De-identification, explained for owners
Before a buyer trains on your records, the people in them are removed. Here is what that means in practice, what a de-identified ticket looks like next to the original, and what belongs in the contract.
When an owner hears that an AI company wants to train on their support tickets, the first question is usually about the people in them. Customers who wrote in with a problem. Employees whose names are on every reply. Suppliers copied on a thread. None of them signed up to appear in a training set, and no owner wants to be the reason they did.
De-identification is the answer to that question, and it is less mysterious than it sounds. It is the process of removing or replacing the information that identifies a person before a buyer uses the records, and it is written into the contract rather than left to goodwill. Here is what it means in practice.
Before and after, on one ticket
Here is an illustrative support ticket as it might sit in a helpdesk today. Every detail is invented.
From: Dana Whitfield, dana.whitfield@example-client.com, +1 415 555 0142
Subject: Invoice 48213 charged twice
Body: Hi, we were billed twice for the March renewal on account AC-7781. Can you reverse one? Our CFO Mark needs this fixed before the board meeting on the 14th.
Agent: Priya Raman
Resolution: Duplicate charge confirmed, refund issued, root cause was a retry in the payment gateway.
The same ticket after de-identification:
From: [CUSTOMER_0412], [EMAIL], [PHONE]
Subject: Invoice [INVOICE_ID] charged twice
Body: Hi, we were billed twice for the [MONTH] renewal on account [ACCOUNT_ID]. Can you reverse one? Our CFO [PERSON] needs this fixed before the board meeting on the [DATE].
Agent: [AGENT_07]
Resolution: Duplicate charge confirmed, refund issued, root cause was a retry in the payment gateway.
Everything a model learns from is still there: a billing problem, a clear request, an urgency signal, a diagnosis, and an outcome. Everything that points to a person is gone. Tokens like [CUSTOMER_0412] stay consistent across that customer's tickets, so the buyer can still see a history without knowing whose it is.
What gets removed
Direct identifiers. Names, email addresses, phone numbers, account and invoice numbers, physical addresses, payment details, employee IDs. These are replaced with consistent tokens or dropped entirely.
Quasi-identifiers. Information that is harmless alone and revealing in combination: exact dates, small locations, rare job titles, the names of customer companies. These are generalized. A date becomes a month or is shifted by a fixed offset; a town becomes a region; the only pediatric dentist in a small town becomes a dental practice.
Free text. The hard part. Names and numbers appear inside sentences, signatures, quoted replies, and attachments, not just in fields. This step uses pattern matching plus models trained to spot personal information in prose, followed by a human review of a sample.
Attachments and metadata. Documents carry author fields and tracked changes; images carry location data; calendar invites carry attendee lists. Either the attachment is excluded or it goes through the same pass.
Who does the work
The seller-side agent does not, ever. The agent never receives the dataset, so it cannot process it. That is a structural fact about chain of custody, not a policy.
In most licenses, the buyer de-identifies under contract before any training run, using tooling built for this. Some owners prefer to run a first pass on the export themselves, with their IT team or a vendor, so that what leaves the building is already reduced. Both are legitimate. The contract should say which, and should name the method.
What de-identified does not mean
It does not mean perfect. No method reduces re-identification risk to zero, and you should be suspicious of anyone who writes 100 percent anonymized into a proposal. It means the direct identifiers are gone, the quasi-identifiers are generalized to an agreed standard, and the buyer is contractually barred from trying to reverse it.
It also does not automatically cover employees. The reported Spirit Airlines data sale drew an objection from a pilots' union on exactly this point: the privacy terms, by the union's reported account, protected passengers more clearly than staff. Whatever the court decides, the lesson for an owner is simple. Ask whether the de-identification standard applies to the people who wrote the records, not only the people they wrote about.
What belongs in the contract
A workable de-identification clause has six parts. Ask for all of them.
- Method. What is removed, what is generalized, what tooling is used, and whether tokens are consistent.
- Scope. Customers, employees, and third parties alike.
- Verification. Who checks a sample before training, and what failure rate sends it back.
- Prohibition. No attempt to re-identify, by the buyer or by anyone it shares derived work with.
- Deletion. A date by which the supplied copy is destroyed, and written confirmation when it is.
- Audit and notice. A right to confirm the above, and a short deadline for telling you if anything goes wrong.
Every buyer we work with contracts to de-identify records before use and to put these terms in writing. We will not introduce a buyer who will not.
The practical sequence
Nothing is de-identified at the start, because nothing moves at the start. Buyers price from a one-page manifest that contains no records. After a first offer, the buyers you approve see a small sample, which is de-identified before it is shared and is covered by a signed agreement. Only after binding terms and your signature does the dataset move, directly from your systems to that buyer, with the method and the deletion date already fixed.
If you want to see the standard before you decide anything, the privacy and custody page sets it out in full. The manifest is still the place to start, and it is the only thing that leaves your building before you sign.
Questions worth asking
Who actually does the de-identification work?
In most licenses the buyer does it, under contract, before any training run, using tooling built for exactly this. Some owners prefer to run a first pass themselves on the export, with their own IT team or a vendor, so that what leaves the building is already reduced. Either way the method, the verification step, and the deletion date are written into the agreement. A seller-side agent never performs it, because the agent never holds the data.
Can de-identified records ever be traced back to a person?
The honest answer is that no method reduces the risk to zero, which is why no serious buyer or agent will put the words one hundred percent in writing. Direct identifiers can be removed completely; quasi-identifiers like dates, locations, and unusual job titles are generalized so that combinations no longer point at one person. The contract then adds a prohibition on attempting re-identification, an audit right, and breach notice. Ask any buyer to describe their method and their verification step. Vagueness there is a red flag.
Does de-identification remove the value buyers are paying for?
Mostly no. Buyers want the structure of the work: the problem, the questions asked, the decision, the outcome. None of that depends on knowing the customer's name or phone number. What can reduce value is over-aggressive scrubbing of the free text, which is why the method matters and why a sample is reviewed before terms are final. A good de-identification pass removes who while keeping what happened.
Measured from a complete data manifest. A first offer is a process target, not a guarantee; buyers decide eligibility, price, and terms. Indicative value ranges mentioned on this site are illustrative, not quotes. Nothing here is legal advice; have your own counsel review any agreement before data moves.