What data AI companies buy, by type.
Workplace chat, email, CRM history, support tickets, call audio, code and internal wikis: why labs want each, what makes a set valuable, and the deals on the public record. Checked 3 November 2026. If your company has 30 or more office employees, Allied arranges a direct license with an AI lab and gets you a written offer.
The short answer
AI companies now buy the records of how work gets done inside a company. The public web holds finished articles and posts. It rarely holds the thread where a team argued over a fix, the ticket that got escalated, or the deal that stalled at legal review. Labs training models to act as agents need those steps.
Mercor, which supplies data to AI labs, puts it plainly on its company data page: frontier models are increasingly being trained to work inside companies, navigating real tools, workflows and decisions, and public and synthetic data cannot replicate that.4
The money behind this is large. On 9 October 2026 micro1, another lab supplier, committed to spend $1 billion over 12 months acquiring and licensing enterprise operational data, to build reinforcement learning environments where models and agents learn to complete business tasks.3
Much of the buying runs through suppliers that turn raw records and expert work into training sets. Tencent Technology, citing documents obtained by The Information, reported that Mercor had revenue of $614 million in the first half of 2026, about 91% of it from foundation model companies including OpenAI, Anthropic and Google DeepMind. The same report notes that post-training also needs demonstration answers, preference judgments and complex tasks, the human feedback labs use to tune a model.10
The older market, licensing published text, images and video from publishers and stock libraries, still runs. It is covered in section VIII. This page is about the data. For the companies doing the buying, see who buys AI training data. For every public deal with its date and price, see the AI data licensing deals list.
| Data type | Why labs want it | Public example | Source |
|---|---|---|---|
| Workplace chat and email (Slack, Teams, Gmail, Outlook) | Shows people coordinating, disagreeing and deciding across days; raw material for simulated workplaces where agents practice | Google’s winning $10 million bid for Spirit Airlines’ records, including 100 million emails and 500 million Teams messages1 | MIT Technology Review1 |
| CRM and sales records (Salesforce, HubSpot) | Deal stages, activity logs and outcomes train sales agents and give a grader a right answer | micro1 lists CRM data, sales processes and playbooks among the data it licenses, at $100K to over $3M per company2 | micro12 |
| Support tickets and chat transcripts (Zendesk, Intercom) | Question, reasoning and resolution in one record, the shape of a support agent’s task | micro1 lists ticket workflows, QA processes and customer operations2; no priced support-ticket deal is public | micro12 |
| Call recordings and speech | Voice agents need real conversations; audio paired with transcripts is scarcer than text | Reuters: Defined.ai, which sells to Google, Meta, Apple, Amazon and Microsoft, quoted $100 to $300 per hour of longer film7; no priced business-call deal is public | Reuters7 |
| Code repositories (GitHub, GitLab) | Commits, reviews and linked issues train coding agents on how software changes over time | Stack Overflow’s OverflowAPI partnership with OpenAI6; a closing startup netted about $25,000 for its code base1 | Stack Overflow6, MIT Technology Review1 |
| Documents and wikis (Confluence, Notion, Drive, Jira) | SOPs, specs and bug trackers show how a company is supposed to work, next to how it does | Mercor connects Confluence, Notion and Drive4; micro1 lists bug tracking, product planning and knowledge bases2 | Mercor4, micro12 |
| Publisher, image and video archives | Volume and breadth for pretraining, plus licensed content shown inside chatbots | Reddit’s data licensing arrangements with an aggregate contract value of $203.0 million5 | Reddit Form S-15 |
Workplace communications: Slack, Teams and email
Chat and email are where a company’s reasoning is written down. A channel shows who raised a problem, who pushed back, and what the team agreed to do. A long email thread shows a vendor dispute or an approval moving up the chain. That is the material labs use to teach models how work is coordinated between people, not just how a single answer is written.
The biggest public example is an airline. In August 2026, as part of Spirit Airlines’ bankruptcy, Google outbid the AI data company Mercor for the carrier’s records, including 100 million emails, 500 million Teams messages and 176,000 employee records. The price was $10 million.1
Buyers say depth matters more than tool count. Nick Heiner, who leads Surge AI’s work on the reinforcement-learning gyms where agents practice tasks, told MIT Technology Review that one company’s Slack history, if it captures conversations with variety and depth, can be worth more than another’s Slack, email and calendars combined.1
Mercor’s intake page starts with exactly these systems: Slack, Microsoft Teams, Gmail and Outlook, and it says compensation rises with data volume, the number of tools connected and the depth of the data.4
One limit applies to how a chat archive is pulled. Slack’s API terms say an application offered for use outside your organization may not use API Data to train a large language model, and may not bulk export Slack message and file data except where an additional agreement allows it. The export path has to be settled before a license is signed. Our Slack and Teams page and email page cover what is included and excluded.8
CRM and sales records
A CRM is a record of decisions with outcomes attached. Each opportunity has stages, dates, logged calls and notes, and it ends as won or lost. Outcomes matter to a buyer because a training environment needs a way to score whether an agent did the task well. A closed deal is a scored example.
micro1’s data partnerships page lists CRM data, sales processes, playbooks, customer lifecycle documentation and revenue operations among the operational data it licenses. It says payments typically range from $100K to over $3M, and that it looks for established companies with 30+ full-time employees and substantial historical operational data.2
Mercor names Salesforce among the tools it connects, describing the data as accounts, contacts, opportunities and reports.4
What makes a CRM worth more: years of history, activity logged against each record rather than a bare pipeline, and custom fields that show how the sales process actually runs. Customer and contact names come out before anything leaves. See selling CRM data.
Support tickets and chat transcripts
A resolved ticket holds a customer’s problem, the agent’s reasoning in internal notes, any escalation, and the fix. That is close to the exact task an AI support agent is built to do, which is why help desk exports from Zendesk, Intercom, Freshdesk and Salesforce Service Cloud are among the first records buyers ask about.
micro1 lists support documentation, ticket workflows, QA processes, customer operations and knowledge management as data it licenses, alongside CRM and engineering records.2
No support-ticket deal with a named buyer and price has been made public as of 3 November 2026. Satisfaction scores and resolution times help a buyer tell good answers from bad ones. Tickets that mention health, payment or government ID details are usually excluded. See selling support ticket data.
Call recordings and speech
Voice agents for sales and support need real conversations: interruptions, objections, silences and the way a call reaches an outcome. Recorded calls from Gong, Zoom or a dialer, with speaker-separated transcripts, are the closest match to that work.
Reuters reported that dedicated AI data firms were securing rights to podcasts, short-form video and interactions with digital assistants, and paying contract workers to produce voice samples. Defined.ai, which licenses data to Google, Meta, Apple, Amazon and Microsoft, quoted $2 to $4 per short-form video and $100 to $300 per hour of longer film, against a market rate for text of $0.001 per word.7
Those prices are for media, not business calls. No lab has published a deal for a company’s sales or support recordings. Two rules shape any call archive: every call must have been recorded with the notice the caller’s state requires, and spoken names and numbers come out of the audio as well as the transcript. See selling call recordings.
Code repositories
Coding agents are one of the main things labs are training, and public open-source code is already widely used for training. Private repositories add what public code rarely shows in full: why a change was made, how a reviewer pushed back, which issue it fixed and whether the tests passed.
The best-known licensed code source is Stack Overflow. In May 2024 it and OpenAI announced an OverflowAPI partnership to improve OpenAI’s models for developers. Stack Overflow said it holds more than 59 million questions and answers.6
Private code from ordinary companies sells too. A founder winding down a crypto startup told MIT Technology Review he netted roughly $25,000 for his code base. SimpleClosure, which arranged that sale, said 47% of its deals now include code or data sales.1
Turing, which supplies training data to AI labs, announced Project Lazarus on 25 March 2026. It is a way for founders and engineering leaders to share private codebases and the work around them, such as Jira tickets, product specs, architecture docs and support threads, for payouts on eligible data. For how a code sale works, see selling GitHub repositories for AI training.11
Your team should scan every commit for keys and credentials before export, and a company can keep any repository out of scope. See selling code repository data.
Documents, wikis and project tools
Confluence spaces, Notion pages, Google Drive folders and Jira or Linear boards show how a company is supposed to work: the runbook, the spec, the SOP, the bug as filed. Next to chat and tickets, they show how it actually works. Buyers building simulated workplaces want both.
micro1 says the most valuable datasets capture real operational expertise, decision-making processes and repeatable workflows. Its list runs from source code documentation, bug tracking, product planning and technical knowledge bases to SOPs, reconciliations, approvals and contract workflows.2
Mercor connects Confluence, Notion and Google Drive, and its industry list asks for things like contracts and case files from law firms, client decks and SOWs from professional services firms, and work orders and QA logs from manufacturers.4
The founder in MIT Technology Review’s report expected about $15,000 more for his startup’s workspace data in Google, Notion, Figma and Slack, still pending when the article ran.1
Allied’s six system pages do not have a separate wiki page. Documents are usually licensed alongside the chat, ticket or code archive they belong to.
Publisher, image and video licensing, for context
Before labs came looking for workplace records, they paid for published content. Reddit disclosed in its 2024 IPO filing that in January 2024 it entered data licensing arrangements with an aggregate contract value of $203.0 million and terms of two to three years, delivered through API access and quarterly data transfers.5
Stock libraries were early sellers. Shutterstock’s Data, Distribution, and Services revenues rose 256% to $137.3 million in 2023, driven primarily by growth in its data offering.12
Reuters reported that Shutterstock’s first deals with large tech companies ran from $25 million to $50 million each, and that Photobucket was discussing 5 cents to $1 per photo for its 13 billion photos and videos.7
These deals sell volume and breadth. A company’s internal records sell something else: the work in between published outputs. The full list of publisher, media and video deals is on the deals list.
What makes a dataset valuable
Years of history
Sunset, which sells the records of closing companies, told MIT Technology Review it values a client’s assets by head count, years in operation, volume of data, and how the team used its tools. micro1 says companies with several years of records across multiple business systems are generally a stronger fit.21
Real decisions and disagreement
Surge AI looks for archives that show a range of situations, from disagreements to tricky judgment calls. A channel where people argue and settle on a fix teaches more than an announcements feed.1
Outcomes attached
A closed-won deal, a resolved ticket and a merged pull request each end with a result a grader can check. Records without an outcome are harder to turn into training tasks.
Connected systems
A ticket that links to a pull request, which links to a Slack thread, shows how work moves across tools. Mercor says connecting more tools generally results in a higher payout.4
Volume and structure
micro1 prices on dataset size and volume, workflow complexity, data quality, domain expertise and uniqueness. It sets no fixed minimum number of files.2
Clean rights
Records the company can license without breaking a customer contract, a vendor’s terms or a privacy promise. A large archive with doubtful rights is worth less than a small one with clear rights. See what company data is worth.
What stays out, and how a license happens
Buyers ask for identifiers to come out first. micro1 says personal and identifying information is removed or transformed before data is shared with third parties. In the Spirit sale, the flight attendants’ union objected on privacy grounds, and MIT Technology Review reported that suppliers exclude HR records.12
The terms a company wrote matter too. The FTC has warned that a company may be acting unfairly or deceptively if it tries to change the rules by surreptitiously re-writing its privacy policies or terms of service to use consumer data for product development. Records gathered under old promises are licensed under those promises. Read whether it is legal to sell company data and employee consent for AI training data.9
With Allied, a company licenses directly to an AI lab. You answer a short questionnaire, take a discovery call and sign with Allied. Allied aims to have a written offer naming the lab within two weeks of signing. Your own team exports the systems you chose and removes names and identifiers inside your own systems, and the copy goes straight to the lab. Allied never sees it. Payment arrives by wire 30 to 60 days after the lab accepts delivery. Licenses are non-exclusive, so the same records can be licensed again. The licensing agreement guide covers the clauses.
Closing a company? The archive can still be licensed, but only while the accounts are paid up. See selling startup data when shutting down.
Questions
What data do AI companies buy most?
In 2026, the records of how work gets done inside companies: Slack, Teams and email, CRM history, support tickets, call recordings, code and internal documents. Lab suppliers such as micro1 and Mercor list these systems on their own intake pages. Published text, images and video are still licensed as well.
Why do AI labs want internal company data instead of public data?
Labs are training models to act as agents inside companies. The public web shows finished work, not the steps, disagreements and decisions in between. Internal records show those steps, and many end with an outcome a grader can check.
Which kind of company data is worth the most?
There is no public price list by type. Buyers say value comes from years of history, depth of real decisions, outcomes attached to records, and how many systems connect. A deep Slack archive can be worth more than shallow chat, email and calendars together.
Do AI companies buy data from small businesses?
Some do. Several buyers set a minimum headcount, and closing startups have sold code and workspace data through shutdown firms. Allied works with companies that have 30 or more office employees.
Does selling company data to AI companies mean giving up ownership?
Usually not. Most deals are licenses, where the company keeps its records and grants the buyer the right to train on a copy. Allied’s licenses are non-exclusive, so the same records can be licensed again.
What data will AI companies not buy?
Buyers generally exclude HR files, health records, payment card data, consumer personal data, privileged legal material and anything the company has no right to license under its contracts or privacy policy.
Sources
- MIT Technology Review, “AI labs are buying data from dead startups to train their models,” 13 October 2026
- micro1, “Monetize your company’s operational data” (data partnerships page), read 3 November 2026
- micro1, “micro1 announces commitment to spend $1B on enterprise data acquisition over the next 12 months,” 9 October 2026
- Mercor, “Monetize your data” (company data page), read 3 November 2026
- Reddit, Inc., Form S-1 registration statement, SEC EDGAR, 2024
- Stack Overflow, “Stack Overflow and OpenAI Partner to Strengthen the World’s Most Popular Large Language Models,” 6 May 2024
- Reuters, “Inside Big Tech’s underground race to buy AI training data,” April 2024, syndicated on thestar.com.my
- Slack, API Terms of Service
- Federal Trade Commission, Technology Blog, “AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive,” 13 February 2024
- 36Kr, from Tencent Technology, “AI Companies: Are They Raking in Sky-High Priced Training Data?” 26 October 2026
- Turing, “AGI Advance: Weekly AI & AGI Insights (Mar 24, 2026),” published 25 March 2026
- Shutterstock, Inc., Form 10-K for fiscal year 2023, SEC EDGAR
This page is general information, not legal advice. Whether your company can license a given set of records depends on your contracts, policies and the law where your employees and customers are. Ask your own counsel.
Find out what your records are worth.
A few questions. About two minutes. How Allied works.