How do companies license their internal data to AI labs?

Seven steps from first call to payment. Read every term before step 4; after you sign, the contract decides.

The seven steps at a glance

1. Check that you qualify and own the records. Size and history rules, and whether you may license what the records contain.
2. Describe the data, without sending it. Which systems, how many years, how many people.
3. Get a price after the buyer reviews it. A figure for your records, made after a sample, an audit or a scoping pass.
4. Sign before the full copy moves. Price, scope, exclusivity, who de-identifies and what triggers payment.
5. Deliver the copy. Your own export, the buyer’s read-only connector or its API, with personal information removed.
6. Pass the buyer’s acceptance review. At several buyers, payment waits for acceptance.
7. Get paid and close out. Payment on the agreed trigger, then deletion of the raw copy, any refresh and the end of access.

Who you sign with matters at every step. The buyers in the index are mostly companies that supply AI labs rather than labs themselves: 27 AI data companies and 17 brokers or intermediaries among 51 listings. Broker or direct buyer explains how each kind gets paid.

1. Check that you qualify and own the records

15 of the 51 buyers publish a minimum company size, and the rules vary. Handshake AI wants “20+ full time employees & 3+ years in operation”. Mode asks for “6+ for legal; 10+ for accounting, finance, and medical; 30+ for all other businesses or nonprofits.” SuperAnnotate takes “Any size, any industry, anywhere” as long as you have “records built up over time and the rights to license them.” The size guide lists every published minimum.

You can license only what the company owns and may lawfully hand over. Employee email and chat from EU, UK or Canada-based staff, and client records from law, accounting, tax or health firms, need extra care, usually consent or anonymization. Customer contracts, privacy notices and vendor licenses can rule records out. If the company is closing, the board, investors or a trustee may have to approve the deal (who decides). Have counsel review the rights before you sign anything.

2. Describe the data, without sending it

The first contact is about the company, not its files. Maryah asks for “Your team size, sector, tools and record history. No files or system connections.” Polyshares: “Intake asks for record counts, not files or exports.”

Send a description: which systems you used, for how many years, how many people worked in them, and what you will not license. Keep credentials, raw exports and files with personal information back until a contract says who receives them. Our free value check asks the same questions and returns the buyers that fit; we never ask for data.

3. Get a price after the buyer reviews it

A figure on a buyer’s website is what it says it may pay. A price for your company comes after the buyer has looked at what you have. Ooak Data: “We never quote a number before counting. The estimator gives a range; the audit gives a figure.” Anthroly: “Every record is priced individually, after we have read a de-identified sample.” Aribase says “the final price is set in the written license agreement”, and Grepped’s terms say “Estimates on the site are indications, not offers.”

Ask for the figure in writing, with the scope it covers. What company data is worth sets the published ranges beside the deals that have been reported.

4. Sign before the full copy moves

Several buyers put the written agreement before the export. Appen: “Terms are agreed in writing before extraction begins”. Mode: “Your written agreement sets the amount and payment schedule before export.” Polyshares: “Nothing leaves your systems until Polyshares has made an offer and you have accepted it.” A de-identified sample can move earlier: Anthroly prices after reading one, and at Troveo “Buyers evaluate your data anonymously and are confidentially introduced only after an agreement is signed.” Ask any buyer whose pages do not say so which comes first.

Read four terms before you sign:

• Scope. Which systems, which years, what is excluded.
• Right granted. A license or a sale. Appen says “You keep ownership”, while FileYield also offers a “one-time sale” and Centific offers companies winding down “a dedicated asset-purchase agreement”.
• Exclusivity. SourceX licenses are “exclusive for AI training”, Labelbox Prism’s are “exclusive by default”, and pre.dev is “Non-exclusive by default”.
• Payment trigger. Signature, delivery or acceptance (step 6).

Until you sign, the buyers that address it let you walk away. Labelbox Prism: “No commitment required. Scoping is free and non-binding”. SuperAnnotate lets you approve each buyer of your data, “and can block competitors permanently.” Some commitments come before any license. AI Training Data Broker: “When we find an interested buyer or AI lab for your data, we ask for a 3-month exclusive broker agreement”. FileYield’s seller agreement says you “may not circumvent FileYield’s fees by agreeing to a deal off-platform with a buyer you first met through FileYield.” Read a broker or platform agreement as closely as the license. The offer checklist covers every clause, and our free offer review reads an offer for you.

5. Deliver the copy

Data reaches a buyer in one of three ways: you export the files and hand over a copy, the buyer connects to your tools, or you transfer through its API. Among the 25 buyers in the index that say which:

In the buyers’ words:

• Your export. Troveo: “No OAuth into your systems. No platform to install. We guide you through the export and handle the anonymization, packaging, and every buyer conversation.”
• Their connector. Centific: “Read-only access across 188 supported enterprise applications. You choose which ones, and you set the exclusions.” Mercor: “Choose from 100+ supported tools and link your accounts.” Mode: “Your team reviews it and signs in to the approved systems; no passwords are handed over.”
• Their API. micro1: “You securely transfer your data through our APIs or other approved methods.”

Platform terms can decide the route. Google’s Workspace API policy bars apps from using user data from its restricted and sensitive scopes to train a machine learning or AI model beyond that user’s personalized model. Slack’s API terms bar the provider of an app offered outside its own organization from using API data to train a large language model or bulk exporting message and file data, except where an additional agreement allows it. A company’s own admin export is a different route, and whether licensing that export is allowed under your agreement with Google or Slack is a question for counsel. The index explains the platform terms. Our guides to exporting Slack and Google Workspace cover the export side.

Who removes personal information, in the buyers’ words:

Korra’s page states the limit: “Removing names does not make records anonymous”. Agree the standard and who carries the risk if it falls short.

6. Pass the buyer’s acceptance review

At Appen, Centific, micro1, Nyne, Polyshares and SourceX, payment follows the buyer’s review or acceptance rather than delivery. Centific: “Paid on acceptance rather than submission.” micro1 “has 30 days to review and approve” what you transfer. Nyne: “Once your dataset is delivered and accepted, you receive payment as outlined in your partnership agreement.” SourceX: “Data that does not reach Closing earns no payment.”

Before you sign, get in writing what the acceptance test is, how long the buyer has to run it, and what happens if part of the data is rejected. The offer checklist covers payment triggers.

7. Get paid and close out

When the seller is paid, in the buyers’ own words:

End to end, Mode “generally expects about three months from the first conversation through payment.” micro1’s pages give two triggers, “within 30 days of acceptance” in its process and “within 30 days of invoicing” in its answers, so confirm which one the contract uses.

Settle three things at close-out. First, a deadline for deleting the raw copy: Ooak Data commits to deletion “within 30 days of delivery” (retention and deletion clauses). Second, whether deliveries recur: Centific says “Most partners then move to a quarterly refresh”, while Turing offers “One payment. No retainer, no recurring commitment.” Third, the end of any access: Grepped says its read-only access “is revoked when the project ends.”

Where we fit

The free value check covers steps 1 and 2 with you: an estimated range, the buyers that fit and the terms to watch. No buyer hears from you until you approve it, and we never touch your data. A buyer may pay us a referral fee when a company we introduced signs, and any buyer that does is named before an introduction (how we make money).

How this guide was built

Every buyer quote comes from the buyer records behind the index and was re-read on the buyer’s live page on November 2, 2026; the platform lines come from Google’s and Slack’s own terms. The counts and the ingestion table are computed from the index’s ingestion field at build time, and a buyer that does not say how it takes in data counts as not published. We have not verified any buyer’s process, and the order of steps at a given buyer is set by its contract.