How much is a company's data worth to an AI lab?

Why no general figure exists

Two companies in the same trade can hold records that are worth very different amounts to a lab. One keeps a clean history in a single system. The other keeps fragments across many tools, with gaps and mixed ownership. A lab pays for what it can use, and what it can use depends on the records themselves.

Research on valuing data says the same thing in general terms. A UK government review of data valuation methods describes three families: cost-based, market-based and use-based. It notes that market prices are often missing because data is produced and analysed inside organisations rather than traded. A Frontier Economics report summarised by the same department says the value of a data asset depends on how it is used and who uses it. For company records, that means value is set by the buyer's use as much as by the data.

What raises value

Continuous history is the first factor. Records that run without long gaps let a lab see how work changes through busy periods, staff turnover and new rules. A long, steady record of one kind of work is harder to find than a large pile of unrelated files.

The second is records that tie decisions to outcomes. A quote that can be matched to the job that followed, or a ticket that shows what was tried before it was solved, shows cause and result. Specialized work adds to this, because a lab can already find general text and wants examples of tasks done by people who know a trade.

Linked evidence matters too. A dispatch note joined to a photo, an invoice and a customer reply shows a whole job, not a fragment. Clear rights complete the list. If the company owns the records, holds the permission needed to license them and can describe how personal details were removed, the lab has less to worry about, and that lowers the cost of a deal for both sides.

What lowers value

A short history gives a lab little to learn from, because patterns need time to show. Records the company does not own, such as content licensed from a software vendor or material that belongs to a customer, cannot be licensed and are left out. If much of a set falls in that group, what remains may be thin.

Unclear consent lowers value as well. Where records contain personal details and the company cannot show what people were told, a lab may decline them or ask for more work first. Poor structure, duplicated rows and missing fields can also reduce use, though they are sometimes fixable before a sample is built. Each of these is a question about the records, not about how much space they take up.

Why file size does not set price

Storage size measures how much space files take. A folder of large scans or videos can be big and say little about how a business decides things. A small, well-kept table of decisions and results can carry more of what a lab wants. Counting gigabytes, rows or years says how much there is, not how useful or how clean it is.

Size can matter when it reflects real coverage, for example many cases of the same kind of job. Even then the buyer asks whether the records are varied, linked and licensable. Size is one input to that question and rarely the main one.

An estimate and an offer are different things

The data value calculator on this site asks about the kind of records, how long they run and how they are kept, and it gives an estimate. An estimate is a way to think about where a set might sit. It is not an offer, it does not bind any lab, and it cannot account for what a lab needs at the moment.

A real number comes later. A sample is chosen, the owner approves it, a lab reviews it and then makes an offer in writing. The owner decides whether to accept. Before signing, the owner can use the check-an-offer page to compare what the offer covers against what was discussed.