Geoff CuddLast updated: October 29, 2026
To anonymize company data for AI training, you remove or replace everything that identifies a person, including names, contact details, ID numbers, faces, voices and file metadata, so the records can’t be linked back to anyone. Most AI data buyers do part of the job, but what leaves your company is still your responsibility.
This guide is the hands-on version. For the full selling process, start with our guide to how to sell your company’s data to AI companies.
This isn’t legal advice. Have a lawyer confirm which privacy laws apply to your records.
Key takeaways
- Anonymized data can’t be traced to a person. Pseudonymized data, where names become codes, still counts as personal data under GDPR.
- HIPAA’s Safe Harbor list of 18 identifiers is the most concrete checklist, even for businesses outside health care.
- micro1 and Mercor remove personal information for you. AIData.Broker makes you do it first.
- Free and paid tools find most personal information, but none of them promise to find all of it. A person still has to check samples.
- File metadata, like a document’s author, is the step owners skip most.
What does it mean to anonymize data for AI training?
AI buyers want your records for how work gets done: how a support ticket gets resolved, how a contract gets negotiated, how a project gets scheduled. They don’t need to know who the customer was.
That’s why anonymization usually costs a buyer little. A support ticket that says “[CUSTOMER] reported the invoice was wrong” teaches a model almost as much as one with the real name.
What’s the difference between anonymization, de-identification and pseudonymization?
| Term | What it means | Still personal data? |
|---|---|---|
| Anonymization | No one can link the data back to a person, even with other information | No. GDPR Recital 26 says data protection rules don’t apply to anonymous information |
| De-identification | Identifiers removed to a set standard, such as HIPAA Safe Harbor or California’s CCPA definition | Not under that law, if the standard is met |
| Pseudonymization | Names and IDs swapped for codes, with a key kept somewhere | Yes. GDPR treats it as information on an identifiable person |
When a buyer says it “anonymizes” your data, ask which of these it means and which standard it follows. Our guide to whether selling company data to AI is safe covers why the difference matters.
Do you have to anonymize data yourself before you sell it?
| Buyer | Who removes personal information | When |
|---|---|---|
| micro1 | micro1 | After transfer, before data goes to third parties |
| Mercor | Mercor, at its 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 |
| AIData.Broker | You, to its published standard | Before you send anything |
Even when the buyer does it, you decide what goes in. The safest data is data you never send, so the first steps below apply to every deal. Our list of companies that buy business data for AI has each buyer’s full terms.
What personal information do you need to remove?
The 18 identifiers come from HHS’s de-identification guidance. They cover the person, plus their relatives, employers and household members.
| Group | Safe Harbor identifiers | Where they hide in business records |
|---|---|---|
| Names and contact details | Names; addresses smaller than a state; phone and fax numbers; email addresses | Email signatures, CRM records, invoices, delivery notes |
| Dates | All parts of a date except the year, when tied to a person, and ages over 89 | Birth dates in HR files, appointment and service dates |
| Government and account numbers | Social Security, medical record, health plan, account, and certificate or license numbers | Payroll, billing, contractor files, insurance claims |
| Devices and online IDs | Vehicle IDs and plates, device serial numbers, URLs, IP addresses | Fleet logs, IT tickets, web analytics exports |
| Images and biometrics | Full-face photos and comparable images; fingerprints and voice prints | Site photos, ID scans, call recordings |
| Anything else unique | Any other unique identifying number, characteristic or code | Employee IDs, case numbers, customer account codes |
Addresses can keep the first three digits of a ZIP code, but only where that area holds more than 20,000 people. Outside health care, nobody requires the full list, but it’s the most specific one any regulator has published.
How do you anonymize company data, step by step?
- Cut whole systems first. HR files, payroll, health records, payment data and privileged client files usually shouldn’t go at all. Removing a folder is safer than cleaning it.
- Export only what the buyer needs. Pick folders, channels and date ranges instead of a full archive.
- Run a detection tool. Let software flag names, numbers, addresses and emails across the export. The tools section below compares options.
- Replace, don’t just delete. Swap each item for a placeholder like [CUSTOMER] or [PHONE], so the text still reads naturally. Don’t keep a key that maps placeholders back to real names, or the data is only pseudonymized.
- Strip file metadata. Remove document authors, edit history, comments and photo GPS tags.
- Check samples by hand. Read a random sample from each source. Tools miss nicknames, misspellings and details like “the CFO’s wife”.
- Write down what you did. Record which systems you excluded, which tool you used and what the sample check found. California’s CCPA expects “reasonable measures”, and a record shows you took them.
What tools can find and remove personal information?
| Tool | What it does | Limits |
|---|---|---|
| Presidio | Open-source detection and anonymization for text, images and structured data. It started at Microsoft and is now community-run | Says itself that automated detection can’t guarantee it finds everything |
| Amazon Comprehend | Finds personal information such as names, addresses, SSNs and card numbers, and returns redacted copies through batch jobs | English and Spanish text only |
| Google Cloud Sensitive Data Protection | Finds sensitive data and de-identifies it by masking, redaction, bucketing, date shifting or tokenization, including in images | Needs a Google Cloud account and some setup |
| ExifTool | Reads and deletes metadata in photos, documents and other files | Its own site says it isn’t guaranteed to remove metadata completely, and it doesn’t remove a PDF’s original metadata |
| Microsoft Office Document Inspector | Removes comments, tracked changes, authors and hidden content from Word, Excel and PowerPoint files | Can’t remove embedded objects, macros or Excel’s external links |
None of these replace the human check in step 6. They catch the patterns, like phone numbers and email addresses, and miss the context.
How do you remove metadata from files?
Metadata is the information stored inside a file rather than on the page. A cleaned contract can still name its author, and a site photo can still carry the GPS location and the phone that took it.
- Office files: Document Inspector removes authors, comments, tracked changes, hidden rows and sheets, and presentation notes. Microsoft recommends running it on a copy.
- Photos: ExifTool removes camera, date and GPS data.
- PDFs: Re-export them from the cleaned source file, or use a PDF editor’s sanitize function, rather than relying on ExifTool.
Is anonymized data really anonymous?
The 2019 study is why combinations matter more than single fields. “The only female site supervisor in our Boise office in 2021” identifies someone without a single name in it.
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 hashing identifiers doesn’t make data anonymous. Swapping names for scrambled codes isn’t enough.
The practical fix is to leave the highest-risk records out entirely, rather than trying to clean them.
Which anonymization standard counts legally?
| Law | Standard for de-identified data | What it means for you |
|---|---|---|
| HIPAA | Safe Harbor (remove 18 identifiers) or Expert Determination (a qualified expert certifies the risk is very small) | Applies to medical practices and their business associates |
| CCPA (California) | Reasonable measures, a public commitment not to re-identify, and a contract requiring recipients to comply | Your license should include a no-re-identification clause that binds the buyer |
| GDPR (EU) | Only anonymous data is outside the law. Pseudonymized data is still personal data | Records about people in the EU need a lawful basis unless fully anonymized |
California’s wording is in Civil Code 1798.140(m), and the GDPR test is in Recital 26. The AI data licensing agreement guide covers the contract clauses that carry these duties to the buyer.
Law firms, accountants and medical practices have extra confidentiality rules on top. Our guide to selling law firm data to AI covers ABA Model Rule 1.6.
Common mistakes when anonymizing data for AI
- Cleaning a whole email or Slack archive when you could leave most of it out
- Swapping names for codes and keeping the key, which leaves the data pseudonymized, not anonymous
- Trusting a detection tool without reading samples
- Forgetting metadata, photo GPS tags and Excel’s hidden sheets
- Leaving combinations of job title, place and date that point to one person
- Assuming a broker or marketplace removes personal information for you
- Not writing down what you did, which makes “reasonable measures” hard to prove
Frequently asked questions
How do you anonymize data for AI training?
List what you’ll send, cut the systems that don’t need to go, run a detection tool to find names, numbers and addresses, replace or remove them, strip file metadata, then have a person check samples. Most AI data buyers do some or all of this for you.
What’s the difference between anonymized and pseudonymized data?
Anonymized data can’t be linked back to a person by anyone. Pseudonymized data swaps names for codes but can be linked back with a key. GDPR still treats pseudonymized data as personal data.
Do I have to anonymize my data before selling it to an AI company?
Not always. micro1 and Mercor remove personal information themselves, and Scale AI and Endless Data do it inside your systems. AIData.Broker makes you do it before you send anything.
What personal information has to be removed?
Names, addresses, phone numbers, emails, ID and account numbers, dates tied to a person, faces, voices and file metadata. HIPAA’s Safe Harbor method lists 18 identifiers, and it’s a useful checklist even outside health care.
What tools remove personal information from documents?
Microsoft’s open-source Presidio, Amazon Comprehend and Google Cloud Sensitive Data Protection find and redact personal information in text. ExifTool and Microsoft Office’s Document Inspector remove file metadata.
Is anonymized data really anonymous?
Not always. A 2019 Nature Communications study found 99.98% of Americans could be re-identified from 15 demographic attributes. The FTC says hashed identifiers can still identify people.
