Guides · Published
Best AI training data licensing companies (2026)
Compare the best AI training data companies in 2026: licensing services, expert data vendors and annotation providers, judged on the same criteria.
The best option depends on what you have. If your company holds years of internal records and wants to license a copy to AI buyers, Smelted is built for that. If you need new labeled data or expert feedback, Mercor, micro1 and Surge AI are common picks in published rankings.
We run Smelted, so it's on this list. We judged it on the same criteria as the others, using only what we could check on each company's own pages and the published guides listed under Sources.
How we picked
Every entry is judged on the same six criteria:
- Type of data. Does the company license existing business records, or produce new labeled data and human feedback?
- Who pays the company. Is the company paid by the buyer, or does it buy or commission the work?
- Ownership and control. Does the company keep ownership, and choose what is shared and with whom?
- Fit by company profile. Which companies, sizes and industries is the service built for?
- Privacy handling. What is removed before data is shared?
- Next step. How quickly can a company move from first contact to a decision?
We only list companies whose descriptions we could check. Where a price or payment figure was not published, we leave it out.
The list
1. Smelted
Smelted is a free matching service that licenses anonymized internal business records, such as support tickets, CRM history, SOPs, project files, code, CAD and lab data, to companies that train AI agents. It finds buyers that fit a company, gives a ballpark range, and introduces the company to two or three matching buyers.
Best for: US companies with 50 to 500 employees and years of records in tools like Zendesk, Jira/GitHub, Salesforce/HubSpot, Bullhorn, AutoCAD/Revit or Benchling.
- Type of data: existing records, not new labeling work.
- Who pays: the buyer pays Smelted's fee. The company pays nothing.
- Ownership and control: the company keeps ownership, licenses a copy, and chooses what is included and which buyers it meets.
- Privacy: names, contact details and IDs are removed before anything is shared.
- Watch for: the value depends on headcount, years of records and how connected the data is, and the real figure comes after a buyer reviews a sample. Payment usually arrives a few weeks to two months after signing, so plan on one to three months from first call to payment. Companies under 30 employees should expect buyers to pay less.
2. Opendatabay
Opendatabay publishes a guide to licensed AI training data types that compares commercial and general licenses for both buyers and sellers.
Best for: teams that want to understand license types before they sign anything.
- Type of data: licensed datasets, described by type and license terms.
- Strength: the guide sets out the difference between commercial and general licenses clearly.
- Watch for: the guide explains licensing; check the terms of any specific dataset deal yourself.
3. Mercor
Mercor is named in published rankings for expert data and human feedback. Analytics Insight places it third in its 2026 list of AI training data providers.
Best for: expert evaluation and human feedback work for AI models.
- Type of data: expert-generated data and feedback, not licensing of existing company records.
- Strength: a common pick when the task needs skilled reviewers.
- Watch for: you are buying or commissioning expert work, so it is a different model from licensing your own records.
4. micro1
micro1 is listed alongside Mercor and Surge AI for expert data and human feedback.
Best for: teams that need expert review of model outputs.
- Type of data: expert data and human feedback.
- Strength: included in several 2026 rankings of providers.
- Watch for: like Mercor, it serves model builders who need new work done, not companies selling existing records.
5. Surge AI
Surge AI is named for expert data and human feedback, and Analytics Insight ranks it second among 2026 providers.
Best for: large programs that need expert human feedback at scale.
- Type of data: expert data and human feedback.
- Strength: frequently named as a leading provider.
- Watch for: the scale suits large buyers; check fit before you start.
6. Scale AI
Scale AI is the name most often given for annotation, according to Troveo's 2026 comparison. That guide also reports that Meta invested in the company in June 2025.
Best for: high-volume annotation and labeling.
- Type of data: annotation of new data, not licensing of existing records.
- Strength: long-standing choice for labeling at volume.
- Watch for: Troveo reports that some major customers were winding down work with the company after the investment, so check current availability.
7. Appen
Appen is named for annotation in Troveo's 2026 comparison and in CleverX's list of leading companies.
Best for: annotation and data collection projects.
- Type of data: annotation and collection of new data.
- Strength: named across several 2026 guides.
- Watch for: it does new labeling work, so it does not suit a company that wants to license records it already has.
8. CleverX
CleverX describes itself as an on-demand B2B research and expert platform with more than 10 million verified professionals across 150-plus countries.
Best for: teams that need quick access to verified experts for evaluation or preference data.
- Type of data: expert engagement on a pay-as-you-go basis.
- Strength: CleverX says delivery takes two to five days, and its professionals are verified through work email and LinkedIn.
- Watch for: it provides expert time, not licensing of company records.
Comparison
| Company | Type of data | Who pays the company | Ownership and control | Best for |
|---|---|---|---|---|
| Smelted | Existing internal business records | Buyer pays Smelted's fee; company pays nothing | Company keeps ownership and chooses what is shared | US companies with 50 to 500 employees and years of records |
| Opendatabay | Licensed dataset types (guide) | Not covered by the guide reviewed | Depends on the license | Understanding license types |
| Mercor | Expert data and human feedback | Buyer commissions the work | Not applicable to records licensing | Expert evaluation |
| micro1 | Expert data and human feedback | Buyer commissions the work | Not applicable to records licensing | Expert review of outputs |
| Surge AI | Expert data and human feedback | Buyer commissions the work | Not applicable to records licensing | Large expert programs |
| Scale AI | Annotation of new data | Buyer commissions the work | Not applicable to records licensing | High-volume labeling |
| Appen | Annotation and collection | Buyer commissions the work | Not applicable to records licensing | Annotation projects |
| CleverX | Expert engagement | Pay-as-you-go | Not applicable to records licensing | Fast expert access |
The market in 2026
The market is growing. According to Troveo, the global AI training dataset market is projected to reach 16.3 billion dollars by 2033, growing 22.6 percent a year from 2026, based on Grand View Research. Image and video data held the largest share, at 41.9 percent of revenue in 2025.
Supply of public text is a concern too. Epoch AI has projected that the stock of high-quality public text data could be effectively used up for training between 2026 and 2032, which is one reason labs look for non-public data (Troveo).
FAQ
How do companies license their internal data for AI training?
A company can work with a licensing service that matches its records to buyers. The company keeps ownership of its data, licenses a copy, and chooses what is included. Names, contact details and IDs are removed before anything is shared. Smelted works this way and is free for the company.
How much is company data worth for AI training?
It depends on headcount, the number of years of records and how connected the data is across tools. Smelted gives a ballpark range after it has matched a company to buyers, and the real figure comes after a buyer reviews a sample.
How long does it take to get paid?
With Smelted, plan on one to three months from the first call to payment. Buyers usually pay a few weeks to two months after signing.
Do I need to label data to license it?
No. Licensing existing records is different from annotation. Annotation providers such as Appen and Scale AI create new labeled data, which is a separate service.