Who buys AI training data, and what each buyer wants.

MIT Technology Review, reporting on startups that sell their records when they close, lists the buyers as foundation AI labs, vendors that build simulated workspaces for AI training, and mature tech companies building their own models.¹ A fourth group sits in the middle: data companies that license records and resell them as training sets.

Foundation labs. The companies training the largest models. OpenAI said in 2023 that it wanted large-scale data that is not easily found online, such as long-form writing and conversations, and that it would remove sensitive or personal information where needed.³

Data and environment vendors. Companies such as Surge AI and Mercor sell training data and practice environments to the labs. Surge builds what it calls reinforcement-learning gyms, where AI agents practice real tasks, and looks for archives that include disagreements and tricky judgment calls.¹

Scale AI is another. Meta’s 2025 investment valued it at over $29 billion.²

Data brokers. Firms like Defined.ai license data from owners and sell it on. Reuters reported that Defined.ai sells to Google, Meta, Apple, Amazon and Microsoft, and that text sold for about $0.001 per word.⁴

Tech companies training their own models. Larger software companies now train models for their own products. They are likely to want records from their own market: support queues for a help-desk product, code for a developer tool.

The public web is mostly finished writing. Labs building agents for support, sales and engineering need the work that happens before a result is published: the argument in a thread, the ticket that got escalated, the pull request that was sent back.

Slack and Teams channels where a team reaches a decision. Email threads with vendors and customers. CRM history from Salesforce or HubSpot with deal outcomes. Support tickets from Zendesk or Intercom through to resolution. Sales and support call recordings. Code from GitHub with review threads. All six systems.

Buyers pay most for records with an outcome attached, because a model can be graded against it. A closed deal, a resolved ticket and a merged commit all count.

Most deals are private. Reuters found that the first image-library deals with large tech companies ran from $25 million to $50 million each, and that market researchers had not yet sized the market because so few deals are disclosed.⁴

Small sellers are in the market too. SimpleClosure, which helps closing startups sell their code and workspace data, told MIT Technology Review it expects to return over $10 million to startups by the end of 2026.¹

A typical license through Allied runs $100K–$2M, set by volume, years of history and depth. What sets the price.

A company can approach a data vendor, which resells the records at a price it sets, or work through a broker that negotiates for the seller. Allied is a broker: you license an anonymized copy of your records to an AI lab, and Allied arranges the deal. The lab pays Allied, so there is no cost to the seller.

The steps: answer a short questionnaire, take a discovery call, and receive a written offer naming the lab one to two weeks later. You choose the systems, channels and dates. Names, contact details and customer identities are removed, and you approve a sample before anything is delivered. Payment arrives by wire 30 to 60 days after the lab accepts delivery. The license is non-exclusive, so you can license the same records again. Standard terms.

The usual fit is 25 or more employees and more than a year of records.

Every buyer will ask whether you had the right to license the records. The FTC has warned AI companies that they must honor their privacy and confidentiality commitments to customers and must not use customer data for secret purposes, such as to train or update their models.⁵

The same logic applies to the seller. Customer contracts, employee policies and privacy notices set what you can license, and some need consent or changes first. Allied reviews them before making an offer.