# Data anonymization before a company data license

> Searching for anonymization before a company data sale? Licensing needs it too. What the legal tests ask, why email and chat are hard, and who does the work.

By Matt Burns, Vaultfish. Canonical: https://vaultfish.com/guides/data-anonymization  
Published: 2026-07-24  
Updated: 2026-10-14

Data anonymization means changing records so no one can be picked out from them. Three facts shape the job when a company licenses an email, chat or ticket archive to an AI buyer.

1. The legal tests differ by place.
2. Free text is much harder to scrub than a database.
3. In the deals I read, the buyer or its chosen vendor does the work.

So the company's job is to ask who does it, to what standard and with what certification.

Most search results for this phrase come from software vendors. This guide is the data owner's view. It is research, not legal advice.

## What does data anonymization mean when you license an archive?

Three words get used loosely.

- **Anonymization** means a person can no longer be identified from the data.
- **Pseudonymization** swaps identifiers for labels and keeps a way back.
- **De-identification** is the American term. Its meaning depends on the statute.

Our page on [pseudonymization vs anonymization](/answers/pseudonymization-vs-anonymization) covers the first two. [What de-identification means in a data license](/answers/data-de-identification) covers the third.

> **Anonymizing records means removing what points to a person.** Think of the library. Anonymizing pulls the borrower's name card out of every book for good. Swapping names for labels keeps the card in a locked drawer.

This matters because AI buyers often want linked records. A name swapped for the same stand-in in Slack and in a spreadsheet keeps the trail from request to deliverable. That trail is pseudonymization.

_Swapping "Dana" for "Person 4" feels like privacy. Ask who holds the key._

## Which legal tests apply?

| Test | Source | The question it asks |
|---|---|---|
| Anonymous data | GDPR Recital 26 | Is the person no longer identifiable, counting all means reasonably likely to be used? |
| Deidentified | California Civ. Code 1798.140(m) | Can it be linked to a consumer, and are measures, a public commitment and recipient contracts in place? |
| Safe harbor | 45 CFR 164.514(b), health data | Are 18 listed identifier types removed, with no actual knowledge of re-identification? |
| Hashed data | FTC | Hashes can still identify users, so hashing alone is not anonymization |

GDPR Recital 26 asks what means are reasonably likely to be used. It counts cost, time and available technology ([Recital 26](https://gdpr-info.eu/recitals/no-26/)). The ICO says pseudonymised data is personal data in the hands of anyone holding the additional information. It marks its guidance as under review ([ICO](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/pseudonymisation/)).

The EDPB's draft Guidelines 02/2026 on anonymisation are open for comment until 30 Oct 2026. They address singling out, linkability and inference ([EDPB](https://www.edpb.europa.eu/public-consultations/guidelines-022026-on-anonymisation_en)).

California's test has three parts: reasonable measures, a public commitment not to re-identify, and contracts that bind every recipient ([statute text](https://cppa.ca.gov/regulations/pdf/cppa_act.pdf)). Without them, the statute treats a transfer of personal information for money as a sale. Our [CCPA page](/answers/is-selling-data-to-an-ai-lab-a-sale-under-ccpa) explains.

HIPAA's safe harbor lists 18 identifier types. It bars reliance on it where the entity has actual knowledge that the data could still identify someone ([45 CFR 164.514](https://www.law.cornell.edu/cfr/text/45/164.514)).

## Why is an email and chat archive harder than a database?

Names sit inside sentences, signatures and attachments. Context identifies people even without a name.

The law firm Frankfurt Kurnit Klein and Selz wrote that removing identifying details from millions of emails and chats is much harder than removing names from structured fields. It also wrote that removing names says whether a record links to a person. It says nothing about whether the content is confidential ([FKKS](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes)).

Research points the same way. Staab and co-authors reported that language models inferred personal attributes from text with up to 85% top-1 accuracy. They found that text anonymization was ineffective against that inference ([arXiv 2310.07298](https://arxiv.org/abs/2310.07298)).

Lermen and co-authors reported up to 68% recall at 90% precision when matching pseudonymous accounts across platforms ([arXiv 2602.16800](https://arxiv.org/abs/2602.16800)). Both are research papers on arXiv. Neither tests a company archive. They show what a determined model can do.

The Spirit Airlines consumer privacy ombudsman wrote in a footnote that technology has "far outpaced the law" on re-identification ([Doc 1684](https://storage.courtlistener.com/recap/gov.uscourts.nysb.331039/gov.uscourts.nysb.331039.1684.0.pdf)). Our [re-identification page](/answers/can-ai-re-identify-anonymized-emails-and-slack-messages) goes further.

## Who does the data anonymization in a deal?

In the deals I read, the receiving company, the marketplace or a vendor they accept.

**[micro1](/buyers/micro1).** Its CFO said in an undated recorded forum talk that every corpus goes through a scrubbing step for personal identifiable information, sampled at several points.

micro1's David Revland, who works on human data, said it turns names into consistent stand-ins. micro1 staff said companies cannot pre-scrub, because blanking names out would destroy the cross-platform trail.

**Spirit Airlines.** The sale agreement has the buyer accept or designate the de-identification agent ([Doc 1489](https://storage.courtlistener.com/recap/gov.uscourts.nysb.331039/gov.uscourts.nysb.331039.1489.0.pdf)). The ombudsman says Tonic.ai will strip identifying details from unstructured data and file a certification before each transfer. It also says micro1 proposed to do that work itself ([Doc 1684](https://storage.courtlistener.com/recap/gov.uscourts.nysb.331039/gov.uscourts.nysb.331039.1684.0.pdf)).

**[Handshake AI](/buyers/handshake-ai).** Its page says it removes direct and indirect identifiers before delivery. It also says a third party may still be able to re-identify personal information in combination with other data ([Handshake AI](https://joinhandshake.com/ai/data-partnerships/)).

## Should you scrub the data yourself first?

Ask before you touch anything. micro1 staff said in an undated recorded forum talk that a company's own scrub gets in its way.

FKKS advises you to stay in control. That means selecting or jointly selecting the provider, approving the protocol and requiring certification to your satisfaction.

Both can be true in one deal. The company chooses and approves the process, and the receiving company's side runs it. Settle which of those a buyer accepts before any export.

A scrub also leaves a second problem untouched. FKKS recommends leaving out privileged material, HR and personnel files, payroll, health and benefits data, union communications and customer or partner confidential information entirely.

In FKKS's view, anonymization does not make confidential content safe to license.

## What should you do before the first call?

1. Inventory systems and record types, with rough sizes.
2. Mark the categories to exclude, and read contracts and NDAs for third-party content.
3. Collect the dated privacy notices and handbooks people were given.
4. Ask each buyer who transforms the data, to what legal standard, and who certifies it.
5. Ask whether links between records survive, why, and who can reverse them.
6. Ask whether the buyer can pass the data on, and on what terms.

Our [rights check](/rights-check) covers the contract, NDA and personal-data questions in steps 2 and 3, and the [risk map](/risks) shows where deals have broken. A [managed readiness review](/managed-readiness) gathers the facts for your lawyer. Talk to a privacy or commercial lawyer in your state before you sign.

Figure: steps Exclude high-risk records -> Agree who strips names -> Certify the result. Get all three settled in writing before any data leaves.

micro1 may pay us a referral fee if a company we introduce goes ahead with it. [How we are paid](https://vaultfish.com/disclosure)

[Apply to micro1's data partnership (paid link)](https://req.micro1.ai/post/74520e8b-4673-4da5-b855-53609aaf6710?referralCode=f82e519e-5d64-4e0e-9574-760c84371cf2&utm_source=referral&utm_medium=share&utm_campaign=data_referral) · [micro1's own data-partnerships page](https://www.micro1.ai/data-partnerships)

_Paid link. [How we are paid](https://vaultfish.com/disclosure)_

_Vaultfish helps you run the process. We are not lawyers and this is not legal advice. If you need counsel who knows this area, we can point you to one._

## Sources

- [Recital 26](https://gdpr-info.eu/recitals/no-26/)
- [ICO](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/pseudonymisation/)
- [EDPB](https://www.edpb.europa.eu/public-consultations/guidelines-022026-on-anonymisation_en)
- [statute text](https://cppa.ca.gov/regulations/pdf/cppa_act.pdf)
- [45 CFR 164.514](https://www.law.cornell.edu/cfr/text/45/164.514)
- [FKKS](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes)
- [arXiv 2310.07298](https://arxiv.org/abs/2310.07298)
- [arXiv 2602.16800](https://arxiv.org/abs/2602.16800)
- [Doc 1684](https://storage.courtlistener.com/recap/gov.uscourts.nysb.331039/gov.uscourts.nysb.331039.1684.0.pdf)
- [Doc 1489](https://storage.courtlistener.com/recap/gov.uscourts.nysb.331039/gov.uscourts.nysb.331039.1489.0.pdf)
- [Handshake AI](https://joinhandshake.com/ai/data-partnerships/)