How much is company data worth to AI companies?

Why there’s no price list

Company data isn’t sold like a commodity. Each archive is different, and each buyer wants it for something specific. Many buyers today use company records to build training environments where AI agents practice real work, like answering tickets or fixing bugs (Gizmodo). What a buyer will pay depends on how easily your records turn into that kind of practice, and what it costs them to get there.

Buyers describe it the same way. Brendan Mahony, CEO of Sunset, which helps startups shut down and also buys their data, told Forbes that price depends on a company’s size, its age and what he called data richness: how well records trace and link across tools. A Jira ticket tied to a specific code commit is worth more than a standalone document, he said (Forbes).

So the honest answer to “how much is my data worth” is a range, narrowed by the factors below. The reported deals show where the ranges sit.

What drives the price

1. Volume

More records give a buyer more to train on. The Spirit Airlines archive that Google won at a bankruptcy auction held roughly 100 million emails, 500 million Microsoft Teams messages and about 30 million lines of code (AI Weekly). Volume alone doesn’t set the price, though. A smaller archive of linked, useful records can beat a bigger pile of noise.

2. Years of history

Long, continuous history shows how work changes over time. Spirit’s corporate records went back to 1986 and its transaction records to 2008 (AI Weekly). Cielo24’s founder sold 13 years of accumulated Slack, Jira and email (Mezha). Gaps in the record lower the value, because a buyer can’t follow a case from start to finish.

3. Specialization

Generic office chatter is easy to find or to generate. Records from a specialized industry, with its own rules, tools and edge cases, are not. An airline’s operations data, for example, drew three bidders (Fortune). The same logic applies to any specialized field, though regulated ones also carry more sensitive data to remove.

4. Outcomes and labels

Buyers value records that show what someone did and whether it worked: a ticket with its resolution, a code change with its review, a deal with its result. That turns raw data into a task an AI can be graded on. Finished training tasks built from real work sell for about $200 to $2,000 each, and complex ones occasionally reach $20,000 (Epoch AI). Your archive is the raw material for those tasks, so the closer it is to task-ready, the more it’s worth. Mercor, a data vendor that supplies AI labs, sees a large market for that kind of work:

5. Cleanliness

Every hour a buyer spends removing personal data, fixing broken exports or untangling duplicates comes off the price. Well-organized records with consistent fields, intact links between systems and identities already removed are cheaper to use, so buyers pay more for them.

6. Exclusivity

A buyer pays more to be the only one with your data. That premium has to beat what you could earn by licensing the same data to several buyers. Limiting exclusivity to a set time or a narrow use can be a middle ground.

7. Buyer demand

Demand shifts with what models are weak at that quarter. Competition moves the number too. At the Spirit auction, Mercor bid $7.5 million and Google won with $10 million (AI Weekly). Then micro1 made a late offer of $12.5 million (Fortune). The top offer rose by $5 million as more buyers competed for the same data.

Reported deal ranges by size of company

Few deals are public, and most of those come from companies closing down. Treat these as reference points, not quotes.

Platform and publisher deals are a different market. They license huge volumes of public content, often renewed every year. Company archives are private records of how work gets done, priced one company at a time. For an operating company, licenses typically land between $20K and $5M.

How buyers price data

Buyers use a handful of pricing models. Which one you’re offered depends on the data and the buyer.

• Flat license. One price for a defined scope: these systems, these years, this use. This is how the company archive deals above were reported, as a single amount per company (Gizmodo).
• Annual license. A yearly fee for ongoing access. Reddit’s deal with Google and the New York Times’ deal with Amazon were both reported as annual amounts (Quartz). Fresh data on a schedule usually earns more over time than a one-off.
• Per unit. Common where a unit is easy to count. HarperCollins’ deal with Microsoft offered $5,000 per nonfiction title, split between author and publisher (eMarketer). AI companies have paid creators $1 to $4 per minute of unpublished video (The Decoder). Finished RL tasks sell per task (Epoch AI).
• Revenue share. Some intermediaries pay owners a share of what each license earns. Troveo, which licenses video to AI companies, pays creators a share of gross revenue once content sells (AIM Research).
• Per token. We haven’t found a reported company archive deal priced per token. Token counts help a buyer judge volume, but the price usually reflects quality and fit, not size alone.

To compare offers, line up four things: guaranteed cash, any extra you might earn later, how long until you’re paid, and what you give up, such as exclusivity. The biggest headline isn’t always the best deal.

What lowers the value

• Heavy personal data. HR, health, payment and customer details all have to come out. The more of it there is, the more cleanup and risk a buyer takes on. Spirit’s sale excluded passenger profiles and loyalty records entirely (AI Weekly).
• Data you can’t clearly sell. Customer and partner information covered by NDAs, licensed content and vendor code can’t always be included. A law firm’s review found most companies can’t sell everything a buyer asks for (Frankfurt Kurnit).
• Hard or restricted exports. Some tools limit how data leaves them. Slack’s API terms, for example, bar apps from using API data to train large language models (Hunton), so the export route matters.
• Short or broken history. A year or two of records, or isolated files with no links between systems, is hard to turn into tasks.
• Objections from people with a stake. Spirit’s sale drew objections from its flight attendants’ union and a software vendor (Frankfurt Kurnit). Disputes slow deals and add risk.
• Tight restrictions. A very short term, a narrow permitted use or a ban on any onward licensing leaves a buyer less to work with.
• Time pressure. A company selling days before shutdown has little room to negotiate. Starting early keeps your options open.

Most of these are rights questions as much as pricing ones. Have a lawyer review what you can include before you share samples. This guide isn’t legal advice.

How to get a valuation for your data

1. Take inventory. List each tool, the years of history in it, and rough volumes: tickets, repos, docs, deals, messages.
2. Note what links together. A ticket that ties to a code change and a customer conversation is worth more than three unrelated records.
3. Flag what has to stay out. Client-owned material, sensitive projects, HR and legal files.
4. Get an estimate. Our 60-second estimate asks four questions and gives you a range. A short call turns that into an exact valuation.
5. Compare before you commit. A firm price comes after a buyer reviews a redacted sample. You approve the buyer and the terms before anything is licensed.

If you’re still deciding whether to sell at all, start with how to sell your company’s data to AI companies, which covers what sells, the ways to sell and the rights to check.

Questions founders ask