How to anonymize company data before licensing it for AI training.
A working guide for the IT and legal staff who do the job: which identifiers to remove, what the law counts as anonymous, the techniques and tools for each system, the re-identification risk and the sample review. Your team does all of it inside your own systems. Allied never sees the data. To learn what the records could fetch first, request an offer.
The short answer
You remove or replace everything that points to a person, in every system you plan to license, and you prove it on a sample before anything leaves. Names and email addresses are the easy part. Phone numbers in signatures, account numbers in ticket bodies, faces in screenshots and voices in call recordings are where most misses happen.
Labs want this data in bulk. In August 2026, in Spirit Airlines’ bankruptcy, Google outbid Mercor for the carrier’s data, including 100 million emails, 500 million Teams messages and 176,000 employee records, for $10 million. The same report says suppliers employ data scientists to scrub personally identifiable information and exclude HR records, and that attorney Greg Ewing has called such data “legally radioactive.”1
The work happens on your side. Your admins export the records, your team removes identifiers in your own environment, your reviewers approve a sample, and the copy goes from you to the lab. Allied never sees, stores or moves it. If you have not yet settled whether the license itself is allowed, start with is it legal to sell company data and employee consent for AI training data.
Anonymous, pseudonymous, deidentified: what each law means
The words are not interchangeable, and the license will use one of them. NIST notes that some publications use de-identification and anonymization interchangeably, while others reserve anonymization for de-identification that cannot be reversed. Pick the legal test that applies to your records and write the license to it.6
GDPR: anonymous vs pseudonymous
Recital 26 says data protection principles do not apply to anonymous information, meaning data rendered anonymous so the person is “not or no longer identifiable.” Data that has undergone pseudonymisation and could be attributed to a person with additional information is still information on an identifiable person. Identifiability is judged by “all the means reasonably likely to be used,” including singling out.2 Article 4(5) defines pseudonymisation as data that cannot be attributed to a person without additional information that is kept separately.3 If your team keeps a lookup table from placeholder to real name, the copy is pseudonymous and the GDPR still applies.
California: “deidentified” under Cal. Civ. Code § 1798.140(m)
Information that cannot reasonably be used to infer information about, or be linked to, a particular consumer, provided the business does three things: takes reasonable measures so it cannot be associated with a consumer or household, publicly commits to keep it deidentified and not to try to reidentify it, and contractually obligates every recipient to comply. The third requirement belongs in the license with the lab.4
HIPAA: two methods, 45 CFR 164.514(b)
Expert determination: a qualified expert finds the risk is “very small” that an anticipated recipient could identify a person, and documents the methods. Safe Harbor: remove 18 listed identifiers of the individual and of relatives, employers or household members, and have no actual knowledge that what remains could identify someone. Most companies are not covered entities, but the Safe Harbor list is the most concrete checklist in US law.5
NIST: the federal playbook
NIST IR 8053 (October 2015) surveys de-identification and re-identification techniques for structured data, free text, multimedia and images.6 NIST SP 800-188 (September 2023) builds on it with specific guidance for agencies that de-identify data before sharing it.7
Direct identifiers and quasi-identifiers
NIST splits identifiers into two kinds. Direct identifiers identify one person on their own: names, Social Security numbers, email addresses. Quasi-identifiers, also called indirect identifiers, do not identify anyone alone but can be combined and linked with other information to do so. Its example: birthday, ZIP code and sex.6
Work records are full of quasi-identifiers. A job title, an office, a team name and a start date can point to one person even after every name is gone. Treat them as identifiers when they are rare.
The Safe Harbor list in 45 CFR 164.514(b)(2) is a useful checklist even outside health care. Here is each identifier and where it tends to sit in company systems. ZIP codes may keep their first three digits only where that three-digit area holds more than 20,000 people, and ages over 89 must be grouped into 90 or older.5
| Safe Harbor identifier (45 CFR 164.514(b)(2)) | Where it hides in work records |
|---|---|
| Names | Slack display names, @mentions, email From/To lines, greetings, signatures, CRM contacts |
| Geographic units smaller than a state | Customer and shipping addresses in CRM and tickets, office locations in signatures |
| Dates tied to a person (except year) | Birthdays, hire and leave dates in HR threads, appointment dates in calendars |
| Telephone numbers | Email signatures, ticket bodies, CRM fields, call metadata |
| Fax numbers | Older signatures and vendor records |
| Email addresses | Headers, CC lists, forwarded threads, CRM, ticket requester fields |
| Social Security numbers | Payroll, benefits and onboarding threads |
| Medical record numbers | Leave and accommodation threads, health-sector customer data |
| Health plan beneficiary numbers | Benefits enrollment emails |
| Account numbers | Invoices, billing tickets, CRM account fields |
| Certificate and license numbers | Driver’s license copies, professional license records |
| Vehicle identifiers and license plates | Fleet, insurance and parking tickets |
| Device identifiers and serial numbers | IT help-desk tickets, asset inventories |
| URLs | Personal LinkedIn or social links in signatures and CRM |
| IP addresses | Email headers, login logs, security tickets |
| Biometric identifiers, including voice prints | Call recordings, voicemail, meeting recordings |
| Full face photos and comparable images | Profile pictures, screenshots, video-call recordings |
| Any other unique number, characteristic or code | Employee IDs, Slack user IDs, customer IDs, ticket requester IDs |
Techniques, system by system
Start by cutting whole sources, then clean what is left. NIST lists the methods for quasi-identifiers: suppression (remove the value), generalization (ZIP code 12345 becomes a range of 12000 to 12999), perturbation (dates moved by the same amount for each person, its example a shift of up to 1,000 days), swapping and sub-sampling. Suppression gives the most privacy and the least utility.6
Exclude before you clean
Leave out HR, payroll, legal, board and benefits channels, direct messages unless you opt in, and any customer data a contract bars from reuse. Removing a source is safer than scrubbing it, and it shrinks the review.
Redaction
Replace each detected value with a type label such as [PHONE] or [ACCOUNT]. Google Cloud Sensitive Data Protection describes masking characters with a symbol, replacing each value with a token or surrogate string, and encrypting values with a key.10
Consistent replacement
Map each real person to one placeholder, such as EMPLOYEE_0412, used in Slack, email and CRM alike, so the model still sees who replied to whom. Labs value those relationships. Destroy the mapping when the run is done. A kept key makes the copy pseudonymous under the GDPR, and a key derived from the person’s own data fails HIPAA’s re-identification code rule.35
Generalization
Turn exact values into ranges: a city into a region, a date of birth into a year, a deal size into a band. Apply it to rare titles too, so “VP of Treasury, Boise” becomes “finance leader.”6
Free-text PII
Signatures, phone numbers typed into ticket bodies, home addresses in shipping disputes, pasted passwords and API keys. Run detection over message bodies, not only structured fields, and strip signature blocks with a rule before detection runs.
Attachments and file metadata
Office files carry comments, tracked changes, document properties, email headers and hidden text. Microsoft’s Document Inspector finds and removes these, recommends working on a copy, and cannot detect text hidden by other methods such as white text on a white background.13 Screenshots need image redaction: Google’s service covers detected text in an image with opaque rectangles.11
Call recordings and voice
A voice print is a Safe Harbor identifier.5 Amazon Transcribe can redact PII from transcripts, and warns that if only a redacted transcript is generated, the original media is the only place the complete conversation is stored.12 NIST notes that audio distortion only prevents identification by the speaker’s voice and cannot remove other cues in the audio track.6 The lower-risk choice is to license redacted transcripts and keep the audio.
Code repositories
Commit history carries author names and email addresses, and code can hold credentials and customer data in test fixtures. Rewrite author fields to placeholders and run a secret scanner before the copy is made.
Tools your team can run
No tool finds everything. Presidio’s own documentation warns that because it uses automated detection, “there is no guarantee that Presidio will find all sensitive information,” and that other protections should be used alongside it.8
| Tool | What it does | Where it falls short |
|---|---|---|
| Presidio (open source) | Detects and anonymizes PII in text, images and structured data; started at Microsoft and is now community-governed89 | No guarantee of full detection8 |
| Google Cloud Sensitive Data Protection | Masking, surrogate tokens, encryption, date shifting, bucketing and pseudonymization; redacts images1011 | Runs in Google Cloud; you choose the infoTypes to scan for10 |
| Amazon Transcribe PII redaction | Redacts PII in call and meeting transcripts12 | Redacts the transcript, not the voice |
| Microsoft Document Inspector | Removes comments, revisions, properties and hidden text from Office files13 | Misses text hidden by other methods13 |
Run the tools inside your own environment, on a copy, and log which version and settings ran on which export. That log is your evidence for the “reasonable measures” in California’s definition.4
Re-identification risk is real
Latanya Sweeney found that 87% of the US population (216 million of 248 million) likely had a unique combination of 5-digit ZIP code, gender and date of birth.14
Rocher, Hendrickx and de Montjoye, writing in Nature Communications in 2019, estimated that 99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes, and concluded that even heavily sampled anonymized datasets are unlikely to meet GDPR standards.15
Work records leak the same way. MIT Technology Review’s sources point to private details in emails and calendars: doctor appointments, school pickup arrangements, date-night plans. A detector sees none of those as PII, and each can identify a person to a colleague.1
The FTC’s view, from its July 24, 2024 post: “data is only anonymous when it can never be associated back to a person.” Companies should not act or claim as if hashing personal information renders it anonymized, and FTC staff will “take action when the privacy claims they make are deceptive.” So describe the copy in the license as what it is, deidentified to a named method, and do not call hashed IDs anonymous.16
Sample review and sign-off
Draw a random sample from every system
Include the hard cases on purpose: long email threads, tickets with attachments, call transcripts, channels with external guests.
Two reviewers read it
One from IT, one from legal or HR. They look for names, numbers and quasi-identifiers the tools missed, and for sensitive content such as health, pay and discipline.
Fix the rule, not the record
Each miss becomes a new detection rule or exclusion. Re-run the whole export and draw a fresh sample.
Approve in writing
Sign off the sample before delivery and keep the record: sources excluded, tools and settings, misses found and fixed. HIPAA’s expert method requires documented methods and results, and the same file supports a CCPA reasonable-measures position.45
Bind the buyer
Put a no-reidentification clause and onward-transfer limits in the license, as California’s third requirement asks.4 See AI data licensing agreement terms.
Where Allied fits
Allied arranges a non-exclusive license between your company and an AI lab. Allied never sees, reads or stores the records. Your team exports them, removes identifiers in your own systems and approves the sample; the copy goes straight from you to the lab. Allied sees only a description of the data that you approve. See data handling.
Old operational records from companies with 30 or more office employees could be worth $100K–$2M. Allied aims to have a written offer within two weeks of the company signing with Allied, and payment arrives 30 to 60 days after the lab accepts delivery. See the systems labs license, including Slack and Teams, email and call recordings, read how Allied works, or request an offer.
Questions
Is removing names enough to anonymize company data?
No. Quasi-identifiers such as job title, location and dates can single out a person, and free text holds phone numbers, account numbers and personal details that name-removal misses.
Is pseudonymized data anonymous under the GDPR?
No. If anyone holds a key that links a placeholder back to a person, the data is pseudonymous and still personal data. Destroy the mapping if the goal is anonymous data.
What does California require for deidentified data?
Reasonable measures so the data cannot be linked to a consumer or household, a public commitment not to reidentify it, and a contract that binds every recipient to the same rules.
Can we license call recordings with the voices left in?
A voice is a biometric identifier, so the safer route is redacted transcripts. If a buyer needs audio, treat it as identifiable data and get legal advice first.
Who removes the identifiers, our team or Allied?
Your team. Allied never sees the data. Your own staff remove identifiers in your systems and approve a sample, and the copy goes from you to the lab.
Which tool should we use?
Any of Presidio, Google Cloud Sensitive Data Protection or Amazon Transcribe for call transcripts can do the detection. None finds everything, so the human sample review decides when the copy is ready.
Is this legal advice?
No. It is general information on common methods. Have counsel confirm which standard applies to your records.
Sources
- MIT Technology Review, “AI labs are buying data from dead startups to train their models,” October 13, 2026
- GDPR Recital 26, not applicable to anonymous data (gdpr-info.eu)
- GDPR Article 4, definitions, including (5) pseudonymisation (gdpr-info.eu)
- California Consumer Privacy Act, Cal. Civ. Code § 1798.140(m), “Deidentified” (California Privacy Protection Agency)
- 45 CFR § 164.514, de-identification of protected health information (Cornell LII)
- Simson Garfinkel, NIST IR 8053, De-Identification of Personal Information, October 2015
- NIST SP 800-188, De-Identifying Government Datasets: Techniques and Governance, September 2023
- Presidio documentation, home
- Presidio documentation, project transition update
- Google Cloud Sensitive Data Protection, De-identifying sensitive data
- Google Cloud Sensitive Data Protection, Image redaction
- Amazon Transcribe Developer Guide, Redacting or identifying personally identifiable information
- Microsoft Support, Remove hidden data and personal information by inspecting documents, presentations, or workbooks
- Latanya Sweeney, Simple Demographics Often Identify People Uniquely (Data Privacy Lab)
- Rocher, Hendrickx and de Montjoye, “Estimating the success of re-identifications in incomplete datasets using generative models,” Nature Communications, 2019
- Federal Trade Commission, Technology Blog, “No, hashing still doesn’t make your data anonymous,” July 24, 2024
General information, not legal advice. Which standard applies depends on where the people in your records live and what your contracts say; have counsel review your method before any copy leaves your systems.
Find out what your records are worth.
A few questions. About two minutes. How Allied works.