The economics of company data for AI

Company data for AI does not behave like most goods. It already exists, costs the seller very little to license, cannot be manufactured on demand, and is worth very different amounts depending on how its parts connect. Understanding those economics helps both sides set sensible expectations.

This is a general explanation of how the market works, not financial or tax advice.

For the company that licenses its data, the inputs were paid for long ago. The salaries, the tool subscriptions and the years of work that produced the archive were spent running the business. The data is a by-product.

That makes the marginal cost of licensing low. The seller's time is about an hour across an intro call, a read-only evaluation and a review of an anonymized sample. Evaluation, anonymization and delivery are handled on the licensing side. The main costs to the seller are attention and the judgment calls about scope.

There is also a cost of doing nothing, and it is easy to overlook. Operational data decays as an asset. Admin access disappears when people leave, subscriptions get cancelled, and workspaces are deleted when a company closes. A dataset that could have supported a licence one year may be gone the next. Timing matters explores that in more depth.

AI labs need examples of real work to train and evaluate agents. They can generate synthetic data, and they do, but synthetic work tends to be cleaner and more predictable than the real thing. Synthetic data vs. real data covers that trade-off.

Real operational data has a supply problem. Each company's history is unique, it was produced once, and nobody can create more of it on request. The most useful version, where the same project can be followed across chat, tickets, code and email for several years, is rarer still. Most of it sits inside companies that have never thought of it as an asset.

Evaluation adds another source of value. A dataset that has never been on the public web is useful precisely because models have not seen it. That value depends on the data staying private, which is one reason licence terms restrict use to training and evaluation and forbid resale for marketing or people search.

In many data markets, price scales with volume. Here, volume matters less than structure. The main drivers are:

• Connection depth: whether work can be traced across several tools
• Years of history: how much of the arc of each project is captured
• Headcount: how many roles and handoffs appear
• Number of systems: how many angles exist on the same work

The reason is what the buyer is trying to teach. An agent that has to take a support ticket to a fix needs to see the ticket, the discussion, the code change and the follow-up, linked together. A large volume of email with no surrounding context does not teach that, however many messages it contains.

This is why typical licences range so widely, from about $20K to $400K and up, and why no firm number is possible before the data is counted. Two companies of the same size can land far apart if one ran its work through connected tools and the other did not. How company data is valued goes through the factors one by one.

Between a company with useful data and a lab that needs it, there is real work, and that work shapes the economics of the whole market. Someone has to find the right companies, reach the person who can sign, agree scope, evaluate the data, anonymize it carefully and deliver it in a usable form.

The hardest part is often the first one. The companies with the most useful data are rarely looking for a buyer, and the people who can sign are busy founders, CFOs, directors or liquidators. That is why sourcing is a cost in its own right, and why we approach it through outbound: 17 years of B2B outbound, more than 200 onshore sales reps across the US, Canada, the EU and LATAM, and cold calling, email and LinkedIn aimed at the people who make the decision.

Anonymization is the other significant cost. Consistent tokens for every person, company and account across every tool, removal of credentials, keys, card numbers and government IDs, blacking out faces, logos and signatures, and default exclusions for HR, payroll, medical and privileged folders all take careful work. Done poorly, anonymization either destroys the connections that make the data valuable or leaves risk in place. Done well, it preserves the structure and removes the identities.

For sellers, the economics favor acting while the data is intact. The cost is small, the asset is unique, and payment goes to the company entity. For a company that is winding down, a licence can contribute to a recovery for creditors and shareholders. For an operating company, it can mean licensing a scoped slice of historical records without touching day-to-day work.

For buyers, the economics favor specificity. Paying for connected, multi-year, multi-tool data from the right sectors goes further than paying for volume. A clear brief that names sector, size, tool stack and years of history also makes sourcing more efficient, because effort goes into the companies most likely to fit.

For both sides, the market works best when expectations are set by what the data actually contains rather than by how large it is. If you want to see where your own company sits in this picture, book an intro call.