How companies license their internal data to AI labs.

Nobody sells the company’s records outright. The company signs a license that lets one buyer train and evaluate models on a copy of defined records, for a defined use, while it keeps ownership. micro1 states on its program page that companies retain ownership of their underlying data, and Appen says a partnership grants a defined licence for a defined use without transferring ownership.²¹

There are three ways to reach a buyer. A company can go to a lab’s own data-partnerships team. It can license to a data vendor or marketplace that resells or packages the records for labs. Or it can use an arranger that finds the lab and sets up a license the company signs directly. The steps after that are much the same: scope, rights check, offer, contract, export, identifier removal, sample approval, delivery and payment.

Allied Advantage Assets is an arranger. Allied arranges the license, and the company licenses directly to the AI lab. Allied never sees, reads or stores the records: the company’s own team exports them and removes identifiers in its own systems, and the copy goes straight to the lab.

Labs want records of how work gets done inside a company, which the public web does not contain. Mercor says frontier models are being trained to work inside companies and that public and synthetic data cannot replicate that.³ micro1 lists SOPs, ticket workflows, CRM data, engineering workflows, contract workflows and internal communications, from every industry.²

In practice that means Slack and Teams channels, email, CRM history in Salesforce or HubSpot, support tickets in Zendesk or Intercom, call recordings and code with its review history. Appen says the records labs ask it for most are decision and escalation records: the reopened ticket, the rejected approval, the deal that needed a second signature.¹

Size and history matter more than any one file. micro1 asks for 30+ full-time employees, records primarily in English, and says companies with several years of records across several systems are a stronger fit; it sets no fixed minimum number of files.²

For a direct license arranged by Allied, the usual fit is 30 or more office employees and more than a year of records, mostly in English. See the standard terms.

Most sellers leave some records out. Direct licenses arranged by Allied exclude HR, legal and board material by default, and companies whose core data is health records, payment card data or children’s data are not a fit. Client work held under an NDA is a common exclusion on any route. Records a customer gave you under a contract that limits their use are not yours to license until that contract allows it. Is it legal to sell company data? covers the rules.

Going direct to a lab. OpenAI announced a Data Partnerships program in 2023, seeking “large-scale” data sets that are not easily accessible online, particularly data that “expresses human intention,” such as long-form writing or conversations. It offered two tracks, an open-source data set and private data sets for training its own models, and said it would remove sensitive or personal information if necessary. TechCrunch noted the announcement mentioned no compensation for data owners.⁶

Labs also buy at auction. In Spirit Airlines’ bankruptcy, Google outbid the AI data company Mercor for the carrier’s records, including 100 million emails and 500 million Teams messages. The reported price: $10 million.⁷

Direct deals with labs tend to be large, rare and slow to start, and a lab rarely publishes a price or a process for an ordinary operating company.

Data vendors and marketplaces. These firms license your records, package them and supply them to labs. Defined.ai runs what it calls the world’s largest AI data marketplace, with more than 100 partners, and pays data partners licensing fees from marketplace sales.⁴ Reuters reported that Defined.ai sells to Google, Meta, Apple, Amazon and Microsoft.⁹ micro1, Mercor and Appen run enterprise programs aimed at operating companies.²³¹ Most of them connect to your systems or take a transfer and do the extraction and de-identification on their side.

Arrangers. These firms find the lab and set up a license, or help a company compare offers. Handshake AI’s data-partnerships page lays out three stages, from exploring the opportunity to agreeing scope and terms to delivery and payment, and lists “Which AI labs do you work with?” among its questions.⁵ Allied arranges a license the company signs directly with an AI lab, or gets several buyers bidding through competing bids.

The table below is built only from each firm’s own public page, read on November 3, 2026. The exception is OpenAI, whose page refused automated access, so its row cites TechCrunch’s report of the announcement. It is not a ranking. Allied’s row uses only the terms Allied publishes.

The main trade-off is who touches the raw records. With a vendor, an outside pipeline reads them and the vendor licenses the result onward. With an arranger such as Allied, your team does the export and identifier removal, and the copy goes from you to the lab. Neither is right for every company. A company with a busy IT team may prefer the vendor route. One whose counsel will not let raw records leave may prefer to do the work in-house.

Three people usually carry the work: an admin who runs the exports, someone who approves the scope and the sample, and counsel who reads the license. The time goes into the export and the identifier removal, not the calls.

The export depends on your plan. On Slack’s Free and Pro plans, owners and admins can export messages and file links from public channels only. Exporting private channels and direct messages needs a Business+ or Enterprise plan, and owners must apply to use those export types. Free workspaces can only export links to files from the last 90 days.¹⁰

Slack describes the full export as being for legal and regulatory compliance, so check its terms with counsel before using it for a license.¹⁰

Identifier removal happens before the copy leaves your systems. Names, email addresses, phone numbers, account numbers and customer names are replaced or removed, and roles, timestamps and the work itself stay, because that is what the buyer is paying for. The data handling page describes how this runs on a direct license. Staff notice is its own step; see employee consent for AI training.

The vendor route moves this work outside. Mercor connects to more than 100 supported tools and its platform extracts and de-identifies the data, and says no data leaves your systems until you approve it.³ Appen describes read-only pipelines scoped per engagement.¹ That saves your team hours. It also means a third party’s pipeline reads the raw records.

Published timelines vary by route. Appen says its partnerships team responds within two business days.¹ Mercor says payment typically comes within 2-4 weeks of the signed agreement, depending on extraction and processing time.³ micro1 allows up to 30 days to approve delivered data and pays within 30 days of the invoice, so about 60 days can pass between transfer and payment.²

Through Allied, a written offer naming the lab arrives one to two weeks after the discovery call. The lab pays by wire 30 to 60 days after it accepts delivery. The time between those two points depends mostly on how fast your own team runs the export and approves the sample.

Before any of that, budget time for the rights check. Counsel reading customer contracts and the old privacy notice is usually the slowest step and the one no buyer can speed up.

Most published programs for operating companies pay a lump sum after the buyer accepts the data. Mercor pays by wire on delivery of the anonymized dataset.³ micro1 pays in full within 30 days of acceptance.² Defined.ai’s partners earn licensing fees as the marketplace sells their data, so income arrives over time rather than at once.⁴

Large platform deals often run for several years. Reddit disclosed that in January 2024 it entered data licensing arrangements with an aggregate contract value of $203.0 million and terms ranging from two to three years, delivering content through continuous API access and quarterly transfers of Reddit data.⁸

Archive deals have been sized in the tens of millions. Shutterstock’s CFO told Reuters its first deals with large tech companies ranged from $25 million to $50 million each. Defined.ai’s CEO told Reuters the market rate for text was $0.001 per word.⁹

Small deals are real too. MIT Technology Review reported a founder winding down a crypto startup netted roughly $25,000 for the code base, with about $15,000 more pending for workspace data in Google, Notion, Figma and Slack.⁷

Published ranges for operating companies run higher. micro1 says payments typically range from $150K to over $2M, and Handshake AI advertises $100K to $4M+.²⁵ A typical direct license arranged by Allied runs $100K–⁠$2M, set by volume, years of history and depth. What sets the price and the sourced list of public deals go further.

Licensing income is business income. Ask your accountant how it will be reported before you sign.