Why AI agents are trained on Slack, email and support tickets

In recent years, AI has gone from writing text to doing work. Modern AI agents are expected to handle a support case from start to finish, pull together materials before a meeting or follow up on a quote. To learn that, they need training data that shows how work actually gets done. That is why AI labs have started buying workplace data – Slack, email and support tickets – to an increasing extent. This guide explains why, how that kind of training works and what the trend can mean for European companies.

From chatbot to AI agent

A classic chatbot receives a question and gives an answer. An AI agent does more than that. It can:

• break a task down into several steps,
• use tools such as search, calendars or ticketing systems,
• make decisions based on what it finds,
• follow up and adjust when something doesn't go as planned.

That calls for a different kind of ability from writing a good sentence. The agent needs to understand processes: in what order things are done, who should be involved, when a case should be escalated and how you know a problem is really solved.

How agents are trained

Large language models are first pre-trained on enormous amounts of text to learn language and general knowledge. To become a useful agent, further training stages are needed. Two concepts are central.

Demonstrations

The model is shown examples of how tasks are carried out by people – a customer service agent who reads a case, asks a follow-up question, checks an order and solves the problem. By studying many such sequences, the model learns which steps tend to lead to a good outcome.

Reinforcement learning

In reinforcement learning, often abbreviated RL, the model tries to solve tasks itself and is rewarded when it succeeds. For that to work, you need realistic training environments and clear criteria for what counts as a successful outcome. Real workflows from companies are an excellent basis for building such environments: the case, the actions and the outcome are already documented.

Why web text isn't enough

The open web consists largely of finished texts – articles, product pages, forums. It rarely shows how something came about. A blog post about solving a delivery problem says little about the ten emails, two calls and three internal chat messages it actually took.

On top of that, the largest models have already been trained on most of what is freely available. Further improvements increasingly come from data that isn't public, and most of that sits inside companies' own systems.

Why workplace data is so sought after

Slack and Teams

Chat channels capture collaboration in real time. You can see how a team divides work, asks each other questions, makes decisions and follows up. For an agent meant to work within a team, that is invaluable. Read more about Slack and Teams data.

Email

Email archives often have a long history and contain complete sequences: inquiry, negotiation, agreement, follow-up. They also show how people phrase things to customers, suppliers and colleagues in different situations.

Support tickets and chat logs

Ticketing systems are perhaps the clearest source of all. Every case has a problem, a series of actions and an outcome. That is close to the structure needed to train and evaluate agents. More on our page about customer service data.

What makes particular workplace data extra useful?

Not all archives are equally interesting. For agent training, these usually weigh most:

• Complete sequences – being able to follow a case or discussion from start to finish, not just isolated messages.
• Clear outcomes – it is evident whether the problem was solved, the deal went through or the decision was made.
• Long history – more years give more and more varied examples.
• Real reasoning – discussions where people weigh alternatives against each other, rather than automated notifications.
• Consistent structure – timestamps, threads and metadata that allow the sequences to be reconstructed.

So a support system where every case has a clear status and resolution can be more useful than a much larger email archive full of newsletters.

The US trend: closed startups sell their archives

In the US, several media outlets, including Fast Company, Gizmodo and Forbes, reported in 2025–26 on startups that shut down and sold their Slack and email archives to AI companies. For a company being wound down anyway, the archives have become an asset that can bring in one last source of revenue, instead of something that is simply deleted.

The reporting shows two things. First, that demand for workplace data is real. Second, that the market is still young, with varying practices around personal data, consent and agreements. What works in the US can't be copied straight across to Europe.

What does this mean for European companies?

Data in local languages is scarce

Nordic and other smaller European languages make up a small share of global training data. AI Sweden built GPT-SW3 on a Nordic corpus, and OpenEuroLLM has had to generate synthetic data because of the shortage. In Norway, the National Library showed that licensed books and newspapers improved Norwegian models. Real workplace communication in these languages is even rarer than published text.

Governments are paying attention too. In Sweden, for example, a national coordinator for Swedish language models has been appointed, with a remit that includes compensation for rights holders. The question of how local data should benefit AI development, and how those who own it should be paid, is high on the agenda.

But the rules are stricter

Workplace data is full of personal data. In the EU, among other things, the following apply:

• GDPR – the European Data Protection Board made clear in Opinion 28/2024 that legitimate interest is not an automatic basis for AI training, and data protection authorities stress that reuse must be compatible with the original purpose.
• Employee privacy – emails and chats from staff may be covered by workplace privacy rules, and depending on your country, employee representatives or a works council may need to be informed or consulted.
• Platform terms – Salesforce has reportedly restricted bulk export and LLM training through Slack's API terms. How you can export your own archives needs legal review.

That doesn't mean European workplace data is ruled out, but every dataset requires a careful assessment and almost always de-identification before anything can be licensed. This guide is general information, not legal advice.

The EU AI Act places requirements on buyers

Under Article 53 of the AI Act, providers of general-purpose AI models must respect copyright reservations and publish a summary of their training content. That means serious buyers look for data with clear provenance and documentation – which favors licensed, well-documented deals over gray markets.

How you can think about it now

You don't need to make any decisions today, but it can be wise to get an overview:

1. Inventory your sources – which chat, email and ticketing systems do you have, and how far back does the history go?
2. Think about archiving – are you planning to switch systems or delete old archives? Then it may be worth looking into the question before anything disappears.
3. Identify sensitive parts – is there health data, HR cases or anything else that should never be included?
4. Involve the right people – your data protection officer, a lawyer and, where employee communication is concerned, employee representatives.

With us, the first step is a free assessment. You describe the data in a short form without sending any files, and we get back to you within 48 hours on business days. You keep ownership, approve every deal, and the license is by default non-exclusive and limited. To learn more about what training data is and how it is used, there is an overview under AI training data.

Summary

AI agents are trained on Slack, email and support tickets because that is where real work shows – step by step, with decisions and outcomes. In the US, the trend has already led closed companies to sell their archives. For European companies the picture is twofold: data in local languages is in demand, but GDPR, employment rules and platform terms mean every deal has to be built carefully. Those who understand both sides are best prepared.