Skip to content
Outrich Data

What data do AI labs buy?

Updated 2026-11-16 · 7 min read · By the Outrich Data team

AI labs buy records of real work that cannot be collected from the public web. Four kinds matter most: workflow traces that show a task done step by step, expert work product with its corrections, data linked across systems, and unpublished material held out for evaluation. Public, scraped and thin data does not sell.

Why labs buy company data at all

The large language models were first trained on public text: websites, books, forums and open code. That supply is close to used up. Labs have collected most of what is freely reachable, and what remains is increasingly restricted by publishers or mixed with machine-generated text.

Public text has a second limit. It shows finished results. An article shows the final version, not the drafts. A public code repository shows the merged commit, not the discussion that shaped it. Models trained on results learn to produce plausible results. To act as an assistant or an agent inside a business, a model has to learn the process: what a person looked at, what they tried, what was rejected and why.

That process is recorded inside operating companies. It sits in ticketing systems, chat tools, document stores, repositories and line-of-business software. It was never published, so it has to be licensed from the company that holds it. This is the supply that Outrich Data buys as a direct buyer.

Workflow traces

A workflow trace is a record of a task from start to finish. A support ticket with its full thread is one: the customer's problem, the agent's questions, the internal notes, the escalation and the resolution. A change request with its approvals is another. So is an insurance claim file or a purchase order moving through review.

Labs want traces because they are building agents, meaning systems that complete multi-step tasks across tools. An agent needs examples of sequences, not single answers. The most useful traces include the state at each step, the action taken and the outcome, including the cases that went wrong and were corrected.

Value rises with completeness. A trace that ends with a recorded outcome, such as resolved, approved or refunded, can be used to teach and to test. A trace that stops midway is worth less. Timestamps, role labels and links to the documents that were consulted all add to it.

Expert work product

Expert work product is material created by people with training that is costly to acquire: engineers, analysts, clinicians, underwriters, lawyers, accountants. Examples include design documents, code with review comments, financial models, technical reports and reviewed drafts.

The strongest signal is in the corrections. A document alone shows one expert's output. A document with tracked changes and reviewer comments shows a second expert's judgment about the first. Code review threads work the same way. Labs use this material to teach models the standards of a profession, which rarely appear in public text.

Professional data also carries the most restrictions. It often belongs to a client, is covered by confidentiality duties or falls under sector rules. Much of it can only be licensed after a rights review removes client-owned material and personal details.

Cross-system linkage

Most company data lives in separate systems. The sales record is in one tool, the support history in another, the contract in a third and the engineering fix in a fourth. Each export is useful alone. Linked together, they are worth more.

Linkage means the records share a key, such as an account or case identifier, so one customer problem can be followed from the first call through the ticket, the internal discussion and the code change that fixed it. Real work crosses tools, and models meant to do real work need examples that cross tools.

This is the hardest kind of data for a lab to assemble from outside, and one reason data from a single company with long history can be valued above a larger pile of disconnected records. Linkage also raises re-identification risk, since joined records reveal more. Keys are normally replaced with salted hashes that keep the join and hide the original value.

Held-out evaluation data

Labs do not only train models. They test them, and a test is only meaningful if the model has never seen the material. Public benchmarks lose value over time, because their contents leak into training sets. Labs therefore pay for data that has never been published and keep it out of training.

A company's internal records are well suited to this. If the systems were never open to the internet and the data was never licensed before, the lab can be reasonably confident it is unseen. Evaluation sets are usually smaller than training sets but need precise documentation: the task, the correct outcome and proof of the data's history.

The seller's side of this is disclosure. A buyer will ask whether any part of the dataset was ever public or shared, and the answer goes into the license as a representation.

The modalities at a glance

The four categories above cut across formats. The table lists the formats labs most often ask for and the use each serves.

ModalityTypical recordsWhy a lab wants it
Support and service textTickets, chat transcripts, resolutionsMulti-turn problem solving with known outcomes
VoiceRecorded calls with transcriptsSpeech models; natural conversation under real conditions
Internal documentsSpecifications, reports, procedures, reviewed draftsDomain knowledge and professional standards
CodeRepositories with history, reviews and issuesHow software is changed, reviewed and fixed over time
Workplace communicationChannel threads and work emailCoordination, decisions and context around tasks
Workflow and system logsApprovals, state changes, tool actionsAgent training on step-by-step task completion
Structured business recordsClaims, orders, cases with outcomesDecision making on real cases; evaluation sets

What does not sell

A large share of what companies first offer has little value to a lab. Knowing this early saves effort.

  • Public content: website pages, published marketing and anything a crawler could already reach
  • Scraped data: material collected from other sites, which carries rights problems and is not unique
  • Purchased lists: contact databases and third-party data the company does not own
  • Thin records: a few months of history, a handful of users or tables with most fields empty
  • Templated output: automated notifications, canned replies and system-generated text counted as conversations
  • Machine-generated text presented as human work
  • Data with no clear right to license, however rich it is
  • Raw personal data that has not been de-identified

What public prices show

Published figures are scarce, and they describe specific cases, not a market rate. According to MIT Technology Review, AI labs buying data from closed startups have paid about $25,000 for a company code base and about $15,000 for workspace data such as chat and documents. Those are small companies selling once.

At the other end, Neudata reports that Reddit's arrangement with Google has a total contract value of about $203M for forum posts and comments. That is a large public archive with continuing access. Operating company data sits between these, and its price depends on type, depth, exclusivity and how clean the rights are. No public figure should be read as a prediction for any one dataset.

Judging your own data

A quick test is to ask four questions of any dataset. It should record a process, not only a result. It should have been created by people doing real work. It should be something that was never public. And the company should be able to show it has the right to license it.

Data that meets all four is worth a closer look. Depth of history and the number of people who contributed matter next, which is why buyers set size thresholds. Outrich Data works with companies that have 50 or more employees and at least two years of records.

Common questions

Do AI labs buy customer lists or contact data?
No. Contact lists teach a model nothing about how work is done, and they are personal data. Labs want records of tasks, decisions and expert output with personal details removed.
Is older data still valuable?
Often, yes. Depth of history is a value driver because it shows more cases and more variety. Very old records in obsolete formats or about discontinued products are worth less, but several years of continuous history is a strength.
Do labs want raw or cleaned data?
They want de-identified data in a consistent format, with the original structure and context kept. Over-cleaning that strips timestamps, roles or thread order removes the signal that made the data useful.
Why is unpublished data worth more?
Because a model has not seen it. Unseen data adds new information in training and gives a fair test in evaluation. Data that was ever public may already be in a model's training set.

Find out if your data qualifies.

Eleven questions, about five minutes, no call required. You get an answer on screen and a written follow-up.

esc