The short answer
AI models have largely consumed the public web. Epoch AI estimates the stock of usable public human text could be fully used between 2026 and 2032, which is why developers are turning to non-public data (Epoch AI).
Ordinary companies hold exactly that: years of records showing how skilled people solve real problems. Several programs now pay businesses for a licence to use a copy of those records. The business stays the owner and agrees, in writing, what the buyer may do with the copy.
Who buys business data for AI?
Three kinds of buyer are active in 2026:
- AI labs building general models, which need examples of professional work for training and evaluation.
- Data partners, which license company data and supply it to labs.
- Large technology companies. Google won a bankruptcy auction for Spirit Airlines’ internal records with a $10M bid in August 2026 (Skift).
Calto Data is not a buyer. We are an independent introducer: we check whether your records fit what buyers ask for and introduce you when they do.
What do they actually buy?
Buyers want records that capture a complete task, with the reasoning and the outcome. Published licensing programs list source systems including AutoCAD, NetSuite, QuickBooks, ServiceTitan and Slack.
For technical businesses, that typically means engineering drawings with revision history, inspection and test results, maintenance logs, production data and support tickets. See what data AI companies buy for a full breakdown.
How much does it pay?
| Source | Published figure |
|---|---|
| Published licensing programs | $100K to $4M+ |
| Spirit Airlines records (Google) | $10M winning bid, Aug 2026 |
No program publishes a per-record price. Value depends on volume, uniqueness, quality and the rights granted. Read how much company data is worth for the factors that move price.
How does a deal work?
- Describe your business, systems and record history. No files change hands.
- Discovery: the buyer assesses fit and gives an indicative range.
- Scope: agree which systems, teams and years are included, and what is excluded.
- Contract: a licence sets permitted uses, retention, security, price and payment.
- Preparation: agreed specialists extract and de-identify the copy.
- Delivery and payment: the buyer checks the copy and pays.
Is it legal?
Licensing records your company owns is generally lawful, but two things need checking first. Privacy law applies to any personal information in the records, under GDPR, Japan’s APPI, Hong Kong’s PDPO, Taiwan’s PDPA or China’s PIPL. Contracts with customers and suppliers may also restrict how their information is used.
See how to prepare records for licensing. This is general information, not legal advice.
Is your business a good fit?
The strongest candidates are owner-operated firms with roughly 20–500 staff, five or more years of digital records they created themselves, and specialist work: manufacturing, engineering, inspection, maintenance, semiconductors, energy and technical services.
Firms whose records mainly belong to clients, such as accountants, law firms and payroll providers, are usually a poor fit.
Programs differ on eligibility. Published programs typically ask for 20+ employees and 3+ years in operation. Few programs publish terms for businesses in Japan, Taiwan, Hong Kong or China.
Frequently asked questions
Do I lose ownership of my data if I license it?
No. A licence grants limited rights to use a copy. Ownership of the originals stays with your company.
Can a small business sell data to AI companies?
Yes, if it has deep, well-kept records. Published programs set minimums of 20–30 employees, but the depth and uniqueness of the records matter more than headcount.
Will competitors see our records?
Licences normally restrict onward sharing and require confidentiality. Negotiate who can access the copy and whether the buyer may resell it.
How long does a deal take?
Programs don’t publish timelines. Expect several weeks to a few months for discovery, contract, preparation and acceptance.
Is the payment one-off or recurring?
Both exist. Most published business-data programs quote a one-off licence fee; ongoing data, such as agent trajectories, can also be licensed for a recurring annual fee.