Skip to content

Privacy and preparation

Redaction consistency across systems: a test for email, CRM and files

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

To transform the same identity consistently across email, CRM and attachments, resolve every name variant to one person, then apply one replacement function with one key and one configuration to every system, including attachment text, images and file metadata. Prove it with a test: trace a panel of known people through the output and check for splits, merges and leaks before anything is shared.

Redaction

Redaction consistency

This page covers redaction consistency across systems: a test for email, CRM and files.Illustrative

Key takeaways

  • Redaction consistency across systems means one person gets one placeholder in the CRM, email and attachments, and no original value survives anywhere.
  • Test for three failures: splits, merges and leaks, using a panel of known people traced through every system.
  • Most failures come from different tools, keys or settings per system, unresolved name variants, and unprocessed attachment metadata.
  • Consistency makes records more useful and more linkable; keep it only for the joins the task needs.
  • Store the key or mapping apart from the data, never ship it, and destroy it when processing ends.

What does redaction consistency across systems mean?

Redaction consistency across systems means one real person receives exactly one replacement value everywhere they appear, whether in a CRM contact record, an email header, a signature block or a PDF attached to that email, and no original identifier survives in any of them. You prove it with a test, not by trusting the tool settings: pick a panel of known people, trace each one through every system, and check the output for splits, merges and leaks.

Those three failure types are worth naming because they have different consequences. A split gives one person several placeholders, so a buyer sees two strangers where there was one customer and the thread loses its meaning. A merge gives two different people the same placeholder, which corrupts the record and can attach one person's history to another. A leak is an original name, email address or phone number that survives in one system while it was removed in the others.

NIST describes pseudonymization as a de-identification technique that replaces an identifier for a data principal with a pseudonym. Consistency is the property that makes the pseudonym useful across sources. It is also the property that makes the output easier to link, which is why the test below checks both utility and residual risk.

Why do email, CRM and attachments drift apart?

They drift apart because each system stores the same person in a different shape and is usually processed by a different tool. The CRM holds structured fields such as name, email, phone and account owner. Email holds display names, addresses in headers, salutations, signatures and quoted replies. Attachments hold rendered text, scanned images, embedded pictures and file metadata such as the author field.

Different shapes invite different transformations. Structured fields are often replaced with field rules, email bodies with an entity recognizer, and attachments with OCR followed by redaction. Unless all three call the same replacement function with the same key and the same normalization rules, the same person will come out differently.

Tool defaults can make this worse. Microsoft Presidio's documentation says its hash operator uses a random salt by default since version 2.2.361, so the same value receives different hashes across entities and calls unless you supply a consistent salt. Presidio's own pseudonymization sample, which keeps a shared mapping of values to placeholders, warns that the logic is not thread-safe and may produce incorrect results when run concurrently, because the mapping has to be shared between workers. Both are exactly the kind of setting that differs between a CRM job and an attachment job run weeks apart.

The cross-system identity transformation consistency test

The test below is the reusable asset. Run it on the processed output of every system in scope, before anything leaves your control, and record the result for each row. Each test uses a reference panel of real identities whose original values you hold separately and securely.

Run the tests in order. T1 comes first because a configuration difference explains most later failures and saves a long search. T2 to T5 test utility: whether the same person can still be followed where the task needs it. T6 to T8 test residual exposure: whether anything original survived. T9 and T10 test whether the result can be repeated and whether the means to reverse it stays with you.

Record each result with the date, the batch or export it covers, the reviewer and any fix applied. A failed row is not a reason to patch one record and move on; find the cause, fix it in the pipeline, and rerun every row the fix could affect. Keep the panel list itself in the same restricted location as the key, because it pairs real identities with their placeholders and is itself a re-identification aid.

The cross-system identity transformation consistency test
TestWhat you checkHow to run itPass condition
T1 Configuration parityEvery system was processed with the same key identifier, normalization rules, entity types and tool versionCompare the run logs or configuration files for the CRM, email and attachment jobs side by sideOne configuration record covers every run; any difference is explained in writing
T2 Structured joinA CRM contact and the same address in email headers receive the same placeholderFor each panel person, look up the placeholder in the CRM output and in From, To and Cc of their emailsEvery panel person joins across the two systems
T3 Variant resolution (no splits)Full name, first name, initials, nickname, email local part and signature form map to one placeholderList each panel person's known variants and search the output for the placeholder each receivedOne placeholder per panel person in every system
T4 Collision check (no merges)Different people who share a name or initials receive different placeholdersInclude at least two panel people with the same or similar names and compare their outputsNo placeholder is shared by two panel people
T5 Attachment text agreementNames found in attachment text layers or OCR receive the same placeholder as the email bodyOpen processed PDFs, spreadsheets and documents attached to panel emails and compareAttachment placeholders match the body and CRM placeholders
T6 Pixel layerImages, screenshots and scanned pages no longer show the person's name, face, signature or contact detailsView every processed image attached to panel emails at full sizeNo original identifier is visible in any image
T7 Metadata residueFile author, last-modified-by, CRM owner and audit fields, and email header fields are transformed or removedRead the metadata of processed files and the full header set of processed emailsNo original panel value appears in metadata
T8 Leak searchNo original panel value survives anywhere in the outputSearch all outputs for every original name, email, phone and account number on the panel listZero hits
T9 Re-run determinismA later batch gives the same people the same placeholdersReprocess a small slice of already processed records with the stored configuration and compareIdentical placeholders, or a documented decision to rotate keys
T10 Key and mapping custodyThe key or mapping table is stored apart from the output and is not shipped with itInspect the delivery package and the storage location of the key or mappingKey and mapping absent from the package and held under restricted access

How do you build the reference panel?

Build the panel from the CRM, because it is the one system that already ties a name to an email address, phone number and account. Choose people who appear in all three places: a contact record, email threads and at least one attachment. A panel of 25 to 50 people is a practical starting point for a mid-sized archive; the aim is coverage of hard cases, not a statistical estimate. The panel complements, and does not replace, a random sample review of the whole output for missed identifiers.

Deliberately include the people most likely to break the transform. The list below covers the cases that cause most splits and merges in practice.

  • Two or more people with the same first and last name, or the same initials, at different accounts.
  • A contact who changed email address or company, so older threads use a different address.
  • A contact usually addressed by a nickname or short form, such as Bob for Robert.
  • Someone who appears only in signature blocks or forwarded chains, never in the CRM owner fields.
  • An internal employee who is also a CRM record owner, so their name sits in audit and owner fields.
  • A contact whose name also appears inside an attachment, such as a quote, order form or screenshot of a CRM screen.

Illustrative worked example: one contact across three systems

Illustrative example: a fictional industrial parts distributor prepares CRM contact records, a sales mailbox export and the PDF quotes attached to sales emails. One panel person is a fictional buyer named Dana Ortiz, who also shares a first name with a second fictional contact, Dana Lee. The table at the end of this section shows where she appears, what the test expected and what the first run produced; the four fixes listed first answer the failures it records.

  • Fix for the merge: resolve first-name-only mentions using thread context, and when context cannot decide, replace them with a generic token rather than guessing a person.
  • Fix for the split: add signature forms and initials to the entity resolution step before replacement.
  • Fix for the attachment job: run it with the same key and configuration as the email job, then repeat T1, T3 and T5.
  • Fix for the metadata leak: strip file metadata as a separate step, then repeat T7 and T8 across all attachments, not only the panel's.
Illustrative worked example: one contact across three systems
LocationOriginal formExpected outputFirst-run output
CRM contact recordDana Ortiz, d.ortiz at the customer's domainPERSON_0412 and EMAIL_0412PERSON_0412 and EMAIL_0412 (pass)
Email From headerDana Ortiz with the same addressPERSON_0412 and EMAIL_0412PERSON_0412 and EMAIL_0412 (pass)
Email salutationHi DanaPERSON_0412PERSON_0877, the placeholder for Dana Lee (merge, T4 fail)
Signature blockD. Ortiz, PurchasingPERSON_0412PERSON_1290, a new placeholder (split, T3 fail)
PDF quote text layerAttn: Dana OrtizPERSON_0412PERSON_3301, from a separate OCR job with its own salt (split, T1 and T5 fail)
PDF author metadatadortizRemoveddortiz left in place (leak, T7 and T8 fail)

What identifying combinations survive consistent placeholders?

Consistent placeholders remove names but keep everything that was linked to them, so the combination can still point to one person. A placeholder that appears in the CRM with a job title, a company size band and a city, and in email with a distinctive project and dates, builds a profile that someone with outside knowledge may recognize. Consistency across systems makes that profile richer than any one system alone.

The UK ICO's anonymization guidance asks organizations to consider the means reasonably likely to be used by a motivated intruder, someone reasonably competent with access to resources such as the internet and public documents. The EDPB's Guidelines 01/2025 on pseudonymization, adopted in January 2025 as a version for public consultation, say pseudonymized data remains personal data where it can be attributed to a person using additional information. A consistency test proves the transform is clean and stable; it does not prove the output is anonymous.

Check the residue that survives the transform: job titles combined with small accounts, unique events described in email, writing style in long messages, and images of people. Treat what you find as an input to a separate re-identification review rather than as a failure of the consistency test.

Attachments deserve a second look here. A scanned purchase order or a screenshot of a CRM screen can carry a full name, a signature or a face that the text pipeline never saw, and the Presidio Image Redactor, one open-source option for the pixel layer, works by running OCR on images and is still labeled beta by its maintainers. Images of people and handwriting are the cases where excluding the attachment type is often safer than redacting it.

When should you break consistency on purpose?

Break consistency on purpose when the intended use does not need to follow a person across systems. If a buyer wants single emails for classification, or CRM records for field-level modeling, a different placeholder per system or per record removes the cross-system link and lowers linkage risk at little cost to utility.

This is the main exception to the method. Consistency is a utility feature that carries a privacy cost, so justify it with the task. Write down which joins the task needs, such as customer to email thread to quote, and keep consistency only for those entity types. Staff names, for example, may need consistency for a routing task and none for a summarization task.

Where should the key and mapping table live?

The key or mapping table should live outside the dataset, under restricted access, and be destroyed when no further processing is planned. The EDPB guidelines describe such additional information as tables matching pseudonyms to the identifiers they replace, or cryptographic keys, and say it must be kept separately and protected by technical and organizational measures.

Presidio's documentation gives practical advice for its salted hash: discard the salt after processing unless you need to process more data later, store it in a key vault or secrets manager if you must keep it, use at least 128 bits, and never include it in the output. For health records in the United States, HHS guidance on the HIPAA de-identification standard says a re-identification code should not be derived from information about the individual, which rules out placeholders built from initials or birth dates.

Decide early whether you will ever process more records for the same buyer. If you will, keep the key under restricted access with a named owner and a review date, because a later batch has to reuse it for consistency. If you will not, destroying the key after the final run removes one route back to the original identities. Either way, write the decision down so a reviewer can see what was done and why.

Do not build placeholders from the original values in any readable way. PERSON_DOrtiz or a plain unkeyed hash of an email address can be reversed by anyone who can guess the inputs.

Which test failures should pause a delivery?

Stop and narrow the scope when a test fails twice after a fix, or when a failure cannot be traced to a cause. Ask a privacy specialist when the data includes sensitive categories, records about minors, health or financial details, or when the output will be shared outside your organization.

  • T8 finds an original value after a fix: pause delivery, trace the source system and rerun the full test.
  • T4 shows merges you cannot resolve from context: replace the ambiguous mentions with a generic token.
  • Attachments include scanned forms or images of people: consider excluding that attachment type rather than redacting it.
  • The buyer asks for the mapping or key: refuse, and route the question to counsel.
  • Detection depends on automated tools alone: add human review, since Presidio's documentation says automated detection may not find all sensitive information and other protections should be used.

Does a passed consistency test clear records for licensing?

Consistency is a privacy property, not a permission. Access to the systems, a privacy basis for using the data, ownership of the records and contractual rights to license them are four separate questions, and a clean consistency test answers none of them. Counsel reviews those facts for the specific dataset.

When CRM, mailbox and attachment records are prepared for an AI license through SourceX, the rights review and contracting sit alongside the data preparation, and the de-identification itself is done by SourceX through approved processing routes rather than left to the seller's own scripts. The company's own sign-off on scope and terms comes before any delivery, and its eligibility is checked first in a brief assessment that needs no data.

Two kinds of link do the damage. A copy is one passage stored in more than one file: a quoted reply, a forwarded chain, an attachment saved to both mailbox and CRM, or version 2 of a contract beside version 3. A pointer is a value that leads from one file to another: an order, invoice, ticket or contract number, a file name or folder path, a link, a message or thread identifier, or the words see attached.

Each fails two ways. It restores an identity when copies were masked differently, or when a pointer leads from a clean file to one where the person or customer is still named. It breaks the task when the pointer is dropped or rewritten differently per file, so nobody can tell which quote an email is chasing. T1 to T10 follow one person through each system; this test follows each document against its neighbors.

Regulators judge identifiability across records, not file by file. The UK ICO's anonymisation guidance treats singling out one person across records as a mark of identifiability and asks whether others' specific knowledge, colleagues' included, supports inferences that link data to someone (paraphrase). The EDPB's Guidelines 01/2025 on pseudonymisation (EU; consultation version, January 2025) call "the context in which pseudonymisation is to preclude attribution of data to specific data subjects" the pseudonymisation domain, often "the set of all authorised recipients": files delivered together share one, so a clue in any counts against all.

Do default detector settings catch the keys that join one file to another?

Mostly not, and the keys they do catch become one shared label, not a consistent token. We read the shipped defaults of two open-source redaction engines at pinned commits without running them. In Presidio (data-privacy-stack/presidio, commit 211afc3, 27 October 2026), 17 of the 95 entries in the default recognizer list load for English, and the default spaCy model adds people, locations, nationality, religious or political groups, and dates or times; organization names are mapped but ignored. Its only recognizers for document reference numbers, four US healthcare-administration types, are disabled. Phileas (philterd/phileas, commit 8c93b96, 23 October 2026) documents 32 predefined filters that apply only when a policy names them; its one business reference type is UPS, FedEx and USPS tracking numbers, and its custom identifier filter defaults to any word of four or more capitals, digits, hyphens or underscores, so four-digit years and four-letter acronyms match too.

Both write one label per type unless told otherwise: Presidio the entity name in angle brackets, Phileas a REDACTED tag with the filter type. So a default run leaves most join keys and company names in place, keeping restore paths open, and turns the keys it catches, such as links and UUIDs, into one placeholder, cutting links the task may need. Patterns built for other values fire unevenly too: Presidio's US bank-number pattern accepts any run of 8 to 17 digits, so a long numeric invoice number may be replaced while a short one survives. Each key type needs a decision.

How we counted: files read, never executed. A Presidio entry is on for English when it lacks enabled: false and lists English or no language (the engine then uses English); default spaCy labels count unless ignored; entity types are each recognizer's declared default. Key types match whole underscore-separated tokens of entity names, or whole words of Phileas filter names and descriptions (final s dropped from words over three letters), against the table's tokens. Not measured: accuracy on real text, other engines, team settings.

What each engine does with each key type as shipped:

Do default detector settings catch the keys that join one file to another?
Key typeTokens searched (whole tokens)Presidio, default English setup (211afc3)Phileas, predefined filters (8c93b96)What a default run does to cross-file links
Business reference numbersORDER, INVOICE, PURCHASE, PO, QUOTE, TICKET, CASE, CONTRACT, CLAIM, TRACKING, SHIPMENT, REFERRAL, AUTHORIZATION, PRESCRIPTIONOff: US claim, prior authorization, prescription and referral numbers, all disabled; no order, invoice, ticket or contract typeTracking numbers (UPS, FedEx, USPS) when a policy lists the filterNo type aimed at order, invoice, ticket or contract numbers, so they survive unless a broader pattern fires: an email still joins the invoice it cites, and any file where the customer is still named
File names and pathsFILE, FILENAME, PATH, DOCUMENT, ATTACHMENTNoneNoneNo filter aimed at them; a person or customer folder in a file name or path survives unless another pattern happens to match
Message and thread identifiersMESSAGE, THREAD, CONVERSATIONNoneNoneNo identifier filter in either engine; whether an email pattern also catches some identifiers was not tested
URLs and linksURL, URI, LINKOn: URLURLs, when a policy lists the filterWhere detected, every link gets the same label, so a reader cannot tell which document a message pointed to
Record identifiersUUID, GUID, RECORDOn: UUIDNonePresidio turns every UUID into one label, collapsing record links; Phileas has no UUID filter, so they survive unless a custom pattern catches them
Organization namesORGANIZATION, ORG, COMPANYOff: ORGANIZATION, mapped from spaCy but ignored, and elsewhere in the list emitted only by two disabled model-based recognizers; a South African company registration number, disabledNo general company filter; a hospital-name list existsCustomer and vendor names survive in every file
Email addresses and phone numbersEMAIL, PHONEOn: EMAIL_ADDRESS, PHONE_NUMBEREmail addresses, phone numbers and phone number extensions, when listedOne label per type, so sender and recipient look alike unless a consistent operator is set

Which checks prove a linked set is both private and still usable?

A pair of checks per link: a restore check, asking whether a neighbor undoes the redaction, and a break check, asking whether the link the task needs still works. Run them on the processed output of the whole linked set after T1 to T10 pass and before anything leaves your control. A row that passes one check and fails the other has failed.

The rows are editable defaults for a mailbox, CRM and shared-drive export; a privacy specialist should confirm them against the actual records:

Which checks prove a linked set is both private and still usable?
LinkRestore check (privacy)Break check (task meaning)Pass condition
Quoted and forwarded copiesAlign each quoted block with the message it copies and compare masks span by span; a span masked in one copy and readable in another failsPlaceholders in the quote match the original, so the reply still answers something legibleIdentical masks and placeholders in every copy
One attachment stored in several placesMatch copies by hash or near-duplicate text across mailbox, CRM and drive, then compare their redactionsThe CRM record or email still points to a delivered rendition of the fileOne redacted rendition used everywhere, or every copy withheld together
Successive versions and draftsCompare each version with the next, tracked changes and comments included; a value removed from one and visible in another failsChanges the task depends on, such as amounts and terms, stay visible between versionsEvery version redacted under the same rules, or only the final version kept
Business reference numbersFor each surviving number, list every delivered file that carries it; if any still names the customer or a person, the number failsAn email and the order, invoice or ticket it cites carry the same value or the same tokenA written keep, consistent-token or drop decision per number type, applied in every file
File names, folder paths and linksSearch names, paths and link text for person names, customer folders and user account foldersA see-attached line resolves to a delivered file under its new namePaths and names carry the same tokens as body text; a withheld file leaves a stub naming its type
Message and thread identifiersCheck what each identifier contains, such as a mail domain, before deciding to keep itReply chains rebuild from the delivered headersOne remapping applied across all messages, so threads rebuild and no domain remains
Documents left out of the setSearch the remaining files for quotes from, or summaries of, each withheld documentReferences to a withheld document are marked as withheld, not left danglingNo content of a withheld file survives elsewhere, and every reference is marked
Counterparty namesCheck whether a company name plus a role, date or deal size in one file singles out a person pseudonymized in anotherThe customer or vendor the task depends on remains the same party in every fileA recorded decision per counterparty: keep, consistent token or generalize

What must the final check on a linked set prove before delivery?

It must show that every copy carries the same masks, every surviving key is a recorded decision, and no reference points at something the reader cannot see. Deliver when every row passes. Narrow the scope when one link type keeps failing, for example by dropping drafts or a folder of scans, rather than patching files singly. Hold the set for a privacy specialist when a surviving key reaches a file you cannot clean, or the task needs a join that also singles someone out. Limits: this finds clues inside the delivered set, not in outside data a recipient holds; the next evidence is a count of copy disagreements and surviving keys on a sample of your own records.

Through SourceX, the supplier decides which systems, records and date ranges are in scope and the rest stays out, so documents that cannot be cleaned together are best left out together. SourceX de-identifies the data through approved processing routes and records the method per dataset; its supplier agreement states that "re-identification may still be possible using the deidentified data alone or combined with other sources". The estimator takes a few questions and no uploads; its estimate is indicative, not an offer or a payout promise.

The checklist, kept with the T1 to T10 results:

What must the final check on a linked set prove before delivery?
Checklist itemWhat the record showsBlocks delivery when
Link inventoryEvery link type in the set and how many files carry each, listed before processing startsA link type turns up after processing that the inventory missed
Copy agreementEvery quoted, forwarded, duplicated and versioned pair compared span by spanAny span is masked in one copy and readable in another
Key decisionsOne written keep, token or drop decision per reference number, path, link and identifier type, with a search showing it was applied in every fileA key type has no decision, or one file applies it differently
Neighbor searchFor a sample of each surviving key, every file carrying it opened side by sideAny file names the person or customer the others hide
Task replayA reviewer who never saw the originals follows cross-file references chosen before the run, using the delivered files aloneA reference cannot be followed, or leads to the wrong document
Withheld-file sweepThe remaining files searched for quotes or summaries of each excluded documentExcluded content survives, or a reference to it is unmarked
Detector settingsEngine, commit or version, enabled entity types and operators, and the key types the defaults left aloneSettings are unrecorded, or differ between batches without a written reason
Sign-offReviewer, date, batch and every fix, with affected rows rerunA fix went in without rerunning the rows it could affect

Sources for this section

  • The UK ICO's guidance on effective anonymisation treats singling out one person across records as a mark of identifiability and asks whether others' specific knowledge, such as colleagues', allows inferences linked to an identifiable person (paraphrase). Source
  • The EDPB's Guidelines 01/2025 on pseudonymisation (version for public consultation) call the context in which pseudonymisation is to preclude attribution the pseudonymisation domain, which may be restricted to defined entities, most often the set of all authorised recipients. Source
  • Presidio's default recognizer list at commit 211afc3 has 95 entries, 25 enabled and 17 loading for English; its only reference-number recognizers (US claim, prior authorization, prescription and referral numbers) are disabled. Source
  • Presidio's US bank-number recognizer, enabled for English by default, matches any run of 8 to 17 digits (pattern \b[0-9]{8,17}\b), and the default analyzer score threshold is 0. Source
  • Presidio's default NLP configuration uses spaCy en_core_web_lg, maps model labels to PERSON, NRP (a person's nationality, religious or political group), LOCATION, ORGANIZATION and DATE_TIME, and lists ORGANIZATION among the labels to ignore. Source
  • Presidio's anonymizer defaults to the replace operator, which writes the entity type in angle brackets when no replacement value is given. Source
  • Phileas documents 32 predefined filters at commit 8c93b96, enabled in combinations called policies; its tracking-number filter covers UPS, FedEx and USPS, and no predefined filter targets file names, message or thread identifiers, UUIDs or company names in general (a hospital-name list exists). Source
  • Phileas's custom identifier filter has a default pattern matching any word of four or more capital letters, digits, hyphens or underscores. Source
  • In Phileas, a filter with no strategies defaults to REDACT with a redaction format that inserts the filter type. Source

Frequently asked questions

How do you transform the same identity consistently across email, CRM and attachments?

Resolve every variant of a person to one entity, then apply one replacement function with one key and one configuration to every system. Prove the result with a test: trace a panel of known people through the output and check for splits, merges and leaks, including attachment text, images and metadata.

Is a hash of the email address enough to keep placeholders consistent?

No. An unkeyed hash is consistent but can be reversed by hashing a list of likely addresses and comparing. A keyed hash or a protected mapping table keeps consistency without that weakness, as long as the key is stored apart from the data.

What is the difference between a split and a merge?

A split gives one person several placeholders, so their records no longer connect. A merge gives two different people one placeholder, so their records are wrongly joined. Splits mostly hurt utility; merges hurt both utility and accuracy, and can attribute one person's details to another.

Does a passed consistency test make the data anonymous?

No. The test shows that names and contact details were replaced the same way everywhere. Linked details such as job titles, events and dates can still identify someone, so a separate re-identification review is needed, and pseudonymized data can remain personal data under GDPR.

Should staff names be pseudonymized the same way as customer names?

Usually yes, with their own entity type, because staff names appear in CRM owner and audit fields as well as in email. Whether they need to be consistent depends on the task; if no one needs to follow a staff member across systems, separate placeholders per system lower linkage risk.

Sources

  • NIST's glossary defines pseudonymization as a de-identification technique that replaces an identifier for a data principal with a pseudonym, citing SP 800-188 and ISO/IEC 20889. Source
  • Presidio's hash operator uses a random salt by default since version 2.2.361; consistent hashes require a supplied salt, which should be discarded after processing or kept in secure storage, be at least 128 bits and never appear in output. Source
  • Presidio's pseudonymization sample keeps a shared value-to-placeholder mapping and warns the logic is not thread-safe and may give incorrect results when run concurrently. Source
  • Presidio's documentation says automated detection may not find all sensitive information and additional systems and protections should be employed. Source
  • The Presidio Image Redactor detects and redacts PII text in images using OCR, and is described as beta. Source
  • EDPB Guidelines 01/2025 say additional information such as pseudonym mapping tables or keys must be kept separately and protected, and pseudonymized data remains personal data where it can be attributed using such information. Source
  • HHS guidance on the HIPAA de-identification standard says a re-identification code should not be derived from or related to information about the individual. Source
  • The ICO's anonymization guidance uses the motivated intruder test: a reasonably competent person with access to resources such as the internet and public documents. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies