How do you sell AI training data?

Companies sell AI training data by licensing de-identified operational records, either to a training-data company such as micro1, Scale AI or Mercor, or through an agent or revenue-share program such as Troveo. The work runs in order: data audit, rights review, a written scope, de-identification, then delivery and acceptance before payment.

How do I download Slack channel data?

Workspace Owners and Admins export from Admin, then Workspace settings, Security and Import & export data, per Slack's help center (read October 25, 2026). Free and Pro plans export public channels only, and Free includes file links from the last 90 days. Private channels and DMs need Business+ or Enterprise and an application to Slack.

How do you license business data for AI training?

You grant a buyer the right to use a defined set of records for a defined purpose and term, while keeping ownership. Write the systems, years and exclusions into the agreement, then settle exclusivity, deletion, the position on trained models and the payment date. Labelbox Prism, for example, is exclusive by default and pays an upfront fee plus a revenue share.

Where can I find AI training data?

AI labs get much of their company workflow data from training-data companies that license it from operating businesses. micro1, Scale AI, Mercor, Handshake AI and Labelbox Prism all run programs that license company records, and micro1 says it licenses only to labs and enterprises in the US and allied countries (read October 25, 2026).

Does my company need to anonymize data before selling it?

Not always. micro1 says companies don't need to anonymize before transferring because it de-identifies sensitive information itself (read October 25, 2026), while Scale AI anonymizes inside your environment before extraction. Either way, agree in writing who removes identifiers, to what standard, and whether that happens before data leaves your systems.

Can AI sell your data?

An AI company can pass your company data on only as far as your license allows. Training-data companies exist to supply AI labs, so onward licensing is normal: micro1 commits not to resell or distribute data until identifying information is anonymized, and Endless Data licenses pooled data to several buyers. Read the onward-transfer clause before signing.

How to Sell Company Data to AI: 2026 Audit and Licensing Process

How to sell data to AI companies, run as an ops project: data audit, rights review, offers, de-identification spec, license terms and delivery, checked October 25, 2026

Selling company data to AI companies is a licensing project, not a file sale. An ops lead audits which systems hold firm-owned records, clears client, employee and platform rights, collects two or three offers on one written scope, has the buyer de-identify that scope, and delivers only after the license fixes exclusivity, deletion and payment dates.

Payment usually follows delivery and acceptance. On October 25, 2026, micro1 said it pays within 30 days of invoicing once it has approved the data, Polyshares said net 30 to 60 days, and Mercor said its wire typically arrives 2 to 4 weeks after signing. Published offers ranged from $10K to “$4M+”.

Onfolio may earn a referral fee if you sign up to Polyshares through our links. This does not affect our editorial assessment.

This guide is written for the person who would run the project, usually an ops, RevOps or IT lead. Each phase below has an output and a gate, so the project can stop cleanly at any point before data leaves your systems.

The process at a glance

Seven phases, in order. The owner column is our suggestion, not a buyer requirement.

Phase	Suggested owner	Output	Gate before the next phase
1. Data audit	Ops or IT	System inventory with years and record counts	Every system has a named owner and a count
2. Rights review	Legal, with ops	Exclusion list covering clients, people and platforms	Client-owned material is tagged
3. Buyer model	Owner or CEO	Shortlist of 2 to 3 buyers	You know which entity signs and pays
4. Indicative offers	Owner, with ops	Written offers on the same scope	Offers compare like for like
5. Scope and de-identification spec	Ops and legal	Data manifest and anonymization standard	You’ve reviewed a de-identified sample
6. License terms	Legal and owner	Signed license	Exclusivity, deletion and payment dates are in writing
7. Delivery and payment	Ops or IT	Transfer log, acceptance, invoice	Acceptance confirmed in writing

Phase 1: Run the data audit

Count before you export. Buyers price from systems, years and record counts, and the audit is also where you find the client and personal data you’ll need to exclude.

Record type	Typical systems	What to count	Why buyers care
Team chat	Slack, Teams, Google Chat	Years, channels, messages, share in shared client channels	Scale AI wants “full conversation histories, threads, participants, and reply structure preserved intact”
Email and calendar	Gmail, Outlook, Calendar	Mailboxes, years	Named by Scale AI, Mercor, Polyshares and Troveo
Docs and knowledge base	Drive, SharePoint, Confluence, Notion	SOPs, playbooks, templates, versions	micro1 asks for SOPs, knowledge bases and process documentation
Projects and tickets	Jira, Asana, Linear, Zendesk, ServiceNow	Projects, tickets, years	Named by Scale AI and Troveo; Polyshares lists support tickets
CRM	Salesforce, HubSpot	Deals, activities, notes, years	micro1’s William Almeida told Business Insider that without Slack, CRM or email records “the rest of the corpus is less valuable”
Finance and ERP	Accounting, ERP, invoices	Years, entities	Named by Polyshares and Troveo
Data warehouse	BigQuery, Snowflake, Databricks	Tables, years	Named by Scale AI and Troveo
Code and design	GitHub, Figma	Repositories, files	Mercor names GitHub; SimpleClosure AssetHub handles source code and Figma

For each system, also record the date of the first record, active users over time, the share of content in English, and whether client material sits in the same place. micro1 and Endless Data both want mainly English documentation, and micro1 says several years of records across multiple systems is “a stronger fit.” Mercor says connecting more tools generally raises the payout.

Slack is the system most buyers ask about, and your plan limits what an admin can export. From Slack’s help center, read October 25, 2026:

Slack plan	What admins can export
Free	Public channels, with file links from the last 90 days only
Pro	Public channels
Business+	Public channels; private channels and DMs only after Workspace Owners apply for a self-serve export tool, which Slack frames as “for legal and regulatory compliance”
Enterprise	Custom exports by conversation type, member or workspace, after Org Owners request access

Slack exports contain messages and file links, not the files themselves, so files sit in a separate step.

Phase 2: Clear rights and write the exclusion list

Tag everything your firm doesn’t own or can’t strip of people, and write it down as an exclusion list. It goes into every offer request and, later, into the license.

Exclude	Where it hides	Rule or source
Client deliverables and client files	Shared drives, client folders, project tools	Polyshares’ license “does not cover what you hold on behalf of your own clients”; RICHT law firm (October 18, 2026) notes many agency contracts carry a “no client data used to train AI” warranty
Shared client channels	Slack Connect, shared Teams channels	Client-owned conversations, with the client’s own people in them
Contact-level CRM data	Contacts, email signatures	CCPA/CPRA has covered B2B contact and employee data since January 1, 2023
HR and payroll	HR systems, payroll, performance reviews	Excluded by default at Labelbox Prism, Endless Data and SimpleClosure
Employee talk about pay and conditions	Chat, email	Protected under NLRA Section 7
Recorded calls	Call recorders, sales tools	State consent rules; Brewer v. Otter.ai (August 2025) alleges wiretap violations
Credentials and secrets	Code, tickets, chat	SimpleClosure (October 1, 2026) flags live credentials
Third-party licensed data and open-source code	Warehouse, repositories	Governed by someone else’s license terms
Regulated content	Health records, privileged legal matters	Endless Data never licenses protected health information or privileged client matter content

Platform terms need their own check. Under Slack’s API terms (effective October 10, 2025, read October 25, 2026), an app offered to other organizations may not “use API Data to train a large language model,” may not bulk export message and file data unless an additional agreement allows it, and may not use data from one organization to benefit a different one. That clause covers app providers, which is what a buyer’s connector is. It doesn’t settle whether your own admin export can be licensed, so read your Slack customer agreement as well. Google Workspace API policies restrict using data to train non-personalized models, and the FTC said in February 2024 that retroactively changing privacy terms to allow AI training can be deceptive.

Phase 3: Choose a buyer model

Pick the model by how you want to be paid and where you want de-identification to happen. Direct buyers pay a fixed fee, agents and marketplaces take a cut, and revenue-share programs pay as your data is licensed.

Model	Examples	How data leaves your systems, as published	Payment pattern
Direct buyer	Polyshares, micro1, Scale AI, Mercor, Endless Data	Mercor connects to your tools; micro1 takes transfers “through our APIs or other approved methods”; Scale AI anonymizes inside your environment; Polyshares routes data through a specialist PII partner	A fixed fee after delivery and acceptance
Agent or marketplace	SimpleClosure AssetHub; Handshake AI, which says it connects businesses with AI labs	AssetHub starts from a short form or a GitHub connection and processes every dataset for PII before licensing	Price minus a fee; AssetHub’s percentage isn’t published
Revenue share	Troveo; Labelbox Prism (upfront fee plus revenue share)	Troveo works from your exports, with no OAuth or integrations; Prism removes PII, then transfers with customer-managed keys	Recurring, with no fixed total

Two points shape the shortlist. Endless Data says “AI labs don’t license data from individual companies,” and the training-data companies above are how most operating companies in our sources reached the labs. And introducers are often paid. Referral bonuses published by the buyers themselves include up to $50,000 from Scale AI if it signs the company, $10K from Handshake AI at closing, and, from micro1, “$50,000 for deals up to $700,000, or 7.5% of the deal value for deals above $700,000,” paid after the company’s data is transferred and approved (read October 25, 2026). Our nine-buyer comparison covers each program’s data handling side by side.

Phase 4: Request indicative offers on one scope

Send two or three buyers the same scope sheet and compare their written offers line by line. Without a common scope, the prices aren’t comparable.

Prices for similar records vary because, as MIT Technology Review put it on October 13, 2026, valuations rest on “research theses.” The scope sheet should include:

List each system with its first and last record dates.
Give record counts per system.
State headcount history and the share of records in English.
Attach the exclusion list from phase 2.
Say whether you’d accept exclusivity, and for how long.
Name the delivery routes you’ll allow: connector, your own export or in-environment processing.

Published turnaround is short: one business day for a Labelbox Prism preliminary valuation, one to two weeks from intake for a Polyshares offer, and days or hours for a SimpleClosure viability assessment. Endless Data starts with a 20-minute call, then a data manifest you approve.

Put the answers in one comparison sheet with these columns: price, fixed fee or revenue share, exclusivity, term, deletion, trained-model position, payment trigger and date, the signing entity, and the named end buyers or buyer rules.

Phase 5: Write the scope and de-identification spec

Turn the chosen offer into a data manifest: what’s in, what’s out, which identifiers are removed, what’s kept, where the work happens and how long raw data is retained.

Spec item	What to write down	What buyers publish
Systems and dates	Each system and its date range	Endless Data’s manifest lists systems, years, workflows in scope and exclusions, for the seller to approve
Identifiers removed	Names, emails, phone numbers, account numbers, addresses	Polyshares says these “are removed or replaced”
What’s kept	Roles, timestamps, thread structure	Polyshares keeps “Roles, timestamps, and the work itself”; Scale AI keeps threads, participants and reply structure
Method	Tokenization, pseudonyms or synthetic rewrites	Prism tokenizes names, emails and internal IDs; Troveo uses pseudonyms such as PERSON-014; micro1 says some datasets “may undergo synthetic rewrites”
Where it happens	Inside your environment, or after transfer	Scale AI: inside your environment before extraction. Endless Data: before anything leaves your company. Polyshares: a PII partner, before Polyshares or a lab sees it. micro1: you transfer, then micro1 de-identifies
Sample review	A de-identified sample of your own data	micro1 and Endless Data offer representative samples
Raw-data retention	A deletion date or schedule	micro1: raw data deleted on an agreed schedule. Prism: deletion within 30 days on defined triggers
Security evidence	Reports and sub-processors	Mercor states SOC 2 Type II; Prism uses customer-managed keys; Scale AI keeps an audit trail

Phase 6: Negotiate the license terms

Get these points in writing before any export. Each one has a published position from at least one buyer to anchor the conversation.

Price exclusivity separately. Epoch AI (January 12, 2026) reports that exclusive AI training environments cost buyers 4 to 5 times as much as non-exclusive ones. Polyshares and Prism are exclusive by default, Endless Data is non-exclusive, and Troveo’s exclusivity is time-limited.
Fix the term. Polyshares’ data-handling page says “a defined term,” while its four published case studies are “perpetual exclusive” licenses.
Set a deletion deadline and a certificate. Prism’s kill clause forces deletion within 30 days for material breach, regulatory change or a change-of-control trigger you define up front.
Limit future training, because past training can’t be reversed. Prism says models already trained “cannot be unlearned,” and Endless Data’s buyer terms let buyers keep using models trained during the license.
Name the onward buyers, or the rules for them. micro1 licenses only to US and allied buyers. Polyshares doesn’t disclose which labs take which data. Troveo introduces buyers confidentially after signing.
Tie payment to a date. micro1 allows itself 30 days to approve the data, then pays within 30 days of your invoice, so the outer bound is about 60 days after transfer by our arithmetic.
Check M&A effects. A prior data license and its exclusivity or change-of-control terms surface in due diligence (Frankfurt Kurnit via Mondaq, October 2026).
Get the seller-side contract early. Neither Polyshares nor Endless Data publishes seller terms we could read: Polyshares’ standard terms link led to a not-found page on both October 21 and October 25, 2026, and the Endless Data terms online cover its buyers.

Phase 7: Deliver, confirm acceptance and invoice

Deliver only what the manifest covers, through the agreed route, and log it. Payment clocks usually start at acceptance, so get acceptance in writing.

Delivery route	Buyer examples	Ops checks
Tool connectors	Mercor (Slack, Teams, Gmail, Salesforce, GitHub, Drive, Notion and “100+ supported tools”)	Limit access scopes to the manifest; revoke access after delivery; ask about Slack’s API terms
In-environment processing	Scale AI (on-premises and VPC options)	Who gets access, and the audit trail
API or approved transfer	micro1	Encryption in transit and a transfer log
Your own exports	Troveo, Endless Data	Export only the manifest’s scope; scan for credentials first
Form or code connection	SimpleClosure AssetHub	Which repositories are selected; secrets removed

Buyer	What starts the payment clock	Published timing
micro1	Approval, then your invoice	Up to 30 days to approve, then within 30 days of invoice (read October 25, 2026)
Polyshares	Approval, anonymization and delivery	Net 30 to 60 days
Mercor	Delivery of the anonymized dataset to the buyer	Typically 2 to 4 weeks after signing
Labelbox Prism	Signing, then downstream use	Upfront fee at signing, plus a revenue share
Endless Data	A fixed sum agreed before anything is shared	Timing after signing not published
Troveo	Each sale	Not published
Scale AI, Handshake AI, SimpleClosure AssetHub	Not published	Not published

Close the project with three records: written acceptance, confirmation that connector access is revoked, and, at the end of the term, a deletion certificate.

What sells and what doesn’t, system by system

The licensable part of each system is the firm’s own record of how work gets done. The excluded part is anything a client owns or that identifies a person.

Source	Usually licensable	Usually excluded
Slack or Teams	Internal channels about how work gets done, de-identified	Slack Connect and shared client channels; HR channels
Email	Internal process threads	Client correspondence under NDA; contact details
Drive or SharePoint	SOPs, playbooks, templates, training material	Client deliverables; contracts with confidentiality clauses
Project tools	Tasks and tickets with client names removed	Tickets holding credentials or client environment details
CRM	Pipeline stages, activity patterns, notes with contacts stripped	Contact-level personal data
Finance	Process documentation and de-identified workflows	Client books and proprietary pricing logic, both on Endless Data’s exclusion list
Design	Assets made for your own business; MyPortal Marketing licensed its own Figma and Canva files (Business Insider)	Client brand work
Call recordings	Rarely licensable	Recordings without the consent your state requires

Price ranges, with sources

Advertised ranges run from $10K to “$4M+”, but the deals you can check are mostly in the five and low six figures. Figures below were read on official pages on October 21, 2026 and re-checked October 25 unless noted.

Buyer	Published range	Pay structure
Scale AI	$10K to $1M+, “illustrative value per data partnership”	Not published; says recurring revenue is possible
Labelbox Prism	About $50K for targeted datasets, up to $1M+	Upfront fee at signing plus revenue share
SimpleClosure AssetHub	“Tens to hundreds of thousands of dollars” (October 21 only)	Percentage on success
Polyshares	$100K to $2M+; four published deals of $110,000 to $450,000	Fixed license fee, net 30 to 60 days
micro1	“$200k-2M+”; payments “typically ranging from $150K to over $2M” (October 25)	Fixed fee after approval and invoice
Handshake AI	“$100K to $4M+”	Not published
Endless Data	“Up to $4M,” no typical range	Fixed sum agreed before sharing
Troveo	Full-company deals “start at six figures”	Revenue share on every sale
Mercor	No range published	Wire on delivery to the buyer

Checkable deals sit lower. The Information reported (headline only) about $150,000 for a 14-person HVAC company, MIT Technology Review (October 13, 2026) described a wound-down startup getting about $25,000 for its code with about $15,000 pending for workspace data, and AIData.Broker, a sell-side broker, puts the median of the few public sales at $31.6K in its own compilation. For context, Epoch AI reports AI labs pay environment vendors $200 to $2,000 per training task. The full price evidence is on our company data value page.

Qualification screen

Screen yourself on the four signals buyers publish: headcount, years of history, language and region.

Buyer	Headcount	History	Language and region
micro1	30+ full-time employees	“Substantial historical operational data”	Primarily English; US first, then other Western markets
Scale AI	40+ employees preferred, case by case	Established data systems and real workflows	Not published
Handshake AI	20+ full-time employees (referral page)	3+ years in operation	Not published
Endless Data	Roughly 30+ employees	Several years	English; US, UK and Canada prioritized
Polyshares	None published	“Real operating history,” any sector	Not published
Labelbox Prism, Troveo, Mercor	None published	Not published	Not published
SimpleClosure AssetHub	Startups	Operating or closing	Not published

Red flags for the ops team

Stop the project if a buyer wants raw exports before pricing, sets a 48-hour deadline, or can’t say who signs, who pays and who receives the data.

Full export before an offer. SimpleClosure (October 1, 2026) warns against anyone wanting a full export before a meaningful offer. Counts and samples are enough to price; Polyshares says nothing leaves your systems until you accept its offer.
Raw data leaving before de-identification, without the contract saying so. Where scrubbing happens differs by buyer (phase 5). If raw data moves first, the license should say who holds it and for how long.
Connector access wider than the manifest. A connector that can read every channel can reach the client and HR material you excluded.
48-hour deadlines. Business Insider (October 13, 2026) reports short “exploding” offers have become common in this market. It doesn’t attribute them to any company named here.
An unclear buyer chain. SimpleClosure’s checklist asks four questions: which company signs, who funds the purchase, who gets the files, and who owes you the money. A buyer that can’t answer all four isn’t ready.
Headline-only pitches. SimpleClosure also warns about new buyers advertising $1M to $5M.
Weak provenance on your side. Buyers have met sellers without rights: Surge AI’s Nick Heiner described to Business Insider a group trying to sell a codebase it didn’t own. Expect rights warranties, and keep the phase 2 records that back yours.

Deciding whether to sell at all? onfolio.com’s owner’s checklist for selling company data to AI labs covers the same decision from the owner’s side: rights, buyer types, red flags and what you might get.

Sources

micro1 data partnership page, read October 21 and October 25, 2026: https://www.micro1.ai/data-partnerships
micro1 company referral page, read October 21 and October 25, 2026: https://www.micro1.ai/company-referral
Scale AI data partnership page, read October 21 and October 25, 2026: https://scale.com/data-partnership
Mercor data page, read October 21 and October 25, 2026: https://www.mercor.com/data/
Handshake AI data partnerships and referral pages, read October 21, 2026 (main page re-checked October 25): https://joinhandshake.com/ai/data-partnerships
Labelbox Prism FAQ, read October 21 and October 25, 2026: https://prism.labelbox.com/faq
Polyshares homepage and data-handling page, read October 21, 2026 (homepage and terms link re-checked October 25): https://polyshares.com/data-handling
Endless Data homepage and Terms and Conditions, read October 21, 2026 (homepage re-checked October 25): https://endless-data.com
Troveo business data page, read October 21 and October 25, 2026: https://www.troveo.ai/business-data
SimpleClosure AssetHub, read October 21, 2026: https://simpleclosure.com/product/assethub/
SimpleClosure, “How Real Are Those Million-Dollar Data Offers?”, October 1, 2026: https://simpleclosure.com/blog/guides/how-real-are-those-million-dollar-data-offers/
Slack API Terms of Service, effective October 10, 2025, read October 25, 2026: https://slack.com/terms-of-service/api
Slack Help Center, “Export your workspace data,” read October 25, 2026: https://slack.com/help/articles/201658943-Export-your-workspace-data
Business Insider, “AI labs are fueling a wild market for your Slack DMs,” October 13, 2026: https://www.businessinsider.com/ai-training-employee-data-emails-slack-dms-2026-10
MIT Technology Review, “AI labs are buying data from dead startups to train their models,” October 13, 2026: https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models/
The Next Web, “micro1 commits $1bn to buy company data for training AI agents,” October 9, 2026: https://thenextweb.com/news/micro1-1bn-enterprise-data-ai-training-citi-hercules
Epoch AI, “An FAQ on reinforcement learning environments,” January 12, 2026 (interview-based): https://epoch.ai/gradient-updates/state-of-rl-envs
AIData.Broker public price claims, read October 21, 2026: https://aidata.broker
The Information, headline on an HVAC company data deal, 2026 (headline only)
RICHT law firm on agency client data, October 18, 2026
Frankfurt Kurnit via Mondaq on data licenses in M&A, October 2026
FTC guidance on retroactive privacy policy changes, February 2024
Brewer v. Otter.ai complaint, August 2025
Google Workspace API user data policy

FAQ

How do you sell AI training data?: Companies sell AI training data by licensing de-identified operational records, either to a training-data company such as micro1, Scale AI or Mercor, or through an agent or revenue-share program such as Troveo. The work runs in order: data audit, rights review, a written scope, de-identification, then delivery and acceptance before payment.
How do I download Slack channel data?: Workspace Owners and Admins export from Admin, then Workspace settings, Security and Import & export data, per Slack's help center (read October 25, 2026). Free and Pro plans export public channels only, and Free includes file links from the last 90 days. Private channels and DMs need Business+ or Enterprise and an application to Slack.
How do you license business data for AI training?: You grant a buyer the right to use a defined set of records for a defined purpose and term, while keeping ownership. Write the systems, years and exclusions into the agreement, then settle exclusivity, deletion, the position on trained models and the payment date. Labelbox Prism, for example, is exclusive by default and pays an upfront fee plus a revenue share.
Where can I find AI training data?: AI labs get much of their company workflow data from training-data companies that license it from operating businesses. micro1, Scale AI, Mercor, Handshake AI and Labelbox Prism all run programs that license company records, and micro1 says it licenses only to labs and enterprises in the US and allied countries (read October 25, 2026).
Does my company need to anonymize data before selling it?: Not always. micro1 says companies don't need to anonymize before transferring because it de-identifies sensitive information itself (read October 25, 2026), while Scale AI anonymizes inside your environment before extraction. Either way, agree in writing who removes identifiers, to what standard, and whether that happens before data leaves your systems.
Can AI sell your data?: An AI company can pass your company data on only as far as your license allows. Training-data companies exist to supply AI labs, so onward licensing is normal: micro1 commits not to resell or distribute data until identifying information is anonymized, and Endless Data licenses pooled data to several buyers. Read the onward-transfer clause before signing.