Skip to content
AI Data Payout

Guides

How AI labs use business emails and Slack to train models, explained simply

AI labs do not read your Slack the way a person would. First the records are cleaned of names and personal details. Then they are turned into practice tasks, like answering a supplier or fixing a ticket. An AI agent tries each task in a pretend workspace, gets graded, and slowly learns how real work is done.1,2,3,4

By Lead analyst (to be named) · reviewed by Reviewing editor (to be named) · published 14 Oct 2026 · updated 14 Oct 2026

The six steps, from your inbox to a trained model

Buyers do not all work the same way, and few publish every step. But what micro1, Mercor, Surge AI and the court papers in the Spirit Airlines case describe fits one pattern.4,5,2,6

  1. Collect. The buyer gets a copy, either by linking to your tools, as Mercor does with 100+ apps, or by an export your team prepares, as in micro1's data manifest.5,7
  2. Clean. Names, emails, phone numbers and account details are removed or swapped for codes. In the Spirit case, an outside firm, Tonic.ai, de-identifies the messages and files a certificate with the court before each transfer.8,6
  3. Rewrite, sometimes. micro1 says some datasets get synthetic rewrites, meaning the text is reworded so it no longer ties directly to the original records.9
  4. Turn it into tasks. A thread where a customer complains and your manager fixes it becomes a task: here is the complaint, now handle it. Surge AI values an archive by how many useful tasks it could produce.2
  5. Practise and grade. The tasks go into a training environment, a pretend version of the software where an AI agent tries the task and is rewarded when it gets it right. TechCrunch's example is an agent buying socks in a fake web browser.3
  6. Test. Some records are kept back to check how well agents do on work they have not seen. micro1 sells a separate product, Cortex, for testing and improving AI agents.10

Why Slack and email in particular

Chat and email are where decisions happen in writing. A Slack channel shows who asked, who answered, what went wrong and how it got sorted. Surge AI says it wants archives covering disagreements and tricky judgment calls, because that is the hard part of real work.11

They are also common. SimpleClosure found Slack in 95% of the closing startups' workspaces it looked at. Across Mercor's partners, Slack shows up in 55% and Gmail in 50%. A buyer that can read those two formats can use records from almost any company.12,13

What stays in the model afterwards

The data itself can be deleted. micro1's terms say it deletes a seller's data within 120 days of final delivery. But the terms also say models already trained on it are not deleted. What the model learned stays.14

Can a model repeat what it learned word for word? Sometimes. Researchers have pulled hundreds of exact passages out of a language model's training data, including names, phone numbers and email addresses. That is why the cleaning step matters more than any promise made after the fact.15

Pros and cons

Pros

  • Cleaning happens before training: buyers remove or swap names, and some use outside firms.1,6
  • The use is narrow: teaching agents to handle work tasks, not publishing your messages.4,2
  • Some buyers rewrite text so it no longer ties directly to your records.9

Cons

  • What a model learns from your records is never deleted.14
  • Models can sometimes repeat training text word for word, including contact details.15
  • Cleaning is imperfect: 99.98% of Americans could be re-identified from 15 traits in one study.16
  • You often cannot see which AI lab ends up training on it.17

What it means for a business like yours

Picture your service desk Slack channel at a 50-person IT support company. A lab would never care who your client was. It cares that a printer stopped working, a tech asked three questions, tried two fixes and closed the ticket. That pattern is what gets turned into a task.2,18

So the questions to ask a buyer are about the cleaning step, not the training step. Who removes personal details, with what tool, do you see a sample first, and is there an outside check? In the Spirit case the court got a certificate before each transfer. You can ask for something similar.19,6

Should your business get in now?

Many businesses like yours are already licensing their records to these firms. micro1 has committed $1B over 12 months to buying company data. At least 18 brokers started advertising for business data between September and mid-October 2026. SimpleClosure has done over 150 data deals, and Google bid $10M for Spirit Airlines' archive, though the court has not ruled on that sale yet.20,21,22,23

Buyers also publish deals by industry. Polyshares, for example, lists a $110,000 license for a marketing agency. That is its own claim, like every payout a buyer or broker publishes about itself.24

Our view, not a promise: the money will not be around forever. AI labs are racing to buy now, budgets are concentrated in today's build-out, and as more businesses sell, prices tend to come down. There is no deadline, but waiting has a cost.20,25

It is not for everyone

Think twice if any of these sound like you:

  • Most of your records are client files you hold in confidence.26
  • Your team would feel their messages were sold out from under them.27
  • Your data is regulated or sensitive, which buyers ask about up front.28
  • Your archive is small: most buyers want 20-30+ staff and a few years of records.29,30,31,32
  • You cannot spare the weeks a deal takes.33,34,35
  • You are not comfortable losing control of how a licensed copy is used.14

Next step

Take the quiz to see which firms fit your business. Then talk to the brokerages: they can answer the questions unique to your business. Legal questions go to the licensing companies.

Take the quizWho the brokerages are

We feature companies that pay us a referral fee. Our scores and reviews don't change for it. How we earn.

FAQ

Questions, answered with sources

  • They clean the records, turn them into practice tasks, and let AI agents learn by trying those tasks in a simulated workspace and being graded.3,2,4

  • Buyers describe different set-ups. Mercor says its processing is automated, Polyshares uses an outside scrubbing firm, and micro1 and Goldmined say nothing is released without the owner's review or approval.5,1,19,36

  • A practice world that copies a piece of software, where an AI agent tries a task and is rewarded when it succeeds. RL stands for reinforcement learning.3

  • It can happen. Researchers extracted hundreds of exact passages, including names and email addresses, from one model's training data.15

  • The copy may be. micro1 says it deletes data within 120 days of final delivery, but models trained on it are not deleted.14

  • Buyers remove or swap them for codes, sometimes through an outside firm. In the Spirit Airlines case, Tonic.ai does it and certifies the result to the court.8,1,6

  • We do not answer legal questions. Ask the licensing company: their team reviews this with you before anything is signed.

Before you sign

Legal questions go to the licensing company

We do not answer legal questions. Ask the licensing company before you sign: their team reviews these with you.

Questions for this kind of business

  • Do Slack's or Google's terms limit how we export this data for a buyer?
  • How do CCPA, GDPR, UK GDPR and state privacy laws apply, and is de-identified data still personal information?
  • Do customer, vendor or platform contracts, NDAs or API terms restrict this?

Questions for any buyer

  • Is the license exclusive, and for how long?
  • Can you sublicense the data, and to whom?
  • Which AI labs will receive it, and will you tell us?
  • How long do you keep it, and how is it deleted?
  • What is your liability cap, and who indemnifies whom?
  • Will the contract ban re-identification of people in the data?
  • Who removes personal data, by what method, and do we approve it first?

How this page was made

Every figure comes from a published source listed below, each with the date we checked it. Buyer statements are the buyer's own claims and are marked so in the source list. Firm counts come from the US Census Bureau's Statistics of US Businesses. Anything marked as our estimate is ours, not a buyer quote.

Drafted with AI assistance from our research files, then checked against each source. Not yet reviewed by a person, so this page is not indexed by search engines. Last updated 14 Oct 2026.

We feature companies that pay us a referral fee. Our scores and reviews don't change for it. See how we earn and how we rank. This site does not give legal, tax or financial advice.

Sources (36)

  1. 1.Polyshares website. Accessed 14 Oct 2026. Company's own claim.
  2. 2.MIT Technology Review, published 13 Oct 2026. Accessed 14 Oct 2026.
  3. 3.TechCrunch, published 16 Sep 2025. Accessed 14 Oct 2026.
  4. 4.The Next Web, published 9 Oct 2026. Accessed 14 Oct 2026.
  5. 5.Mercor Data page. Accessed 14 Oct 2026. Company's own claim.
  6. 6.The Next Web, published 9 Oct 2026. Accessed 14 Oct 2026.
  7. 7.Summary of micro1's video "How to earn up to $2M training AI models with anonymized company data". Accessed 14 Oct 2026. Company's own claim.
  8. 8.Meta Ad Library, Nyne.ai. Accessed 14 Oct 2026. Company's own claim.
  9. 9.micro1 data partnerships page. Accessed 14 Oct 2026. Company's own claim.
  10. 10.micro1 website. Accessed 14 Oct 2026. Company's own claim.
  11. 11.MIT Technology Review, published 13 Oct 2026. Accessed 14 Oct 2026.
  12. 12.SimpleClosure, What Your Assets Are Worth, published 3 Sep 2026. Accessed 14 Oct 2026. Company's own claim.
  13. 13.LinkedIn Ad Library, Mercor. Accessed 14 Oct 2026. Company's own claim.
  14. 14.micro1 Data License Master Terms, published 7 Oct 2026. Accessed 14 Oct 2026. Company's own claim.
  15. 15.arXiv (Carlini et al.), published 14 Dec 2020. Accessed 14 Oct 2026.
  16. 16.Rocher, Hendrickx and de Montjoye, Nature Communications, published 23 Jul 2019. Accessed 14 Oct 2026.
  17. 17.MIT Technology Review, published 13 Oct 2026. Accessed 14 Oct 2026.
  18. 18.micro1 Enterprise Data Partnership page. Accessed 14 Oct 2026. Company's own claim.
  19. 19.beri.net, published 10 Oct 2026. Accessed 14 Oct 2026.
  20. 20.The Next Web, published 9 Oct 2026. Accessed 14 Oct 2026.
  21. 21.Meta Ad Library (our review). Accessed 14 Oct 2026.
  22. 22.Forbes, published 16 Apr 2026. Accessed 14 Oct 2026.
  23. 23.Bloomberg Law, published 17 Aug 2026. Accessed 14 Oct 2026.
  24. 24.Polyshares track record page. Accessed 14 Oct 2026. Company's own claim.
  25. 25.TechCrunch, published 16 Sep 2025. Accessed 14 Oct 2026.
  26. 26.Polyshares website. Accessed 14 Oct 2026. Company's own claim.
  27. 27.Slashdot discussion, published 18 Apr 2026. Accessed 14 Oct 2026.
  28. 28.micro1 data payout estimator. Accessed 14 Oct 2026. Company's own claim.
  29. 29.micro1 company referral page. Accessed 14 Oct 2026. Company's own claim.
  30. 30.Nyne sell-company-data page. Accessed 14 Oct 2026. Company's own claim.
  31. 31.Cognor Labs partners page. Accessed 14 Oct 2026. Company's own claim.
  32. 32.Return My Time Sell My Data page. Accessed 14 Oct 2026. Company's own claim.
  33. 33.RemoWork, published 29 Jun 2026. Accessed 14 Oct 2026.
  34. 34.Your Data Value website. Accessed 14 Oct 2026. Company's own claim.
  35. 35.Cognor Labs manufacturing page. Accessed 14 Oct 2026. Company's own claim.
  36. 36.Goldmined.ai website. Accessed 14 Oct 2026. Company's own claim.