Comparison · Comparison
Manual or automated anonymisation: which to use and when
Anonimatum Team
In short
Manual anonymisation works for a handful of simple documents; automated tools bring speed and consistency at volume. Neither removes the need for human review, so the most effective approach is usually hybrid: the tool proposes, a person decides.
When a team starts anonymising documents, the first question is usually whether it is enough to open the PDF and redact by hand, or whether a tool is worth it. The answer depends on volume, document type and the risk of publishing something that should have been removed.
This comparison does not quote accuracy figures, because they depend on the documents and the tool. It focuses on what can be said with confidence: how each method goes wrong, what it takes to stay consistent and why a person has to review the result either way.
What we mean by manual and automated
Manual means a person reads the document and marks each item: with the redaction tool in a PDF editor, with tracked changes in Word or, at worst, with black rectangles placed on top that leave the text underneath. Quality depends on the attention of whoever does it and on the editor actually removing the content.
Automated means rules and models propose what to hide: patterns (National Insurance numbers, IBANs, phone numbers, emails), AI entity recognition (names, addresses, organisations, dates) and, for scans, OCR first. A good tool does not apply changes blindly: it shows a proposal you can correct before the redaction becomes irreversible.
Side-by-side comparison
| Criterion | Manual | Automated (with review) |
|---|---|---|
| Time per document | Grows with length: every page has to be read | Detection is quick; the time goes into reviewing the proposal |
| Consistency | Depends on the person and the day; name variants slip through | Same rules across all documents; consistent pseudonyms |
| Typical error | Misses due to fatigue; redactions that do not delete text | Misses on unusual cases and false positives that hide too much |
| Scanned documents | You can see the data but search cannot find it | Needs OCR; unreadable parts are covered with drawn areas |
| Context and nuance | Strong: a person knows who is acting in an official role | Needs an allow list and human judgement for exceptions |
| Audit trail | Decisions must be documented by hand | Can record what was detected, who reviewed it and when |
| Scaling to hundreds of files | Needs more people and more hours | Batch processing; review is still human |
| Human review | Required (second reviewer) | Required (approve the proposal before applying) |
Two kinds of error: misses and false positives
It helps to separate errors because they do not carry the same weight. A miss (personal data left visible) can become a personal data breach if the document is published. A false positive (hiding something that was not personal data, such as a street name in the title of an Act or the organisation’s own name) makes the document less useful, but rarely harms anyone.
- In manual work, misses cluster at the end of long documents, in appendices, tables and footnotes, and in variants of the same name (“Mrs Jones”, “the appellant”, “A.J.”).
- Another classic manual failure is cosmetic redaction: a black box over text that is still there and can be copied. We cover this in common mistakes when redacting PDFs by hand.
- In automated work, misses tend to come from unexpected formats (an ID number split across two lines, an unusual foreign name) or poor-quality scans.
- Automated false positives are reduced with an allow list of terms that must stay visible and corrected during review.
When manual is good enough
Manual redaction is reasonable when volume is low and documents are simple. Even then, use a tool that really removes content, such as one that lets you redact a PDF online irreversibly, and run the search and copy tests afterwards. The ICO’s guidance on disclosing documents securely is a useful reference for those checks.
- A few short documents a month with selectable text.
- Few data items per document, easy to find (a name and an ID number in the header).
- No special category data and no open publication, or a second reviewer available.
- Exceptional cases that need fine judgement no rule captures well.
A hybrid workflow that works
For medium or high volumes, or for documents that will be published, combining both methods is usually most effective. The tool does the repetitive work and the person brings judgement. That fits what data protection regulators expect: a documented, repeatable and verified process.
Hybrid workflow steps
- Define rules per document type: what is removed, what is generalised and what stays visible.
- Run OCR on scanned documents before detection.
- Run automated detection with patterns, AI and a prepared allow list.
- Review the proposal: add what is missing, remove false positives and draw areas over signatures, stamps or handwriting.
- Apply irreversible redaction or consistent pseudonymisation and strip metadata.
- Verify the output (search, copy, file name) and log decisions; our GDPR anonymisation checklist includes a template.
Anonimatum is built for that workflow: it detects with its own AI on EU servers and with configurable patterns, supports allow lists, areas and shapes, runs OCR, processes batches and shows a preview so a person can review before applying. Try it with the PDF anonymiser.
Frequently asked questions
Is automated anonymisation enough without review?
It is not advisable. Any detector can miss an unexpected format or flag too much. Human review before redaction is applied is the practical safeguard that the result is right.
Is manual redaction safer?
Not necessarily. People are good with context but tire and become inconsistent on long documents, and it is easy to use a method that covers text without deleting it.
How many documents before automation pays off?
There is no universal number. Length, data density, scans and whether documents are published all matter. If manual redaction takes hours every week or consistency errors appear, it is time to look at tools.
What about signatures and stamps?
Both methods need care. Manually, draw an area over them; automated tools may detect stamps and digital signatures, but scanned handwritten signatures usually still need a manually drawn area.
Spending hours redacting by hand?
See the plans and start combining automated detection with human review on your own documents.