The short answer
Every workflow data buyer we track de-identifies records, but only Scale AI says where: inside your own environment, before anything is extracted, with your approval on each package. micro1 and Labelbox Prism do it themselves without saying where; Nyne's page only promises de-identified data by the time it ships. Remove obvious personal details yourself first either way.
On this page
Why does it matter who does it?
Workflow records are full of people. Tickets carry customer names and email signatures. Project logs carry staff names in every assignment. CRM notes carry phone numbers and account details. Before any of it is useful to a lab, those identities have to go, while the sequence of who did what has to stay readable.
Where that happens changes your exposure. If the buyer de-identifies inside your environment, raw records never leave your systems. If it happens after transfer, raw records sit with the buyer for some period, under whatever security the agreement requires.
Who does it changes your control. If you do a first pass yourself, you decide what is removed. If the buyer does all of it, you rely on their process and on your right to review the result.
What does each buyer publish?
| Buyer | Who de-identifies | Where or when | What you can review |
|---|---|---|---|
| Scale AI | Scale AI | On your side, pre-extraction | Sign-off per package |
| micro1 | micro1 | Not published | Representative samples; synthetic rewording on some sets |
| Labelbox Prism | Labelbox | Not published | Not published |
| Nyne | Not published | Prior to release | You hold back chosen categories |
Condensed from each buyer’s wording, opened 11/07/2026. Disclosure: an introduction that ends in a deal can earn us a referral payment from one of these buyers. Nothing in the comparison hinges on that.
Two of the four leave “where” unstated. That is a question for the first call, and an answer to write into the agreement.
What is a synthetic rewrite?
micro1 mentions rewriting some datasets synthetically. In plain terms, text is reworded so the process and outcome survive while the original wording does not. It reduces the chance that a phrase in a ticket points back to a real person. Ask which of your record sets it would apply to, and whether you see the rewritten version before use.
What about records from a company that has closed?
SimpleClosure’s AssetHub, which sells on behalf of wound-down startups, says each dataset is screened for personal information before a buyer licenses it; who does the screening goes unsaid. And in the Spirit Airlines case, Reuters noted in August 2026 that the records behind Google’s $10 million bid would be de-identified before any sale closed.
What should you remove before anyone sees a sample?
A first pass on your side costs little and narrows what can go wrong. It is also evidence of care if a question ever arises about what left your systems.
| Field | First-pass treatment |
|---|---|
| Staff names and emails | Replace with stable tokens (agent_07) so the sequence still reads |
| Customer names, emails, addresses | Replace with tokens; strip email signatures and footers |
| Account and payment numbers | Remove entirely |
| Client names in project records | Tokenize, or exclude the project if the client contract forbids licensing |
| Attachments and screenshots | Usually exclude; they hide details that text filters miss |
| Credentials and internal URLs | Scan and strip before any export |
Stable tokens matter more than people expect. If “agent_07” in one ticket is the same person as “agent_07” in the next, a buyer can still see handoffs, workloads and escalation patterns. Random replacement per record destroys that.
Does de-identification make the records yours to share?
No. De-identification answers “can this record identify someone?” It does not answer “did we have the right to license it?”
Two categories stay out unless the contract and the law allow them in:
- Client-owned material. Work made for a client often belongs to the client under the services agreement, de-identified or not.
- Customers’ personal data. A privacy notice written for delivering a service rarely covers licensing records for AI training. De-identification lowers the risk; it does not create a lawful basis that was not there.
Have your own lawyer confirm which record sets clear both tests before you export anything.
What should the agreement say about it?
- Method. Which fields are removed or tokenized, and by what process.
- Location. Inside your environment, or after transfer, and under what security.
- Review. Your right to see a representative sample of the de-identified output, and to reject a package.
- Re-identification. A promise from the buyer not to attempt it, and not to let its customers attempt it.
- Onward use. micro1 promises no resale or distribution while identifying details remain; get the same promise from any buyer, in writing.
Where does this fit in the deal?
De-identification is agreed after the scoping call and before the license is signed: step four of seven in the runbook on the overview of how to sell workflow data. The de-identification sheet keeps each buyer’s published terms in one table.
The safest record is the one that never left your systems with a name still in it, whoever's process runs afterward.