What company data qualifies for AI data licensing

What kinds of data do AI buyers ask for?

Each program publishes its own list. The lists overlap more than they differ: almost every buyer wants the written record of people doing their jobs.

• micro1 lists SOPs, process documentation, knowledge bases, workflows, templates, internal communications, project histories, CRM records and support tickets, with agent trajectories as an optional add-on.
• Mercor lists industry records such as contracts, claims and work orders, plus Slack threads, Gmail messages and metadata, Salesforce records, GitHub repos and pull requests, and Notion pages.
• Polyshares lists Slack and Teams, Drive and SharePoint, email, CRM history, support tickets, design files, call recordings, and accounting and ERP exports.
• Endless Data asks for quotes, tickets, work orders, SOPs and handover notes “and the decisions behind them”.
• Turing Project Lazarus takes chat, email, tickets, wikis and docs, CRM and support history, databases and source code, and accepts code on its own.
• Korra lists team conversations, business email, sales records, customer support, documents and guides, approved call transcripts, operational workflows, and AI review and feedback.

Tool coverage is broad. Handshake AI names AutoCAD, Dropbox, Figma, Gmail, Google Drive, Outlook, Monday.com, NetSuite, Notion, Paychex, QuickBooks, ServiceTitan and Slack, then “+1K more” (source). Mercor cites “100+ supported tools” (source), and Polyshares publishes about 100 per-tool guides (source). If a system you use is not named, ask before ruling it out.

Which systems hold the most useful records?

The table maps common systems to the buyers that list them, what makes that data stronger, and what is usually cut. Buyer names link to our sourced buyer pages. The exclusions come from our industry pages, where each one is explained with its legal or contractual reason.

System	Listed by	Stronger when	Usually excluded
Email (Gmail, Outlook)	Mercor, Polyshares, Turing, Korra	Operational mailboxes with long reply chains between staff, customers and suppliers	Client-confidential correspondence; HR, pay and health matters
Team chat (Slack, Teams)	Mercor, Polyshares, Turing, Korra	Internal channels where work is planned and decided	Shared channels with clients (agencies)
CRM (Salesforce, HubSpot)	micro1, Mercor, Polyshares, Turing, Korra	Activity notes and stage changes kept over years	Customer personal data unless removed; pricing clients treat as confidential (logistics)
Support tickets (Zendesk, Intercom)	micro1, Polyshares, Endless Data, Turing, Korra	Request, back-and-forth and resolution in one record	Customer content and account numbers inside tickets (SaaS)
SOPs, wikis, templates (Confluence, Notion)	micro1, Mercor, Endless Data, Turing, Korra	Written by your own staff, with earlier versions kept	Material that belongs to clients or vendors
Projects and work orders	micro1, Mercor, Endless Data	Tasks with status changes, comments and outcomes	Client-owned deliverables (professional services)
Code (GitHub, GitLab)	Mercor, Turing, SimpleClosure AssetHub	Commit history with pull-request discussion	Secrets, vendored and third-party code (SaaS)
Accounting and ERP (QuickBooks, NetSuite)	Polyshares, Handshake AI	Approval and workflow history	Bank, card and payroll account data; client books (accounting firms)
Call recordings and transcripts	Polyshares, Korra	Transcripts that the law and your clients allow you to use	Audio, spoken card numbers, voiceprints (call centers)
Design and CAD (Figma, AutoCAD)	Polyshares, Handshake AI, AIData.Broker, SimpleClosure AssetHub	Revision history and review comments	Customer drawings under NDA, security-sensitive plans (manufacturers, construction)

What never qualifies?

Some data stays out whatever the offer. The buyers’ own exclusions are a good guide, and the law sets a floor under them. Talk to your lawyer before scoping anything in the categories below.

Data you do not own. Polyshares says data a company holds for its own clients is out of scope (source), and Korra takes only records the business owns (source). In practice that means leaving out client deliverables at agencies, customer content at software companies and MLS listing data at real estate brokerages.

Health information. Endless Data says PHI is never licensed (source). Korra excludes patient medical records (source), AIData.Broker excludes health information (source), and SimpleClosure AssetHub excludes sensitive employee, compensation and health data (source). The HIPAA Privacy Rule governs protected health information held by covered entities and their business associates.

Tax returns and financial accounts. 26 U.S.C. 7216 makes it a crime for tax return preparers to knowingly or recklessly disclose or use return information beyond what the law allows. AIData.Broker excludes financial account data and clients’ legal and tax files, and Endless Data excludes client books (AIData.Broker, Endless Data).

Privileged legal material. Endless Data never licenses privileged matter content (source). Our law firms page explains why client confidentiality reaches well beyond privilege.

Children’s data and bulk sensitive data. Korra and AIData.Broker both exclude children’s data, and AIData.Broker also excludes bulk US sensitive data under 28 CFR 202 (Korra, AIData.Broker).

Export-controlled technical data. Technical data covered by the International Traffic in Arms Regulations cannot go to a party that may give foreign persons access without authorization. See manufacturers.

Secrets and credentials. Turing says personal data and secrets are removed before anything reaches a model (source). Rotate keys found in code or chat anyway; removal by a buyer is not a substitute.

Records already shown to other buyers. Korra excludes them (source). If you plan to approach more than one program, read exclusivity explained first.

De-identification narrows the risk but does not end it. Handshake AI says “no de-identification process eliminates that risk entirely” (source). Our is it safe page, the legal and privacy checklist and the consent guide cover the rest.

What makes company data stronger?

Buyers do not publish scoring formulas. They do publish entry criteria and preferences, and these point in the same direction.

Years of history

Handshake AI looks for companies with “20+ full time employees & 3+ years in operation” (source). Korra wants “Two or more years of records, still in their tools” (source). Endless Data asks for several years of history (source), and micro1 says several years of records across multiple business systems are generally a stronger fit, with no fixed minimum volume (micro1 buyer page). Polyshares lists a short operating history among the things that cap value (source).

Korra’s phrase “still in their tools” matters. If old workspaces were deleted or exported to an archive nobody can search, the usable history is shorter than the company’s age. If you are winding down, preserve the systems before you cancel them; see closing a company.

Enough people

micro1 asks for 30 or more full-time employees (source) and Endless Data for roughly 30 or more (source). Korra lists fewer than about 20 staff as not a fit, prefers 50 or more and says “100 or more is best” (source). Polyshares takes the other view: “The length and texture of the record matter more than headcount or revenue” (Polyshares buyer page). Headcount is a proxy for how many conversations, handoffs and decisions the record contains.

English

micro1 wants English data and English-speaking representatives (source). Korra asks for 90% or more of work in English (source). Endless Data requires documentation in English, and the Polyshares intake asks what share of your data is in English (Endless Data buyer page, Polyshares buyer page).

Threads that show work being done

Endless Data asks for records “and the decisions behind them” (source). Mercor lists Slack threads and “Gmail messages and metadata” (source). One person’s notes show what was done. A thread where staff, a customer and a supplier work a problem through to an outcome shows how it was done, and the metadata (who replied to whom, and when) keeps that structure intact.

Several systems describing the same work

A ticket that links to a code change and to the chat where the fix was agreed tells a fuller story than any of the three alone. micro1’s preference for records “across multiple business systems” points the same way (micro1 buyer page).

How do you check your own data?

Before you talk to any buyer, write down four things for each system: how many years it covers, how many people used it, what share is in English, and what must stay out. Then run the value check. It asks ten questions, never asks for data or logins, and returns an estimated fit and the programs that match. It is not a price.

Compare entry criteria side by side on the tracker and the compare page, then read how to prepare your data. For the full deal process, start with how licensing company data to AI works.