To anonymize company data, inventory every identifier, replace direct identifiers with consistent pseudonyms, hash key columns with a secret salt, scrub free text, generalize quasi-identifiers, measure re-identification risk, scan for secrets and document the method. Automated tools give a first pass. Manual review and legal sign-off are still needed.
What anonymizing means in practice
In everyday use, anonymizing means changing records so they no longer point to a particular person or client. In law, the words anonymized, de-identified and pseudonymized each have their own test, and the tests differ by regime. This guide covers the practical work. It does not tell you which legal standard your output meets. That is a question for counsel.
The work has a fixed order: find the identifiers, treat the direct ones, treat the indirect ones, measure what risk remains, and write down what was done. Skipping the first step is the usual cause of failure. Identifiers cannot be removed from fields nobody knew were there.
Work on a copy, never on the source system. Keep the copy inside the company until the process is finished. Each of the tools mentioned below runs in the browser.
Decide the goal before starting. The aim is a dataset that is still useful for its purpose with the identifying detail gone. Removing too little leaves risk. Removing too much leaves records that teach nothing. Each choice below is a trade between the two.
Step 1: Inventory the identifiers
List every field in every table or file and say what it contains. Do not trust column names. A field called notes may hold phone numbers, and a field called reference may hold a national ID. Look at real values.
The profiler at /tools/dataset-profiler reads a CSV, JSON or JSONL file and reports each column's fill rate, distinct count and first value, and marks columns that appear to hold personal data. Columns with a distinct value on almost every row are candidates for identifiers. Free-text columns need separate handling.
Record the result as a table: field, what it holds, identifier class and planned treatment. This inventory becomes the core of the documentation.
Direct and quasi-identifiers
Identifiers come in two kinds, and they need different treatment. A direct identifier points to a person or company without help. A quasi-identifier does not, but several together can. A job title, an office location and a start date may describe exactly one employee.
A third category deserves its own line: client identifiers. Company names, account numbers, domain names and project code names identify a business customer. Privacy law may not protect them, but confidentiality clauses usually do.
| Class | Examples | Usual treatment |
|---|---|---|
| Direct, personal | Name, email, phone, government ID, employee ID, home address | Remove, or replace with a consistent pseudonym |
| Direct, client | Company name, account number, domain, contract number, project code name | Remove, or replace with a consistent pseudonym |
| Quasi-identifier | Job title, department, city, exact dates, age, rare product or condition | Generalize, or suppress where groups are small |
| Technical | IP address, device ID, hostname, session token, URL with parameters | Remove or hash |
| Embedded in free text | Any of the above inside messages, notes and transcripts | Pattern scrubbing followed by manual review |
Step 2: Consistent pseudonyms and column hashing
Deleting identifiers outright destroys structure a buyer needs. If every customer ID is blanked, nobody can tell that five tickets came from the same customer. The answer is a consistent pseudonym. The same input always maps to the same replacement, so relationships survive while the real value does not.
For key columns, the standard method is a salted hash. The tool at /tools/column-hasher applies one to the columns you choose. The salt is a secret value mixed into every hash. Without it, anyone can hash a list of known emails and match the results. Use one salt across all files in a delivery so that joins still work, keep it out of the delivery, and decide whether to destroy it afterward.
A hashed value is a pseudonym, not anonymity. While the salt exists, the mapping can be rebuilt. Under the GDPR, pseudonymized data is still personal data.
Not every identifier should be hashed. Names and email addresses that appear only as attributes, and are not needed to link records, are better dropped or replaced with a label. Hash the keys that hold the dataset together, such as customer, ticket and employee IDs.
Step 3: Scrub free text
Free text is where anonymization most often fails. Messages, ticket bodies, notes and transcripts carry identifiers in any position and any form: a signature block, a quoted email header, a name in a greeting, an order number in the middle of a sentence.
The scrubber at /tools/pii-scrubber finds identifiers such as email addresses, phone numbers, card numbers, Social Security numbers and IP addresses, and replaces them with typed labels. The same value always gets the same label. Add your own patterns for the formats your company uses, such as account and ticket numbers.
Pattern matching does not reliably find names, and it does not understand context. A sentence about the only customer in a small town who runs a bakery contains no pattern and still identifies someone. After the automated pass, read a random sample by hand, record what was missed, add patterns or word lists, and repeat until the miss rate on a fresh sample is acceptable.
Step 4: Generalize, then check re-identification risk
Quasi-identifiers are treated by reducing precision. Replace a birth date with a year or an age band. Replace a city with a region. Replace an exact timestamp with a date, or shift all dates for one person by the same random offset so that intervals are preserved. Group rare job titles into broader families.
Then measure. The checker at /tools/reidentification-risk-checker groups records by the quasi-identifier columns you select and reports the size of each group. A group of one means a record is unique on those fields and is the easiest to link back to a person. Generalize further or drop the outliers, and run the check again.
Group size is a guide, not a legal threshold. It also covers only the structured columns you chose. It says nothing about what is in the free text.
Small populations need particular care. In a company of sixty people, a department and a job title can identify one person however coarse the other fields are. Where the population is small, suppressing the column is often the only reliable treatment.
Step 5: Find secrets in code and configuration
Secrets are a separate problem from personal data, and they are common in code repositories, runbooks and chat. API keys, passwords, tokens, private keys and connection strings give access to live systems. Removing them from the current version of a file is not enough when version history is part of the delivery, because earlier commits still hold them.
The scanner at /tools/secrets-scanner looks for cloud keys, API tokens, private keys, database connection strings and passwords in configuration files. It reads the current files and skips the .git folder, so earlier commits need a separate check. Treat every finding as compromised and rotate the credential, whether or not the data is ever delivered. Then remove it from the dataset.
Step 6: Document the process and state its limits
Write down what was done while it is fresh: the inventory, the treatment of each field, the tools and patterns used, the sample sizes reviewed by hand and what was found, the risk measurements, and who did the work. Buyers ask for this, and several legal standards expect it. When the files are final, the tool at /tools/delivery-manifest lists each file with a checksum, so both sides can confirm that what was approved is what was delivered. Outrich Data records the method in the data card for each dataset it buys.
Be plain about the limits. They apply to these tools and to any others.
- Pattern-based tools are a first pass. They find formats, not meaning.
- Names, nicknames and descriptions of people in free text need manual review or a dedicated model.
- Images, screenshots, audio and attachments are not covered by text scanning.
- A low re-identification score on chosen columns does not cover columns you did not choose.
- Hashing is pseudonymization. It does not make data anonymous.
- No tool decides whether a legal standard is met. Counsel does.
- Anonymization does not remove contract limits on sharing client information.