contractual clauses.
• Practical point. If a Team or Enterprise user rates a reply, that conversation may be used for training. Administrators can switch this off with the «Rate chats» setting.

Google (Gemini)

• Gemini app with a personal Google account. With «Keep Activity» on, Google uses conversations to improve its services. That includes training models. In addition, human reviewers read a sample of them, disconnected from the account, and reviewed chats are kept for up to three years. Google’s own Gemini privacy hub advises against entering confidential information.
• Gemini in Google Workspace (eligible editions, including Education). Google does not use the data for training without the customer’s permission, and the content is not reviewed by humans. Gemini also respects the organisation’s permissions and controls, and processing is governed by the Cloud Data Processing Addendum.
• Exceptions within Workspace. Gemini Notebook and Gemini in Chrome do not hold the ISO or SOC certifications of the rest, according to the Workspace privacy hub. Gemini Notebook also does not follow the organisation’s data region or data loss prevention (DLP) settings.

xAI (Grok)

• Grok with a personal account (grok.com, the app or X). xAI acknowledges that it may use conversations for training. Users can turn this off in Settings, and a private mode deletes the chat within 30 days. Its own FAQ advises against sharing personal or sensitive data.
• Grok for business and the API. xAI states that it does not train on this data and that it signs a DPA with business customers. It also deletes conversations after 30 days. However, its enterprise FAQ does not say where data is stored or offer a European option.
• Regulatory context. In April 2025, the Irish Data Protection Commission (DPC) opened an inquiry into X. It concerns the use of European users’ posts to train Grok.

Microsoft (Copilot)

• Copilot with a personal Microsoft account. Its privacy controls page lets users exclude conversations from training; so, unless the user does so, they may be used. Organisational accounts are excluded.
• Copilot Chat and Microsoft 365 Copilot with a work account (Entra ID). Enterprise data protection applies (Microsoft Learn). Microsoft acts as a processor under its DPA. In addition, it does not use prompts or responses to train foundation models.
• Two exceptions few people read. Web searches that Copilot sends to Bing fall outside that regime. For those, Microsoft acts as an independent controller and the EU Data Boundary does not cover them. Nor, for now, does it cover the Anthropic models Microsoft integrates into Copilot.

Summary

Why paying for a personal account does not protect customer data in AI

ChatGPT Plus or Pro, Claude Pro or Max, and paid Gemini or Grok subscriptions are still personal accounts. Paying improves the features, but it does not change the privacy terms. Nor is it enough to switch off training, because that does not create a processing agreement or give the company control over the accounts.

From a legal standpoint, when an employee pastes customer data into a personal account, three things happen:

1. Confidentiality may be breached. The GDPR requires data to be processed with «appropriate security» (Article 5(1)(f)). In Spain, the LOPDGDD extends that duty to everyone involved in the processing.
2. There may be a personal data breach. The GDPR defines a breach as including «unauthorised disclosure of, or access to» personal data. The company must therefore record it and assess whether the risk requires notifying the supervisory authority within 72 hours.
3. The company is accountable. The GDPR requires anyone acting under the controller’s authority to process data only «on instructions from the controller» (Article 32(4)). Without instructions or approved tools, the failure is organisational, not just the employee’s.

What almost nobody mentions: personal data is not the only thing at stake

The analysis does not end with the GDPR. Many companies upload contracts, bids, prices or code to AI tools. That information may be a trade secret. However, under Spain’s Trade Secrets Act (Ley 1/2019), mirroring the EU Trade Secrets Directive, protection depends on the holder having taken «reasonable steps» to keep it secret. If a company lets that information flow into personal accounts unchecked, it may weaken its own protection in litigation.

In addition, confidentiality agreements with customers often prohibit sharing their information with third parties without consent. So, before using AI on a customer’s documents, check your contract with that customer too.

Customer data in AI business tools: what to review first

A business account is necessary, but not sufficient. Before approving it, review at least the following:

• Contract and transfers. Check the contracting entity and the DPA. Also check the basis for transfers to the United States: standard contractual clauses or the EU-US Data Privacy Framework. The EU General Court dismissed the challenge to this framework in September 2025. That judgment was appealed to the Court of Justice in October 2025.
• Admin console settings. Disable what you do not need: ratings that feed training, shared links, connectors to other apps and excessive retention.
• Record of processing. Add the provider as a processor and, where needed, update the information you give customers and employees.
• Impact assessment. Consider whether a DPIA is needed when health data, children’s data or large-scale processing is involved.
• Decisions about people. If AI plays a part in hiring, performance reviews or customer decisions, also review the rules on automated decision-making.
• Data minimisation. Even with a business account, upload only what is needed. In its January 2026 guidance for citizens, the Spanish Data Protection Agency (AEPD) advises describing a fictitious case. At work, it also asks people not to include confidential information about the organisation, its staff or its customers.

The school example

A real case we see often: a teacher wants to upload a class list to Gemini to form groups. With the school’s Google Workspace for Education account and Gemini enabled by the administrator, the data falls under the same contract as the school’s Drive. Even so, the school must record that use in its record of processing. With the teacher’s personal Gmail, by contrast, there is no contract at all. Data minimisation still applies: to form groups, initials or class numbers are enough. We cover this in more detail in our article on Spain’s first GDPR fine against a school over Google Workspace.

What the AEPD itself does with AI

In January 2026 the AEPD published its summary of obligations for managing generative AI. It is an internal document, but it sets the standard the authority applies to itself. Among other measures, it provides for:

• Contracting external services «in business mode».
• Banning AI systems not listed in the corporate inventory.
• Including no-reuse-of-data clauses in contracts with providers.
• Restricting use to trained and authorised users.

It is therefore a good benchmark of what the supervisory authority itself considers good practice.

Training: the obligation changed in July 2026, accountability did not

Since February 2025, the EU AI Act had required organisations to ensure, «to their best extent», a sufficient level of AI literacy among their staff. However, Regulation (EU) 2026/1744, in force since late July 2026, softened that Article 4. Companies must now take measures «to support the promotion» of AI literacy, without guaranteeing any specific level.

Even so, training remains essential. The GDPR requires clear instructions for anyone handling data, and the AEPD limits AI use to trained people. Moreover, if an incident occurs, being able to show that staff were trained is one of the best proofs of the company’s diligence.

Five-step action plan

1. Inventory. Find out which AI tools your team already uses, including personal accounts.
2. Procurement. Choose one or two providers with a business plan, a DPA and clear transfer terms.
3. Usage policy. Define which data may go into each tool, what is prohibited and who approves borderline cases.
4. Training. Explain to staff the difference between personal and business accounts, and what to do if someone makes a mistake.
5. Breach protocol. Plan how to respond if someone has already uploaded customer data to a personal account.

In short, customer data in AI tools should only go to business accounts with a contract, proper settings and a clear usage policy.

Need help putting AI use in order at your company?

At Auratech Legal we help companies and schools use AI with legal certainty. We draft the AI usage policy, review provider contracts, update the record of processing and train staff. To review your case, see our AI legal services or contact us.

Frequently asked questions

Can I put customer data in AI tools on a free plan if I turn off training?

It is not advisable. Turning off training does not create a processing agreement or give the company control over the account. For customer data, use a business plan with a DPA.

Is temporary chat or private mode enough?

No. It reduces use for training, but the data still reaches the provider under consumer terms and is kept for a while.

Is Gemini in Google Workspace as safe as Drive?

In eligible editions, Gemini is governed by the same data processing addendum as the rest of Workspace. There are exceptions, such as Gemini Notebook, that you should check.

What should I do if an employee has already uploaded customer data to a personal account?

Treat it as a possible security incident. Record what data was uploaded and ask for the conversation to be deleted. Then assess the risk and decide whether to notify the supervisory authority within 72 hours.

This article provides general information and does not replace legal advice on a specific case. Provider terms reviewed in October 2026.