Home / Blog / Is it legal to sell employee Slack and email data for AI training?
Is it legal to sell employee Slack and email data for AI training?
It can be lawful, but it is not automatic. Owning the workspace does not give a company unlimited rights over the personal data in it. In the EU and UK, using employee messages to train AI is a new purpose that needs a lawful basis and transparency, employee consent is rarely reliable, and pseudonymised data is generally still personal data. In California, employee data has been covered by the CCPA since 2023. The careful route is to check what people were told, remove sensitive content, de-identify thoroughly, and bind the buyer by contract.
Reports that closed startups have sold their Slack archives and email to AI developers have raised a fair question from employees and founders alike: is that allowed? The honest answer is that it depends on where the people in the data are, what they were told, and how well the data is de-identified. This post sets out the main rules in the EU, UK and US, and what a careful process looks like. It is general information, not legal advice, and some of these questions have not yet been tested by regulators or courts.
Who is in the data
A workplace archive is not just about employees. It usually contains messages by and about contractors, job candidates, customers, suppliers and partners. Each group can have rights, and some of what they wrote may be covered by confidentiality agreements. Any analysis has to start with who is in the data, not just who owns the workspace. If the company has closed, also check who now controls the records; our post on selling data and IP during a shutdown covers that.
EU and UK: the GDPR
In the EU the rules are in the General Data Protection Regulation. The UK has its own version, the UK GDPR, alongside the Data Protection Act 2018; it has been amended by the Data (Use and Access) Act 2025, so check current guidance from the ICO. The main points are similar in both.
- Messages are personal data. Anything that relates to an identifiable person counts (Article 4(1)), including what they wrote and what others wrote about them.
- New purpose, new questions. Data collected to run a business can't simply be reused for something else. Under purpose limitation (Article 5(1)(b)) and the compatibility test in Article 6(4), licensing messages for AI training has to be assessed as a further purpose.
- A lawful basis is needed. Consent is a weak basis at work: the European Data Protection Board's guidelines on consent say employees are rarely in a position to give it freely. Legitimate interests (Article 6(1)(f)) is the more common candidate. In Opinion 28/2024, the EDPB said it can be a possible basis for AI development, but only after a balancing test in which people's reasonable expectations carry weight. Most employees are unlikely to have expected their chats to become training data.
- People must be told. Before using data for a new purpose, the controller must give people information about it (Article 13(3); Article 14 for people whose data came from someone else). When people can object, as they can to processing based on legitimate interests (Article 21), they need a real chance to do so.
- Sensitive data has stricter rules. Chat is full of special-category information: sick days, medical appointments, religion, union activity. Processing it needs an exception under Article 9. In practice, remove it.
- Assess the risk. Large-scale processing of employee communications for a new purpose is the kind of case where a data protection impact assessment (Article 35) is expected.
- National employment rules. Article 88 lets member states set extra rules for employee data, and some countries give works councils a say over employee data use.
Anonymised is not the same as pseudonymised
This distinction decides whether the GDPR applies at all.
- Anonymous data can't be linked to a person by any means reasonably likely to be used. It falls outside the GDPR (Recital 26).
- Pseudonymised data has names replaced with codes, but could be linked back with extra information. For the company that holds the archive, and anyone else who could reasonably link it back, it is still personal data (Article 4(5) and Recital 26), so all the rules above still apply.
Free-text chat and email are hard to make fully anonymous. People are identifiable from context ("the only designer in the Berlin office"), from writing style, from job titles and from events. Replacing names is a necessary step, not the finish line. The UK ICO's anonymisation guidance is a practical reference for judging identifiability risk.
United States
There is no general federal privacy law covering workplace messages, but several rules can still apply.
- California (CCPA). Since 1 January 2023, the California Consumer Privacy Act has covered employee and business-contact personal information, not just consumers. Licensing personal information for something of value can count as a "sale", which brings notice and opt-out rights (Cal. Civ. Code sections 1798.100, 1798.120 and 1798.140). The CCPA applies only to businesses above certain size or data-volume thresholds.
- De-identified data under the CCPA. Information that meets the CCPA's definition of "deidentified" (section 1798.140(m)) falls outside personal information. The business must take reasonable measures so the data can't be linked to a person, publicly commit not to re-identify it, and contractually bar recipients from re-identifying it.
- The FTC. The Federal Trade Commission can act against unfair or deceptive practices. It has warned that quietly changing a privacy policy to allow AI training on data collected under earlier promises may be unfair or deceptive (FTC, February 2024). This matters most for customer data inside support tickets and email.
- Other states. Most other state privacy laws exclude data handled in an employment context, but they may cover customer data in the same archive.
- Contracts. Confidentiality agreements with customers and partners, and promises in employee handbooks, can restrict use regardless of privacy law.
What a careful process looks like
- Read what people were told. Privacy policy, employee privacy notice, handbook and acceptable-use policy, in the versions in force when the data was collected.
- Narrow the scope. Many companies exclude direct messages, private channels and HR, legal and medical channels altogether, and license only work channels and shared inboxes.
- Remove what shouldn't be there. Special-category data, payment data, credentials, and customer data covered by contracts.
- De-identify thoroughly. Consistent pseudonyms across systems, removal of contact details and IDs, coarsening of rare details, then human review of free text. See our methodology.
- Document the assessment. A legitimate interests assessment and, where it applies, an impact assessment, kept with the licence.
- Tell people where you can. Notifying current and former staff and offering a way to opt out reduces legal and reputational risk, even where the law may not strictly require it.
- Bind the buyer. The licence should ban re-identification and contact with individuals, limit use to training and evaluation, and require deletion at the end. Our rights checklist covers the other terms.
Some archives should not be licensed at all, for example where the content is mostly personal or the notices clearly promised something else. A good partner will tell you that.
Talk it through before you decide
Datadhara licenses business records only under a signed agreement with the company that controls them, with people pseudonymised, sensitive content removed and contracts checked first. If you're unsure whether your company's Slack or email can be licensed, book 15 minutes and we'll walk through what is likely in scope and what should stay out. More on the data itself is on our business operations data page, and the full process is in how to license your company's data.
This post is general information, not legal advice. Privacy law differs by country and state and is changing; take advice from a qualified lawyer before licensing personal data.
Questions
Can a company sell employees' Slack messages to train AI?
Sometimes, but not automatically. Under the GDPR it needs a lawful basis, a compatibility assessment and notice to the people involved; in California the CCPA covers employee data. Thorough de-identification and removal of sensitive content are usually essential.
Is pseudonymised data still personal data under the GDPR?
Yes. Data where names are replaced with codes but could be linked back with extra information is generally still personal data under Article 4(5) and Recital 26, certainly for anyone who can link it back. Only anonymous data falls outside the GDPR.
Does the CCPA cover employee data?
Yes. Since 1 January 2023, the CCPA has applied to personal information of employees, job applicants and business contacts held by businesses that meet its thresholds.
Do employees have to consent to their emails being used for AI training?
Under the GDPR, consent is rarely a reliable basis at work because of the power imbalance, so companies usually rely on legitimate interests with a balancing test and clear notice. Rules differ by country, so take legal advice.
Find out what your data could be worth
Tell us roughly what you have: repos, years of history, the systems you used. We'll give you a straight answer in 15 minutes.