
De-identified data is information stripped of anything that points back to a specific person, so it can be studied or licensed without exposing anyone. Here is how de-identification works, which standards matter, and how operational data is cleaned before it leaves your hands.
If you are thinking about licensing your company's operational data to AI and machine-learning builders, de-identification is the step that makes the whole idea workable. It is also the step most owners misunderstand. Deleting a name column is not de-identification, and neither is scrambling an email address with a hash. This guide explains what regulators and standards bodies actually mean by the term, the main methods in use, and what that looks like for email, tickets, CRM records and accounting ledgers.
What de-identified data actually means
De-identified data is data that can no longer reasonably be linked to a particular person, household or device. The goal is not to make records meaningless; it is to keep the patterns (how work flows, how long things take, what gets escalated) while removing the hooks that would let someone figure out who a record is about.
Three ideas sit underneath almost every definition you will meet:
- Direct identifiers are fields that name or contact someone outright: names, phone numbers, street addresses, account numbers, government ID numbers.
- Quasi-identifiers are fields that are harmless alone but identifying in combination: a rare job title plus a small town plus an exact date can single out one person.
- Re-identification risk is the chance that someone, using the released data plus other information they can get, works out an identity. De-identification is about pushing that risk down to an acceptable level, not pretending it is zero.
You will also see the term anonymized used loosely. In many legal contexts it implies a stricter, effectively irreversible standard. And pseudonymized data (where identifiers are swapped for tokens that someone could map back) is generally not treated as de-identified. For definitions in plain English, see our glossary entry on de-identification and the related entries for PII and PHI.
The HIPAA benchmark: Safe Harbor and Expert Determination
The most widely cited de-identification standard in the United States comes from health care. The U.S. Department of Health and Human Services (HHS) guidance on the HIPAA Privacy Rule describes two methods for de-identifying protected health information: Safe Harbor and Expert Determination.
Safe Harbor
Safe Harbor is a checklist. A specified list of identifiers about the individual (and their relatives, employers and household members) must be removed. The list includes names, geographic subdivisions smaller than a state, all elements of dates except the year, phone and fax numbers, email addresses, Social Security and account numbers, device identifiers, web URLs, IP addresses, biometric identifiers, full-face photographs, and a catch-all for any other unique identifying number, characteristic or code.
The detail matters. For example, HHS guidance explains that the first three digits of a ZIP code can only be kept when the geographic unit they form contains more than 20,000 people. The organization must also have no actual knowledge that what remains could identify someone.
Expert Determination
Expert Determination is a judgment. A qualified expert applies statistical and scientific methods and documents that the risk of identifying anyone is very small, given the data and the context in which it will be shared. This route allows more useful detail to survive (for instance, generalized dates) as long as the analysis supports it.
Why this matters if you are not a health care business
Most operational data sellers are not covered by HIPAA at all. But buyers, attorneys and auditors borrow this vocabulary because it is concrete. And to be clear about our own line: We never take HIPAA/PHI data or personal customer data. Everything is de-identified before it leaves your control. If you run a clinic or a dental or veterinary practice, only admin-side operational data (scheduling and billing operations) is in scope; patient records are never included. See healthcare practices for how that boundary works.
What NIST says about de-identification techniques
The National Institute of Standards and Technology (NIST) has published two documents that practitioners lean on heavily:
- NIST IR 8053, De-Identification of Personal Information, which surveys research on removing identifying information and openly discusses the risk that de-identified data can sometimes be re-identified. It covers structured data, free-form text and multimedia, not just tidy tables.
- NIST SP 800-188, De-Identifying Government Datasets: Techniques and Governance, which lays out a full program: set goals, assess risk, pick a sharing model, apply techniques, and test.
The practical takeaways from NIST that apply directly to business data:
- Remove direct identifiers entirely or replace them with tokens that are never shared alongside the key.
- Transform quasi-identifiers: generalize (exact date to month, exact dollar amount to a range), suppress rare values, or add noise.
- Consider synthetic data for the most sensitive fields, where a model generates realistic stand-in values.
- Test for re-identification rather than assuming success. Simple masking often falls short.
- Govern it: write down the standard you are applying and who signs off.
NIST is written for government agencies, so treat it as a strong reference model, not a regulation that binds a private business.
The FTC's position: hashing is not anonymization
For businesses outside health care, the Federal Trade Commission is the regulator most likely to care about how you describe de-identified data. Two FTC positions are worth knowing.
First, the three-part test. In its 2012 privacy report, the FTC said data is not "reasonably linkable" to a consumer to the extent a company (1) takes reasonable measures to de-identify the data, (2) publicly commits not to try to re-identify it, and (3) contractually prohibits downstream recipients from trying to re-identify it. That third prong is why every serious data license includes a re-identification ban. More on that in our guide to data licensing agreements.
Second, hashing alone does not count. In a July 2024 technology blog post titled "No, hashing still doesn't make your data anonymous," FTC staff explained that a hashed email or phone number still produces a consistent, unique value that can track a person over time. The post states plainly that companies should not act or claim as if hashing personal information renders it anonymized.
The lesson for sellers: if a field identifies someone, the answer is to remove or generalize it, not to scramble it and call the job done. This is general information, not legal advice; talk to your attorney about how these principles apply to your business.
How de-identification works for operational business data
Operational data is different from a patient chart or a consumer marketing list. What buyers want is the shape of the work: request, response, hand-off, escalation, resolution. That shape usually survives de-identification well. Here is how it typically looks by source.
Email and team chat
- Names, signatures, phone numbers, addresses and account numbers are detected and replaced with role tokens (for example, "[CUSTOMER]", "[TECHNICIAN]", "[INVOICE_ID]").
- Customer personal data and personal messages are excluded.
- Timestamps are kept or shifted consistently so response-time patterns stay intact.
Relevant sources: Microsoft 365 / Outlook, Google Workspace / Gmail, Slack.
Support tickets and CRM
- Requester identities and contact fields are dropped; categories, priorities, stages, durations and outcomes are kept.
- Free-text fields are scanned for identifiers buried in prose, which is where most leaks hide.
Relevant sources: Zendesk, Salesforce, HubSpot.
Accounting
- Vendor and customer names become generic categories; amounts can be banded; dates can be generalized.
- What remains is the rhythm of the business: billing cycles, payment lag, seasonality.
Relevant sources: QuickBooks, Xero.
A practical de-identification workflow
Whatever tools are used, a defensible process tends to follow the same arc.
- Inventory. List every system, its fields and its volume. A header sample (column names only, no records) is enough at this stage.
- Classify fields. Mark each column as direct identifier, quasi-identifier, sensitive, or operational signal.
- Exclude out-of-scope data. Anything that is PHI, customer personal data, or personal employee matters is removed, not transformed.
- Transform. Remove or tokenize direct identifiers; generalize quasi-identifiers; scan free text.
- Test. Sample the output and try to re-identify. Check for rare combinations.
- Document and contract. Record the method, then license the data with a re-identification prohibition and use limits.
- Deliver from a controlled environment. Only the cleaned output leaves; the raw data and any token keys stay with you.
The order matters: cleaning happens on your side, before delivery. That is what "before it leaves your control" means in practice.
What de-identification does to the value of your data
Owners often worry that removing names will make their data worthless. Usually the opposite is true. AI builders are not trying to learn who your customers are; they want to learn how a plumbing dispatcher triages a no-heat call, how a law firm moves a matter from intake to engagement, or how an accounting team closes the month. Those patterns are what make operational data valuable for training data, and they do not depend on identities.
What does reduce value is over-aggressive cleaning: deleting all timestamps, collapsing every category into "other," or stripping all free text. A good process keeps sequence, timing, categories and outcomes intact.
According to founder Jeff Cline, clean, well-documented data is what gets serious bids, and small companies have gotten over $100,000 for their data. Results vary; not a guarantee. For the factors that move price, see how to value your data or try the valuation calculator.
Common de-identification mistakes to avoid
- Hashing and calling it anonymous. As the FTC has said, a consistent hash is still a unique identifier.
- Forgetting free text. Ticket descriptions, email bodies and notes fields carry names, phone numbers and addresses in prose.
- Leaving rare combinations. A one-of-a-kind job title or a tiny location can identify someone even with names removed.
- Shipping the token key. If the mapping table travels with the data, it is pseudonymized, not de-identified.
- Skipping the contract. Without a re-identification ban, you lose one of the strongest protections available.
- Including data you should never include. Patient records, personal customer data and HR files belong out of scope entirely.
If any of this feels uncertain for your situation, involve your attorney early. It is cheaper than fixing a mistake after delivery.
Ready to see what your de-identified data is worth?
You do not need to clean anything to get a first read. Start by telling us what systems you use, how many years of history you have, your seat count and rough volume. Then share a header sample: column names only, never records. Buyers review it and bid, so you can compare competing offers side by side. Only after you accept does the de-identification, licensing and delivery happen, and then you get paid.
In Jeff Cline's experience, bids come back within 72 hours when the data is correct and a header sample is provided, and deals typically close in about 90 days. Results vary; not a guarantee. The window matters: once the models in a given domain are trained, fresh copies of the same kind of data are worth less. See how it works for the full process.
How to prepare operational data for de-identification
- Inventory your systems. List each tool you use, the years of history it holds, seat count and approximate record volume.
- Export a header sample. Pull column names only, with no records, from each system so fields can be classified.
- Classify every field. Label each column as a direct identifier, quasi-identifier, sensitive field to exclude, or operational signal.
- Exclude out-of-scope data. Remove PHI, patient records, personal customer data and HR files entirely rather than transforming them.
- Transform and test. Remove or tokenize identifiers, generalize dates and amounts, scan free text, then test samples for re-identification.
- License with protections. Deliver only cleaned output under a license that prohibits re-identification and limits use.
Frequently asked questions
What is de-identified data?
De-identified data is information that has had direct identifiers removed and indirect identifiers transformed so it can no longer reasonably be linked to a specific person. It keeps useful patterns, such as timing, categories and outcomes, while removing the details that would reveal who a record is about.
Is de-identified data the same as anonymized data?
Not exactly. The terms overlap, but anonymized usually implies a stricter, effectively irreversible standard in many legal contexts, while de-identified describes data processed to a recognized standard with low re-identification risk. Pseudonymized data, where identifiers are swapped for reversible tokens, is generally treated as neither.
What are the two HIPAA de-identification methods?
HHS recognizes Safe Harbor and Expert Determination. Safe Harbor removes a specified list of identifiers about the individual and their relatives, employers and household members; Expert Determination relies on a qualified expert documenting that the risk of identification is very small. Neither applies to us directly because we never take HIPAA/PHI data or personal customer data. Everything is de-identified before it leaves your control.
Does hashing an email address make data anonymous?
No. FTC staff have stated that hashing creates a consistent, unique value that can still track a person, and that companies should not claim hashed personal information is anonymized. Identifying fields should be removed or generalized, not just hashed.
Can de-identified data be re-identified?
Sometimes, which is why the risk has to be managed rather than assumed away. NIST research documents re-identification risks, so good practice combines technical cleaning, testing, and a license that contractually prohibits buyers from trying to re-identify the data.
Does de-identification make my business data less valuable?
Usually not, because AI builders want the patterns of how work gets done, not the identities of the people involved. Value drops only when cleaning is so aggressive that it destroys timing, sequence, categories or outcomes. Our valuation calculator gives an illustrative range.
Who performs the de-identification, and where?
The cleaning happens on your side before delivery, so raw data and any token keys stay under your control. Only the de-identified output is licensed and delivered to a buyer, under a contract that limits use and bans re-identification.
Do I need a lawyer to sell de-identified data?
You should talk to your attorney before signing any data license. This guide is general information, not legal advice, and the right approach depends on your contracts, your industry and the laws that apply to you. Our guide on whether it is legal to sell business data covers the common questions.
Sources
- HHS: Guidance Regarding Methods for De-identification of Protected Health Information (HIPAA Privacy Rule)
- NIST IR 8053: De-Identification of Personal Information (2015)
- NIST SP 800-188: De-Identifying Government Datasets: Techniques and Governance (2023)
- FTC Technology Blog: No, hashing still doesn't make your data anonymous (July 2024)
- FTC: Protecting Consumer Privacy in an Era of Rapid Change (2012 report)
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.