
De-identification removes or transforms the details that point to a real person so data can be used without exposing anyone. This plain-English guide explains the main techniques, from HIPAA's Safe Harbor model to k-anonymity and differential privacy, and how they apply to business data.
If you are thinking about licensing your company's operational data, the first question is usually: how do you make sure nobody can be identified? The answer is de-identification, a family of techniques with decades of research and regulation behind them. This article explains the main approaches in plain English. It is informational, not legal advice.
By OpenDataPolicies Editorial · October 24, 2026 · 7 min read
First, what ODP does and does not take
Before getting into techniques, the ground rule. We never take HIPAA/PHI data or personal customer data. Everything is de-identified before it leaves your control.
We reference HIPAA below because its de-identification rules are among the most detailed and widely cited in the United States. They are a useful benchmark for thinking about any dataset. But ODP does not handle protected health information at all; for clinics and practices, only admin-side operational workflows are in scope. See healthcare practices for how that line is drawn.
What buyers want from operational data is the shape of the work: how a request arrives, how it is routed, how long it takes, how it resolves. They do not need to know who the people were. That is why de-identification can remove a great deal of detail while keeping the data useful for training AI on how operations work.
Why removing names is not enough
The classic lesson in this field comes from privacy researcher Latanya Sweeney. Using U.S. Census data, her lab found that 87% of the U.S. population (216 million of 248 million people) likely had characteristics that made them unique based on just three fields: 5-digit ZIP code, gender and date of birth (Data Privacy Lab).
The takeaway: a record with no name can still identify someone when several ordinary details are combined. Those details are called quasi-identifiers. Good de-identification deals with direct identifiers (names, emails, phone numbers, account numbers) and quasi-identifiers (dates, locations, rare job titles, unusual amounts).
In business data, quasi-identifiers hide in places people do not expect:
- An email signature with a title and town.
- A support ticket mentioning a rare product configuration.
- A free-text note that says "the owner's daughter called."
- An invoice with an unusually precise amount on a specific date.
That is why de-identification of operational data combines structured rules for known fields with careful handling of free text. See /glossary#quasi-identifier for the term.
Technique 1: Rule-based removal (the Safe Harbor model)
HIPAA's Privacy Rule sets out two ways to de-identify health information. The first, often called Safe Harbor, is rule-based. Under 45 CFR 164.514(b)(2), a covered entity removes 18 categories of identifiers about the individual and their relatives, employers or household members, and has no actual knowledge that what remains could identify someone (eCFR, 45 CFR 164.514).
The 18 categories include names, geographic subdivisions smaller than a state, all elements of dates (except year) directly related to an individual, phone and fax numbers, email addresses, Social Security and account numbers, vehicle and device identifiers, URLs, IP addresses, biometric identifiers, full-face photos, and "any other unique identifying number, characteristic, or code."
The rule also shows how precise these standards get. ZIP codes may be kept only to their first three digits, and only if the area formed by those three digits contains more than 20,000 people; otherwise the three digits become 000. Ages over 89 are grouped together.
Why it is useful outside healthcare: Safe Harbor is a clear, checkable list. Even though ODP never handles PHI, the same categories make a sensible starting checklist for scrubbing email, CRM and ticket data: if a field matches one of those categories, it should be removed, generalized or replaced.
Technique 2: Expert Determination
HIPAA's second method is Expert Determination. A person with appropriate knowledge of "generally accepted statistical and scientific principles and methods" analyzes the data and determines that the risk is very small that the information could be used, alone or with other reasonably available information, to identify an individual, and documents the methods and results (eCFR, 45 CFR 164.514(b)(1)).
The idea is risk-based rather than list-based. An expert might keep some details that Safe Harbor would strip, such as month and year, because the overall risk in context is still very small, while removing others the list does not mention.
For operational data, the same thinking applies. A data set's risk depends on who will receive it, what else they could combine it with, and the contractual controls around it. A purchase agreement that forbids re-identification attempts is part of the protection, not an afterthought.
Technique 3: Pseudonymization (and its limits)
Pseudonymization replaces identifiers with consistent tokens. "Maria Lopez" becomes "Customer_4821" everywhere she appears. That keeps the structure of conversations intact, so a model can still learn that the same customer wrote three times before a ticket was resolved.
Regulators treat pseudonymized data carefully. Under GDPR, Recital 26 says personal data that has undergone pseudonymization, which could be attributed to a person using additional information, should still be considered information about an identifiable person; only truly anonymous data falls outside the regulation (GDPR Recital 26). California's privacy law similarly defines pseudonymization as processing that makes information no longer attributable to a specific consumer without additional information that is kept separately (CCPA text, CPPA).
The practical rule: pseudonymization is a useful step, but it is not the finish line. Any key that maps tokens back to real names must stay with the data owner and never travel with the data.
Technique 4: Generalization and k-anonymity
Generalization makes details less precise: a full date becomes a month, a street address becomes a region, an exact invoice amount becomes a range.
k-anonymity is a formal way to measure whether generalization went far enough. Sweeney defined it this way: a release provides k-anonymity protection if the information for each person cannot be distinguished from at least k−1 other individuals who also appear in the release (Data Privacy Lab, k-Anonymity). If k is 10, every combination of quasi-identifiers matches at least 10 records.
k-anonymity is intuitive and works well on tabular data such as CRM exports or transaction tables. It is weaker on its own when the records in a group share the same sensitive value, which is why practitioners often combine it with related checks and with suppression of rare records.
Technique 5: Differential privacy
Differential privacy is the most mathematically rigorous approach. NIST describes it as "a mathematical framework that quantifies privacy loss to entities when their data appears in a dataset" (NIST SP 800-226). In practice, it adds carefully calibrated random noise so that the results of an analysis look almost the same whether or not any one person's data is included.
The best-known deployment is the U.S. Census Bureau, which applied a disclosure avoidance system based on differential privacy to 2020 Census data (U.S. Census Bureau).
Differential privacy is strongest for statistics and aggregate releases. It is harder to apply to raw text such as emails, which is why operational-data pipelines usually rely on removal, pseudonymization and generalization for the records themselves, and reserve differential privacy for summary statistics.
Technique 6: Free-text scrubbing
Most operational value lives in free text: email bodies, chat threads, ticket descriptions, notes. De-identifying it usually combines:
- Pattern matching for emails, phone numbers, card and account numbers, URLs and IP addresses.
- Named-entity recognition to find names of people, places and organizations.
- Allow-lists and deny-lists specific to the business (staff names, client names, product codes).
- Human review of samples to catch what automated passes miss.
No method is perfect on its own, which is why layered approaches, plus contractual bans on re-identification, are the norm.
How this fits together for a data seller
For a business licensing operational data, the process typically looks like this:
- Scope out what is never included. PHI, patient records, payment card data, and personal customer data stay out. We never take HIPAA/PHI data or personal customer data. Everything is de-identified before it leaves your control.
- Share a header sample (column names and counts, no content) so buyers can judge fit. See what a header sample is.
- Apply layered de-identification to the in-scope records: removal, pseudonymization, generalization and free-text scrubbing.
- Check the result against a list like Safe Harbor's categories and, where appropriate, measure risk with methods such as k-anonymity.
- Back it with contract terms that forbid re-identification. California's definition of "deidentified" itself expects the business to publicly commit not to re-identify and to contractually bind recipients to the same (CCPA text, CPPA).
For a deeper walk-through aimed at sellers, read our pillar guide on de-identified data, and before you sell, run through the data rights checklist.
Want to know what your de-identified data could be worth? Request a free bid.
Frequently asked questions
What is the difference between de-identification and anonymization?
De-identification is the process of removing or transforming identifying details. Anonymization usually means the result cannot reasonably be linked back to a person by anyone. Regulators such as GDPR treat only truly anonymous data as outside their scope.
Does ODP use HIPAA Safe Harbor?
We reference Safe Harbor's 18 identifier categories as a benchmark checklist, but ODP never handles protected health information. We never take HIPAA/PHI data or personal customer data. Everything is de-identified before it leaves your control.
Is pseudonymized data anonymous?
Not on its own. Under GDPR Recital 26, pseudonymized data that could be re-attributed with additional information is still personal data. The key linking tokens to identities must stay with the data owner.
What is k-anonymity in simple terms?
It means each record looks the same as at least k−1 others on the details someone could use to identify a person, so no individual stands out.
Sources
- eCFR — 45 CFR 164.514 (de-identification standard)
- Data Privacy Lab — Simple Demographics Often Identify People Uniquely
- Data Privacy Lab — k-Anonymity (Sweeney)
- NIST SP 800-226 — Guidelines for Evaluating Differential Privacy Guarantees
- U.S. Census Bureau — Differential privacy and the 2020 Census
- GDPR Recital 26 — Not applicable to anonymous data
- California Privacy Protection Agency — CCPA statute text
Updated November 19, 2026.
OpenDataPolicies.org connects businesses with buyers of de-identified operational data. We are not a law firm, broker-dealer or investment adviser, and nothing on this site is legal, tax or financial advice. Statements about amounts, timing and outcomes are Jeff Cline's statements based on past transactions. Results vary; nothing here is a guarantee of any bid, price, buyer or closing. Every deal depends on the data, its quality, its rights and buyer demand at the time.