Who buys company data for AI training: the 2026 buyer index

How to read the index

Each row below is one company that says, on its own website, that it buys, brokers or lists internal operating data from businesses: team chat, email, support tickets, code, CRM records and documents. We record what the company publishes, link the page we read and date the check. Every row was last checked on October 18, 2026. We do not grade offers we have not seen, and we do not estimate what your data is worth. The columns, types and tiers are explained under the table.

Columns, types and tiers

The columns

The four types

• AI data company. Licenses data from businesses and supplies it to AI labs, usually after removing names and identifiers. It is the counterparty on the contract you sign.
• Broker or intermediary. Finds a buyer for your data and manages the deal. The company that ends up with your data may not be named.
• Marketplace. You list a description of your data and buyers license it through the platform. The marketplace takes a commission or charges buyers for access.
• Shutdown service. Helps companies wind down and sells or licenses what is left, including code and workspace data.

Some companies describe more than one role. Anthroly, for example, calls itself “Data lab · Buyer · Broker”. When that happens we use the role the company states for the contract you would sign.

Tiers

“Established” and “new” describe how much public track record a company’s seller program has, not how good its offers are. “New” is not a warning in itself. Each row gives its reason in a few words, such as the year the domain was registered or whether a legal entity is named. The full test is on our methodology page.

Every range is self-reported

Every figure in the Published range column is published by the company. We have not seen the contracts or bank records behind them. Several come from calculators that the companies themselves call non-binding. Mode says its calculator “is illustrative, not a guaranteed offer”, and Nyne says of its estimate: “This figure is indicative and not an offer.” Nyne’s estimator also shows a different range each time the page loads, even for the same inputs, so we do not list its output as a range. Handshake AI’s estimator, for the same default company of 100 employees and 10 years, shows bands from “up to $100,000” to “$684,000–$2,174,000+” depending on the country selected.

“Not published” means we looked for the figure on the company’s own pages on the check date and did not find it. It does not mean the company refuses to say. Ask.

Payout models compared

Payout models in the index fall into three groups, plus mixes of them. The headline number matters less than three other things: when the money arrives, what you give up to get it, and what has to happen between signing and payment.

One-time license

You agree a price for a defined set of records, deliver it, and get paid once. Mode states: “Mode buys the agreed copy directly and charges no broker fee. Your written agreement sets the amount and payment schedule before export.” Mercor says payment “is issued via wire transfer upon delivery of the anonymized dataset to the buyer” and that Mercor covers extraction, anonymization and transfer, with “no upfront costs, setup fees, or deductions from the payment.” SourceX “usually pays a one-time fee for the dataset in the agreement” and adds: “There are no royalties or usage fees.” Polyshares states that “Typical deals run $100K to $2M and above.”

The fixed sum is fixed only for what the buyer accepts. Nyne states: “Once your dataset is delivered and accepted, you receive payment as outlined in your partnership agreement.” SourceX is explicit about the downside: “Unselected submissions earn no payment,” and “Partial selection requires a mutually agreed reduced fee.” Read the acceptance clause before you read the price.

Revenue share

An intermediary licenses your data to several buyers and pays you a share of each sale. Troveo describes its model as “paying you on every sale, not just the first”, and states that “ownership never transfers”. The page we read does not give the percentage split.

Maryah offers both models and describes the trade-off in its own words. A one-time payment is “Usually a lower total than the potential from revenue share.” A revenue share has “Potential to earn more, but no guaranteed amount or timing.”

If you are offered a share, get these in writing: the percentage, what counts as revenue, who sets the price of each resale, how often you receive statements, whether you can audit them, and how long the share lasts.

Hybrid models

Some programs combine an upfront fee with later payments. Labelbox Prism states: “You receive an upfront licensing fee at signing plus a revenue share on downstream use.” SuperAnnotate pays “on delivery, and on each relicense.” Scale AI answers its own question “Is this revenue recurring?” with “Yes, Scale designs its partnerships so there’s opportunity for recurring revenue.” Labelbox Prism and SuperAnnotate appear as “mixed” in the index. Scale’s row stays “Not published” because it does not say how the recurring part is paid.

Marketplaces and listing boards

On a marketplace or listing board you describe what you hold, and the platform puts it in front of buyers. Who pays the platform varies, and several do not publish a percentage:

• Aribase: “Free until you license. Aribase is paid out of a closed license.” No percentage on the page we read.
• SimpleClosure AssetHub: “There’s no upfront cost to use AssetHub. SimpleClosure only takes a percentage if your assets are successfully sold or licensed.” No percentage stated.
• Korra: “No fees to you. The buyer pays.”
• BuyMyDeadStartup: “No fee, no commission” for sellers, and it says no sale money passes through it. Buyers pay “$1,500 per month” for access to listings.

General dataset marketplaces outside this index publish rate cards that give a reference point. Opendatabay’s schedule runs from 30% on sales of $1 to $100 down to 5% on sales of $250,001 and above, and Datarade’s commission-only plan is “$0 / year + 30% commission”.

Exclusivity changes the price

Two offers for the same records are not comparable if one is exclusive and the other is not. SourceX states: “As standard, exclusivity is perpetual,” adds that if you push back “a 24-month or 12-month term can be agreed instead,” and says “non-exclusive licensing is available at a substantially lower price.” Labelbox Prism licenses are “exclusive by default.” pre.dev takes the opposite default for code: “Non-exclusive by default,” so a repository “can pay more than once.” Aribase puts the trade-off plainly: non-exclusive licenses “can go to several buyers, and each deal pays you again,” while “Exclusive terms command higher prices.”

Before comparing prices, line up exclusivity, term, territory and what happens after termination. Our offer checklist walks through each clause.

Who fits which buyer

Buyers publish entry criteria: headcount, years of history, country, language and company status. Check yourself against them before you book a call.

A match is not a recommendation. Check the tier and its reason in the index before you share anything.

Mode also lists lower minimums for regulated professions: “6+ for legal; 10+ for accounting, finance, and medical” (source). Read the next section before you act on that.

Three cautions apply to every row in this table.

1. A fit on paper is not an offer. Published criteria are a filter, not a promise. Maryah, for one, says it reviews each business before approaching buyers “and an offer is not guaranteed.”
2. Different intermediaries can lead to the same end buyer. Three offers may trace back to one AI lab or data company. Polyshares states that “we do not disclose which labs take which data”. Ask each party who will end up with the data before you treat their offers as competing.
3. In a wind-down, the company signs and the company is paid. Ooak Data states: “Payment goes to the entity, not to individuals.” Our shutdown guide covers the order of steps.

What disqualifies data

Some data should not be licensed at any price, and some needs work before it can be. Employee email or chat from staff based in the EU, the UK or Canada, and client data from law, accounting, tax or health firms, are the hardest cases. The rest is law and contract that apply to you no matter who you work with. This section explains what we flag. It is not legal advice.

Employee communications from staff in the EU, the UK or Canada

A Slack thread or an email chain holds the personal data of everyone in it, including people who never agreed to its reuse. In the EU, reusing personal data for a purpose other than the one it was collected for has to pass a compatibility test under Article 6(4) of the GDPR. The European Data Protection Board’s Opinion 28/2024 addresses when legitimate interest can support AI development and whether models trained on personal data can be treated as anonymous. The UK applies its own version of the GDPR, with specific guidance on employment data from the ICO. In Canada, private-sector privacy is split between federal and provincial law, as the Office of the Privacy Commissioner explains.

We have not found a regulator decision that clears licensing workplace messages for AI training in any of these places, so we leave them out. Code, product documentation and records that contain no personal data are a separate question, and we do route those.

Client data in law, accounting, tax and health firms

These firms hold client information under duties that do not end because the data is old:

• Tax. Under 26 U.S.C. § 7216, a tax preparer who “knowingly or recklessly” “discloses any information furnished to him for, or in connection with, the preparation of any such return” is guilty of a misdemeanor.
• Health. Under 45 CFR § 164.502, a covered entity or business associate “may not use or disclose protected health information, except as permitted or required by this subpart”.
• Law. ABA Model Rule 1.6(a) states that “A lawyer shall not reveal information relating to the representation of a client unless the client gives informed consent”, subject to narrow exceptions.

From these firms we route only the firm’s own internal operations data, never client files, client correspondence or return information.

Privileged material in any company

Legal advice also sits in ordinary inboxes, board threads and Slack channels. Sharing privileged communications with a third party can waive the privilege. Federal Rule of Evidence 502 limits the scope of a waiver only for disclosures made in a federal proceeding or to a federal agency, so it offers no protection for a data sale. Carve out legal, HR investigation and board material before anything is exported.

Data that is not yours to license

• Customer data under contract. If your customer agreements or data processing agreements limit how you may use their data, those limits apply to every support ticket and shared file that contains it.
• Shared channels. Channels shared with other organizations contain their messages, not only yours.
• Client-owned work. Code and deliverables built for clients under work-for-hire or assignment terms may belong to the client.
• Bankruptcy estates. Under 11 U.S.C. § 363(b)(1), if the debtor’s privacy policy prohibited transferring personally identifiable information, the trustee “may not sell or lease personally identifiable information” unless the sale is consistent with that policy or a court approves it after a consumer privacy ombudsman is appointed.

Platform terms: ask how the buyer ingests your data

How data leaves your systems matters as much as who receives it. Two widely used sources, Google Workspace and Slack, restrict how third-party apps may use data obtained through their APIs for AI training.

Google’s Workspace API User Data and Developer Policy, last updated September 3, 2026, lists among prohibited uses:

“Transferring, selling, or using user data to create, train, or improve a machine learning or artificial intelligence model beyond that specific user’s personalized model for the appropriate use case or user-facing feature.”

Google’s help page for apps under OAuth verification adds that the restriction covers “a foundational machine learning or artificial intelligence model” and that it “applies equally to the raw data obtained from the scopes AND data aggregated, anonymized, or derived from the scopes.” Anonymizing data obtained through these APIs does not take it outside the rule.

Slack’s API Terms of Service, effective October 10, 2025, say that a provider of an app offered for use outside its own organization may not “use API Data to train a large language model” or “bulk export Slack message and file data except where expressly allowed by an additional agreement”. For the Data Access and Real-Time Search APIs, a third-party app provider “may not create persistent copies, archives, indexes, or long-term data stores of other organizations’ API Data.”

What these terms cover

Both documents govern apps that reach your data through the platforms’ APIs. A company’s own admin export, such as Google’s Data Export tool for Workspace or Slack’s workspace export, is a different route. Whether licensing that export for AI training is permitted under your own customer agreement with Google or Slack is a question for your counsel. We have not found a public statement from either company that answers it.

Exports also have plan limits. On Slack’s Free and Pro plans, the standard export covers messages and file links from public channels. Private channels and direct messages require Business+ or Enterprise, and the owner must apply to Slack to use those exports. On the Free plan, file links cover only the last 90 days (Slack). If you are closing a Google Workspace account, our guide covers what to export before you cancel.

How buyers describe ingestion

Buyers describe their pipelines in different ways:

• Mercor: “Choose from 100+ supported tools and link your accounts.”
• Grepped: “Your team’s tools, connected read-only.”
• Labelbox Prism: “Read-only connectors to your existing stack”, and also “Most teams can export the needed operational data with existing tools.”
• Scale AI: “Scale anonymizes everything inside your environment, before any data is extracted.”

A connector alone does not tell you whether a platform term is affected. That depends on which API it calls, under which terms, and whether the buyer holds a separate agreement with the platform. None of that is on these pages, which is why the index has an Ingestion column. We are not saying that any listed buyer breaches these terms. We have not seen their agreements with Google or Slack.

We have not yet reviewed the API terms of Microsoft 365 and Teams, Atlassian, Zendesk, HubSpot or Salesforce for similar clauses. Until we have, treat each as an open question.

Questions to ask any buyer before you sign

1. Will our data reach you through our own admin export or through a connector you operate? Which systems go which way?
2. If a connector: whose credentials, which permission scopes, and do you have an agreement with that platform that allows AI training use?
3. Who receives the data after you, and do they get raw or de-identified records?
4. Is the license exclusive, for how long, and does it survive termination? SourceX, for example, states that its license “is permanent and survives termination of the agreement” once the fee is paid (source).
5. What happens to the raw copy after delivery? Ooak Data says it is “deleted within 30 days of delivery” (source).
6. Are we paid on signature, on delivery, on acceptance, or when your end buyer pays you?

If you want help putting these questions to several buyers at once, tell us what you have. We compare the buyers that fit, you choose who hears from you, and we never ask for data, credentials or exports. How we pick and check rows is on our methodology page, and who pays us, which today is no one, is on our disclosure page.