Short answer
There's no fixed price list. The value of business data for AI depends mainly on how unique it is, how much of it exists, its quality and history, the industry it covers, what uses are permitted, and whether Acquisition Partners currently need it. Specialized, hard-to-find datasets with years of history tend to be worth the most. The only reliable way to know is an assessment against current Acquisition Partner demand.
Value drivers
Essentials
The factors that drive value
Uniqueness: data that can't be found on the public internet or from other vendors.
Volume and history: years of records are more useful than a few months.
Quality: consistent, well-structured, accurately labeled information.
Specialization: data that captures expert decisions in a specific field.
Allowed use and exclusivity: broader or exclusive rights usually cost more.
Demand: what AI teams are actively trying to build right now.
Types of data that often attract interest
Customer support conversations and resolutions, sales and call recordings, internal documents and knowledge bases, code repositories, domain-specific records (legal, medical, financial, industrial) and workflow data showing step-by-step expert processes.
Why estimates vary so much
Deals are negotiated individually, and the same dataset can be worth very different amounts to different Acquisition Partners. Treat any online estimate as a starting point, not a quote.
Getting an estimate
SourceX offers a free five-question estimate on its homepage, followed by a review of your systems. No data is shared during the initial assessment.
Which drivers move a dataset's value up or down?
Eleven drivers move a dataset's value, and the worksheet below turns each one into something you can measure in your own systems. History depth and freshness share a row because both describe time coverage. Several drivers go beyond the factor list above: outcome labels, consistency, rights clarity and the cost of de-identification often decide how much of your volume a buyer can actually use. Fill the second column with your own measurements; the last two columns show which way each finding pushes value.
Machine learning research supports reading the table this way. Data Shapley, introduced by Ghorbani and Zou at ICML 2019, values each training example by its contribution to a model's performance for a given learning algorithm and performance metric, and in their experiments low-value points tended to be outliers and corrupted records. That is why consistency and buyer fit sit next to volume in the worksheet: a record only adds value if it helps the buyer's model on the buyer's task.
| Driver | What to measure | How buyers judge it | Raises value when | Lowers value when |
|---|---|---|---|---|
| Uniqueness and scarcity | Whether comparable records exist publicly or at many other firms | Can they get this anywhere else, including synthetic generation | The work is specialized and rarely recorded, such as expert judgment calls in a niche trade | Similar data is public, scraped or held by many competitors |
| History depth and freshness | Years of continuous records per system, gaps, date of the latest records, and whether new records can be delivered on a schedule | Enough time to cover seasonality, rare events and process changes, and data that is current or can be kept current | Several years of unbroken history in one system, including records created before generative AI, plus recent records and an ongoing feed | Short windows, migrations that dropped history, long gaps, or a snapshot from a process that no longer runs |
| Volume | Record counts by type, per year and per system | Enough examples per task to train or evaluate, after filtering | Tens of thousands of comparable records or more in one system | Large raw volume that collapses to a small usable set after cleaning |
| Outcome labels | Fields that record what happened: resolution, approval, defect cause, win or loss, payment | Can each record be checked against a real result | Outcomes are recorded consistently and linked to the work that produced them | Outcomes are missing, free text only or stored in a different system with no join key |
| Expert work content | Share of records that capture reasoning, steps or decisions by skilled staff | Does it show how professionals actually do the job | Job notes, review comments, escalation threads and approvals with rationale | Mostly automated logs, templates or boilerplate |
| Consistency | Schema stability, field fill rates, code lists, duplicate rate | How much cleaning and normalization the buyer must do | Stable fields and codes across years, documented changes | Frequent schema changes, inconsistent categories, heavy duplication |
| Exclusivity | Whether you will license to one buyer, a few or many, and for how long | What competitive advantage the license gives them | A buyer gets exclusive or limited access for a defined period | Data is already licensed broadly or the buyer expects competitors to have it |
| Rights clarity | Who created the data, under what customer, vendor and employee terms, and what the contracts allow | Can you grant the license without third-party claims | Records are your own business records and contracts do not restrict reuse | Customer contracts restrict use, data belongs to clients or a software vendor's terms limit export |
| Privacy and de-identification cost | Personal data fields, sensitive categories, free-text exposure, jurisdictions | How much it costs and risks to make the data shareable | Personal data is limited to a few structured fields that can be removed cleanly | Sensitive records, identifiers buried in free text or audio, or consent questions that need legal review |
| Buyer fit | Which AI tasks the data teaches or tests, such as agents for support, sales or field service | Does it close a known gap in their model or evaluation suite | The records map to a task a prospective buyer has said it is building or testing for | The data is interesting but does not match a task a buyer is training or testing for |
How do you score your own dataset with the worksheet?
Score each driver as strong, adequate or weak, with one line of evidence, and treat rights clarity as a gate rather than a score. A dataset with strong scarcity and outcomes but unresolved rights is not yet licensable at any value.
The OECD's 2022 report on measuring the value of data makes a related point at a broader level: the value of data depends to a large extent on the governance framework that determines how it can be created, shared and used. For a seller, four questions stay separate: who can technically access the records, whether privacy law and notices permit the use, who owns them, and what your contracts allow you to grant. De-identification reduces privacy exposure, but it does not create a right to license data a contract bars you from sharing. Work through those questions with counsel on the specific facts.
Check the volume bar early. As a general guide, buyers look for years of consistent history, typically two or more, and tens of thousands of records or more in one system; smaller sets can qualify when the work is rare, and each case is assessed deal by deal.
Here is how scoring might look for a hypothetical field-service company running ServiceTitan. Job records with technician notes, callbacks and warranty outcomes score strong on expert work content and outcome labels. Six years in one system scores strong on history depth. Customer names and addresses in notes score as a privacy cost that de-identification must address. Rights depend on the company's customer terms and its software agreement, which counsel reviews before anything else is scored.
- Pick one system and one record type first, such as Zendesk tickets, Salesforce opportunities, Jira issues or NetSuite transactions.
- Pull counts, date ranges and field fill rates from the system's own reporting, without exporting record contents.
- Write one sentence of evidence for each driver and mark it strong, adequate or weak.
- List every contract that touches the data: customer terms, vendor terms, employee and contractor agreements.
- Mark rights clarity as clear, needs review or blocked; stop and get counsel's view on anything not clear.
- Name the two or three drivers that would most change a buyer's view, and decide whether fixing them is worth the effort.
Sources for this section
- Data Shapley (Ghorbani and Zou, ICML 2019) values each training point by its contribution to predictor performance; low-value points tend to be outliers and corruptions, and high-value points indicate what data to acquire. Source
- The OECD's Measuring the value of data and data flows (Digital Economy Papers No. 345, 2022) argues that the value of data depends to a large extent on the data governance framework for how data is created, shared and used. Source
Frequently asked questions
Is there a standard price per record or per gigabyte?
No. Pricing depends on the Acquisition Partner, the use case and the data's uniqueness rather than raw size.
Can small businesses earn money from data?
Sometimes. For US small businesses, size matters less than uniqueness and specialization.
See if your company qualifies
A short company assessment. No data uploads are needed.