Home / Guides / Sell company data to AI
How to sell your company’s data to AI companies
AI labs and the vendors that supply them license established companies’ operating records, such as email, documents, chat, CRM and accounting history, to train AI agents to do office work. Buyers publish headline ranges sized mostly by head count, but offers are set after each buyer’s own review, terms vary a lot, and many companies get no offer.
What is being bought, and why
Buyers want the private trail of how a real business runs: who asked for what, what went wrong, how it was fixed, and what it cost.
AI labs are training agents to do real office work, such as answering a customer, updating a file or raising an invoice. To practise, an agent needs realistic workplaces, which the industry calls RL environments: simulated companies built from real companies’ records. Public text is plentiful; years of one company’s connected email, tickets, documents and invoices are not.
Variety and connections matter more than volume. A firm that resells company data told MIT Technology Review that a ticket linked to the work that resolved it is worth more than an isolated document, and that one company’s varied, deep conversations can be worth more than another’s Slack, email and calendars combined.
What buyers typically ask about:
- Email and calendar (Google Workspace, Microsoft 365)
- Documents and shared drives, wikis and SOPs
- Chat (Slack, Teams)
- CRM and sales records; accounting and billing
- Project and ticketing tools; call and meeting recordings
- Industry systems (ERP, TMS, CAD, property management) and, for software firms, code history
Who buys
Four kinds of organisation buy or broker company records. Most programmes open to operating companies are run by training-data vendors, not by the labs themselves.
- Training-data vendors and RL-environment builders. Firms that build training environments and datasets and sell them to labs. Most published programmes for operating companies come from this group. Builders charge labs from about $20k for a simple website replica to about $300k for a Slack-grade replica (Epoch AI), which is why realistic records are worth paying for.
- Foundation labs. They buy some data directly, usually in large lots. MIT Technology Review reported that one lab outbid a vendor for a bankrupt airline’s internal records.
- Brokers and marketplaces. They list or package companies’ records and sell them on to labs and vendors, usually taking a cut or charging the buyer. Read whose contract you would sign and who pays them.
- Wind-down services. Firms that close down startups and sell their remaining assets, now often including code and workspace data (MIT Technology Review).
Published buyer minimums: Nyne asks for 20+ employees and 3+ years in business; micro1 wants 30+ employees; Scale prefers 40+; Mode asks for 30+ in most industries. (each buyer’s own page, read 2026-10-18). Most also look for three or more years of operating history and US companies first.
What it pays
Published headline ranges run from the tens of thousands of dollars to the low millions; reported deals are smaller and few.
| What was sold or offered | Published figure | Source |
|---|---|---|
| Company records, priced by staff sizeReplay’s headline bands; it says head count is only a proxy. A buyer’s claim, not a reported deal. | $10K–$100K for 20–50 employees; $100K–$1M for 50–250; $1M+ for 250+ | Replay’s own page, read 2026-10-18 |
| Data programme for operating companiesDescribed by the buyer as indicative, subject to a full review. | $100K–$5M | Mode’s own page, read 2026-10-18 |
| Licence of an operating company’s dataPublished by the buyer. The top tier is for “highly unique” data. | Tiers of $100k+, $500k+ and $1M+ | micro1’s own page, read 2026-10-17 |
| Data licence from an operating companyDescribed by the buyer as illustrative. | “$10K–$1M+” per agreement | Scale AI’s own page, read 2026-10-17 |
| One data point: code and workspace data of shut-down startupsTypical payouts across 150+ deals by a wind-down service. | $10k–$100k per company | SimpleClosure, reported by MIT Technology Review |
| 13 years of Slack, email and Jira from one company | “Hundreds of thousands” of dollars | Forbes, via Slashdot |
Headline ranges are the top of a buyer’s marketing, not a likely offer. Size is only a proxy: buyers also price on how far back records go, how many systems hold them, how unusual the work is and who owns it. The full set, with the change log, is on the price index; how our own estimate is built is on the methodology page.
How a deal works
Describe the records, get a valuation, sign an NDA, agree scope and terms, then de-identify and deliver a copy. Payment follows delivery or signing, depending on the contract.
Describe what you hold
Which systems, how many years, how many people, what share is in English, and who owns the records of client work. No data changes hands at this stage.
Valuation and NDA
The buyer (or a matchmaker such as us) gives an indicative range. Before any detail is shared, both sides sign a mutual NDA.
Scope and contract
You choose which systems and years go into the copy and what is left out. The contract sets the price, exclusivity, permitted uses, de-identification standard, warranties and payment.
De-identification and export
Personal data, credentials and anything excluded are removed, and people are replaced with consistent pseudonyms. Ask who runs the export, where the copy is processed and who checks the result before it leaves.
Delivery and payment
Published timelines range from about 30–60 days from intake to payment (Corpus’s own page, read 2026-10-18) to about three months (Mode’s own page, read 2026-10-18).
What to keep out
Leave out anything that belongs to someone else or would harm someone if it leaked, whatever the buyer says it can scrub.
- Personal data of staff, clients and contacts that cannot be reliably removed, and anything identifying a private individual
- HR, payroll, medical and disciplinary records; private and direct messages
- Passwords, API keys, tokens and other credentials
- Payment card, bank and government ID numbers
- Client work your contracts say the client owns, or that is under a confidentiality or NDA obligation
- Privileged legal communications. Above the Law argued that removing names does not remove privilege, which belongs to the client (Above the Law).
- Third-party content you only licence, such as paid reports, stock assets or software you don’t own
Our data-handling pledge lists what must be removed before any transfer we are involved in.
Questions for your lawyer
We don’t give legal advice. These are the questions owners take to their own counsel before signing.
- Do our client contracts give clients ownership of, or confidentiality over, the records of work we did for them?
- Do our employee handbook, privacy notice or employment contracts cover licensing a de-identified copy of work email and chat? Do we need to tell staff, or ask them?
- Which privacy laws apply to the people in these records (staff, clients, contacts), given where they live? Would this licence count as a “sale” or “sharing” of personal information under any of them?
- Is the buyer’s de-identification standard adequate for those laws, and who carries the risk if someone is re-identified?
- Do any of our software agreements (email, chat, CRM, accounting) restrict exporting our own data or using it this way?
- What warranties and indemnities are we giving, and are they capped?
- Is this a licence or a sale, is it exclusive, for how long, and can the buyer pass the data on?
- How is the payment taxed (ask your accountant as well)?
Platform terms to check
Some software platforms restrict what apps built on their APIs may do with your data. Read the current terms for every tool in scope.
- Slack. Slack’s Marketplace guidelines list apps that “use Slack data to train Large Language Models (LLMs)” and apps that “export or backup message data” among those it does not accept. (Slack Marketplace guidelines, read 2026-10-18). Slack’s API terms changed in May 2025 to restrict bulk export and LLM training on data pulled through the API (Cloud Wars).
- Google Workspace. Google’s Workspace API user data policy bars apps from “transferring or selling user data to third parties like advertising platforms, data brokers, or any information resellers”, and from using it to create, train or improve a machine-learning or AI model beyond that user’s own personalised model. (Google Workspace API user data policy, read 2026-10-18).
- Everything else. CRM, accounting and project tools have their own terms on exports and API use. A buyer’s app that pulls data through an API is not the same route as an owner exporting the company’s own data, but how each set of terms applies to you is a question for your lawyer.
Go direct or run a round?
Going direct is fine when one buyer clearly fits. When several might, decide before you contact any of them, because applying alone gives some things up.
Applying to one buyer
- One bid, with nothing to compare it against
- That buyer’s own contract terms and price review
- Possibly exclusivity, which can rule out licensing the same records elsewhere
- Once a buyer has been in contact with you, some buyers’ referral rules mean a later introduction to them no longer counts, and your position in any round changes
A round
- One valuation, described to every buyer that fits, at the same time
- Several bids and terms side by side: price, exclusivity, de-identification, payment
- You choose which buyers see your profile; nothing is shared without your OK to each one
- No data leaves your systems until you sign with the buyer you pick
Exclusivity is worth pricing carefully: in the related market for training tasks, exclusive deals cost labs 4–5 times as much (Epoch AI).
If a buyer licenses your records, Quiet Asset may receive a referral fee from that buyer. It is paid by the buyer, not deducted from your offer, and you never pay us. TODO(owner): confirm the fee model and the written disclosure with a lawyer before launch
Run a round for your records
Start with a free, indicative valuation. About 4 minutes; nothing is uploaded.
FAQ
Can a company sell its data to AI companies?
Yes. Several AI labs, training-data vendors and brokers publicly license operating records from established companies: email, documents, chat, CRM, accounting and project history. Whether your company may license its own records depends on its contracts and the law where it operates, which is a question for your own lawyer.
How much is company data worth to AI buyers?
Buyers’ published headline ranges for an operating company start in the tens of thousands of dollars and reach the low millions, sized mostly by head count. They are claims, not deals. Reported deals so far are mostly five and six figures. Many companies get no offer. Our price index lists every public figure with its source.
Do I have to give a buyer access to our systems?
Not in the published programmes we have read. The usual route is an agreed scope, then a de-identified copy exported by your own team or under your control. A valuation from us needs no access at all: it uses only your answers.
What happens to staff and client names?
Personal data, private messages, HR records and credentials should be removed before any transfer, and people replaced with consistent pseudonyms. Ask each buyer for its de-identification standard in writing, and have your own lawyer review it.
Is a licence exclusive?
It depends on the contract. Published programmes include non-exclusive licences of a copy, where you keep your records and can license them again, and exclusive deals, which pay more but limit what else you can do with the same records.
Should I contact a buyer directly?
That is fine when one buyer clearly fits and you have compared its terms. If several might fit, decide before you contact any of them: a single application gets one bid on that buyer’s contract, and once a buyer has been in touch, some introductions and referral rules no longer apply.
How does Quiet Asset get paid?
If a buyer licenses your records, Quiet Asset may receive a referral fee from that buyer. It is paid by the buyer, not deducted from your offer, and you never pay us.