Sell customer service data to AI

Why customer service data is interesting for AI

AI labs are increasingly buying workplace data such as Slack, email and support tickets to train AI agents. Customer service is one of the clearest use cases for such agents, and to learn they need to see real cases: how a customer describes their problem, what follow-up questions are asked, how the fault is found and how the case is closed.

So if you want to sell customer service data, you often have something in demand. That is especially true for cases in Nordic and other smaller European languages, which make up a small share of global training data.

Selling customer service data – what might be relevant?

• Support tickets from helpdesk systems, with the full conversation and the resolution.
• Chat logs from web chat and messaging services.
• Call recordings and transcripts from phone support.
• Knowledge bases and internal instructions that agents use.

The value lies mainly in the sequences. A case that goes from problem to resolution carries more weight than short "where is my parcel" questions. Categories, resolution codes and status make the data easier to use.

What makes customer service data more valuable?

• Complete sequences – cases where both the problem and the resolution are included.
• Long history – several years give more case types and more unusual cases.
• Structure – categories, status and resolution codes that make the data easy to sort.
• Local languages – customer communication in Nordic and other European languages is rare in training data.
• Low share of personal data – technical cases about products and systems are often easier to license than cases about individual customers' finances or contracts.

De-identification – the core of every deal

Customer service data is full of personal data. Customers write their names, addresses, order numbers and sometimes details about their finances or lives. That is why de-identification is not a final step, but the very foundation for a license to be possible at all.

What needs to be handled

• Direct identifiers – names, emails, phone numbers, national ID numbers, addresses.
• Indirect identifiers – order numbers, customer numbers, unusual products, places and dates that together can single out a person.
• Free text – customers often enter details where you least expect it.
• Voice – a voice can in itself identify a person, which makes call recordings particularly demanding.
• Sensitive data – we don't handle health data, and cases where it appears need to be filtered out.

How we work

1. Assessment first. Before anything else happens, we assess the legal basis and the amount of personal data. Some material cannot be licensed, and then we say so early.
2. Selection. It is often better to select case types with little personal data than to try to clean everything.
3. De-identification and checks. Identifiers are replaced or removed, and the result is spot-checked before delivery.
4. Documentation. What is done is documented, which also helps the buyer meet the requirements of the EU AI Act (Article 53).

Note that generating synthetic data from personal data also counts as processing, as European data protection authorities have pointed out. Anonymizing or synthesizing therefore does not automatically solve the question of legal basis. Read more about GDPR and anonymization.

Legal basis and purpose

Data protection authorities stress that reusing data for AI training must be compatible with the purpose the data was collected for. The European Data Protection Board has made clear that legitimate interest is not automatic support for AI training, but requires a documented balancing test.

For customer service data, this means you need to look at what customers were told, which terms applied and how long the data has been kept. The guide Can you sell customer data? covers these questions in more detail. What we write is general information, not legal advice.

How it works

1. You describe the data in our form – system, number of cases, number of years and whether there are call recordings. No files needed.
2. We assess rights, personal data and demand, and get back to you within 48 hours on business days.
3. You approve a concrete offer. Nothing is shared until a non-disclosure agreement is in place.
4. Delivery and payment only take place after de-identification.

By default the license is non-exclusive and limited in scope, time and use, and you keep ownership. Read about what the data might be worth on what is your data worth.