Insights · 6 min read

Your Company Data Is Not Cheap: Why Data Owners Undersell It

Not cheap does not mean worth millions. It means worth pricing carefully.

Bearish AI · October 21, 2026

Most owners who undersell do not do it out of carelessness. They do it because the first number they see seems generous compared with what the data cost them to create, which was nothing extra. This article is about the five habits that lead to underselling, and the discipline that fixes them. It is also about the opposite error: a small or messy archive is still worth little, and pretending otherwise helps nobody.

Habit 1: pricing against effort, not use

Your chat history and tickets were a by-product of running the business, so it feels like free money at any price. Buyers think differently. In MIT Technology Review’s coverage, one wind-down firm’s chief executive likened such operational exhaust to wood shavings that everyone thought had no value. A buyer then estimates how many tasks the data yields. The price should follow what they can do with it.

Compute shows the stakes. Epoch cites an estimate of about $2,400 of compute spent per task during training. Graded tasks mostly sell for $200 to $2,000, per Epoch AI. The compute bill for training on a task can match or exceed what the task cost to buy, so a good task is not a large expense in that context.

Habit 2: anchoring on the wrong comparison

A headline says an airline’s data sold for $10 million, so a founder expects a fortune. Another headline says a shut-down company got $25,000, so a founder expects pocket change. Both are unrelated to your archive. The airline reportedly involved 100 million emails and 500 million Teams messages. The $10,000 to $100,000 range comes from reports on shut-down companies. A thoughtful price starts with your own yield, not someone else’s headline.

Habit 3: ignoring scarcity

Epoch AI estimates the stock of public human text at about 300 trillion tokens and projects it will be fully used between 2026 and 2032. It also says non-public data, such as instant messages, seems unlikely to be used at scale because of legal issues. Read that carefully. It means that workplace data cleared for use, with rights, privacy and provenance handled, is scarce. The scarce part is the cleared version, not the raw logs.

Habit 4: selling raw when you could sell prepared

A raw dump sells at the buyer’s guess, minus their cost to clean it. Checked, linked material can be sold per unit. We explain the gap in /insights/raw-data-vs-prepared-data-value. The cost to prepare is real, and doing it badly is worse than not doing it, but skipping it entirely is a common reason for low offers.

Habit 5: giving away rights in the contract

The price is only part of the deal. A journalist writing in the Columbia Journalism Review described an AI company’s offer of $1,000 per case of reporting notes, and quoted its draft contract as a non-exclusive, royalty-free, fully paid up, perpetual, irrevocable, transferable and sublicensable license, with ownership staying with her. That is one document, not a market norm, but it shows what a broad grant looks like. Compare it with the structures that practitioners describe as usual: use limited to a named model, extension to successor models at an extra fee, and a perpetual right for models but finite access to the dataset. Broad grants can cost you more than a low headline price.

What underselling looks like in practice

These are illustrations, not reported cases. A founder accepts a flat sum for a decade of code history without asking whether the buyer can build tasks from it, and later learns the buyer built hundreds. A team grants a perpetual, sublicensable license to everything in a workspace for a modest payment, and loses the ability to sell the same material elsewhere. A company ships raw exports to a buyer who then does the cleaning, and the cost of that work comes out of the price. In each case the seller was not wrong about the material. They were wrong about what a buyer could do with it, and about what the contract let the buyer do next. Reading a license as carefully as you read the price is the cheapest protection available, and it is worth paying a lawyer to do it with you.

A pricing discipline

• Count what you have. People, years, kinds of records, language, personal data, rights.
• Estimate the yield. How many checkable tasks could this plausibly make? Even a rough count beats none.
• Get a range, not a number. Buyers quote widely different prices for the same data, so a single offer is a sample of one.
• Seek more than one view. At least one independent estimate, such as the free one at /estimate, before you respond to an offer.
• Read the grant. Scope, term, exclusivity, sublicensing and what survives termination.
• Do not rush. MIT Technology Review reports prices will likely stabilize within a year, unless an AI bubble pops first. That is neither a promise of rising nor of falling prices.

The honest counterweight

The opposite error is just as costly. A founder who rejects every offer because a headline mentioned millions can wait until the archive is stale, the team has scattered and the export path has closed. Underselling and overpricing both come from skipping the same step: working out what your material can really yield, with a lawyer checking what you may sell.

Not every archive deserves a premium. Short histories, tiny teams, unclear ownership, heavy personal data and messy formats lower the value, and the cost of fixing them may exceed what a buyer pays. If that is you, the best outcome may be to bundle with other material or keep the data. A lawyer should review any sale either way. For reported prices to compare against, see /insights/how-much-do-ai-companies-pay-for-training-data.