De-identifying company email and support tickets means removing or replacing everything that could link a record to a person, then checking that what remains cannot reasonably be traced back. Under the CCPA, data counts as de-identified only when it cannot reasonably be linked to a consumer and the holder also takes reasonable measures, publicly commits not to re-identify and binds every recipient by contract (Cal. Civ. Code 1798.140(m), read Nov 3, 2026). In most deals the AI lab or data buyer runs the de-identification. Your part is deciding what goes in, reading a de-identified sample and getting the standard written into the license.
If you want to know what your records are worth before you do any of this, get a confidential estimate.
A note on words: owners usually say they are selling data. In these deals the company licenses a copy for a set use and keeps the original, which is why this guide says “license.” Our guide to what a data licensing agreement contains covers the contract.
This is general information, not legal advice. Talk to your own counsel before you decide what to license.
What counts as de-identified under US law?
US law has no single federal test for de-identified data. Three sources matter for most US companies, as read on Nov 3, 2026.
| Source | What it requires |
|---|---|
| CCPA, Cal. Civ. Code 1798.140(m) | Data cannot reasonably be linked to a consumer or household, plus reasonable measures, a public commitment not to re-identify, and contract terms binding every recipient |
| HIPAA, 45 CFR 164.514(b) | Either an expert determines the re-identification risk is “very small,” or 18 listed identifiers are removed and the holder has no actual knowledge the rest could identify someone |
| FTC staff, July 24, 2024 | Hashing personal information does not make it anonymous |
The CCPA test covers California employees and business contacts as well as customers (FKKS, Oct 1, 2026). Virginia, Colorado and Utah used the same three-part approach in Greenberg Traurig’s comparison of state standards (May 17, 2022).
The HIPAA list is a useful checklist even outside healthcare. It includes names, geography smaller than a state, dates tied to a person except the year, phone and fax numbers, email addresses, account numbers, web URLs, IP addresses, device identifiers and full-face photos. If you are a business associate, your agreements set the limits first; see our page for healthcare administration.
Pseudonymized data is a separate category. The CCPA defines it as data that cannot be tied to a person without additional information kept separately (1798.140(aa), read Nov 3, 2026). Replacing “Dana Ortiz” with “Person 14” is pseudonymization while someone still holds the table that maps one to the other.
Where does personal data hide in support tickets and email?
Personal details in email and tickets sit well beyond the From line. A careful review checks each of these places.
| Where | What it holds |
|---|---|
| Headers and ticket fields | Sender, recipients, CC lists, requester name and email, assignee |
| Signatures | Full name, title, direct phone, office address, sometimes a photo |
| Quoted replies and forwards | The same identifiers again, further down every thread |
| Free text | Customers typing order numbers, account numbers, addresses or health details into a ticket |
| Attachments and screenshots | Invoices, IDs, contracts, photos of people, logos and handwritten signatures |
| Metadata and links | IP addresses, device IDs, URLs carrying account tokens |
| Rare combinations | A job title plus an office plus a date that only fits one person |
The last row is the hardest. FKKS notes that buyers want records to stay linked to each other, and that the same linking is “what makes it possible to draw conclusions about individuals or small groups” (Oct 1, 2026). Removing names does not help much if the thread still says the only payroll manager in a five-person office was disciplined in March.
What should come out entirely?
Some records stay sensitive after every name is gone. FKKS advises excluding high-risk data entirely rather than relying on de-identification, and points out that “a disciplinary thread or pay complaint remains sensitive without a name attached” (Oct 1, 2026). In practice that means HR, payroll, health and legal folders, plus anything a customer contract or NDA covers.
The Spirit Airlines sale shows the same approach at scale. Google’s $10M bid covered about 100 million emails (see our market index). The court-appointed privacy ombudsman’s Oct 5, 2026 report backed the sale only after booking and transaction records, refunds, inflight purchases and WiFi sales were cut out, while corporate email stayed in (The Next Web, Oct 9, 2026). We found no court ruling reported as of Nov 3, 2026; the approval hearing was set for Nov 6 (TravelPulse, Oct 10, 2026).
Who does the de-identification?
The AI data buyer usually de-identifies the records, and programs differ mainly in where that happens: inside your systems, or after your raw records have left. These are the programs’ own descriptions, as of the dates shown.
| Program | Who de-identifies, and where | As of |
|---|---|---|
| Scale AI | Scale, “inside your environment, before any data is extracted” (Scale) | Nov 3, 2026 |
| Endless | Endless, before anything leaves the company (Endless) | Oct 19, 2026 |
| Turing | Turing’s team, from a copy in an isolated environment (Turing) | Oct 16, 2026 |
| Mercor | Mercor, fully automated; its engineering team “does not manually review your data” (Mercor) | Nov 3, 2026 |
| Ooak Data | Ooak’s own pipeline; raw archive deleted within 30 days of delivery (Ooak) | Nov 3, 2026 |
| Collinear | Collinear, after export; names become consistent pseudonyms (Collinear) | Nov 3, 2026 |
| Nyne AI | Nyne, after you “transfer the data as it exists today” (Nyne) | Nov 3, 2026 |
| Polyshares | An unnamed specialist PII partner, before Polyshares or any lab sees a record (Polyshares) | Oct 16, 2026 |
| Handshake AI | Not published (our buyer directory) | Nov 2, 2026 |
If raw email must not leave your systems, that narrows the list before price comes into it. Our comparison of AI data partnership programs sets out the other terms.
Should you scrub the data yourself first?
Most companies do not need to scrub records themselves before licensing them. Collinear’s page, read Nov 3, 2026, says a company does not need to clean anything beyond removing what it cannot share. micro1 says the company does not have to anonymize data first (micro1, observed Oct 26, 2026).
A blunt in-house pass that swaps every name for “[NAME]” breaks the links between the email, the ticket and the follow-up, and those links are what buyers pay for (see what AI labs want from company records). Ooak says it applies “one entity dictionary” across every connected source, so the same person gets the same replacement everywhere (Nov 3, 2026).
If you do run your own pass, treat automated tools as a first filter. The open-source tool Presidio says “there is no guarantee that Presidio will find all sensitive information” (Presidio, read Nov 3, 2026). NIST makes the same point more broadly: “not all tools that merely mask personal information provide sufficient functionality for performing de-identification” (NIST SP 800-188, Sep 2023).
What you should remove yourself is everything you decided to exclude: privileged files, HR and health folders, and records owned by customers or partners.
How do you check the buyer’s work?
A seller can check a buyer’s de-identification with a sample drawn from its own records and a few hours of reading.
- Ask for the written protocol and the standard it aims at, such as CCPA 1798.140(m) or a HIPAA method. FKKS suggests approving the protocol and requiring certification “to the company’s satisfaction” (Oct 1, 2026).
- Get a de-identified sample before you sign, and pick hard cases for it: a long thread with forwards, a ticket with attachments, and an email heavy with signatures.
- Search the sample for your own names: your top customers, your staff, your email domain, your office phone numbers and your account number formats.
- Check that the same person carries the same pseudonym across email and tickets.
- Ask who holds the pseudonym key, where raw data sits during processing and when it is deleted.
- Limit who in your company and at the buyer can touch raw data before de-identification, as FKKS recommends.
- Put the standard, a ban on re-identification, the same terms for every onward recipient and deletion of raw data into the license.
The Spirit ombudsman asked for a version of step 7. Tonic.ai must file a certification with the court before each transfer, and the buyer must publicly commit not to re-identify the data (The Next Web, Oct 9, 2026).
Does de-identification settle the legal questions?
De-identification answers whether a record can be linked to a person. Confidentiality owed to customers and partners, privilege, trade secrets and vendor terms are separate questions, as FKKS sets out (Oct 1, 2026). Our guide on whether it is legal to sell company data to AI covers those, and your own counsel should make the call.
Marin Data checks these questions with you before any buyer sees a sample, and we push for the buyer’s protocol and a sample in writing. It costs the company nothing up front, and nothing at all if no deal closes; the AI lab pays us.
Sources
- California Civil Code 1798.140, definitions (m) and (aa), read Nov 3, 2026
- 45 CFR 164.514, de-identification of protected health information, read Nov 3, 2026
- FTC Office of Technology, No, hashing still doesn’t make your data anonymous, Jul 24, 2024, read Nov 3, 2026
- NIST SP 800-188, De-Identifying Government Datasets, Sep 2023, read Nov 3, 2026
- Frankfurt Kurnit Klein & Selz, selling company data for AI training, Oct 1, 2026, read Nov 3, 2026
- Greenberg Traurig via National Law Review, differences in state deidentification standards, May 17, 2022, read Nov 3, 2026
- The Next Web, Spirit’s privacy ombudsman backs Google’s data deal, Oct 9, 2026, read Nov 3, 2026
- TravelPulse, Spirit data sale delayed as lawmakers raise concerns, Oct 10, 2026
- Presidio, read Nov 3, 2026
- Scale AI, Mercor, Ooak Data, Collinear, Nyne AI, read Nov 3, 2026
- Turing Project Lazarus and Polyshares, observed Oct 16, 2026; Endless, observed Oct 19, 2026; micro1, observed Oct 26, 2026 (per our buyer directory)