How to choose document redaction software: criteria and questions for vendors
Choosing a tool to redact or anonymise documents looks like an IT purchase, but the consequences are legal. If the software leaves text under a black box, sends documents to an AI service in another jurisdiction or cannot show who redacted what, the problem is not the vendor's: it belongs to the controller who published or shared the file.
This guide sets out the criteria worth checking, the specific questions to ask any vendor and a simple way to test several tools on your own document types before you commit.
Start with what you need to solve
Before looking at products, describe your case with three facts: which documents (decisions, contracts, medical records, payslips, emails), what volume and for what purpose (publishing, answering subject access requests, FOI disclosures, sharing with a third party, testing systems). Redacting ten judgments a month is not the same as processing thousands of scanned files a year, and publishing openly is not the same as handing files to an auditor.
Decide too whether you need anonymisation (the data is gone and, if nobody can reasonably re-identify the person, the output falls outside the UK GDPR and EU GDPR) or pseudonymisation (the data is replaced by a consistent code and a key exists to reverse it, so it is still personal data). Many organisations need both, depending on where the document is going. If you are weighing manual against automated work, our comparison of manual and automated anonymisation helps size the effort.
The criteria that matter
| Criterion | What to ask | Red flag |
|---|---|---|
| Where data is processed | Which country are the servers in? Which sub-processors? Is text sent to external large language model APIs? | Vague answers; international transfers without appropriate safeguards |
| Formats | Native and scanned PDF, Word, Excel, PowerPoint, images, email with attachments? Does output keep the format? | Everything flattened to images or PDFs without text |
| OCR | Does it read text in scans and photos? Is the output searchable? | Scans treated as pictures where nothing is detected |
| Detection | AI plus patterns (NI numbers, SSNs, bank details, phone), custom patterns and allow lists? | Fixed keyword lists only, or AI with no user control |
| Human review | Can I see and correct detections before applying? | Blind processing without preview |
| Irreversibility | Is the text removed from the file? Are metadata, comments and layers cleaned? | Overlay boxes with text still underneath |
| Pseudonymisation | Consistent substitutes across documents? Where and how is the key stored? | Different substitutes for the same person or an unencrypted key |
| Batch and API | Server-side processing of folders or ZIPs? Documented API and webhooks? | One file at a time, no integration |
| Audit trail | Per-user, per-operation log? Does the log itself contain personal data? | No log, or logs that copy document content |
| Contract and certifications | Article 28 processor terms? ISO/IEC 27001, SOC 2 report, Cyber Essentials? What is the scope? | Certificates with no verifiable scope or not covering the service |
Where the documents are processed
This is the question that separates tools most clearly. Some send text to a third party's large language model to detect personal data; others run their own models. What matters under data protection law is who processes the data, where, and under what safeguards. Our guide to local AI versus cloud AI goes into the trade-offs.
Certifications: ask for the certificate and its scope
ISO/IEC 27001 is the most widely recognised information security certification. US buyers often ask for a SOC 2 report, and UK public bodies may require Cyber Essentials for some contracts. Whatever the scheme, ask for the certificate or report, check who issued it and, above all, its scope: a certificate for a company's head office that does not cover the service you are buying is of little use.
Running a trial that actually informs the decision
Sales demos use clean documents. Your reality is skewed scans, tables, stamps and annexes. Build a test set from your own document types with fictitious data, or from documents already published, and measure:
- Missed data (false negatives): information that should be gone and is still there. This is the serious error; count it per document.
- Over-redaction (false positives): text removed unnecessarily, such as public officials' names, legislation or figures that must stay visible.
- Review time: how long a person takes to check and correct a typical document.
- The output itself: open it, try to copy the redacted text, check the properties and hidden sheets.
- Scans: check that the tool finds data in a scanned PDF and in a phone photo of a document.
Legal questions you should not skip
- Will the vendor sign processor terms covering Article 28 of the UK GDPR / EU GDPR?
- How long are documents and outputs kept? Can I set automatic deletion?
- Are my documents used to train models? Does the contract say so?
- Which sub-processors are involved and where are they located?
- How and how quickly will a personal data breach be notified to me?
In the US, if you handle protected health information, also ask whether the vendor will sign a business associate agreement and how its output supports HIPAA de-identification (Safe Harbor or Expert Determination).
How Anonimatum measures up
Anonimatum is one option on the market and should be assessed against the same table. It is a SaaS service from Politeia Soft, a Spanish company, with its own AI running on servers in the European Union and no text sent to external language model APIs. It handles native and scanned PDFs with OCR, Word, Excel (including column by column), PowerPoint, images, multi-page TIFF and .eml and .msg emails with attachments, and returns the same format. It combines AI with patterns, custom patterns and an allow list; shows a preview for review before applying; removes text and metadata; and supports consistent pseudonymisation with encrypted maps.
It also offers server-side batch processing, a REST API with signed webhooks, templates per document type, case files with version history and a group panel with an audit log that never stores document names or content, plus configurable retention. Politeia Soft holds Spain's ENS (National Security Framework) certification at MEDIUM level, with a scope that includes Anonimatum; ask us, as you would ask any vendor, for the certificate and scope you need. What it does not do matters too: it does not detect faces; images are converted to PDF and handled with OCR, patterns, AI and user-drawn zones or shapes. Plans and page allowances are on the pricing page, and the security page has the technical detail.
Frequently asked questions
Does redaction software make me GDPR compliant?
Not on its own. It is a technical measure that needs rules on what to remove, human review and a record of decisions.
Is a free online PDF tool good enough?
For documents with personal data you need to know where files are processed, how long they are kept and whether redaction is real. Many free tools do not say.
What is the difference between redaction and anonymisation?
Redaction removes specific content from a document. Anonymisation is the outcome where nobody can reasonably identify the person, which may require removing more than names.