How de-identification works for company data licensed to AI

De-identification replaces the details in your records that point to a person or a company, such as names, emails, phone numbers, addresses and account numbers, with placeholders like [Name], while keeping the work readable. In company data licensed to AI, software usually checks every record, people review samples, and the owner checks a sample before signing. It greatly lowers the risk of identifying someone but does not remove it, especially in free-text email and chat.

Data Licensors, the company that publishes this page, licenses company records and runs this process itself, so we describe our own steps below alongside other buyers'.

What gets removed from company records?

This matches what buyers publish. AfterQuery says names, contact details and other identifiers are removed or replaced with consistent pseudonyms and that payroll, privileged legal material and credentials are never used. Mercor says its automated anonymization masks 60+ categories of sensitive identifiers. Troveo shows an example where a customer becomes PERSON-014 and an email address becomes EMAIL-019 (buyer pages, October 2026).

What does de-identified data look like? A before and after example

Here is an invented but typical email from a roofing company, before and after.

Before:

Hi Dana, Mike approved the Henderson re-roof at 14 Oak St. Crew of 3, two days, $8,400. Invoice INV-20931 goes to dana.h@example.com. Call me at 555-201-7788 if the permit slips. Tom, Brightline Roofing

After:

Hi [Name 1], [Name 2] approved the [Customer 1] re-roof at [Address]. Crew of 3, two days, $8,400. Invoice [Invoice] goes to [Email]. Call me at [Phone] if the permit slips. [Name 3], [Company]

What stays: the job type, the crew size, the duration, the price, the approval step and the permit risk. That is what an AI model learns from, how an approved quote becomes a scheduled job. What leaves: who the people are, which customer it was, where the house is.

Note the numbers in brackets. If [Name 1] is the same person in every email and ticket, the record keeps its structure. This is called referential integrity, and it is what makes the data useful, and also what makes it imperfect (see below).

How is it done in practice?

1. Scope first. You choose systems, date ranges, channels and mailboxes. Anything left out is never processed. Excluding a whole category, such as HR, is safer than trying to scrub it.
2. Automated detection. Software scans every record for names, emails, phone numbers, addresses, ID patterns and secrets, then swaps each for a placeholder. micro1 says it uses its own "flow-transform" anonymization model and that some datasets get synthetic rewrites that keep the workflow but loosen ties to the original records (micro1, October 2026).
3. Consistent pseudonyms. The same person or customer gets the same code everywhere, so a thread can still be followed across email, tickets and code.
4. Human review. People check samples from each batch for anything the tools missed, such as a nickname or a customer named in a file path. We review samples from every batch and fix and re-check any batch that fails (Data Licensors, October 2026).
5. Re-identification testing. Some buyers try to re-identify people in the result before delivery. AfterQuery says every dataset gets a re-identification test "because no process removes all risk on its own" (AfterQuery, October 2026).
6. Owner review. You see a de-identified sample before signing and can ask for any person, customer, project or topic to be removed.

Where it happens varies. Scale AI says it anonymizes inside your environment before any data is extracted. Polyshares says a specialist partner strips identifying fields before Polyshares or any lab sees the record. micro1 says you do not need to anonymize before transfer because it does it. We replace identifying details before any record leaves our hands, and delete the identifiable export within 30 days of delivering the de-identified version (buyer pages and Data Licensors, October 2026).

How was it handled in the biggest public case?

The Spirit Airlines archive is the most detailed public example. Google won the August 2026 bankruptcy auction with a $10 million bid. The sale agreement makes the buyer choose and pay for the de-identification agent, requires the work to be certified against the California Consumer Privacy Act standard, and requires referential integrity across the dataset. Mercor, the backup bidder at $7.5 million, had offered $7 million if it could take the raw data and anonymize it itself (The Next Web, August 2026).

Spirit and Google later excluded passenger databases and hired a third party, Tonic.ai, to de-identify material that could contain consumers' personal information, such as company emails. A court-appointed consumer privacy ombudsman then recommended approval, though her review did not cover employee privacy (Business Insider, October 5, 2026). The sale is not final: micro1 filed a $12.5 million competing offer in September (Bloomberg, September 2026), and the court hearing was moved to November 6, 2026 (TravelPulse, October 2026).

Lawmakers were not fully reassured. More than 120 members of Congress wrote that "conventional de-identification safeguards may not be sufficient to protect employee privacy in the context of modern artificial intelligence," citing a NIST report on AI-assisted re-identification (TravelPulse, October 2026).

What are the limits and what risk remains?

• Free text hides clues. An attorney told MIT Technology Review (October 2026) that emails and calendars hold details like doctor appointments, school pickups and date-night plans. A tool can remove a name but miss "my daughter's recital at Lincoln Elementary."
• Combinations identify people. A 2019 study estimated that 99.98% of Americans would be correctly re-identified in any dataset with 15 demographic attributes (Nature Communications, July 2019). Job title, city and start date together can point to one person in a small company.
• Consistency cuts both ways. Keeping [Name 1] the same everywhere makes a richer profile of that person. The Next Web called this the property that makes any anonymisation fragile (August 2026).
• The law may still treat it as personal data. Under the GDPR, pseudonymised data remains personal data, while truly anonymous data falls outside it (GDPR Recital 26). The EDPB's July 2026 draft guidelines add that data can be anonymous for one party and personal for another (DLA Piper, August 2026). This is general information, not legal advice.
• Small companies are easier to re-identify. With 40 staff, "the operations manager" is one person. Removing role titles or merging small teams can help.

The practical answer is layers: leave out sensitive categories, replace identifiers, review samples, and bind every recipient by contract not to attempt re-identification. SimpleClosure, for example, says sensitive employee, compensation and health information is excluded outright (SimpleClosure, September 2026).

What should I ask a buyer about de-identification?

• Which categories do you remove, and do you keep consistent pseudonyms?
• Is it automated only, or do people review samples? How large are the samples?
• Do you test for re-identification before delivery?
• Where does it happen: in my systems, on your servers, or at a third party?
• Can I review a sample and remove more before signing?
• What happens if something identifying is found after delivery?
• Does every recipient sign a no-re-identification promise?

Put the answers in the contract. Our licensing agreement guide shows where they go, and the legal guide covers what to leave out.