AI Training Data Broker is an independent broker for deals between companies and AI labs. Every deal has two sides: the seller is a company licensing its own operating data, and the buyer is an AI lab or model team licensing it for training. The broker represents the seller, the way a sell-side broker does on a business sale: inventory, rights, presentation, competing offers, terms, close. Every deal has a written contract reviewed by counsel, an insured transaction, and independent third-party de-identification before a buyer sees anything.
Owners, executives and authorized representatives start at /sell. The free check asks for counts and years per system, never files, and returns a straight yes or no, then competing offers.
Who qualifies. 20 or more employees (most buyers want 30 or more full-time, the largest want 50), mature processes that leave a written record, 3 or more years on the same tools (a software switch resets the clock), data in English, and a base in the US first, then the UK, Canada and Australia. Remote-first, white-collar teams have the richest data. Companies under the floor can still ask, because demand rotates by industry month to month.
What pays. Support tickets with resolutions, Slack or Teams threads where decisions were argued out, CRM history with deals won or lost, email with boards and approvals, code with issues and reviews, SOPs and knowledge bases next to QA records, and accounting or ERP records as they ran. Labs pay for the chain from request to outcome.
What does not. An email archive on its own, marketing copy and finished reports (the reasoning is edited out), client files held under a master agreement, public or scraped material, and data whose whole value is who the people are.
What it pays. Approved data partnerships are reported at $100,000 to $2 million USD. A typical operating company lands around $200,000 to $300,000; large, rare or exclusive data reaches $1 million and up; the archive of a closed company runs $10,000 to $100,000. Continuous years on the same tools, visible reasoning, a rare industry and clean rights move the number.
Terms. Never perpetual. The usual buyer ask is 24 months of exclusivity; a shorter term is possible at a lower price, and a company planning a sale can time the term to its exit. The license payment is one-time and the company keeps ownership of its data.
AI labs, model teams and superintelligence programs start at /buy and send a request with data types, industries, volume and exclusivity.
What can be sourced. Support and service tickets (Zendesk, Intercom, Freshdesk, ServiceNow, Jira Service Management), team chat (Slack, Microsoft Teams, Google Chat), CRM pipelines (Salesforce, HubSpot, Pipedrive, Zoho), email and calendar (Google Workspace, Microsoft 365), project and field work (Asana, Jira, Monday, ServiceTitan, Procore, Jobber), code with issues and reviews (GitHub, GitLab, Bitbucket), finance and ERP (QuickBooks, Xero, NetSuite, SAP), and documents and knowledge bases (Google Drive, SharePoint, Notion, Confluence).
Industries. Logistics and freight, construction and trades, professional services and agencies, accounting, IT services, software, manufacturing, insurance brokerages, property management and healthcare administration.
Outcome-bearing data. Tickets closed with a resolution, deals marked won or lost and changes approved or rejected can be rebuilt into workflow trajectories and graders for RL environments. Buyers who build environments say so in the request and get outcome-bearing, multi-system data.
Terms. Exclusive or non-exclusive with a term (two years is common), one-off corpus or ongoing refresh, scope written in (training, evaluation, environment construction or all three), seller sample review then a written acceptance window, and a provenance memo with every package covering ownership, contract clauses and exclusions. The broker never holds raw data: the seller exports to the de-identification vendor, which delivers to the buyer.
The full process, with who does what and how long each step takes, is at /how-it-works for sellers and /how-it-works/?side=buyer for buyers.
For a seller there are nine steps: a ten-minute intake, a thirty-minute discovery call, a rights check with the seller's counsel (2 to 5 days), a one-page summary (2 days), buyers competing on it (1 to 2 weeks), negotiation (a few days), third-party de-identification (about a week), delivery and the buyer's acceptance window (5 to 30 days), and payment on the license terms (1 week to 60 days).
For a buyer there are five: send a request, review the one-page summary and de-identified samples and make an offer, negotiate in writing with counsel, receive the de-identified package with its provenance memo and accept it, then pay.
What kills a deal on the seller side: phone-first operations with nothing written between "opened" and "closed", a software switch in the last three years, a missing core system, client-held data under a master agreement, non-English data, a seller who wants to de-identify the data themselves, and padded data.
A third-party de-identification vendor replaces all personally identifiable information (names, emails, phones, account numbers) with stable pseudonyms such as "employee 632", removes anything the company's own data policy marks confidential, holds the raw files about a week and then deletes them. The seller reviews a sample before anything is accepted. Buyers and the broker see summaries and samples only.
Nothing up front for sellers. The buyer pays a finder's fee when a deal closes. Preparing, presenting and negotiating on the seller's behalf is billed as an advisory fee, agreed in writing before work starts. No listing fees, no retainers and no charge when nothing closes. Both fees are disclosed to both sides before anyone signs. Details are at /about.
The blog has plain answers for both sides, filtered at /blog/?for=seller and /blog/?for=buyer. Useful starting points:
- How much do AI labs pay for company data?
- Is it legal to sell company data to AI labs?
- Exclusive vs non-exclusive data licenses: which fits your company
- What is RL data? What AI labs mean when they ask for it
- Where to buy licensed company workflow data for training
Questions go through /contact; sellers and buyers track their deal in the partner portal at /signin. Terms and privacy policy are at /terms and /privacy.