Don't Do It Yourself

Is Selling Company Data to AI Safe? Privacy, Legal and Security Risks (2026)

Last updated: October 27, 2026

Selling company data to AI is safe if you license only records you fully own, keep personal and client information out, and sign a contract that limits what the buyer can do with the files. It isn’t risk-free: buyers have been breached, de-identified data can sometimes be traced back to people, and nothing comes back out of a trained model.

Most of the risk sits in one place: personal information inside your records. That’s why the biggest section on this page covers how it gets removed, and who does it.

For the full process, see how to sell your company’s data to AI companies. To compare buyers, see our list of companies that buy business data for AI.

This isn’t legal advice. Have a lawyer review any data contract before you sign it.

Key takeaways

  • Personal information in your records is the main risk. Some buyers remove it after transfer, some inside your systems, and some brokers leave it to you.
  • The FTC says data “is only anonymous when it can never be associated back to a person.” Swapping names for codes doesn’t meet that bar.
  • Law firms, accountants and medical practices face extra rules: ABA Model Rule 1.6, IRC Section 7216 and HIPAA.
  • Mercor was hit by the LiteLLM supply chain attack in March 2026. It says only a “very limited subset” of nearly five million experts had sensitive data affected.
  • You can delete a buyer’s raw copies, but you can’t remove your data from a model already trained on it.
Yes, a business can legally license its own data to an AI company if it holds the rights and keeps the privacy promises it made when it collected the data. Selling company data to AI becomes illegal when the records belong partly to clients, or contain personal information you can’t lawfully share.

Our guide covers the main laws in more detail. The rest of this page is about the risks that remain even when a deal is legal.

How is personal information removed before company data reaches AI?

Personal information is removed by the buyer, a specialist partner or you, depending on the buyer. Most use automated detection tools first, then redaction or synthetic rewriting, then human review of samples. None of them can promise the risk is zero.

Removing personal information from company data is the first worry most owners have, and with reason. Your email, CRM and support tickets name customers, staff and suppliers on almost every page.

Who removes personal information: the seller, the buyer or both?

The buyer removes personal information in most deals, but the timing differs. micro1 does it after you transfer the data, Mercor during its own export, Scale AI and Endless Data inside your systems before anything leaves, and AIData.Broker makes you do it yourself.
Buyer Who removes personal information When
micro1 micro1; you don’t need to anonymize first After transfer, before data is shared with third parties
Mercor Mercor, which also covers the cost During extraction
Scale AI Scale AI Inside your environment, before extraction
Endless Data Endless Data Before data leaves your company
Polyshares A specialist partner Before any lab sees it
Troveo Troveo, to its own standard Before licensing
AIData.Broker You, on your own computer, to its published standard Before you send anything

“Inside your systems” is the safer model, because raw personal information never leaves your company. Our buyer comparison has each buyer’s full terms. Either way, ask to review de-identified samples before the data is used.

How do AI data buyers screen out personal information?

AI data buyers screen out personal information in three passes: automated detection finds names, numbers and addresses, redaction or synthetic rewriting replaces them, and people check samples. For health data, the standard is HIPAA’s Safe Harbor list of 18 identifiers or an expert’s sign-off.
  • Automated detection. Software scans text, files and metadata for patterns like names, email addresses, phone and account numbers. micro1 says its own “flow-transform anonymization model” does this step.
  • Redaction or rewriting. Detected details are deleted, replaced with placeholders, or rewritten. micro1 says some datasets get “synthetic rewrites that preserve workflow structure.”
  • Human review. People check samples the tools missed. micro1 and Endless Data both let you review samples first.
  • A named standard. HIPAA allows Safe Harbor, which removes 18 kinds of identifier, or Expert Determination, where a qualified expert certifies the re-identification risk is very small. Outside health care, the Spirit Airlines sale used California’s CCPA standard.

Ask which standard a buyer uses, by name. “We anonymize everything” isn’t a standard. Our step-by-step guide to anonymizing company data for AI training covers the tools and the process.

Do data wholesalers and marketplaces strip personal information before passing data to AI labs?

Some do and some don’t. Endless Data, which aggregates data for several labs, strips identifiers before data leaves your company and never licenses personal data. AIData.Broker, a broker, makes you remove personal information yourself before it pitches your data to buyers.

Marketplaces that simply list datasets, such as Datarade, leave the dataset’s contents to the seller. If a broker or marketplace sits between you and the lab, get in writing who de-identifies the data and to what standard. Don’t assume the middleman does it.

What counts as personal information in business records?

Personal information in business records is anything that identifies a person, directly or combined with other details. That includes far more than names: email addresses, phone numbers, street addresses, account numbers, faces in photos, voices in recordings and the metadata inside files.
Where it hides Examples
Text Names, signatures, email footers, phone numbers, addresses
Numbers Account, policy, invoice, case and employee ID numbers
Images and video Faces in site photos, ID scans, screenshots of other people’s screens
Audio Voices in call recordings and voicemails
File metadata Document authors, “last edited by”, GPS tags on photos
Combinations Job title plus office location plus start date can point to one person

Metadata is the one owners miss most. A de-identified PDF can still list its author in the file properties.

Can redacted data be traced back to a person?

Yes, sometimes. A 2019 Nature Communications study found that 99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes. Removing names doesn’t remove the details around them.

The FTC said in a July 2024 post that data “is only anonymous when it can never be associated back to a person,” and that hashed identifiers “can still be used to identify users.”

Daniel Goldberg of the law firm Frankfurt Kurnit Klein & Selz gives the safest rule in his October 2026 guide: “exclude high-risk data entirely rather than relying on de-identification.”

You need client consent when your contract, NDA or professional rules say client information stays confidential, and most service contracts do. De-identifying client files doesn’t cancel those promises, so selling client files to an AI buyer without consent can breach the contract even if no name survives.

Mercor’s design agency listing rules out “client work you can’t license,” and Polyshares says its license doesn’t cover what you hold on behalf of your clients. Read your client agreements before you count client work as yours to sell.

What about employee emails, Slack messages and HR files?

Leave HR files out entirely. Employee emails and Slack messages can be licensed after de-identification, but they carry two extra risks: California’s CCPA has covered employee data since January 1, 2023, and workplace messages can be protected activity under the National Labor Relations Act.

Goldberg’s guide flags the labor-law point: Section 7 of the NLRA protects employees discussing pay and conditions with each other. Flight attendant and pilot unions objected to the Spirit Airlines sale over employee data.

Check your chat tool’s terms too. Goldberg names Slack’s API terms as one example of platform rules that limit what you can do with exported data.

Do HIPAA, GDPR and state privacy laws like the CCPA apply?

HIPAA applies if you’re a health care provider, plan or their business associate. GDPR applies to records about people in the EU. California’s CCPA applies to businesses that meet its size thresholds, and treats handing personal information to an AI buyer as a sale unless the data is properly de-identified.
Law When it applies What it means for a data sale
HIPAA Health care providers, health plans and their business associates Selling patient information needs the patient’s authorization; properly de-identified data isn’t covered
GDPR Records about people in the EU AI training is a new purpose, so you need a lawful basis for it
CCPA California businesses over its revenue or data thresholds Disclosing personal information for value is a sale, with notice and opt-out rights
Your own privacy policy Always The FTC warned in February 2024 that quietly loosening your terms for AI use can be unfair or deceptive

Other states, including Virginia, Colorado and Texas, have their own privacy laws. Ask your lawyer which apply to the people in your records.

What special rules apply to law firms, accountants and medical practices?

Law firms, accountants and medical practices each have a confidentiality rule on top of privacy law: ABA Model Rule 1.6 for lawyers, IRC Section 7216 for tax preparers and HIPAA for medical practices. Each one generally requires the client’s or patient’s consent before their information is used outside the work itself.
Profession Rule What it means for selling data
Law firms ABA Model Rule 1.6 and state bar rules Client information is confidential. ABA Formal Opinion 512 (July 2024) says lawyers need informed consent before putting client information into AI tools that learn from it
Accountants and tax preparers IRC Section 7216 Using or disclosing tax return information for anything beyond preparing the return needs the taxpayer’s written consent; knowing violations are a crime
Medical practices HIPAA Patient records can’t be sold without authorization; de-identified data must meet Safe Harbor or Expert Determination

Firm templates, internal SOPs and training material you wrote yourself are usually outside these rules. Client and patient files almost never are. Our guide to selling law firm data to AI covers Rule 1.6, former clients and public filings in detail.

What happens if the buyer gets hacked?

If a buyer is hacked, your files can leak from its systems, and you’ll have little control over what happens next. Mercor was hit by a supply chain attack through the open-source LiteLLM library in late March 2026, and said on June 25 that only a “very limited subset” of nearly five million experts had sensitive information affected.

The hacking group Lapsus$ claimed to have taken about 4 TB of Mercor data, Cybernews reported. Mercor hasn’t confirmed that figure. Several proposed class actions from contractors followed within a week of the news.

For companies, the relevant line is Mercor’s own: because many customers work on their own platforms, “the impact to customer information was very limited.” Limited isn’t none. Our Mercor data breach guide covers what was exposed and the lawsuits in full.

Scale AI had a different lapse. Business Insider reported in June 2025 that it tracked confidential work for Google, Meta and xAI in public Google Docs. Send the least data you can, and write a deletion date and a breach-notice deadline into the contract.

Can your data show up in ChatGPT answers or reach a competitor?

It can. A 2023 study by researchers from Google DeepMind and several universities extracted gigabytes of training data from open models and from ChatGPT. Nobody has published a test of one company’s licensed records, but verbatim recall is a known behavior.

The competitor risk is quieter. Your pricing logic, playbooks and processes are what make the data valuable, and the model that learns them may be sold to your competitors.

Goldberg’s guide says data used to train a model “cannot practically be pulled back out.” Leave out anything you’d be harmed to see repeated in someone else’s chatbot. Our explainer on what AI training data is covers how memorization works.

Does selling data break your NDAs or client contracts?

Selling data can break NDAs and client contracts if the files contain the other party’s confidential information, even inside your own email. The Spirit Airlines data sale drew objections from vendors including Navitaire and Amadeus, who asked the court to enforce their confidentiality agreements.

Search your contracts for confidentiality, data use and “no AI training” clauses before you scope the deal. Some newer vendor and client agreements ban AI training outright.

What does a safe data deal look like?

The Spirit Airlines sale is the most detailed public example. Before the court’s privacy ombudsman backed Google’s $10 million bid in an October 5, 2026 report, four customer systems were pulled out and Tonic.ai was hired to de-identify the rest.

Passenger transactions, refunds, inflight purchases and WiFi sales were excluded entirely. Corporate email stayed in, de-identified by Tonic.ai before transfer. The buyer publicly committed not to re-identify the data, and can’t pass it on unless the next recipient makes the same promise.

A small business won’t have an ombudsman, but you can copy the structure: a written list of excluded systems, a named party doing the de-identification and a no-re-identification promise in the contract. Our guide to AI data licensing agreements covers what to ask for in each clause.

What should you check before you sell company data to AI?

Check these before you transfer anything:

  1. List every system you’d include, and the date range for each.
  2. Remove HR files, health records, payment data and privileged or client-confidential files.
  3. Search client, vendor and employee agreements for confidentiality and AI-training clauses.
  4. Check your privacy policy said what this use needs it to say.
  5. Confirm who de-identifies the data, by which named standard, and when.
  6. Review de-identified samples, including file metadata.
  7. Get a no-re-identification promise that passes to anyone the buyer licenses to.
  8. Write a deletion date for raw files and a breach-notice deadline into the license.
  9. Agree on an indemnity: who pays if a client, employee or regulator objects.
  10. Decide who you’ll tell, staff and clients, and when.

micro1, for example, publishes its privacy commitments but not its license terms. Our micro1 review shows what to ask for.

Common mistakes that make selling data unsafe

The most common mistake is assuming the buyer’s de-identification makes any record safe to send. These are the others:

  • Exporting a whole Slack or email archive instead of choosing folders and date ranges.
  • Forgetting file metadata, which can name authors after the text is clean.
  • Counting client work as yours because it sits on your server.
  • Taking a buyer’s website privacy promises as contract terms.
  • Assuming a broker or marketplace removes personal information.
  • Expecting a trained model to forget your data after the deal ends.

Frequently asked questions

Is selling company data to AI safe?

It’s reasonably safe if you license only data you fully own, leave out personal, health, HR and client records, and sign a license with de-identification, deletion and indemnity terms. Most risk comes from personal information you didn’t remove.

Who removes personal information before data reaches an AI company?

Usually the buyer. micro1 does it after transfer, Mercor during its own export, Scale AI and Endless Data do it inside your systems, and some brokers, like AIData.Broker, make you do it yourself.

Is anonymized company data really anonymous?

Not always. The FTC says data is only anonymous when it can never be linked back to a person, and a 2019 study found 99.98% of Americans could be re-identified from 15 demographic attributes.

Can a law firm or accountant sell client data to AI?

Only with client consent in most cases. Lawyers are bound by ABA Model Rule 1.6, and tax preparers need written taxpayer consent under IRC Section 7216 to use return information for anything else.

What happened in the Mercor data breach?

Mercor was hit by a supply chain attack through the LiteLLM library in late March 2026. It said on June 25, 2026 that only a very limited subset of nearly five million experts had sensitive information affected.

Can I get my data back after selling it to AI?

You can ask the buyer to delete its raw copies if the license allows it, but you can’t remove data from a model already trained on it.

Don't Do It Yourself
Privacy Overview

This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.