Customer data in ChatGPT: what reps may paste

Microsoft's Work Trend Index 2024 found that 78 percent of employees who use AI at work bring their own AI tools, and that 52 percent of AI users are reluctant to admit using it for their most important tasks. Sales is no exception. What sets sales apart is that the raw material is almost always personal data: a name, a role, an email address, what the person said in the meeting and what they are hesitating about. What follows is an editorial walkthrough of the rules, not legal advice, but the basics are clearer than most people assume.

Three things happen when you hit paste

The first is that you are processing personal data. IMY's guidance on GDPR and AI, with pages updated in May and June 2026, is blunt about it: GDPR applies whenever personal data is processed in connection with developing or using AI, and among the activities that trigger the rules it explicitly lists data being entered, retrieved or generated when an end user uses a deployed AI system. A customer's name in a prompt is enough.

The second is that the data leaves your organisation without a contract. Under GDPR, anyone who processes personal data on your behalf, the processor, must be bound by a written agreement under Article 28(3). A personal account on a free service has no such agreement with your employer. That leaves you with a vendor who answers to nobody on your side, and no way to show where the data went.

The third is that you no longer control what the data is used for. In May 2023 Samsung banned its staff from using generative AI tools on company devices after internal information had leaked to ChatGPT, TechCrunch reported. The example is old, but the mechanism is unchanged in 2026: what you type into a service without a business agreement can be stored, reviewed and, in some services, used to train the model.

What IMY says about legal basis and roles

IMY's guidance holds AI use to the same standard as any other processing: a clear purpose, one of the six legal bases, and protection of the data at every step. For sales work, legitimate interest is usually the only plausible basis, and IMY describes the test in three steps: there has to be a legitimate interest, the processing has to be necessary for it, and that interest has to be weighed against the data subject's. Drafting a polite follow-up email can be justified. Letting a tool with no agreement keep the customer's entire meeting notes rarely can.

The guidance also states that organisations must clarify their role, as controller or processor, so that responsibility lands in the right place. For customer data in a sales team, the employer is the controller. That means the company, not the rep, has to be able to show which basis and which agreement apply, and it is the company that has to notify IMY of a personal data breach within 72 hours if something goes wrong.

What a business agreement changes, and what it does not

The business tiers of the big AI services settle the processor question. OpenAI states on its enterprise privacy page that business data from ChatGPT Business, Enterprise and the API platform is not used to train its models by default. For Enterprise the page adds that deleted conversations are removed within 30 days unless the law requires otherwise. Anthropic's commercial terms from June 2025 say Anthropic may not train models on customer content from its services, and that the services are not for consumer use. A business agreement also brings a data processing agreement to sign, and a counterparty that is obliged to report incidents to you without undue delay.

What the agreement does not change is the purpose or the legal basis. A vendor that does not train on your data does not make the processing lawful by itself; it only makes lawful processing possible. The company still has to decide what the tool may be used for, how long data stays in it and who may read it. Personal accounts run on different terms from business accounts, and those terms change. Anyone who wants to use a personal account for work should read them rather than assume.

One rule per situation

• Email drafts and help with wording: give the role, the industry and the problem, not the name, the company or the quotes. 'A CFO at a mid-sized construction company who is hesitant about the contract term' gets you text every bit as good as the personal data would.
• Meeting notes and summaries: only in a tool the company has approved, with a data processing agreement in place and a deletion rule for whatever is kept there.
• Research before a call: reading the customer's public pages is one thing; pasting them into a tool with no agreement is a further processing operation. Use the company's tool, or do the reading yourself.
• Recordings and transcripts: follow the rules for AI meeting assistants. They are stricter than the rules for text, because the voice and the whole conversation travel with the file.
• Never, in any tool: personal identity numbers, health data, union membership or anything else in the special categories, whatever the agreement says.

Checklist for the sales manager

1. Take stock of which AI tools the team actually uses, personal accounts included. Ask without punishing; Microsoft's finding that half of users would rather not admit to their AI use applies to your team too.
2. Sign a business agreement, with a data processing agreement, for the tool you choose, and switch off model training wherever that is an option.
3. Write a one-page policy with examples, not clauses: what may be entered, what must be anonymised and what never goes in.
4. Decide on retention: how long conversations containing customer data may stay in the tool.
5. Drill the anonymisation habit. It is a writing technique, and it takes about a week to learn.
6. Have an incident routine: who the rep calls if something sensitive ends up in the wrong place anyway, so the 72-hour clock does not run in silence.

What a rep can do today

The EDPB's draft guidelines on web scraping for AI training, with IMY as lead rapporteur, were out for consultation until 30 October 2026, and the final version will sharpen the picture of what AI vendors may collect. What the individual rep controls is simpler: the habit of never typing a real name into a prompt. The same habit carries over into practice. In Salesprep you talk to fictional AI customers, and when you build your own counterpart from pasted text you describe the role, the industry and the situation, not a real person. The raw audio from practice calls is deleted right after scoring, transcripts and scores are kept in your history, and you can delete them whenever you like. Three free calls are included in the Cold call module, no card needed.