Data Valuation: How to Value Company Data

What data valuation means for a company

Valuing data is harder than valuing a machine or a building. The Harvard Data Science Review puts it plainly: "Today, there is no standard to measure the value of data." (Harvard Data Science Review). A UK government review reached the same view: there is no consensus on the best approach, and most methods fall into cost-based, market-based or use-based groups (GOV.UK).

In practice, a company values data for one of three reasons: to set a price before licensing it, to judge an offer it has received, or to account for it in a sale of the business. The method you lean on depends on the reason. For an AI license, what matters is what a buyer would pay for defined rights to a prepared copy.

This guide covers the methods and how to use them. For published prices by record type and what moves them, see what company data is worth. For the accounting view, see whether data belongs on the balance sheet.

The three methods: cost, market and income

The same three approaches used for patents and other intangible property apply to data. The US Internal Revenue Manual sets them out for intangibles (IRS), and the Harvard review notes that, as with physical assets, data can be valued on its cost, its sale value or its income potential.

Source: IRS Internal Revenue Manual 4.48.5: Intangible Property Valuation Guidelines

Why the methods give a range, not one number

The UK review explains why the answers differ. Cost-based methods "are likely to provide a lower-bound estimate of the value of data", while market-based methods are likely to provide an upper bound, because it is difficult to separate the value of data from other digital and intangible assets (GOV.UK). It also notes that market prices for data are often not available, because data is usually produced and analysed inside the company.

Income methods sit in between and swing with your assumptions. When applied to businesses, the review says, use-based methods are often called "income-based" or "revenue-based" methods. The IRS warns that the income method relies on estimates of future earnings, how long they last and the risk of not realising them (IRS).

So run all three and read them together. If the cost floor is above anything a buyer would pay, licensing is probably not worth the preparation. If comparable deals sit well above the floor, the work is in proving that your records are comparable.

What AI buyers price on

An AI developer is not buying your storage bill. It is buying the ability to train, test or ground a model on records it cannot get elsewhere. In Origin's experience as an independent data partner, five things move the number most.

• Scarcity: records that never appeared on the public web, such as resolved support cases, internal decisions or operational logs.
• History and continuity: years of dated records with few gaps are worth more than a recent snapshot.
• Structure and links: tickets joined to outcomes, or documents joined to the decisions they led to, beat loose exports.
• Rights: clear authority to license, with customer, employee and third-party material handled. Unclear rights cut the price or end the deal.
• Permitted use and terms: training, evaluation and retrieval are priced differently, and exclusivity, duration and territory change the number. The IRS lists the same terms when it tests whether a deal is comparable (IRS).

Worked example: valuing a small company archive

The company below is an illustration, not a client. The comparable deal is real and described without names; the AI data licensing deals tracker lists the parties and sources for every public deal.

Take a software company with eight years of support tickets, linked to product releases and fixes, plus internal chat about the same issues. It is weighing a non-exclusive license to an AI developer for training.

Reading the example

The three answers frame a negotiation rather than set a price. The cost floor tells the owner when to walk away. The comparable deal suggests a small archive can trade in the tens of thousands of dollars rather than millions; larger, longer or rarer records can fetch more. The income view shows why a non-exclusive license can be worth more over time than a one-off sale.

Small sellers also face a weak position. MIT Technology Review notes that in shutdown sales "there's little assurance they're getting a fair offer" (MIT Technology Review). Running the three methods before the first call is the best protection.

For an indicative range for your own records, answer six questions. Published ranges and estimates are not offers.

Common data valuation mistakes

Before a buyer makes an offer, use the offer worksheet to compare proposals on more than the headline number. See also data monetization for the routes to earn from data, and AI licensing for how license deals are structured.

• Using a headline deal as a benchmark. The largest publicly reported licenses are for large archives of published content, not for one company's internal records.
• Pricing by volume. Gigabytes and row counts matter less than history, links between records and rights.
• Valuing data you cannot license. If contracts or privacy rules block the use, the value to a buyer is close to zero until that is fixed.
• Ignoring the terms. The same records are worth different amounts for training, evaluation or retrieval, and exclusive rights should cost the buyer more.
• Treating cost as value. What the data cost to collect sets a floor, not a price.
• Skipping a written scope. Without a defined extract, buyer and seller are valuing different things.