Market

Who buys company data for AI training? The 2026 buyers explained

Charlie Heath · Founder, TydeX14 September 2026

In 2026, about twenty companies run public programmes to buy, broker or license business records for AI training. They want the email, chat, documents, code and tickets that show how real work gets done. This guide sets out who they are, what each one takes, who it accepts and how it pays.

Every figure below comes from the buyer's own website, checked on 19 October 2026, unless another source is linked. These are published ranges, not offers. Programmes change quickly, so check the current terms before you rely on them. TydeX is independent and is not affiliated with any of these companies.

The four kinds of buyer

Data labs. Companies that build training data and practice environments for the large AI labs. Most of the big programmes sit here.
Direct buyers. Newer firms set up to license records from established small and mid-sized companies.
Marketplaces and agents. Firms that list or license your records to many buyers and share the proceeds.
Wind-down specialists. Firms that help startups close and sell their code and records on the way out.

Large technology companies buy too. In August 2026, Google won the US bankruptcy auction for Spirit Airlines' records with a $10m bid. The sale still needs court approval, and the hearing is now set for 6 November 2026.

Data labs

On 9 October 2026, micro1 committed $1bn over 12 months to buying and licensing company data.

Buyer	What it takes	Who it accepts	Published range	How it pays
micro1	Procedures, documentation, CRM, project histories, QA processes	30+ employees, working in English. Demand strongest in the US, UK and Canada	$100k+, $500k+ and $1m+ tiers	Not published
Mercor	Workflow data from tools such as Slack, Gmail, Salesforce and GitHub, with a focus on sectors such as legal, healthcare and finance	Not published	Not published	By bank transfer on delivery, typically 2 to 4 weeks after signing
Scale AI	De-identified records from everyday systems, often as an ongoing partnership	40+ employees preferred	$10k to $1m+, illustrative	Can recur
Handshake AI	Records from tools such as Gmail, Slack, Google Drive, Notion and QuickBooks	20+ full-time employees, 3+ years in operation	$100k to $4m+, indicative	Not published
Appen	Operating documentation, decision and escalation records, systems of record	No criteria published. Wind-downs get a faster route	Not published	When it accepts the data
Centific	Approval chains, exception handling and workflows across business systems	No criteria published. Wind-downs can use an asset purchase agreement	Not published	When it accepts the data
SuperAnnotate	Messages, process documents, decision trails and specialist records	"Any size, any industry, anywhere"	$100k to $1m+, indicative	On delivery, and again on each relicence
Labelbox Prism	Work documents and operating data from systems such as Salesforce, Jira, SAP and CAD	Not published	From about $50k, and $1m+ for large partnerships	A lump sum at signing, plus a revenue share. Exclusive by default
Ooak Data	Documents, Jira and Notion, CRM, finance, wikis, code, email and chat	Operating, winding down or closed. A director, liquidator or administrator can sign	$20k to several hundred thousand dollars so far	Not published. It says deals take 2 to 4 weeks
Turing (Project Lazarus)	Startups' codebases and operating data	Startups	Not published	Not published

Surge AI, which builds practice environments for AI agents, is also active. Its Nick Heiner told MIT Technology Review that one company's Slack history, if it captures conversations with variety and depth, can be worth more than another company's Slack, email and calendars combined.

Direct buyers

Buyer	What it takes	Who it accepts	Published range	How it pays
Mode	Email, documents, chat, CRM and billing records	6+ employees for legal, 10+ for accounting, finance and medical, 30+ for others. Says teams based mainly in the US are the strongest fit	$100k to $5m, illustrative	About three months from first conversation to payment
Replay	Code, decision threads, finance, support, CRM and sales records	Any industry. Prices set by headcount	$10k to $100k (20 to 50 staff), $100k to $1m (50 to 250), $1m+ (250+)	Upfront, plus a revenue share on every licence
Telegraph Lab	Work conversations, business software records and company drives	Across industries and sizes	$100k to $4m target range	Not published
Nyne	Chat, email, documents, CRM, HR, ERP and support records	20+ full-time US W-2 employees, 3+ years of operations	"Six-figure payouts"	On delivery and acceptance

Pewter, a Florida firm that runs licensing deals for business owners with a licensing partner, works only with US-based firms with 20 or more full-time staff, now or at their peak.

Marketplaces and agents

Buyer	What it takes	Published range	How it pays
Grepped	Messages, documents, tickets, code and records	$20k to $5m, as its estimate	In as little as 7 days, it says
Troveo	Company data exports	Says buyers pay "six figures"	A share of revenue on every sale
DataVendor	Private codebases	Quoted per codebase	When buyers purchase

Wind-down specialists

SimpleClosure's AssetHub helps founders who are closing a startup sell its source code, workspace data and other assets. SimpleClosure said it handled 100 sales of startups' workspace data in the year to April 2026, with payouts of $10,000 to $100,000 per deal. Sunset, another wind-down firm, also has what MIT Technology Review calls a "rapidly growing data resale business".

What to compare besides the price

Two buyers can value the same records very differently. Surge AI's Nick Heiner told MIT Technology Review that buyers quote very different prices because valuations rest on "research theses": estimates of how many useful training tasks a dataset can produce. So compare the whole offer:

Who de-identifies, and to what standard. Many buyers do it themselves. Ask for the written standard, and review a sample before you sign.
When you are paid. On signing, on delivery, on acceptance or after a sale are very different promises.
Exclusivity. An exclusive licence stops you licensing the same records elsewhere. It should cost the buyer more.
One payment or a revenue share. A revenue share can earn more over time, but it is less certain.
What happens to the raw data. Ooak Data, for example, says it deletes raw data within 30 days of delivery and confirms this in writing.
Permitted use. Training and testing models is standard. Rights to resell or pass the records on should be limited or priced.

What this means for UK companies

micro1 names the UK, with the US and Canada, as the markets where its demand is strongest, though it says it prioritises US companies. SuperAnnotate says it works with companies "anywhere", and most other buyers publish no country rules. Some programmes are built around US companies, so check which country's law a contract uses before you sign.

UK GDPR still applies to your side of the deal, whoever the buyer is. Our UK guide to selling company data explains the rules, and our valuation guide explains what moves the price.

TydeX presents UK companies to buyers like these, compares their terms and brings you written offers. You sign directly with the buyer you choose, and we never hold your records. Get a free valuation to see which buyers your company fits.

Questions

Which companies buy business data for AI training?

The best known in 2026 include micro1, Mercor, Scale AI, Handshake AI, Appen, Centific, SuperAnnotate, Labelbox, Ooak Data, Mode, Replay and Turing, plus marketplaces such as Grepped and Troveo. Google won a US court auction for Spirit Airlines' records, although the sale still needs court approval.

Can UK companies sell to these buyers?

Some say so. micro1 names the UK as one of the three markets where its demand is strongest, and SuperAnnotate says it works with companies anywhere. Others are built around US companies, such as Pewter, which works only with US-based firms, and Nyne, which asks for US W-2 employees.

How are companies paid?

It depends on the buyer. Mercor pays on delivery, typically 2 to 4 weeks after signing. Appen and Centific pay when they accept the data. Labelbox Prism pays a lump sum at signing plus a revenue share, and Replay pays upfront plus a revenue share.

Should we go to one buyer directly?

You can, but prices differ widely between buyers for the same records. Comparing several written offers is the best way to find the real value.