The Complete Guide to Selling Your Data to AI Companies and AI Labs (November 2026)

The short answer#

Yes, companies can sell or license their own business data to AI companies, and a small number already have. But the market is young, the public evidence is thin, and most of the price figures you will see online are reports from anonymous sources, not published rate cards. This guide explains what is known, what is guesswork, and what a business owner should check before talking to any buyer.

Here is the short version. AI developers want data that is accurate, specific to real work, and legally clean. Public web text is running low and is increasingly restricted, so private data with clear rights has become more attractive.¹² Reported deals range from small payouts for the contents of a failed startup's Slack and email archives to contract values above $200 million for large content owners.³⁴⁵ Whether your own data can sell depends on four things: who owns it, whether personal information can be removed, whether it shows real expert work, and whether the buyer will accept your terms.

Quick answers#

Why AI companies want business data#

The public web is a finite resource#

Large language models learn from text. For years the main source was the public web. Epoch AI, a research group that tracks AI trends, estimated the stock of usable public human-written text at roughly 300 trillion tokens. It projected that if training trends continue, that stock could be fully used between 2026 and 2032.¹ Epoch's estimate covers public text only, and the authors note that techniques such as synthetic data would be needed to keep progress going.¹

At the same time, the open web is closing. The Data Provenance Initiative audited the web domains behind popular training corpora and found that restrictions on AI crawlers and in terms of service rose sharply over a single year. In one widely used corpus, C4, terms-of-service restrictions covered about 45% of the content.² Free, generic web text is also commodity supply. Common Crawl, a nonprofit, publishes a free archive of more than 300 billion pages collected over 15 years.⁶ In our reading, anything a model builder can get for free from the public web is not worth much to pay for.

Synthetic data does not fully replace human data#

A natural question is whether AI can simply generate its own training data. Researchers have studied this. A paper in Nature found that training models on model-generated content can cause "model collapse," in which the rarer parts of the original data distribution disappear over successive generations. The authors added that data from genuine human interaction with systems will become more valuable as AI-generated content spreads online.⁷ Synthetic data is useful in many settings, but this result is one reason developers keep paying for real human and business data.

Quality beats volume#

The other pressure is quality. Several research results show that smaller amounts of carefully chosen data can outperform much larger piles of raw text:

• Microsoft researchers trained a 1.3 billion parameter code model on about 7 billion tokens of filtered and synthetic "textbook quality" data. It reached 50.6% on the HumanEval coding benchmark, and the authors credit data quality.⁸
• The LIMA paper tuned a 65 billion parameter model on only 1,000 carefully curated prompt and response pairs, and argued that most knowledge comes from pretraining, so a small, high-quality tuning set can be enough.⁹
• The DataComp-LM project found that model-based filtering of web data was the key to a strong training set, and its filtered 2.6 trillion token set let a 7 billion parameter model reach 64% on the MMLU benchmark with less compute than the previous open-data best.¹⁰

None of these papers is about buying data from companies. They matter because they explain what a buyer is looking for: specific, correct, well-organized examples of real work, not volume. A support team's resolved tickets, a lawyer's annotated contracts or an engineering team's reviewed code changes may be small by web standards but dense in the kind of signal a developer cannot get from Common Crawl.

Companies have said they want private data#

AI developers have said so directly. In November 2023, OpenAI announced a "Data Partnerships" program seeking large-scale datasets that are not easily accessible online, including human conversations and long-form writing. According to TechCrunch's report, it offered two tracks: open datasets made public, and private datasets that stay confidential.¹¹

There is also reporting on the demand for expert-created work. TechCrunch reported in 2025 that leading AI labs pay Mercor to hire former bankers, consultants and lawyers to write reports and fill out forms used for training, with contractors earning up to $200 an hour, and framed this as a way to get data that companies would not share directly.¹² OpenAI's GDPval benchmark, published in 2025, was built from tasks written by industry professionals averaging 14 years of experience across 44 occupations.¹³ These are signs that labs value examples of real professional output.

Another area is training environments for AI agents. TechCrunch reported in September 2025 that major labs were seeking reinforcement learning environments, which are simulated workspaces where an agent practices multi-step tasks, and that Anthropic leaders had discussed spending more than $1 billion on them over the next year, according to The Information.¹⁴ That figure is second-hand and describes a discussion, not a signed commitment. Epoch AI later described the market from interviews: contracts often run six to seven figures per quarter, and enterprise software workflows are a major growth area.¹⁵

What the reported deals show#

Most data licensing news so far involves publishers and online communities, not ordinary businesses. These deals still show the range of structures and prices. All figures below are reported by the sources named, and several rest on anonymous people. Companies rarely confirm terms.

A few lessons come out of the table. First, headline numbers mix very different things: annual payments, total contract value, revenue segments and projections. Second, the large figures belong to companies with huge, unique archives. Third, the per-unit prices in the Photobucket and Defined.ai reporting are a useful reminder that generic data sells at low unit prices, and that volume alone does not make a business dataset valuable.²¹

Smaller and stranger sales#

At the other end are sales by failing companies. Forbes reported in April 2026 that SimpleClosure, a company that helps businesses shut down, launched a service for startups to sell their code and workspace data, such as Slack messages, emails and Jira tickets, and said it had processed about 100 deals in a year with payouts ranging from $10,000 to $100,000.³ In October 2026, MIT Technology Review reported that the service had handled more than 150 deals, that code or data sales were part of 47% of SimpleClosure's deals, and that buyers included foundation model labs, vendors building simulated workspaces and tech companies building models.⁴

The same report said Google outbid Mercor for Spirit Airlines' data in August 2026 for $10 million, a figure we could find in only that one outlet, and described privacy objections, including opposition from the airline's flight attendants' union and a lawyer calling such data "legally radioactive."⁴ We return to this in the section on bankruptcy sales.

The legal backdrop#

Buyers care about clean rights because the legal risk is real. In September 2025 Anthropic agreed to pay $1.5 billion to settle a lawsuit by book authors, reported as about $3,000 per book across roughly 500,000 books.²⁴ That case concerned books obtained from pirate sources, not licensed business data, but it shows why serious buyers now ask where data came from and who agreed to its use.

What kinds of business data are in demand#

No public price list exists for business data by type, so this section rests on what is publicly known about how models are trained and evaluated. Where we are offering our own judgment, we say so.

Customer support conversations#

Support conversations pair a customer's problem with a resolution written by a trained employee. That is useful for teaching a model how a problem is described and how a good answer is phrased. The caution is price. A published dataset from Bitext offers about 26,872 support question and answer pairs across 27 intents, built in a hybrid synthetic way and released under a sharing license.²⁵ Generic support examples can be produced cheaply. What a real company has that a synthetic set does not is the messy, domain-specific detail of its own product, and that is the part a buyer would value. This is our inference, not a price signal from the market.

Code with reviews and commit history#

Code is the best documented case. The Stack, a large code dataset from the BigCode project, collects permissively licensed source code in 30 languages and offers developers a way to check whether their code is included and to request removal.²⁷ SWE-bench, a benchmark for coding agents, was built from 2,294 real software problems, each pairing a GitHub issue with the pull request that fixed it.²⁶ The practical meaning for a software company is that the history of how problems were found, discussed and fixed has structure a model can learn from. The licensing meaning is that the license on every dependency and every contribution matters.

Expert documents, SOPs and workflows#

A standard operating procedure, a legal template library or a set of annotated reports captures how professionals do their jobs. GDPval and the Mercor reporting above show that expert-produced work is the kind of data labs value.¹³¹² The risk is confidentiality: much of this material belongs partly to clients and is covered by non-disclosure agreements.

Healthcare data#

Healthcare is the most regulated category and also the one with the clearest published prices. Protege, a data licensing platform, describes itself as curating and de-identifying data from many providers and licensing it to AI developers, with a focus on healthcare, media and video.²⁸ Stanford's AIMI center publishes a commercial license price of $70,000 per dataset per year for its de-identified clinical datasets, the rate for agreements through August 2025.²⁹ That is one academic seller's price, not a market rate. We cover HIPAA's two de-identification methods later in this guide.

Who buys: AI labs, data vendors and marketplaces#

"Selling data to AI labs" usually means selling to one of six kinds of buyer. Only some deal with an ordinary company directly; many reach AI labs through the companies in between. This summary reflects how we see these buyers work in practice; requirements change, so confirm them with each buyer.

For most companies the practical route to an AI lab runs through the data and evaluation companies that supply them, a marketplace, or a broker that knows which buyer needs what this quarter. Our buyer guide goes deeper on each type.

What does not sell#

Not every dataset has a buyer. Some of this is judgment, and we mark it as such.

• Data you do not have the right to sell. This is the most common reason. Customer contracts, privacy promises, employee relationships and open-source licenses can each block a sale. The rights section below covers how to check.
• Generic public data. Public web text is plentiful and free, as Common Crawl shows.⁶ A copy of your public website is unlikely to be worth paying for. This is our reading of the evidence.
• Raw, messy, undocumented exports. The Data Provenance Initiative and the Datasheets for Datasets proposal both argue that datasets should come with documentation of where they came from, how they were collected and what they are for.²³⁰ In our view a buyer's lawyers will ask for exactly that. Data with no provenance record is harder to price and harder to buy.
• Data that cannot be de-identified. If personal details are woven through the content and cannot be removed without destroying its value, buyers are likely to pass. Free-text clinical notes are a typical example. This is our judgment.
• Synthetic or model-generated content. The model collapse research is the reason buyers do not want training data that was mostly generated by other models.⁷
• Tiny or one-off datasets. Expert-data contracts reported by Epoch involve per-task prices of $200 to $2,000 and require calibrated difficulty and reliable graders, so a handful of examples without structure is not a product.¹⁵

Who owns your data, and what you can actually license#

Holding a database is not the same as holding the right to sell it. Before a buyer sees anything, a rights review should answer one question: for each source of data, what permission do we have, from whom, and does it cover this use?

Ownership is mostly set by contract#

Vorys, a US law firm, notes that ownership of data between parties is mostly set by contract, because raw facts are generally not protected by copyright. It advises spelling out permitted uses rather than relying on the word "own."³¹ Morgan Lewis makes a related point from the vendor side: what matters is the rights a party holds to use, retain and benefit from data, and broad training rights can let learnings stay in a model indefinitely, so agreements should deal explicitly with aggregated and anonymized data.³² Both are general commentary, not rules. They point to the same practical step: read the contracts that govern where your data came from.

Customer contracts and privacy promises#

If your data came from customers, your agreements with them and your public privacy policy set the limits. Three sources of risk are worth checking:

• Confidentiality clauses. Many enterprise contracts bar using customer content for anything beyond providing the service.
• Promises about AI training. The US Federal Trade Commission has said that companies must keep their privacy and confidentiality promises wherever they were made, that there is "no AI exemption" from existing law, and that a promise not to use customer data for training is enforceable, including against indirect workarounds.³³
• Quiet changes to terms. FTC staff have written that it may be unfair or deceptive to loosen data-use terms, for example to allow AI training, and notify customers only through a quiet retroactive change to a terms of service or privacy policy.³⁴

The FTC has also said it has required companies that unlawfully obtained data to delete the models and algorithms built from it.³³ Examples include the Everalbum settlement over photos and videos, the proposed Rite Aid order over facial recognition images, and the WW International and Kurbo order over children's data.³⁵³⁶³⁷ For a data seller, the lesson is concrete: if a buyer trains a model on data you had no right to sell, the consequences can reach the model itself, and a careful buyer will demand warranties that you cannot give if you have not done the review.

Employee data#

Workplace messages, tickets and call logs contain employees' personal information as well as customers'. Guidance from European regulators, summarized by Irish law firm A&L Goodbody, says employee consent is generally not a workable basis for processing because of the imbalance of power, and that employers relying on legitimate interests must show necessity and proportionality.³⁸ UK guidance on monitoring workers, summarized by Freeths, calls for a defined purpose, the least intrusive means and clear notice to workers, and the regulator's own page notes that its guidance is under review.³⁹ In California, the Berkeley Labor Center notes that privacy law now covers employees, contractors and applicants as well as consumers, because the workforce exemption expired on January 1, 2023.⁴⁰

Call recordings and chat transcripts#

Support calls raise a separate issue. California Penal Code section 632 bars recording a confidential communication without the consent of all parties.⁴¹ Recent litigation targets the way AI vendors use recorded calls. In a February 2025 ruling in Ambriz v. Google, a federal court let wiretap claims proceed over a contact center AI product, applying a test under which a vendor can count as a third party if it has the capability to use call data for its own purposes.⁴² In August 2025, a court allowed similar claims to proceed against an AI phone-ordering vendor in Taylor v. ConverseNow, reasoning that the vendor allegedly used call data to improve its own product.⁴³ These are rulings at the pleading stage, so they do not decide who wins, but they show why a company should know what its callers were told before licensing recordings.

Trade secrets#

Documents, SOPs and workflows may contain trade secrets. Under US federal law, business information qualifies as a trade secret only if its owner takes reasonable measures to keep it secret and it has independent economic value because it is not generally known.⁴⁴ Sharing it with a buyer therefore needs a non-disclosure agreement, and law firms disagree on the details. Holland & Knight cautions that a time-limited NDA can jeopardize trade secret protection and recommends open-ended confidentiality for trade secrets.⁴⁵ A 2026 federal decision reported by Faegre Drinker held that unauthorized disclosure by a recipient is still misappropriation even where an NDA exists.⁴⁶ Treat these as signs that the paperwork around a sale needs a lawyer, not as settled rules.

Code and open-source licenses#

Software companies often assume that all their code is theirs to license. In practice most codebases include open-source components. The GNU project's FAQ says the GPL does not require you to release modified versions kept private inside a company, but if you release a modified version publicly you must make its source available under the GPL.⁴⁷ The AGPL goes further: its section 13 requires offering source to users who interact with a modified version over a network.⁴⁸ The MIT license requires that copyright and permission notices stay with copies.⁴⁹ We did not find a source that settles whether these licenses apply to AI training use, so the right approach is to separate code you wrote and own from third-party code, and to ask counsel about the rest.

Copyright and the licensing market#

In a May 2025 pre-publication report, the US Copyright Office said voluntary licensing for AI training is already happening in some sectors and appears likely in others, and that where licensing is available or feasible, that weighs against a fair use finding. The report is not binding law and the Office said a final version would follow.⁵⁰ The practical effect is that buyers have a reason to prefer licensed data to scraped data, and sellers have a reason to document what they are licensing.

Privacy law, in plain terms#

The privacy rules that apply to you depend on whose data is in the set and where they live. This is a map, not a legal opinion.

GDPR and other European rules#

The EU's General Data Protection Regulation defines personal data as information relating to an identified or identifiable person, directly or indirectly, for example through a name, an identification number, location data or an online identifier.⁵¹ Data that has been pseudonymized, meaning it can no longer be tied to a person without extra information held separately, is still personal data. Only truly anonymous data falls outside the regulation, and anonymity is judged by all the means reasonably likely to be used to identify someone, weighing cost, time and available technology.⁵² The UK regulator says the same about pseudonymized data and adds that if you share it without the key, the recipient may hold anonymous information.⁵³ The European Data Protection Board adopted guidelines on pseudonymization in January 2025, and a law-firm summary notes that they give no clear test for when pseudonymization is effective.⁵⁴

Two further principles matter for a data sale. The first is purpose limitation: personal data must be collected for specified purposes and not further processed in a way that is incompatible with them, and the regulation lists factors for judging compatibility, including the link between the purposes, the context of collection and the consequences for people.⁵⁵⁵⁶ The second is the need for a lawful basis, and the regulation lists six, including consent and legitimate interests. Legitimate interests fail if overridden by the person's rights and freedoms.⁵⁶

In December 2024 the EDPB issued an opinion on AI models and personal data. It said a model trained on personal data is not automatically anonymous, and that anonymity requires that the likelihood of extracting personal data from the model, directly or through queries, is insignificant, assessed case by case.⁵⁷ It also described what happens when a model was built using unlawfully processed personal data, saying the effect on later use depends on the facts, including whether a different organization later deploys the model and whether the model has been anonymized.⁵⁷ For a seller, the takeaway is that a European buyer may ask for proof that your data was lawfully obtained.

When your own company handles personal data for a customer under a data processing agreement, GDPR Article 28 says a processor may act only on the controller's documented instructions, and a processor that decides purposes and means itself is treated as a controller for that processing.⁵⁸ That is our reading of how the article would apply to a data sale, not a statement from a regulator: a vendor that sells its customers' data for its own gain may be taking on more than it realizes.

California and other US states#

The California Consumer Privacy Act defines "sell" broadly as disclosing or making personal information available to a third party for money or other valuable consideration, with exceptions that include some asset transfers in a merger or bankruptcy.⁵⁹ Consumers can opt out of sale or sharing, including through a browser signal, and the law applies to businesses over certain revenue or consumer-count thresholds.⁶⁰ The law also defines "deidentified" information. Data counts as deidentified only if it cannot reasonably be used to infer information about or be linked to a consumer, and the business takes reasonable measures against re-association, publicly commits not to reidentify it, and binds recipients by contract to the same terms.⁵⁹ That last step matters for a seller: your contract with a buyer is part of what makes data "deidentified" in California.

California is not alone. The International Association of Privacy Professionals counted 19 US states with comprehensive privacy laws as of its October 2025 update, and said they commonly give consumers the right to opt out of the sale of personal data.⁶¹ The count has probably grown since then, so check a current tracker.

De-identification: methods and limits#

De-identification means changing data so that it no longer points to a person. It is the most important technical step in preparing business data for sale, and also the one most often oversold.

HIPAA's two methods#

For health data covered by the US Health Insurance Portability and Accountability Act, the regulation recognizes two routes to de-identification. We rely here on the federal regulation text, 45 CFR 164.514(b).⁶²

The identifier list covers relatives, employers and household members of the individual as well as the person.⁶² This table is a summary of regulation text for orientation. Whether a particular dataset meets either standard is a question for a privacy professional.

Removing identifiers is not the same as removing risk#

History offers repeated warnings that "anonymous" data is often not:

• Latanya Sweeney of Harvard found that about 87% of the US population, 216 million of 248 million people, had a unique combination of five-digit ZIP code, gender and date of birth, using 1990 census data. This is an estimate of uniqueness, not of how many people were actually identified.⁶³
• Narayanan and Shmatikov showed in 2006 that anonymous movie ratings from roughly 500,000 Netflix subscribers could be matched to named people with a little background knowledge, using IMDb as the outside source.⁶⁴
• AOL released about 20 million search queries from more than 650,000 users in August 2006. The New York Times identified one user within days, and AOL's chief technology officer resigned.⁶⁵
• A 2019 study in Nature Communications estimated that 99.98% of Americans could be correctly re-identified in any dataset using 15 demographic attributes. The figure comes from a statistical model, not from re-identifying a real dataset.⁶⁶

The lesson for text data is the same as for tables: people are identified by combinations of details, not just names. A support transcript can reveal a person through a rare job title, a street, an order number and a date.

Methods that exist#

• Redaction tools. Microsoft's Presidio, now hosted by the Data Privacy Stack project, is an open-source toolkit that detects and anonymizes personal data in text and images using entity recognition, patterns and rules. Its documentation says there is no guarantee that it finds all sensitive information, so extra safeguards are advised.⁶⁷
• k-anonymity. A dataset is k-anonymous if each person's record cannot be distinguished from at least k minus 1 others. Sweeney introduced the idea in 2002.⁶⁸
• Government guidance. NIST Special Publication 800-188, finalized in 2023, covers de-identifying government datasets, including removing identifiers, transforming quasi-identifiers and using controlled-access sharing such as enclaves. It warns that tools that merely mask personal data may not be enough.⁶⁹ It is written for federal agencies, so it is a benchmark and not a legal requirement for private sellers.
• Differential privacy. A mathematical approach that adds calibrated noise. NIST published guidelines in 2025 for evaluating claims of differential privacy.⁷⁰ It fits statistics better than free text.
• Pseudonymization. Replacing names with codes. As noted above, regulators treat it as still personal data while anyone holds the means to re-link it.⁵³
• Controlled access. Instead of handing over files, the seller lets a buyer train inside an environment it controls, so data never leaves. NIST's guidance describes sharing models of this kind.⁶⁹

Why model memorization matters#

A buyer's model can sometimes repeat what it learned. Researchers extracted hundreds of verbatim sequences from GPT-2's training data in 2021, including personal information, chat logs and code.⁷¹ A 2023 paper by researchers including Nicholas Carlini extracted gigabytes of training data from open and closed models, including ChatGPT, using a "divergence attack," and concluded that alignment does not eliminate memorization.⁷² These attacks target models that have already been trained. They are the reason personal and confidential details should be removed before sale, not after.

How pricing works#

There is no standard price list for business data. What exists is a mix of valuation theory, a few published unit prices, and reported deal sizes.

Three ways to value data#

Deloitte describes three approaches used to value data assets. The income approach discounts the future benefits the data produces. The cost approach estimates what it would cost to recreate or replace the data. The market approach uses prices from comparable transactions. Deloitte notes that reliable comparable transactions are scarce, so the market approach works best as a reasonableness check, and that methods are usually combined.⁷³

For a seller, this means two things. A buyer is likely to think in terms of what the data replaces: the cost of paying experts to produce similar examples, or the benefit a model gains. And because comparables are few, the first offer you see may say more about the buyer's budget than your data's worth.

Reported price points#

The table below collects published or reported figures. Each has a different unit and a different context, and none is a quote for your data.

Analyst reports estimate the overall market for AI training datasets in the billions of dollars. One reseller summary of a Grand View Research report lists about $3.2 billion in 2025 growing to about $16.3 billion by 2033.⁷⁶ We do not recommend relying on these estimates. They are definitions-dependent forecasts from firms that sell reports, and estimates from different firms vary.

What tends to move a price#

Public evidence supports some of the following and the rest is our judgment, marked as such.

• Exclusivity. Epoch's interviews suggest exclusivity commands a premium of several times, in the environments market.¹⁵
• Quality and structure. Buyers of environments value reliable grading and calibrated difficulty.¹⁵ More broadly, the quality research above explains why curated data is worth more than raw volume.⁹ (Our judgment applies this to business data.)
• Rights clarity. Clean provenance and clear rights reduce a buyer's legal risk. (Our judgment, supported by the legal cases above.)
• Uniqueness. Data a developer cannot get elsewhere is worth more. (Our judgment.)
• Preparation cost. De-identification, cleaning and documentation cost money, and someone pays. Ask who.
• Alternatives. Free synthetic datasets, such as the Bitext set, put a ceiling on generic examples.²⁵

Deal structures#

Once a buyer is interested, the term sheet and contract decide what you are actually selling. Law firm commentary on AI data licensing agrees on a handful of points. Some of the sources below are written from the buyer's side, such as Sidley's, or for companies licensing technology from a vendor, such as Best Lawyers', so sellers should read them with that in mind.

Exclusive or non-exclusive#

Exclusive deals pay more but cap your future income from the same data and may limit your own use. A non-exclusive license lets you sell to several buyers but gives each less reason to pay a premium. For data that only you can produce, such as recurring business activity, a non-exclusive or time-limited exclusive license can fit. Seek advice before agreeing to exclusivity for a long term.

Flat fee or revenue share#

A flat fee gives certainty. A revenue share ties your payout to the buyer's success but depends on a buyer you cannot audit easily. Public sources describing actual AI data deals say little about revenue shares, so expect to negotiate this from first principles. Created by Humans, a licensing marketplace for authors, takes a cut of each deal but TechCrunch reported no percentage and we found none published.⁸¹

What a seller should insist on#

From the sources above and standard practice, a seller will usually want a field-of-use limit, a prohibition on re-identification, deletion and audit rights, a confidentiality clause, clear ownership of the original data, and a cap on liability. These are not legal advice. A lawyer who has done data licensing should review the deal.

The process, step by step#

Whether you work with a broker or sell directly, the sequence is similar. These steps are our recommended order, based on the sources above and common practice, not a legal requirement.

1. Inventory what you have. List each source of data: systems, date ranges, volume, formats and who created it. Include what is in third-party tools such as support platforms and chat apps, and note where personal data appears.
2. Trace the rights. For each source, find the contract, policy or consent that governs it. Mark what you own outright, what customers own, what employees or contractors created, and what is third-party or open-source. The readiness checklist later in this guide covers the same ground.
3. Remove or restrict the risky parts. Decide what to exclude entirely, for example data covered by confidentiality clauses or customers who never agreed. Anything left in needs a plan for personal data.
4. De-identify and document. Apply the technique that fits each data type, test the result, and write down how it was done and by whom. A buyer will ask for this record, much as the Datasheets for Datasets proposal argues every dataset should ship with documentation of its origins, composition and intended use.³⁰
5. Package a sample. Prepare a small, representative, de-identified sample and a short description. A sample lets buyers judge quality without seeing everything. Share nothing outside a signed confidentiality agreement.
6. Find and screen buyers. Build a list of possible buyers, check their data-use and security practices (see the next section), and approach them under non-disclosure terms.
7. Get offers and compare them. Compare more than price: exclusivity, term, field of use, deletion rights, indemnity and the buyer's ability to protect your data.
8. Negotiate and sign with counsel. Have a lawyer who has done data licensing review the agreement, the warranties and the indemnity.
9. Deliver and monitor. Transfer through a secure method, keep a record of what was delivered, and calendar the deletion and audit dates.

How to evaluate a buyer#

Selling your data means trusting the buyer with it. A buyer is not just a source of payment. Some questions to ask, and what public sources say about each area:

What will the buyer do with the data?#

Licensing data for training is different from using an AI tool. Major AI providers say that data sent through their business products is not used for training by default: Anthropic says it does not use inputs or outputs from its commercial products, including its API, for model training by default;⁸² OpenAI's API documentation says data sent to the API is not used to train its models unless the customer opts in;⁸³ and Google's Vertex AI documentation says Google will not use customer data to train or fine-tune models without prior permission.⁸⁴ These policies concern using the tools. They are not terms of a data sale, where training is the point. Do not confuse the two when a buyer says they "do not train on customer data."

Can the buyer protect it?#

Ask for the buyer's security evidence and how data is stored, accessed and deleted. The AICPA's SOC reports are one form of independent assurance about a service organization's controls.⁸⁵ NIST's AI Risk Management Framework, released in January 2023 with a generative AI profile in July 2024, is voluntary guidance that a buyer can use to show how it manages risks.⁸⁶ Whether a buyer holds these is our suggested test, not a rule, and no certificate guarantees safe handling.

Real incidents show why it matters. Security Affairs reported in April 2023 that Samsung employees pasted chip source code and meeting notes into ChatGPT on three occasions in under a month, and that Samsung warned staff the data could not be retrieved.⁸⁷ That was about employees using a tool, not licensing, but it illustrates how confidential material can leave a company's control.

Will the buyer expect documentation?#

Probably. The Data & Trust Alliance published Data Provenance Standards in July 2024 with 22 metadata fields in three groups covering source, provenance and use, including consent documentation, license to use and confidentiality classification. The standards are voluntary, but they are the closest thing to a buyer-side diligence checklist we found.⁸⁸ Preparing this information early speeds a deal.

What is the buyer's track record?#

Ask for references and what the buyer has disclosed publicly. Look at its public statements about data practices, and at news. Companies that changed terms to allow AI training without clear notice have faced strong reactions, as the next section shows. Be careful of any buyer that wants broad rights, perpetual access and no deletion.

Risks to weigh#

Reputation#

Customers react badly when they feel their data was used without notice. In August 2023, a blog reported that Zoom's terms seemed to allow AI training on "service generated data." Zoom clarified a day later and amended its terms to say it would not train AI models on audio, video or chat content without consent, though privacy advocates remained skeptical.⁸⁹ In May 2024, Slack's privacy principles said it analyzed customer data to develop AI and machine learning models, with opt-out available only by emailing the company. Slack said it had not trained generative or third-party language models on customer data and that its wording was unclear.⁹⁰⁹¹ In June 2024, Adobe's updated terms drew a backlash over fears of AI training on customer content, and Adobe said it did not train generative AI on customer content.⁹² In all three cases the companies said practice had not changed and the wording was the problem, so we do not state that any of them did train on customer content. The reputational damage came from the wording alone.

Automattic, which owns Tumblr and WordPress.com, was reported by 404 Media in February 2024 to be preparing to share content with OpenAI and Midjourney, and an alleged internal post said the first batch wrongly included private posts and deleted blogs. Automattic said it would share only public content from sites that had not opted out.⁹³ The lesson is operational: errors in selecting data are the sort of thing that become news.

Re-identification#

Even careful de-identification can fail, as the history in the de-identification section shows. A seller should require the buyer to commit not to re-identify and to report any suspected breach. In California, that commitment is part of the legal definition of deidentified data.⁵⁹

Competitive leakage#

If you sell data that reflects how you serve customers, your competitors may eventually use models trained on it. This is a real tradeoff and a reason to negotiate field-of-use limits, exclusivity terms or a restriction on selling to named competitors. The research on memorization shows models can repeat their training data in some conditions, so material that gives away pricing, strategy or customer lists should be excluded before sale rather than relied on to stay hidden.⁷¹⁷² We found no public case of a company's trade secrets leaking from a model it sold data to, so this is a risk by reasoning, not an observed event.

Legal exposure after the sale#

The FTC has required deletion of models and algorithms built from unlawfully obtained data in several cases.³³ If a buyer trains on data you had no right to sell, the buyer may face this. In contracts, that risk flows back to you through warranties and indemnity.⁷⁹ The best protection is the rights review done before anyone signs.

Selling data in a bankruptcy or wind-down#

Companies that are closing sometimes sell their data as an asset. This is a real part of the market, and it carries unique rules.

In the United States, section 363(b)(1) of the Bankruptcy Code says that if a debtor's privacy policy promised not to share personal information with unaffiliated third parties, the trustee may sell or lease it only consistently with that policy, unless a consumer privacy ombudsman is appointed and the court approves.⁹⁴ Section 332 requires the US Trustee to appoint an impartial ombudsman at least a week before the sale hearing, who reports to the court on the privacy policy, the privacy effects and alternatives.⁹⁵ In March 2025 the FTC's chair wrote to the trustee in the 23andMe bankruptcy that its privacy promises must be honored in any sale. A law firm's summary says the FTC cited section 363(b)(1) and urged that buyers expressly agree to follow the privacy policies.⁹⁶⁹⁷

The California privacy law, for its part, excludes from the definition of sale certain transfers as part of a merger or bankruptcy, provided the data is used consistently with the law.⁵⁹

Not every wind-down is a bankruptcy. SimpleClosure's service, as reported, helps shuttered startups sell code and workspace data and strips personal information first, while a privacy advocate cited by Gizmodo called employee privacy a key concern.³ MIT Technology Review's report included a union's objection to the Spirit Airlines sale and a lawyer's comment that such data can be "legally radioactive."⁴ It also noted that there is no federal law banning such sales.⁴

Brokers or selling direct#

Some companies sell directly to AI developers. Others use an intermediary. Both work, and the right choice depends on what you have.

What intermediaries exist#

A small set of platforms publicly describe themselves as intermediaries for AI training data:

• Protege describes itself as aggregating data from many providers, curating and de-identifying it, and licensing it to AI developers, with a focus on healthcare, media and video.²⁸
• Troveo licenses video, audio and business data to AI developers and lists no prices on its site, only a request for an estimate.⁹⁸
• Created by Humans lets authors license rights to AI developers and takes a cut of each deal; the percentage is not public.⁸¹
• Cloud data marketplaces such as Snowflake Marketplace and AWS Data Exchange let providers list datasets, but neither page we reviewed states provider fees.⁹⁹¹⁰⁰
• Extruct's June 2026 directory tracks 15 companies in this category, split into those that collect and license data, rights and IP brokers, and creator marketplaces.¹⁰¹

Fees and revenue splits are almost never public. Ask any intermediary exactly how it is paid before you sign.

Selling direct#

Direct sales can save the intermediary's cut and give you full control. They require that you can find the right buyer teams, run rights and de-identification work, and negotiate with counterparties that have more legal resources. Large content owners such as publishers and online communities have done direct deals.¹⁶¹⁷

Using a broker#

A broker takes on some of the work: finding buyers, running the rights and privacy review, preparing the data, and negotiating. A broker also has a conflict of interest, because most brokers are paid when a deal closes. Our How it works page says SellYourAIData takes a commission only if a deal completes, so our incentive is to see deals made. The questions to ask any broker, us included, are below.

• How are you paid, and what percentage?
• Who owns the buyer relationship and the right to re-sell the data?
• What rights and privacy review do you do, and who signs off?
• How do you de-identify, and what happens if re-identification is found?
• Do you work with a lawyer I choose?
• Can I see your agreement before I send data?
• What exclusivity do you require from me?

Readiness checklist#

Use this list before you contact any buyer or broker. A "no" is not the end, but it shows what to fix first.

• We have a written inventory of data sources, dates and volumes.
• We know which contracts and policies govern each source.
• Our customer contracts allow, or we have obtained permission for, this use of the data.
• Our privacy policy does not promise anything inconsistent with a sale.
• We know whether the data includes employee or contractor personal information, and what employees were told.
• If we recorded calls, we know what callers were told and which laws apply.
• Code in scope is ours, and third-party and open-source components are identified.
• We can remove personal and confidential information, and we can test and document the result.
• The data shows real expert work: decisions, resolutions, reviews or procedures.
• Someone inside the company has authority to approve a sale and sign confidentiality terms.
• We have a lawyer who can review a data license.
• We have decided what we will not sell, such as strategy, pricing or customer lists.

Glossary#

Anonymization. Changing data so people cannot be identified by any means reasonably likely to be used. Stricter than pseudonymization.⁵²

De-identification. Removing or altering identifying details. In HIPAA, it means meeting either Safe Harbor or Expert Determination.⁶²

Differential privacy. A mathematical technique that adds noise so that results reveal little about any one individual.⁷⁰

Expert Determination. A HIPAA route in which a qualified expert documents that re-identification risk is very small.⁶²

Field of use. The limit on what a buyer may do with licensed data, such as training a named model for a named purpose.

Fine-tuning. Further training of an existing model on a smaller, specific dataset. The LIMA paper tuned a large model on 1,000 curated examples.⁹

Model collapse. The degradation seen when models are trained on model-generated data over successive generations.⁷

Provenance. The documented origin and handling history of a dataset.⁸⁸

Pseudonymization. Replacing identifiers with codes while keeping a way to reverse the change. Still personal data under GDPR.⁵²

Reinforcement learning environment. A simulated workspace where an AI agent practices tasks and receives feedback.¹⁴

Safe Harbor. The HIPAA method of removing 18 listed identifier types, with no actual knowledge that the remainder can identify someone.⁶²

Training data. Examples a model learns from, as distinct from data it is asked to process at use time.

What changed in this edition#

We update this guide every month. Changes in each edition:

November 2026

• Retitled as the complete guide to selling your data to AI companies and AI labs, with the month in the title so readers can see it is current.
• Added a quick-answers table at the top and a who-buys table comparing AI labs, data vendors, marketplaces, enterprise AI teams, synthetic-data builders and evaluation builders.
• Added our own observation of current deal ranges in the pricing section, labelled as first-party.
• Added an FAQ on selling directly to an AI lab, and an author byline with credentials.

Where to go next#

If you are still deciding whether your data is a candidate, the quickest step is to list your sources and the contracts behind them. Our free valuation asks for a short profile with no files or raw data and replies with a first view on fit, and you can read how the process works on How it works. For specific data types, see customer service data, code, documents and SOPs and healthcare. If you want to read about named companies in the market, our comparisons cover several vendors with the same source-cited approach.