Last updated: 7 October 2026
Short, plain answers to the questions owners and finance leads ask first, and the deal terms you will meet in any buyer's paperwork. Every number comes from the buyers' own published pages, from named news reports or from practitioners we name as such.
Not all company data is worth the same. Our research shows AI labs pay most for the records of real work, held in the systems below. We identify which of these systems each company runs, and it is one of the most important factors in the AI Data Value Score.
Every company comes with the systems we identified, shown as tags. Company level only: we show which systems a company runs, never what is inside them.
It means licensing a copy of your business records, such as documents, messages, tickets, CRM notes or code, to an AI company or lab so it can train or evaluate models. In most programs you license agreed uses of a copy rather than handing over the records themselves. Mode says it buys an agreed copy and originals stay with the company; micro1 states that companies retain ownership of their underlying data. What you are really selling is permission, on written terms, to learn from how your company works.
The US has more than 30 million businesses, but most are one-person businesses with no staff and no records a data buyer would license. We selected our US companies from our index of 120 million domains with fixed criteria, leaving out the companies least likely to hold data worth licensing: personal sites, parked domains and the smallest businesses. Every company that remained, about 1 million, was checked as active and scored from 0 to 100. Our lists come from that set.
Mostly data companies that run buyer programs, such as micro1, Mode and Miro Advisory, which then supply AI labs. Some labs also take data directly for large or unique datasets: Google runs an intake at contentpilot.google.com and OpenAI has a data partnerships page. For a company of 20 to 500 people, the buyer programs are usually the realistic door, because they publish eligibility rules, handle preparation and deal with many labs at once.
An AI lab builds and trains the models. A data company sources, prepares and packages data, evaluations and training environments, then sells them to labs. Labs buy at scale and tend to run formal procurement with NDAs, master agreements and purchase orders, which suits platforms and very large datasets. Data companies run programs aimed at ordinary businesses, set published entry criteria and do the de-identification and packaging that a single company could not do on its own.
Records of real work. micro1 lists SOPs, knowledge bases, internal documentation, CRM data, project histories and QA processes, plus decision-making patterns and human feedback on AI outputs. Buyers value connected histories most: records where a request, the discussion, the decision and the result can be followed across systems such as email, Slack or Teams, Jira, Zendesk, Salesforce or QuickBooks. A scattered folder of finished files says much less about how the work was done.
Public material that anyone can already collect, files duplicated many times across drives, and very small collections that do not show a process from start to finish. Contact lists and customer lists are also not the product; buyers want the reasoning in the work, not the names. Data you do not clearly own, or that you hold under strict client confidentiality, may be worth nothing to you because it cannot be licensed at all.
As published on 7 October 2026: micro1 lists 30+ employees, and a referral posting describes companies of 30 to 200; Mode lists 20+ full-time US office employees, 10+ for accounting firms and 6+ for law firms, with several years of records the company owns. Size is a proxy for how much connected history exists. The eligibility checker applies these published rules to your answers in your browser and sends nothing.
More: how to sell data to AI companies, step by step.
Sometimes. micro1 says it accepts referrals globally but prioritizes US companies followed by other Western markets, with the strongest demand from the United States, the United Kingdom and Canada, and it asks for documentation and communication in English. Mode says US-based teams are the strongest fit. EU and UK sellers also face GDPR questions on lawful basis and international transfer, which is one reason programs prioritize US companies.
Nobody can price your data without seeing it. Published ranges, checked 7 October 2026: micro1 $100k+, $500k+ and $1M+ tiers, and $100K-$2M+ for approved data packages on its referral page; Mode $100K-$5M; Miro Advisory $100K-$1M+ for operating datasets and $10K-$1M+ for private codebases, both indicative. These are ranges and floors, not averages or offers. Volume, connected history, quality, uniqueness and how clean the rights are move the number.
It varies by deal. Some are one-off payments for a delivered and accepted dataset; others involve ongoing participation paid over time. Many tie payment to milestones such as signing, delivery and acceptance, which makes the acceptance criteria as important as the price. Ask whether payment is fixed or depends on review, what triggers each installment, and whether any amount can be reduced or reclaimed later. Get every answer into the agreement.
Weeks to months. A deal runs through NDA, buyer review, agreement, export, de-identification and acceptance, and practitioners cite 60 to 90 days to close. Legal review on your side, client contract checks and staff communication can add time. Plan for months, not days, and do not let a deadline push you past questions you still need answered.
Practitioners say so. Raw data is the cheapest tier; evaluations built on the data are worth roughly ten times raw; full training environments can reach six to eight figures but need heavy engineering. The jump in value comes from expert work: someone who knows the job defines tasks and grading. For most companies the practical question is whether your staff have time to help build evaluations, and whether that work is paid separately.
In August 2026 Google agreed to pay $10 million in bankruptcy proceedings for Spirit internal data, including emails, Teams messages, spreadsheets and operations files, to train AI, as reported by ABC, TIME and others. micro1 then made a $12.5 million rival offer and later said it pursues non-sensitive, non-consumer data with third-party de-identification. It was a liquidation, not a running company licensing a copy, so it shows that operational records have a market price rather than what a typical company would be paid.
More: how much AI companies pay for data and raw data vs evaluations vs environments.
Forbes reported on 16 April 2026 that startups closing down were selling old Slack and email archives as AI training data, and Fast Company and Gizmodo covered the same trend. Troveo cites about $5,000 per code repository and roughly $10,000 to $100,000 per archive deal in that closure market. Those are prices for archives of companies that no longer operate, often with former employees objecting, and they are not a guide to what a running company would be offered.
It depends on what you own, what your client contracts say, what you have told employees, whether personal data is involved, and sector rules such as GDPR, CCPA/CPRA, HIPAA, GLBA and attorney-client privilege. A company can often license operational records it owns, while material held for clients or about patients may be off the table. This is general information, not legal advice. Talk to your own lawyer before you sign.
Buyers describe their own safeguards. micro1 says sensitive information is scrubbed, originals are deleted after processing and no customer information is exposed; Mode says it de-identifies before onward delivery. Statements on a program page describe intent; the agreement is what binds. Agree an exclusions list, review samples before use, and get de-identification, audit and deletion terms into the contract. Tell staff before any export rather than after.
Anything you hold under a confidentiality duty you cannot waive, such as privileged legal material, client files at accounting, M&A or agency firms, and patient records covered by HIPAA. Also leave out customer personal data you have no basis to share, HR and medical files, and passwords, keys or other secrets that sit in code and chat history. Write these into an exclusions list that is applied before anything leaves your systems.
In the published programs, yes. micro1 states that companies retain ownership of their underlying data, and Mode says it buys an agreed copy while originals stay with the company. Ownership is only part of the picture, though. An exclusive license can still stop you selling the same records again, and a broad scope of use can let the copy travel to many downstream buyers. Read the license terms alongside the ownership statement.
More: is it legal to sell company data and the printable questions to ask a data buyer.
Exclusivity and resale rights, scope of use, indemnities and warranties, consent representations about employees and customers, duties to your own clients, payment structure and acceptance criteria, audit rights over de-identification, deletion of originals and copies, and termination and survival clauses. Ask about each one in writing, compare answers across buyers, and have a lawyer read the final draft. General information, not legal advice.
Practitioners advise it. Share a manifest and a few samples with more than one buyer, compare offers on terms as well as price, and never send a full dataset before a price is agreed. Two offers at the same price can be worth very different amounts once exclusivity, acceptance criteria and liability caps are compared. A time-limited offer is normal; a demand to send everything first is a reason to slow down.
Only if your agreements allow it. A non-exclusive license usually leaves you free to license the same records again; an exclusive or time-limited exclusive license does not, and selling twice could breach the first contract. Some deals also restrict related data or competing buyers. Decide before you sign whether you want the option to sell again, and price exclusivity accordingly rather than giving it away by default.
Usually the owner or CEO, the finance lead, whoever is responsible for IT and security, and your lawyer. If you hold client data, the partners or account leads responsible for those clients should confirm nothing restricted is in scope. HR should see the plan for staff communication. Agree who signs before talking to buyers, so that a good offer is not held up by an internal question nobody owned.
No. Sell Data to AI never receives, stores, views, transfers or processes company data, and it never negotiates or signs anything with AI labs. The eligibility checker runs in your browser and sends nothing. If you apply to a program, you deal with that buyer directly, and the buyer runs its own discovery, contracts, export, de-identification and payment.
Sell Data to AI earns from paid plans, company lists and custom lists, and some guide links are referral links, as explained in our disclosure. API plans are Basic at $99, Pro at $299 and Scale at $799 a month, one-time company lists start at $199, and custom list pilots cost $100. If your company signs with a buyer through a referral link, the buyer may pay us a fee under its own terms. Applying through a referral link costs you nothing, we never share fees or offer cash-back, and we are not a partner, agent or representative of any buyer. We describe each buyer only by what its own pages publish.
More: our disclosure, pricing and privacy policy.
The words you will meet in program pages, NDAs and agreements. Definitions are general; the agreement you are offered decides what each term means for you.