How much do AI labs pay for company data?

Nobody can price your data without looking at them, and no buyer publishes an average. What buyers do publish are ranges. As of October 2026 those run from about $20,000 to a few million dollars per company, and the two buyers who have said anything about averages both put a typical deal around $200,000 to $300,000. Where your company lands depends far more on how many years of connected, outcome-bearing data you hold than on how many gigabytes.

The published ranges

Buyer	What they publish	Notes
micro1	"$100k+ qualified", "$500k+ large-scale", "$1M+ highly unique"; referral page says "$100K-$2M+ for approved data packages"	Tiers tied to their own labels. Top tier is conditional on uniqueness.
Polyshares	"$100K to $2M, and above for large data"	Founder said on a podcast the average is about $300,000 and that they have done seven figures on a single deal.
Grepped	"$20K-$5M"	Founder wrote that the average lands "somewhere around $200K to $300K".

Those are the companies' own words. Treat the top of each range as what the buyer will consider, not what a typical seller receives.

Reported deal prices

• Google agreed to pay about $10 million for Spirit Airlines' operating data in bankruptcy court in August 2026: roughly 100 million emails, hundreds of millions of Teams messages, pricing data and code, with personal information stripped by a third party. Two data companies bid against Google.
• Shut-down startups have sold their Slack, Jira and email archives for $10,000 to $100,000 each through wind-down firms, and in one reported case for several hundred thousand dollars for 13 years of data.
• Polyshares lists four closed licenses on its site: $110,000, $275,000, $300,000 and $450,000, from a marketing agency, a solar company, a general contractor and a roofing company.

The airline is the top of the market. The startups are the bottom. An operating company with a few years of connected data sits in between, which is where the $200,000 to $300,000 averages come from.

What moves the number up

1. Continuous years on the same tools. Twenty years in business with a software switch three years ago means three scorable years.
2. Visible reasoning. Threads where people argued, got it wrong, and fixed it. Finished documents carry the answer with the thinking removed.
3. More than one system. A support ticket that links to a chat thread that links to a CRM note is worth more than any one of them alone.
4. Outcomes. Tickets closed with a resolution, deals marked won or lost, changes approved or rejected. These become tasks and graders for the lab.
5. A rare industry. Chip design, wet labs, freight dispatch, roofing. Buyers told their referrers in October 2026 that they are asking for finance, legal, semiconductors, bio, manufacturing and aerospace, and that generic marketing text is not on the list.
6. Clean rights. Data you own outright, with customer contracts that permit de-identified use.

What caps it or kills it

• Data you hold for clients under a master agreement. They belong to the client.
• An email archive on its own. One buyer's founder put it plainly: if the pitch is "ten years of emails, what's that worth," the answer is less than you would like.
• Phone-first operations, where the data reads "ticket opened, ticket closed."
• Personal information woven through in a way that cannot be stripped.
• Anything already public or scraped.

Why "up to" is not a number

Every range above is a range the buyer chose to advertise. The buyer decides the number after reading a sample, and the buyer's margin to the lab sits inside it. The only way to get a real figure is to put a summary of your data in front of more than one buyer and compare what comes back. That is the brokerage's job.