De-identification

De-Identifying Business Data Before Licensing: The Standards, and the Trade-Off

Sukhrobjon (Rob) Ismoilov
Sukhrobjon (Rob) Ismoilov
Founder & Principal, Main Street Wealth

Removing names is the easy fifth of the problem. In home services the service address is the customer, which makes geography the identifier you cannot simply delete — and makes the real question how to stay inside the legal standard without destroying the outcome labels that made the file worth licensing.

Last Updated November 5, 2026
11 min read
4 cited sources
Share:
Quick facts
The misconceptionRemoving names is not de-identification
EU standardAnonymous only if re-identification is not reasonably likely (GDPR Recital 26)
California standard"Deidentified" is a statutory definition, not a description of effort
Home services problemThe service address is the customer — geography is the identifier
The real trade-offSafe enough to release vs. still carrying the labels that made it valuable
When it happensBefore a buyer sees any sample — never after

Section 1

Removing names is not de-identification

The standard assurance in this market is that a specialist will review the data and remove identifying details before a buyer sees it. That describes an activity, not a standard, and it is the reason a lot of datasets fail diligence.

The legal tests are not about effort. Under the GDPR, information is outside scope only where re-identification is not reasonably likely taking account of all means reasonably available — and pseudonymised data, where a re-identification key exists anywhere, remains personal data.[1] California defines “deidentified” in statute, with ongoing obligations attached to maintaining that state, and defines “sale” of personal information broadly enough to capture transfers an operator would never describe as a sale.[2]

Reference treatments of the problem have made the same point for a decade: the difficulty is not the obvious fields, it is the combinations and the re-identification risk they carry.[3] Strip every name from a field services file and you have done perhaps a fifth of the work.

Section 2

Six kinds of identifier, and only one of them is easy

It helps to stop thinking about “PII” as one category. In an operational file there are six, and they need different treatment.

CategoryExamples in an operational fileUsual treatmentWhy it is hard
Direct identifiersName, phone, email, account number, card numberRemoval or tokenisationStraightforward. This is the part most people mean by "de-identification" — and it is the easy 20%.
Quasi-identifiersService address, parcel, ZIP, install date, equipment serialGeneralisation or suppressionWhere home services files actually break. A street address plus an install date is frequently unique.
Structural identifiersRoute sequence, crew assignment, visit cadenceAggregation or re-indexingA weekly route is a map of a neighbourhood. Order alone can reconstruct geography.
Free-text identifiersTechnician notes, call transcripts, scope descriptionsDetection and redactionThe hardest category. Identifiers appear unpredictably, in narrative, misspelled.
Employee identifiersTechnician name, licence number, performance scorePseudonymisation or exclusionWorkforce data with its own legal basis. Licensing files describe work, not workers.
Commercial identifiersClient name on a commercial contract, negotiated pricingExclusionUsually contractual rather than privacy exposure, and usually non-negotiable.
Framework used in Main Street Wealth engagements. Treatment depends on the applicable legal framework and the scope agreed with a buyer; your counsel determines what governs your file.

Section 3

In home services, the address is the customer

This is the structural difficulty that makes field services harder than most industries, and it has no clean solution.

In many sectors a record can describe a transaction without anchoring it to a physical place. Field services work cannot. You attended a specific building, on a specific date, and did something to a specific piece of equipment installed there. A street address combined with an install date is frequently unique to one household — which means location behaves as an identifier even after every name is gone.

You also cannot simply delete it, because geography is often part of what a buyer wanted: climate affects equipment behaviour, local code affects electrical outcomes, route density is a function of where properties sit. Dropping location entirely can remove the covariate that made the dataset interesting.

Section 4

Techniques, and what each one costs you

Every technique trades privacy protection against model utility. Anyone who presents de-identification as free is not describing the same process.

TechniqueWhat it doesProtectionCost to model utilityWhere it fits
SuppressionDelete the field entirelyTotal for that fieldHigh — the signal goes with itDirect identifiers, and anything that cannot be cleaned
PseudonymisationReplace with a stable tokenPartial — still personal data under GDPRLow — joins surviveLinking a customer across visits without naming them
GeneralisationStreet → ZIP3; exact date → monthGood, if coarse enoughModerate — precision is lostQuasi-identifiers like address and install date
AggregationReport groups, not rowsStrongSevere — row-level training data disappearsBenchmarks and summaries, not training sets
k-anonymityEnsure each combination appears k timesMeasurable rather than assertedModerate, and uneven across the setStructured operational records with repeated patterns
Differential privacyAdd calibrated noiseStrongest formal guaranteeHigh for small datasetsRarely appropriate at single-operator scale
General characterisations consistent with standard de-identification treatments; the appropriate technique depends on the dataset, the applicable legal framework and the agreed scope.

Note the fourth row. Aggregation is the technique that most often destroys the asset. An operator advised to “just send summary data” has been advised to send something no one will train on. Benchmarks are not training sets. If a process pushes you toward aggregation, the honest conclusion is usually that this particular file is not licensable — not that a summary will do.

Section 5

Preserving the label is a design requirement

Here is the failure mode we see most often, and it is worth stating bluntly because it wastes real money.

An operator engages someone to clean the data. That person is measured on whether any identifiers remain, so they are appropriately conservative: anything touching a date, a location, a dollar figure or a free-text note gets suppressed. The result passes a privacy review and is commercially worthless, because the outcome — did the repair hold, did the quote close, did the pest return, did the inspection pass — went out with everything else.

This is also why the pass runs on the seller side, before any buyer sees a sample, with the methodology documented. A buyer subject to Article 53 of the EU AI Act has to describe its training content and maintain a copyright-compliance policy, which means a documented methodology is part of the value you are delivering.[4] A buyer offering to do the cleaning for you is asking for the identified version first.

Section 6

What cannot be cleaned into safety

Some records are excluded by category rather than treated, and a credible process names them at the start rather than discovering them late:

  • Safety incidents, injuries and anything adjacent to them
  • Records covered by third-party confidentiality — carrier, owner, general contractor, franchisor
  • Documents describing a client facility, its systems or its security arrangements
  • Property access information: codes, keys, alarm details, occupancy patterns
  • Employment records — I-9s, SSNs, medical notes, immigration documentation, disciplinary files
  • Anything held under a regulatory custody obligation rather than owned outright

Excluding these is not a concession that weakens the file. It is what makes the rest of it defensible, and it is one of the clearest signals to a buyer that they are dealing with a process rather than an opportunist.

Frequently asked questions

Is removing names and phone numbers enough?
No, and this is the most expensive misunderstanding in data licensing. Stripping direct identifiers handles the easiest fraction of the problem. What remains is the combination data — a service address, an install date, an equipment serial, a visit cadence — and in home services those combinations are frequently unique to one household. The European standard asks whether re-identification is reasonably likely using all means reasonably available, not whether the obvious fields were deleted. A file with no names in it can still fail that test comfortably.
Why is home services harder than other industries than most?
Because the address is the customer. In most industries a record can describe a transaction without pinning it to a physical location. Field services work is inherently geographic: you went to a specific building, on a specific date, and did something to a specific piece of equipment there. That makes location a quasi-identifier you cannot simply drop, because proximity, route and property type are often part of what a buyer wanted. Resolving that tension is the actual work.
What does "deidentified" mean legally?
It depends on which law applies, and both relevant standards are stricter than common usage. California defines "deidentified" in statute, with obligations attached to maintaining that state, and its definition of a "sale" of personal information is broad enough to capture transfers an operator would not think of as sales. Under the GDPR, data is only outside scope if it is genuinely anonymous — pseudonymised data, where a key still exists somewhere, remains personal data. Your counsel determines which framework governs your file; the practical point is that neither standard is met by effort or good intentions.
Does de-identification ruin the dataset?
Done carelessly, yes — and that is the failure mode we see most often. A cleaning pass that suppresses every field touching a date, a location or a dollar figure will produce something safe and worthless, because it has removed the outcome labels that made the records trainable. The design requirement is to preserve the label while breaking the link to the individual. That is usually achievable with generalisation rather than suppression, and it has to be designed deliberately rather than handed to a redaction tool.
Can every record be de-identified into a licensable state?
No, and anyone claiming otherwise is selling something. Some records are unsuitable regardless of technique: safety and injury reports, records subject to third-party confidentiality, documents describing a client facility, anything revealing property access or security arrangements, and employment records. Those are excluded rather than cleaned. A credible process says which categories come out at the start, before anybody counts rows.
Who should do the work?
Not the buyer, and not after a sample has been shared. The de-identification pass runs before any buyer sees anything, on the seller side, with the methodology documented well enough that a buyer can satisfy its own provenance obligations. A buyer who offers to handle the cleaning is asking you to hand over the identified version first, which defeats the purpose entirely.
Why does the buyer care how we did it?
Because since August 2025 providers of general-purpose AI models in the EU must publish a sufficiently detailed summary of their training content and maintain a copyright-compliance policy. A buyer with documentation duties needs a methodology it can describe, not an assurance that a specialist looked at it. Documented de-identification is part of what you are selling, and it is one of the few things that reliably separates an advised seller from a scraped alternative.

References

Sourced from primary industry reports, government data, SEC filings, and peer-reviewed valuation research. Retrieved at time of publication.

  1. 1.European Union — Regulation (EU) 2016/679 (GDPR), Recital 26 — Not applicable to anonymous data — The "reasonably likely to be used to identify" test that governs whether a de-identified record is still personal data (2016)
  2. 2.State of California — California Consumer Privacy Act, Civil Code § 1798.140 — definitions of "deidentified" and "sale" — Statutory standard a de-identified operational record must meet, and the breadth of "sale"
  3. 3.U.S. National Institute of Standards and Technology — NISTIR 8053 — De-Identification of Personal Information — Reference treatment of de-identification techniques and re-identification risk (2015)
  4. 4.European Union — Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 53 — Obligations for providers of general-purpose AI models — Copyright-compliance policy and public training-content summary; in force 2 August 2025 under Article 113(b) (2024)

Want to know what survives de-identification in your file?

We will tell you which record categories clear and which come out — before anybody looks at a sample.

Request the evaluation form
Calendar

This article is provided for general educational purposes and is not legal, tax, investment, or accounting advice. Multiples, fee ranges, and timelines are illustrative ranges compiled from third-party sources current as of November 5, 2026 and will vary by transaction and market conditions. Consult qualified counsel before acting on any information contained here.