Is It Legal to Sell Customer Data? A US Guide for AI Training

Yes, a US company can often sell or license customer data, but no single law says so, and for AI training the lawful route is usually to license de-identified records rather than personal data. The United States has no general federal privacy law. The answer comes from layers: what your own privacy policy promised, which the Federal Trade Commission enforces as a ban on unfair or deceptive practices; state laws such as California's, which treats any transfer of personal information to a third party " for monetary or other valuable consideration" as a sale; sector laws for health, money, children, students, faces and voices; and the contracts under which you hold the data. In the most closely watched sale of company records so far, Spirit Airlines' bankruptcy sale to Google, the agreement pays only on delivery of records certified as de-identified to California's standard.

This guide is for the founder, CEO, CFO or counsel of a US company that has been asked, or is wondering, whether it may license its records to an AI buyer. It walks through each layer in the order a buyer's counsel checks it: your own promises and the FTC, the state privacy laws (with a table of every state law in force), the sector laws, the point about business customers' data that most holders miss, de-identification, data broker registration, and the warranties buyers ask for. It ends with a worked example of a 60-person software company and answers to the questions people ask most.

Throughout, it separates two transactions that are often confused. Selling personal data means handing over information about identifiable people: a customer list, user profiles, contact records, purchase histories. Licensing de-identified operating records means handing over the working history of a business (support tickets, chat threads, CRM activity, documents) with the people in it removed to a legal standard. The first is regulated as a sale almost everywhere. The second is how most AI training licences are now structured, and it is the one this guide spends most time on.

This guide is research from public sources, not legal advice.

Contents

1. Why there is no single answer: five layers
2. Layer one: your privacy policy and the FTC Act
3. Layer two: state privacy laws
4. Layer three: sector laws
5. Layer four: B2B customer data belongs to your customers
6. De-identification: the main lawful path
7. Layer five: data broker registration
8. What buyers will ask you to warrant
9. Worked example: a 60-person B2B software company
10. Questions people ask
11. What this means for you

1. Why there is no single answer: five layers

Most countries answer "may I sell this data?" with one statute. The European Union has the GDPR; the United Kingdom has its own version of it. The United States does not. Congress has passed laws for particular sectors (health care, banks, children online, schools, video rentals) and left everything else to the FTC's general power over unfair and deceptive practices and to the states. Since California's law took effect in 2020, a growing list of states has passed comprehensive privacy laws of their own, each with its own definition of "sale", its own exemptions and its own thresholds. A US company's question is therefore never one question. It is five, and a sale is lawful only when all five come out right.

The five layers also have different enforcers, which matters for risk. The FTC acts on broken promises and harmful practices; state attorneys general (and in California a dedicated agency) enforce the state laws; private plaintiffs can sue under a few sector laws, most notably Illinois' biometric law; and your own business customers enforce their contracts. A sale that clears the statutes can still breach a customer contract, and a sale that clears every contract can still contradict your privacy policy. A buyer's counsel works through all five, and so should you, before any sample leaves the building.

Selling people's data versus licensing a business's records

The distinction between the two transactions is not a technicality; it decides which of the five layers bite. An AI buyer training a model to do office work does not need to know who your customers are. It needs the structure of real work: a request arrives, people discuss it, a ticket is opened, a document changes, the customer is told, the ticket closes. That structure survives when names, email addresses and account numbers are replaced with consistent stand-ins. The buyer's goal and the privacy laws' goal therefore point the same way more often than holders expect.

The Spirit sale shows it. By the first round of the court auction, the most competitive bidders had agreed to bid on an asset schedule that expressly excluded personally identifiable information, and after the auction the parties agreed that passenger booking, refund and in-flight purchase systems would not be transferred in any form, de-identified or not. The price held. If you take one thing from this guide, take this: the question "is it legal to sell customer data?" is usually the wrong question for an AI deal. The better question is "can I license my operating records with the customers taken out, and what does taking them out require?" The rest of this guide answers both. For what buyers actually pay for and why, see what sells and the selling guide; the rights hub lists the 26 questions a buyer's counsel asks, of which 15 block a sale until answered.

2. Layer one: your privacy policy and the FTC Act

The first layer is the one holders most often skip, because it is not a statute you can look up. It is your own words. Section 5 of the FTC Act declares " unfair or deceptive acts or practices in or affecting commerce" unlawful, and for two decades the FTC has used it to treat a privacy policy as a binding promise. Deception means saying one thing and doing another: if your policy says you do not sell or share customer information, selling it is deceptive whatever any other law allows. Unfairness is broader: a practice is unfair when it causes or is likely to cause substantial injury that consumers cannot reasonably avoid and that is not outweighed by benefits. A sale people could neither foresee nor refuse can meet that test even where no promise was broken.

The FTC applied this directly to AI training in February 2024. Its technology staff wrote that companies holding user data now have "powerful business incentives" to turn it into fuel for AI, and warned that any firm that " reneges on its user privacy commitments risks running afoul of the law". It may be unfair or deceptive, the post says, to adopt more permissive data practices (it names AI training) and tell people only "through a surreptitious, retroactive amendment to its terms of service or privacy policy". The same post recalls the case of Gateway Learning, the maker of Hooked on Phonics, which the FTC charged after it changed its policy to share customer data without notice or consent. California wrote the same idea into its statute: the provision that lets data move with a business in a merger states that it does not authorize material, retroactive privacy policy changes.

What this means in practice: old data, old promise

The practical rule is simple to state and tedious to apply. Each record is governed by the policy in force when it was collected, not by the policy you publish today. A company that rewrites its privacy policy in 2026 to allow AI licensing has authority, at most, over data collected after the change (and only if the change was prominent enough to count). Data from 2019 to 2025 stays under whatever those earlier versions said, unless the people it describes give fresh, affirmative consent. The buyer-side rights questions put it the same way: the policies in force when the data was collected must have allowed this kind of transfer, or affected people must have consented again, or the data must be fully anonymised.

So the first task is archival. Collect every version of your privacy policy, terms of service and employee handbook since your earliest data (the Internet Archive's Wayback Machine usually has them), note the dates each was in force, and map each system's data to the version that governed it. Then read each version for three kinds of sentence. A "we do not sell" statement (California requires notice before any sale, so most policies say one way or the other) rules out selling personal information collected under it. A purpose limit ("we use your information to provide and improve the service") rules out handing it to an outside party for an unrelated purpose. A de-identified data clause ("we may share aggregated or de-identified information") is the sentence you hope to find, because it means a de-identified licence fits the promise rather than stretching it. Many policies have all three.

Why buyers fear this layer most: model deletion orders

The FTC's remedy is what makes this layer expensive for buyers, and so what makes buyers careful with holders. In several orders the agency has required companies to delete not only data they had no right to use, but the models trained on it. In the Everalbum case (final order May 2021), the photo app was required to delete users' photos and videos and " the models and algorithms it developed by using the photos and videos uploaded by its users". In the case against WW International and its Kurbo app (March 2022), brought under the children's privacy law, the order required the companies to delete children's data, destroy any algorithms derived from the data, and pay a $1.5 million penalty. In the Rite Aid case (December 2023), the proposed order required deletion of the images collected by its facial recognition system and of any algorithms or other products that were developed using those images and photos, and required Rite Aid to tell third parties that received them to delete them too.

Read those orders from the buyer's side. A lab or vendor that trains on your records cannot easily untrain: one buyer's licence states that deleting the data does not require it to delete or retrain any model, and one law firm advising holders warns that once data trains a model it "cannot practically be pulled back out". If your data turns out to have been collected under a promise that forbade the licence, the buyer's model is the asset at risk. That is why every serious buyer asks you to warrant that you had the right to license (section 8), why they prefer de-identified records (section 6), and why a holder who cannot show its old privacy policies will be offered less, or nothing.

3. Layer two: state privacy laws

The second layer is the state comprehensive privacy laws. They share a common shape: they give residents a right to know what a business collects, to delete it, and to opt out of its sale; they require opt-in consent for sensitive data in most states; and they define personal information broadly (anything reasonably linkable to a person) while excluding de-identified data that meets a defined standard. They differ in who they cover, what counts as a sale, and whether employees and business contacts count as "consumers". California is the strictest and the most important for AI licensing, so it comes first.

California: what "sell" means, and who must comply

The California Consumer Privacy Act (CCPA, as amended by the California Privacy Rights Act) applies to a for-profit business that does business in California and meets any one of three thresholds: annual gross revenue above an inflation-adjusted figure, buying, selling or sharing the personal information of a large number of consumers or households each year, or earning half or more of its revenue from selling or sharing personal information. The statute sets them at $25 million in revenue (adjusted), 100,000 consumers or households, and 50 percent of revenue. Note the third threshold: a small company that starts selling personal information can bring itself inside the Act by doing so.

The definition of a sale is the heart of it. To "sell" means "selling, renting, releasing, disclosing, disseminating, making available, transferring, or otherwise communicating" a consumer's personal information to a third party " for monetary or other valuable consideration". A licence of personal information to an AI buyer for a fee is a sale. So is a licence paid in compute credits, model access, a revenue share or a discount, because each is valuable consideration. The one relevant exception covers personal information transferred as an asset in a merger, acquisition or bankruptcy " in which the third party assumes control of all or part of the business". A data-only licence to an AI buyer, in which the buyer does not take over your business, is not that transaction.

If you do sell personal information, the duties follow. You must tell consumers beforehand that this information may be sold or shared and offer the opt-out; every consumer has the right to opt out of sale or sharing at any time; and you may not sell the personal information of a consumer you know is under 16 unless the consumer (or, under 13, a parent) has affirmatively authorized it. For a holder this means three concrete things: a sale your privacy notice never mentioned is a sale without notice; every person who ever opted out must be removed from what you license; and records about minors come out entirely.

Sensitive information, messages and models

California adds a separate category of sensitive personal information: government ID numbers, account credentials, precise location, racial or ethnic origin, religious beliefs, union membership, genetic and biometric data, health, sex life, and (critically for company records) " the contents of a consumer's mail, email, and text messages unless the business is the intended recipient of the communication". Consumers may direct a business to limit its use of sensitive personal information to what is needed to provide the service they asked for. Licensing sensitive information for AI training is the hardest case under California law, and under most other state laws it needs opt-in consent outright.

Two further California points change how a holder should think. First, since 1 January 2023 California is the one state whose law fully covers employees, job applicants and business contacts: the exemptions that kept workforce and business-to-business data out of the Act expired on that date, so a California employee's messages and a client contact's email address are personal information like any other (the employee messages guide covers this in detail). Second, the Act now says personal information can exist in " artificial intelligence systems that are capable of outputting personal information". A buyer that trains on records still holding personal information is therefore building a model that California may treat as holding it too. That is one more reason buyers ask for records that are de-identified before they arrive.

The other states

Outside California, the state laws follow a model first enacted in Virginia and refined in Colorado and Connecticut. The table lists the comprehensive state privacy laws that apply to businesses generally and were in force on 17 October 2026, with the three features that matter most for an AI licence: what counts as a sale, how sensitive data is treated, and whether the law's de-identified standard has the three conditions (reasonable measures, a public commitment, and contracts binding recipients) that California's has.

Read the table for four patterns. First, a cash licence is a sale everywhere. Six states (Virginia, Utah, Iowa, Tennessee, Indiana and Kentucky) count only an exchange for money; the rest count "monetary or other valuable consideration", which reaches compute credits, model access or a revenue share. Structuring a licence around non-cash payment to avoid a "sale" in a money-only state is a poor plan: California, the FTC and the majority of states would still treat it as one. Second, sensitive data is the hard stop. Almost every state requires opt-in consent before sensitive data (health, biometrics, precise location, children's data, and in several states more) is processed at all, which includes disclosing it for training; only Utah and Iowa settle for notice and an opt-out. Some go further and ban its sale outright: Maryland bars selling sensitive data even with consent, New Jersey's 2026 law does the same, and Oregon and Virginia ban the sale of precise location data.

Third, outside California, employees and business contacts are mostly not "consumers". The Virginia model defines a consumer as a resident acting in an individual or household context and says it "does not include a natural person acting in a commercial or employment context"; Texas says the same in nearly the same words. For a business-to-business company, that takes much of its CRM and all of its staff data outside every state law except California's, which is why California usually sets the standard for company records. Fourth, the de-identification test is nearly uniform. Every state in the table except Iowa uses the same three conditions as California (reasonable measures, a public commitment not to re-identify, and contracts binding recipients), so records de-identified to the California standard meet the others too.

Four more states have enacted comprehensive laws that are not yet in force: Oklahoma from 1 January 2027, Louisiana from 1 January 2027, Alabama from 1 May 2027 and Vermont from 1 January 2028. A licence signed today usually runs for years, often with perpetual rights in trained models, so a holder should apply the standards of the laws coming into force as well as those already in it. Each law also has its own applicability thresholds (most turn on the number of state residents whose data a business handles), which this guide does not reproduce; check them for the states where your customers live.

4. Layer three: sector laws

Congress regulates some kinds of data wherever they sit, whatever state the company is in. Each of these sector laws has its own rule for disclosure, its own idea of consent and, usually, its own de-identification standard. None of them cares whether you call the transaction a sale or a licence: the question is whether identifiable regulated data leaves your hands. Several give the people in the data a right to sue for fixed damages per person, which turns a large dataset into a large class action. A holder should screen for these categories before anything else, because for most of them the practical answer is to exclude the data rather than try to meet the law's terms.

The table sets out the laws a company's records most often touch. The test for each is not what kind of company you are, but what is in the records: a software company whose support inbox holds patients' questions, children's accounts or video viewing histories is carrying regulated data whether it meant to or not.

Two of these deserve a word more. Illinois' biometric law is the strictest privacy statute in the country for this purpose: its bar on profiting from biometric identifiers has no consent exception in its text, and plaintiffs use it. In October 2026, for example, two proposed class actions accused delivery-robot companies of retaining customers' face and hand scans for AI training. Voice recordings, video calls and photos are where biometric data hides in company records. Children's data is the other: the FTC wrote into the amended COPPA rule that disclosure for AI training always needs the separate parental consent, and a proposed class action filed in October 2026 accuses a major AI developer of training on children's information (allegations only).

Sector laws also hide in professional records. An accounting firm's client email carries tax return information, which a tax preparer may not disclose or use for another purpose without consent (the employee messages guide works through that case). And one federal rule cuts across all of them: the Justice Department's bulk data rule restricts transfers of sensitive personal data above set volumes to buyers linked to countries of concern, so any foreign buyer must be screened (section 6 explains why de-identification does not take data outside it).

5. Layer four: B2B customer data belongs to your customers

This is the point most holders miss, and it is the one that most often ends a sale of a software company's records. If your business sells software or services to other businesses, a large part of the data in your systems is not, in law or in contract, yours. It is your customers' data, which you hold and process for them under their contracts. A project-management tool holds its customers' projects; a help-desk product holds its customers' tickets with their customers; an analytics product holds its customers' event streams. You can see all of it, and it sits in your database, but your right to use it is whatever your customer contracts and data processing agreements (DPAs) say. They almost always say: only to provide the service.

California's law makes the point in its definitions. A service provider is a business that processes personal information on another business's behalf under a written contract, and that contract must prohibit the service provider from " (A) Selling or sharing the personal information" and from using it for any purpose other than the business purposes in the contract. The regulations list the purposes a service provider may still use the information for, and they are narrow: only for the listed purposes, which include improving the service it provides to that business, but never selling it on to an outside party. A software company acting as a service provider that licensed its customers' personal information to an AI buyer would be breaching the very contract terms that make it a service provider, and the buyer would be receiving personal information sold without the customer's notice or opt-out. The European rule is starker: a processor that decides for itself what to do with customer data is treated as a controller for that processing, with every duty that brings.

Three buckets of data in a software company

The working method is to sort everything you hold into three buckets before any other review. Getting this sort right removes most of the risk in one pass.

The line between the first two buckets is not always clean. A support ticket in which your customer's administrator describes a bug is your relationship record; the screenshot of their customer list attached to it is customer content. A good rights file treats attachments, uploaded files and anything copied out of the product as customer content by default. The safest lots are built from the third bucket and the de-identified second bucket, which is also where the AI value is: the history of how your own company worked.

Your software vendors' terms are a second contract layer

There is a mirror image of the same problem. Your company is itself the customer of a dozen software vendors (Slack, Microsoft, Google, Salesforce, HubSpot, Zendesk and the rest), and their terms govern how data leaves their systems. Most let you export your own data freely: of the 18 business systems whose terms we read, 13 let the customer export its own data without a limit that matters for a sale. But 6 bar using data pulled through their API to train AI models. Slack's API terms, for example, say an app may not " use API Data to train a large language model"; HubSpot's developer terms bar " using data accessed through the APIs to train, fine-tune, or improve AI or machine learning models". The practical rule is to export each system with its own built-in admin export, not through a buyer's connector app. The per-system pages ( Slack, Microsoft 365, Google Workspace, Salesforce, HubSpot, Zendesk) say what each vendor's terms allow.

6. De-identification: the main lawful path

If personal information is the problem, removing the persons is the solution, and the law recognises it. Under California's Act, " personal information does not include consumer information that is deidentified". Records that meet the standard are outside the definition of personal information, and so outside the rules on selling it: no notice of sale, no opt-out, no sensitive-data limit. Most other state laws have the same exclusion. This is why the AI licences that have been made public are licences of de-identified records, and why the Spirit sale agreement tied the whole payment to certified de-identification.

The standard is demanding, and it has three parts. Information is "deidentified" only if it cannot reasonably be linked to a particular consumer and the business that holds it: " (1) Takes reasonable measures to ensure that the information cannot be associated with a consumer or household. (2) Publicly commits to maintain and use the information in deidentified form and not to attempt to reidentify the information ... (3) Contractually obligates any recipients of the information to comply". The first is technical. The second and third are legal and are often forgotten: you must publish a commitment not to re-identify (a sentence in your privacy policy does it), and your licence must bind the buyer, and anyone the buyer passes the data to, to the same. A licence without a no-re-identification clause and a flow-down clause does not meet the standard, however good the scrubbing.

What "reasonable measures" means for messy records

The technical part is harder than it looks, because removing names is not enough. A study in Nature Communications estimated that 99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes, and Latanya Sweeney's classic work found that 87% of the US population was likely unique on just ZIP code, gender and date of birth. Company records are worse than tables, because the identifying detail sits in free text: "the Denver office manager who was out after her surgery", a signature block, a customer's unusual product configuration. De-identifying a Slack archive means finding people wherever they are mentioned, replacing them consistently so the threads still make sense, coarsening details that identify in combination, and then testing the result as an attacker would.

That is why the Spirit model matters as a template. The sale agreement made delivery of the records, and so payment, conditional on certification by an independent de-identification agent to the California standard, and the ombudsman's report describes quality-control testing against direct and indirect identifiers, free text and record linkage before transfer. Who controls de-identification is also a priced term: in the same auction one bidder offered 35% more for the same records if it could de-identify them itself rather than through a third party, and the estate chose the lower third-party bid. Buyers handle this differently. Scale says it anonymises everything inside your environment before anything is extracted; micro1 says personal information is removed "when relevant"; Handshake says that no de-identification process eliminates the risk entirely. The holder who controls the standard, chooses the provider and sees the test result before anything leaves is the holder who can stand behind its warranties. The de-identification standard page sets out ours, and the de-identification tool runs on your own computer.

Where de-identification does not reach

De-identification solves the state privacy laws. It does not solve everything, and a holder should know the four places it falls short. Contracts: if a customer contract limits use of customer content to providing the service, de-identifying that content does not turn the licence into providing the service (section 5). Your own promises: a privacy policy that says "we never share your data in any form" may cover de-identified derivatives too; read the words (section 2). Federal national-security rules: the Justice Department's bulk data rule restricts transfers of sensitive personal data to buyers linked to countries of concern, and its own guidance says the covered data counts " regardless of whether the data is anonymized, pseudonymized, de-identified, or encrypted", so foreign buyers must be screened whatever the scrub. Other standards: health data has HIPAA's own de-identification test (see the HIPAA de-identification guide), and data about people in the EU or UK must meet the stricter European anonymisation test, under which de-identified data stays personal data unless it is truly anonymous.

7. Layer five: data broker registration

The last layer is the newest. Several states now require data brokers to register with the state, disclose what they do and, in California, process deletion requests from a central list. A company licensing records to an AI buyer needs to know two things: whether it becomes a data broker by doing so, and whether its buyer is one. As of 30 September 2026, California, Oregon, Texas and Vermont had active registration procedures (as reported by one law firm's survey), Connecticut's registry begins on 1 January 2027, and New Jersey enacted its own framework in June 2026.

For most holders the first answer is no, because the definitions turn on people the business has no relationship with. California defines a data broker as a business that " knowingly collects and sells to third parties the personal information of a consumer with whom the business does not have a direct relationship". Vermont's definition expressly includes a business that "sells or licenses" such data, but again only about people with no direct relationship; Oregon's excludes data about a business's own customers, subscribers or users; Connecticut exempts data about people in a contractual relationship with the business. Texas goes further for holders: its definition does not include deidentified data, employee data, or publicly available information, and the chapter applies only to a broker that earns more than half its revenue from processing or transferring personal data. A company licensing records about its own customers and staff is not a broker under these laws.

New Jersey is the exception to watch. Its 2026 law creates a second category, the data collector: a business that collects personal data about people it does have a direct relationship with and sells or licenses such personal data to a data broker. A data collector must register every year and pay a fee scaled to the number of New Jersey consumers. The same law says that a data broker or data collector may "in no case" sell or license sensitive data, and that no controller may sell it, with a civil penalty per record. So a holder licensing identifiable records to a buyer that resells data could become a New Jersey data collector, and identifiable sensitive data about New Jersey residents cannot be sold at all.

The second answer matters for your choice of buyer. A buyer that resells or sublicenses personal information about your customers is, for those people, a business with no direct relationship, and so a data broker in most of these states, with registration, disclosure and California deletion duties. Licences that permit sublicensing through multiple tiers make that more likely. De-identified records avoid both questions: the broker laws, like the privacy laws, are about personal information.

8. What buyers will ask you to warrant

Every layer above ends up in one place: the warranties in the licence. A warranty (or representation) is a contractual statement of fact that the holder makes and stands behind; if it turns out to be false, the buyer can claim damages, demand a refund or end the licence. Buyers write warranties to move the risk of the five layers onto you, because they cannot check your old privacy policies, customer contracts or consent records themselves. The model deletion orders in section 2 are why: a buyer whose model is built on data you had no right to license wants someone to pay for it.

The warranties buyers ask for are published, at least in outline. The table collects what each buyer's own terms or intake pages say, as read in October 2026.

The way to give these warranties without taking unlimited risk is the way experienced holders already do: give title and authority without qualification (you either own the records or you do not), and qualify the rest by knowledge, tied to a disclosed schedule. The schedule is your rights file: the systems, the dates, the privacy policy versions that governed them, what was excluded and why, the de-identification standard and its test result. The warranty then says that the disclosed facts are accurate, not that the data is perfect. Cap the indemnity, limit it to defined claim types, and carve out claims caused by the buyer's own use or by combining your records with other data. And make the buyer's mirror-image promises part of the same document: no re-identification, no onward transfer of the raw records, the same terms for anyone it passes data to. Those are not only good practice; under California's definition they are part of what makes the records de-identified at all.

One more consequence is easy to miss. A training-data licence you sign now becomes a disclosure you must make when your own company is sold. In one public acquisition agreement, the company being sold had to represent that, except as scheduled, it had not let any company or customer data, "including deidentified data", be used as training data by anyone else. Keep the licence, the rights file and the test results together; an acquirer's counsel will ask for all three.

9. Worked example: a 60-person B2B software company

Example (illustrative, not a real company). A 60-person company in Austin sells project-management software to mid-sized construction firms across the US. It has been running for eight years. An AI vendor has offered to license "your company data" for training agents that do construction project work. (The page for software companies describes what buyers look for in this kind of company record.) The company holds five things: its product database (customers' projects, schedules, punch lists and files); eight years of support tickets in Zendesk; eight years of Slack; a HubSpot CRM with contacts at construction firms across the country; and its own engineering wiki and code.

Step 1: sort into buckets. The product database is customer content. The company's master subscription agreement says customer data may be used "solely to provide and support the Services", and its DPA appoints it as a service provider. That bucket is out, de-identified or not, unless each customer signs an amendment, which the company decides not to seek. The Zendesk tickets are mixed: the conversations are the company's own relationship records, but the attachments (screenshots of customers' schedules, exported reports) are customer content, so attachments are excluded. Slack, the wiki and the code are the company's own operating records. HubSpot is a relationship record.

Step 2: read the old promises. The privacy policy has existed in three versions. All three say "we do not sell your personal information"; the 2023 and later versions add that the company "may share de-identified or aggregated information". The company concludes that it may license de-identified records collected at any time (the 2018 to 2022 policy did not forbid it, and the records will contain no personal information), but no personal information at all. It also adds a public commitment not to re-identify, meeting the second part of the California standard.

Step 3: apply the state laws. The company has customers' staff and its own employees in California, so the CCPA applies to the people in its records whether or not the company meets a threshold today, because licensing personal information could itself meet one. Business contacts in HubSpot and its own California employees are personal information under the California Act. Under the other state laws, which exclude people acting in a commercial or employment context, they mostly are not. The company decides to apply the California standard to everything, which also satisfies the others.

Step 4: sector laws and vendor terms. Construction is not a regulated sector, but some tickets mention workers' injuries on job sites, which is health information about identifiable people. Those are flagged and excluded. No children's, student, biometric or video data is present; one Slack channel holds recorded sales calls, which come out (recording consent is covered in the employee messages guide). Each system is exported with its own admin export: Zendesk exports are not enabled by default and need the account owner to ask support; HubSpot's developer terms bar training on API data, so no connector is used.

Step 5: de-identify and contract. The remaining lot (Slack, wiki, code, de-identified tickets without attachments, and CRM activity with contacts replaced by stand-ins) is de-identified by a provider the company chooses, tested, and certified before any sample is shown. The licence binds the buyer not to re-identify and to flow that down, gives knowledge-qualified warranties tied to the rights file, and excludes privileged material. Result: the company does not sell customer data at all. It licenses a de-identified record of how a construction-software company ran for eight years, which is what the buyer wanted, and the customer data that would have ended the deal never leaves.

Questions people ask

Is it legal to sell customer data in the US? Often, but not by default. You must keep the promises in your privacy policy, follow your state privacy laws (in California, a transfer for any valuable consideration is a sale, requiring notice and an opt-out), stay out of sector-regulated data, and hold the data for yourself rather than for customers. For AI training, licensing de-identified records avoids most of these limits.

Can I sell my customer list? A list of identifiable customers is personal information, so selling it is a "sale" under every state privacy law and must match what your privacy policy promised. If your policy says you do not sell personal information, you cannot sell the list without fresh consent. AI buyers rarely want a bare list anyway: they want records of work, which can be licensed with the personal information excluded.

Do I need customer consent to sell their data? For ordinary personal information, most state laws require notice and an opt-out rather than consent; for sensitive data (health, precise location, message contents and the like) most require opt-in consent; and for minors under 16 California requires affirmative authorization. Consent is also the only way to sell data your old privacy policy said you would not sell.

Is selling anonymized data legal? Generally yes, if it truly meets the legal standard. California excludes de-identified information from personal information, but only when you take reasonable measures, publicly commit not to re-identify, and contractually bind every recipient. Removing names alone does not meet it.

Can a SaaS company use customer data to train AI or sell it for AI training? Using it to improve the service you provide to that customer is sometimes allowed; licensing it to an outside AI buyer generally is not. A service provider's contract must prohibit it from selling or sharing the personal information it processes, and customer DPAs usually limit use to providing the service.

Do I have to register as a data broker? Usually not, if you license records about your own customers and staff: the California and Texas definitions turn on personal data about people a business has no direct relationship with, and Texas excludes deidentified data and employee data. A buyer that resells personal information may have to register.

What happens if a company sells data it had no right to sell? The FTC can treat it as deceptive or unfair, and its orders have required deletion of the data and of models built from it, as in the Everalbum and Kurbo cases. Under the licence, the buyer will also claim on your warranties.

What this means for you

If you are considering licensing your records to an AI buyer, work through the layers in this order before you share any sample. First, sort everything into customer content, your relationship records and your operating records, and set the first aside unless each customer agrees in writing. Second, collect every version of your privacy policy, terms and handbook, and map each data set to the version that governed it. Third, check the state and sector layers: who is in the records, where they live, and whether any health, financial, children's, student, biometric or recorded-call data must come out. Fourth, decide the de-identification standard (California's three parts, plus HIPAA or the European test where they apply), choose who does it, and test it before anything leaves. Fifth, write the rights file that your warranties will rest on.

The readiness tool walks through the rights questions in order, the rights hub explains each one, the licence terms page shows the clauses buyers propose, and the legal ledger tracks privacy suits over training data, such as the biometric class actions filed in Illinois in October 2026. To see what your records might fetch once the rights are settled, start with a valuation. AIData.Broker acts as the holder's agent in these sales: the holder licenses its own data to the buyer, and we settle the rights file with the holder before any buyer sees anything. For the other rights questions, see the guides on selling customer data in bankruptcy, employee emails and chats, data licensing agreements and training data provenance.