Who buys company data, and what they do with it

If you run a company, or are winding one down, the idea that someone would pay for your old Slack messages and Jira tickets can sound strange. It makes more sense once you see who the buyers are and what they are trying to build. The short answer: AI teams that need examples of how real work gets done, and cannot find them anywhere else.

Most demand for operational company data comes from two kinds of organization.

Frontier AI labs build general-purpose models and, increasingly, agents meant to carry out multi-step work: answering a customer, triaging a bug, preparing a report, moving a project forward. Public web text covers a great deal, but it shows very little of how work happens inside companies.

Enterprise AI teams build agents for specific functions or industries, such as support, finance operations or software engineering. They need data that resembles the environments their agents will be deployed into, and they often need it for evaluation as much as for training.

Both groups buy licences, not ownership. The company keeps its data; the buyer receives an anonymized copy and the right to use it for defined purposes. If the distinction is new to you, licensing vs. selling data covers it.

Buyers use licensed company data in four main ways.

• Agent training. Sequences of real work, such as a customer complaint traced from first email to shipped fix, become demonstrations that teach a model how to plan, use tools and hand off.
• RL environments. Anonymized histories seed sandboxes where agents practice tasks and get scored, so the sandbox starts from a realistic state rather than an invented one. We explain the concept in what an RL environment actually is.
• Post-training. After a base model is trained, smaller, carefully chosen datasets shape how it behaves on specific kinds of work, such as writing a clear status update or reviewing a pull request.
• Evaluations. Real work makes realistic tests. Because private company data has never been published, evaluations built from it measure skill rather than memory. See evaluations that look like real work.

In each case, the buyer is interested in patterns: how a team coordinated, how an issue moved between tools, how a decision was made and carried out. It is not interested in who the people were. That is why every person, company, address, email, phone and account in the data is replaced with a consistent token before delivery. PERSON_014 is the same person in chat, email and the ticket tracker, so the pattern of the work survives while the identity does not.

This part matters as much as the first. A properly drafted data licence limits use to training and evaluation. Under the licences Martal Group arranges, buyers may never:

• market to, profile or contact anyone who appears in the data;
• attempt to re-identify any person or company;
• resell the data for marketing or people search.

Those limits sit alongside the anonymization itself. Credentials, API keys, card numbers and government IDs are removed outright. Faces, logos and signatures in images and PDFs are blacked out. HR, payroll, medical and legally privileged folders are excluded by default, and direct messages are excluded unless the company chooses to include them and has the right to. No third party touches the raw data, and the raw archive is deleted within 30 days of sign-off as a contract obligation, with a deletion certificate. More detail on contract terms is in what a training-only licence should include.

Buyers do not value company data by the gigabyte. They care about how well it captures real work, which comes down to a few factors.

The first is connection depth: whether the same project can be followed across several tools. A dataset where a ticket links to a pull request, which links to a chat thread, which links to a customer email, is far more useful than the same volume of messages from one tool.

The second is years of history. Longer records contain more complete projects, more seasonal patterns and more variety.

The third is headcount. More people means more roles, more hand-offs and more realistic coordination.

The fourth is the number of systems. Chat, email, docs, tickets, code, CRM, support and finance workflows each capture a different side of the business.

Sector and tool stack matter too, because buyers try to build a balanced mix rather than many copies of the same kind of company. How company data is valued goes further into how these factors interact.

For a seller, the practical picture is simpler than it sounds. You decide the scope and the exclusions, and those are annexed to the contract before any data moves. A read-only evaluation takes under an hour of your time. The data is anonymized in house, you review an anonymized sample before anything is delivered, and payment goes to the company entity, typically within two to four weeks. In total, it is about an hour of your time.

Licences typically range from $20K to more than $400K, depending on the factors above. Nobody can give you a firm number before the data is counted, but the evaluation tells you what you have.

The buyers are real, their uses are specific and the limits on them are written into the contract. If you are curious what your company's history would be worth to them, book a 15-minute intro call.