Selling company data for AI training: who buys, what is in demand and how it works
Luca Weber-SchäferUpdated 9 min read
Key points
- AI companies buy internal company data to train AI agents on real workflows, for example in simulated work environments.
- Buyers are AI labs, data vendors such as Scale AI, Mercor, Handshake AI and micro1, providers of training environments, and companies with their own models.
- In demand are datasets in which cases with decisions and outcomes can be traced: email, chat, tickets, ERP, CRM, process documents, code.
- Buyers check rights, anonymisation, exclusions and completeness. Personnel files are usually excluded.
- No data leave the company until a buyer has been found. Only anonymised data are handed over.
The market for internal company data became visible in 2026. The register Who Sells Data counts 13 buyers that state specific payments for company data on their websites, many of them from US$100,000 per company.[6] micro1 has announced that it will spend US$1 billion on company data over twelve months.[7] This article explains who buys such data, for what, and what sellers should prepare.
Why do AI companies buy internal company data?
To train and test AI agents on real workflows. Publicly available text shows little of how companies actually work, coordinate and decide.
AI labs are developing agents that are meant to carry out office tasks on their own. They are trained, among other things, in reinforcement learning environments, often called “gyms”, in which an agent practises tasks and is rewarded for good results. According to The Information, Anthropic has discussed spending more than US$1 billion on such environments within a year.[3] According to Wired, OpenAI together with Handshake AI asked contractors to upload real work products from previous jobs.[5]
Published case studies show that training on real company data works. Scale AI trained an agent on an insurer's data: the share of correctly executed database queries rose from 21.9% for the best off-the-shelf model to 46.9%. At a law firm, the share of fabricated answers to unanswerable questions fell from 49% to 21%.[4] Whether real archives are generally superior to synthetic data, however, no company has yet published in a comparative study.[1]
Who buys company data?
Four groups: AI labs, specialised data vendors, providers of simulated work environments, and companies that train their own models.
- AI labs
Google, OpenAI, Anthropic[2][5][3]
Training and testing their own models and agents
- Data vendors
Scale AI, Mercor, Surge AI, Handshake AI, micro1[6][1]
Preparation and resale to labs
- Training environments
Specialised providers of RL environments[1]
Simulated workplaces in which agents practise
- Companies with their own models
According to Bloomberg, for example SpaceX's AI unit[1]
Models for internal processes
The data vendors are the direct counterparts for sellers. They buy datasets, prepare them and pass them on to the labs. Many have their own buyer programmes, for example Scale AI, Turing, Replay or Mode.[6] What these programmes pay is shown in the article What is company data worth?
Which data are in demand?
Datasets in which a case can be traced from the request to the outcome, ideally across several systems and many years.
micro1 is looking for work instructions, knowledge bases, CRM data and project histories, as well as records of how teams make decisions.[8] Labelbox rates cross-system data from Salesforce, Jira and Slack with timestamps and roles as urgently needed, as well as support histories of at least six months.[9] Turing is expressly interested in the trade-offs behind decisions.[13]
| Type of data | Typical content | Why it is valuable |
|---|---|---|
| Email and team chat | Coordination, complaints, approvals | Decisions and reasons in their original wording |
| Tickets and customer service | Request, resolution, outcome | Clear task with a verifiable result |
| ERP and accounting | Order, delivery, invoice, payment | Structured outcomes with timestamps |
| Process documents | SOPs, quality manual, approval rules | Show how work is meant to be done |
| CRM and projects | Quotes, negotiations, project histories | Decision patterns over long periods |
| Code and development | Commits, reviews, bug tickets | Tasks such as debugging with a verifiable result |
What do buyers check before buying?
Rights, anonymisation, exclusions, completeness and documentation for their own regulatory obligations.
| Check | What it is about | How buyers handle it |
|---|---|---|
| Rights | May the seller license the data? | Mode checks ownership and access rights before making a firm offer[11] |
| Anonymisation | Are people no longer identifiable? | Scale AI anonymises within the company's environment before data are extracted[12]; Replay removes more than 60 categories of personal information[14] |
| Exclusions | Personnel files and particularly sensitive data | At Ooak Data, exclusions are attached to the contract before any data are moved[10] |
| Completeness | Do cases remain linked across systems? | Ooak replaces names in all sources using the same directory so that references are preserved[10] |
| Documentation | Evidence for the buyer's own regulatory obligations | Since 2 August 2025, providers of general-purpose AI models must publish a summary of their training data (Art. 53(1)(d) AI Act)[15] |
How does a sale work?
First the data are described, then checked, and only then processed. Data only leave the company once a buyer has been found, and then only in anonymised form.
- 1
Initial call and estimate
You answer nine questions on industry, software, data volume and history, or talk to us directly. We do not need access to the data for this and give you a first estimate.
- 2
Data sheet and buyer search
We describe the dataset without names, agree the data sheet with you and approach vetted buyers. Company and proceedings remain confidential.
- 3
Fee and anonymisation
As soon as a buyer has been found, we agree a success-based fee. Under a data processing agreement we anonymise the dataset locally at the seller; the buyer receives a sample under NDA.
- 4
Licence and handover
You approve the buyer, we hand over the anonymised data and then delete the mapping.
With the buyers themselves things can move quickly: Ooak Data quotes two to four weeks with around one hour of effort for the seller, and Turing usually reviews a sample within a week.[10][13] What needs to be considered legally in an insolvency is set out in the article Data in German insolvency proceedings.
What are the risks, and how can they be limited?
The biggest risk is personal references that remain recognisable despite cleaning. It can be reduced considerably through exclusions, multi-stage anonymisation and contractual prohibitions.
The US lawyer Greg Ewing calls such datasets “legally radioactive”, because emails and calendars contain private details such as medical appointments.[1] How seriously employees take this was shown by the sale of Spirit Airlines' data: the flight attendants' union objected because preserved links could allow conclusions about individuals.[16]
For sellers, this results in four measures:
- Exclude personnel files, health data and other special categories from the outset.
- Anonymise in several stages: fixed patterns, names in running text and indirect references from context.
- Local data processing at the seller; delete the mapping table after the handover.
- Prohibit re-identification in the licence agreement and provide for spot checks.
Frequently asked questions
Who buys company data for AI training?
AI labs such as Google, OpenAI and Anthropic, specialised data vendors such as Scale AI, Mercor, Surge AI, Handshake AI and micro1, providers of simulated work environments, and companies that train their own models. They mainly use the data to train and test AI agents on real workflows.
Are the data anonymised before the buyer receives them?
With reputable buyers, yes. Scale AI, for example, says it anonymises within the company's environment before data are extracted; Replay removes more than 60 categories of personal information. In Germany, effective anonymisation is practically a prerequisite, because personal data could otherwise only be transferred with consent.
Does the company have to give up all rights to the data?
No. A licence is customary, governing use, scope, term and exclusivity. Ooak Data, for example, licenses for training purposes only, with exclusions attached to the contract. Exclusive licences achieve higher prices.
Will the buyer learn which company the data come from?
That can be excluded contractually. Company names, domains and people are removed during anonymisation, and the licence agreement prohibits attempts at re-identification.
How long does a sale take?
Ooak Data quotes two to four weeks from the first call to payment, and Turing usually reviews a sample within a week. If a court or committee has to approve, it takes correspondingly longer.
Sources
- [1]MIT Technology Review (Matthew King): “AI labs are buying data from dead startups to train their models”, 13 October 2026
- [2]Fortune (Wen Shao): “Google wants Spirit Airlines' data. Micro1 wants to pay more”, 14 September 2026
- [3]Epoch AI (Jean-Stanislas Denain, Chris Barber): “An FAQ on reinforcement learning environments”, 12 January 2026
- [4]Scale AI Labs: “Scaling Enterprise Agent Performance with Reinforcement Learning via Verifiable Feedback Loops”, 17 November 2025
- [5]TechCrunch (Anthony Ha), citing Wired: “OpenAI is reportedly asking contractors to upload real work from past jobs”, 10 January 2026
- [6]Who Sells Data (operated by Secludy AI): “Who's buying”, as of 14–16 October 2026
- [7]The Next Web (Ana Maria Constantin): “micro1 commits $1bn to buy company data for training AI agents”, 9 October 2026
- [8]micro1: “Monetize your company's operational data”, accessed 20 October 2026
- [9]Labelbox: “Prism: Data licensing”, accessed 20 October 2026
- [10]Ooak Data: “Partner with us”, accessed 20 October 2026
- [11]Mode: “Company data partnerships”, accessed 20 October 2026
- [12]Scale AI: “Data Partnerships”, accessed 20 October 2026
- [13]Turing: “Project Lazarus”, accessed 20 October 2026
- [14]Replay: “Value my data”, accessed 20 October 2026
- [15]EUR-Lex: “Regulation (EU) 2024/1689 (AI Act), Art. 53”
- [16]Reuters, via Investing.com: “US court delays hearing on Google's purchase of Spirit Airlines data as union objects”, 18 August 2026
Luca Weber-Schäfer
Founder of European Trainingdata
Luca Weber-Schäfer advises on the realisation and licensing of company data from insolvency, restructuring and company sales, from the first estimate to the anonymised handover to AI buyers.
What is your dataset worth?
Nine questions, no data needed, no upfront costs. You receive a first estimate.