Why AI Companies Buy Slack and Email Data

Two different claims

Labs describe training inside software. Reporters describe sales of workplace archives. Keep those claims apart when a buyer mentions both. Read Forbes for the April news, and Anthropic and OpenAI for what the labs say agents need. Your contract still has to name who receives the copy, and which systems it includes.

Ask three things before you treat a headline as a reason to sign. Which systems does the buyer want: messages, tickets, mail, files, or code? Is the buyer paying for a practice environment, a text corpus, or a named purchase of your archive? Who is named on the contract as the recipient? The posts can help you ask. They cannot answer for your company.

For the prices paid in these deals, read how much do AI companies pay for company data. Bring one question from those figures into the conversation: which record types are priced, and whether a Slack archive is priced like a crawl of the public web. The April reports say why the records are discussed. Your price, if you license, is a term in your own contract.

What reporters wrote in April 2026

The April 2026 reporting says archives were being sold. It is not, on its own, a list of which lab bought which company.

On April 16, 2026, Anna Tong at Forbes reported that defunct startups were being liquidated for their Slack archives, Jira tickets, and email threads, which the article calls operational exhaust. The article's subject line is old work Slacks and emails. Forbes names Slack archives, Jira tickets, and email threads, and it also names code and files kept in Google Drive. For a founder, that list is the inventory a buyer may ask to see. It is not a price, and it is not a roster of buyers. If your archive is mostly one of those systems, say so. Do not assume a buyer wants every system the article names.

On April 17, 2026, Fast Company reported that shuttered startups were selling old Slack chats and emails to AI companies, and it attributed that report to Forbes. A second newsroom, on the next day, pointing back to Forbes, corroborates the report. It does not add a separate measurement of the market. If you are weighing a license, count this as the same story told again, not as a second census of deals.

On April 17, 2026, Gizmodo also reported that failed companies were selling old Slack chats and email archives, and it attributed the account to Forbes. Gizmodo also wrote that a shutdown product lets companies license source code and workplace messages, including Slack and email. The account still does not name which lab bought which archive. If a product offers those records, the offer is the product's. Your buyer is whoever signs your contract.

On April 25, 2026, Futurism reported that data from a workplace such as Slack is valued because researchers use it to build a realistic sandbox where agents can practice work tasks. The article follows the Forbes report. It does not add a new count of deals. When a buyer says they want a sandbox, ask which tasks they expect the archive to show, and which of your systems hold those tasks.

What lab posts say about environments

The lab posts describe practice environments. The news posts describe archives. You can take both seriously and still ask the buyer which one they are paying for. An archive is not, by that pairing alone, the material inside the environment.

Anthropic's safety-research post Teaching Claude why says the lab trains Claude on constitutionally aligned documents, high-quality chat data, and a diverse set of environments. It also says that adding tool definitions to the training data improved results even when the tools were not used. The post is a safety-research note about alignment. It walks through training that teaches a model why an action fits a constitution, not only which action to copy. Nothing in that description says Anthropic bought a company's Slack archive. If a buyer cites the post as the reason they want your messages, ask for a contract line that says whether they want a practice environment, a text corpus, or both, and which of your systems they will actually receive.

OpenAI's October 23, 2026 post says models need computing environments where they can write and run code, search for information, and try different approaches. It says frontier reinforcement-learning runs use an offline cache of web content instead of the live internet. The post is about securing the environments used to train and evaluate frontier models. Inside the boundary, agents can have browsers, code execution, and computer-use tools. That is a contained workspace. It is not a statement that OpenAI bought a company's email. If a buyer says they need an environment like that, ask whether your mail would be an input to it, or whether they are describing infrastructure that is already theirs.

TechCrunch, dated September 21, 2025, and first run on September 16, 2025, reported that labs and vendors were building simulated software environments, called reinforcement learning environments, so agents can practice multi-step tasks. Labs were building those environments in-house and were also looking at vendors. An environment is a simulated workspace with a task and feedback. A static scrape of the public web is a pile of documents. The article is about the workspace, not the pile of documents. If you hold the archive, ask whether the buyer is building a workspace or collecting documents. The contract should say which.

None of those posts is a purchase order. A purchase you can rely on names who bought what. Use the posts to prepare questions: what the practice room contains, whether your records would go into it, and which clause says so.

Why workplace records are not web text

A web page is a document someone put on the public internet. A ticket thread is a record of work. When the April articles talk about agents, they are talking about the second kind.

Forbes says labs started on the public internet: Reddit threads, Wikipedia entries, and digitized books. The article then turns to operational leftovers from companies that closed. Slack jokes, Jira tickets, and email threads are the examples. Forbes says that kind of trail matters for agentic models because the model needs examples of what doing the work looks like. The article attributes a further claim to Ilya Sutskever: that the public internet text used for early training was exhausted by late 2024. Treat that as his remark about public internet text. A lab post about training environments is a different document.

Futurism makes the work-task point in its own words. Workplace data is valuable, the article says, because the industry is focused on agents that carry out work, and researchers want a realistic sandbox. A sandbox built from internal messages is a setting for practice. A Wikipedia page is a document a person chose to release. Slack, email, and tickets are the records those articles name, because the setting is the work.

The difference is the kind of record, not a claim that one lab has run out of text. A founder reading the news can hold two ideas at once. Pages on the open web were the early diet. Workplace traces are what the 2026 articles say agents now get trained against, because the traces show work as it happened. The lab posts describe the practice room. Whether your archive is what goes into that room is a question for the buyer, and the answer belongs in the contract.

What to use when you decide whether to license

Use the reporting to decide what to ask. It will not fill in your buyer or your price.

• A news report that archives were sold tells you the records were discussed as something a closing company could license. Your own license still has to name the buyer. If the contract does not name who receives the copy, you do not yet have that term.
• The figures for these deals are on the company data price index. Use those figures when you want a number. A note you write from the news alone should not add a dollar amount.
• Court filings for one airline sale are explained in Spirit Airlines data auction. An ordinary startup archive is not that docket, and those bids are not a quote for your Slack.
• Before you tell staff anything about a license, read employee privacy when companies sell data to AI. The commercial reason archives are discussed is a different conversation from the objection employees can raise.
• Read who buys company data for AI before you assume the lab in a headline is the buyer in your contract. Match the buyer in front of you to a practice environment, a text corpus, or a named purchase, and ask which one they are offering.

If you are closing and you want the process, read how to sell company data to AI companies. Eligibility is a separate check. A news report is not an offer, and it is not the list of records your own license would include.

Related

• Price index

  Public, sourced prices for company records and for publisher data licenses, in separate tables.

• Who buys company data

  Four kinds of AI data buyer, what each wants, and what a seller should check first.

• Spirit data auction

  What the Spirit bankruptcy filings say about the data auction, the bids, and whether a court approved them.

• Employee privacy when companies sell data to AI

  Critics, re-identification research, and what a careful program tells staff before any license.

• How to sell company data

  How to sell or license company data to AI companies, and which route fits.

Sources

1. Forbes. AI's New Training Data: Your Old Work Slacks And Emails.
2. Fast Company. Shuttered startups are selling old Slack chats and emails to AI companies.
3. Gizmodo. Failed Companies Are Selling Old Slack Chats and Email Archives to Train AI.
4. Futurism. Your Former Employer Is Selling Your Slacks and Emails to Train AI.
5. Anthropic. Teaching Claude why.
6. OpenAI. Securing our frontier model training environment.
7. TechCrunch. Silicon Valley bets big on environments to train AI agents.