Unstructured data masking+ Synthetic documents
PDFs, scans, mailboxes, chats and notes hold the personal data that table masking never sees. Synthesized brings them under the same policy as your tables: found, replaced consistently and logged, or generated as synthetic documents from the start, so test copies, AI training sets and data rooms are clean end to end.
On the Synthesized platform, in production at banks, insurers and health companies. Runs in your cloud, on-premises or air-gapped.
Mr Hale called on 3 October to report water damage from the flat above. Refund to the account ending 8830. Send updates to jordan.hale@example.net.
Ms Lopez called on 3 October to report water damage from the flat above. Refund to the account ending 4417. Send updates to maria.lopez@example.com.
Documents and text use the same policies, the same engine and the same deployment as these results.
Published results from database and synthetic data work on the Synthesized platform.
Start with your team
What works today
Two ways it's delivered. Available: in the product today, self-serve. Engineer-led: runs on the same platform, set up and run by our engineers inside your environment.
| Format | Masking | Synthetic generation | How it's delivered |
|---|---|---|---|
| PDF, DOCX and PPTX documents | Yes | Yes, from templates | Engineer-led |
| ZIP archives of documents | Yes | — | Engineer-led |
| Scanned documents | Yes, after OCR | — | Engineer-led |
| XML | Engineer-led | Yes, schema-valid XML from an XSD | Engineer-led |
| Mailboxes with attachments, Slack exports | Yes, delivered with our engineers | Engineer-led | Engineer-led |
| XLSX, CSV, HTML, plain text and calendar files | Yes, delivered with our engineers | CSV: yes, with keys kept consistent across files | Engineer-led |
| Text inside images (PNG, JPEG, and images embedded in documents) | Yes, read with OCR and masked in place | — | Engineer-led |
| Free-text database fields, live chat and tickets, call transcripts, JSON and logs | Engineer-led | Engineer-led | Engineer-led |
Mixed archives de-identified in a single pass: Slack exports, mailboxes with attachments, spreadsheets, Word and PowerPoint files, text and scanned PDFs, HTML, calendar files and images, with spelling variants of each name resolved to one replacement across every file.
By format
By use case
How it works
How it compares
When personal data turns up in files, teams usually reach for one of three things. Each does part of the job.
| What you need | SynthesizedUnstructured data masking | Redaction tools | Data-security scanners | In-house scripts |
|---|---|---|---|---|
| Finds personal data in files and text | Patterns, context and local models, checked against a labelled sample of your files | Often page by page, with a person reviewing | Built to find and report it | Only the patterns someone wrote |
| Replaces it consistently | One realistic replacement per person, in every file of a run | Blacks values out, so nothing is replaced | Reports findings and leaves the files as they are | Only with a mapping someone builds and maintains |
| Keeps documents usable | Layout, formats and relationships kept, so parsers, tests and models still work | Black boxes break extraction, search and tests | The files still contain the data | Formats and layouts often break |
| Runs in your environment | Your cloud, on-premises or air-gapped | Varies; many are hosted or desktop tools | Varies by product | Yes, and you own every fix |
| Evidence for every run | What was found, how it was treated and what was checked, per run | A redaction log per document | Findings reports, not proof of treatment | Whatever someone remembers to log |
Yes Partly or it depends NoCategories, not named products. Ask any tool the same five questions.
Where it runs
Extraction, detection and reconstruction run where you configure them. Nothing about this needs your files to visit our cloud.
Consistency
Tests break when a ticket names a customer the system has never heard of. So each person gets one replacement, used in every document, email and note in the run.
Files can be grouped so they share replacements across a whole archive.
Built for regulated teams
The answers, before they ask.
Free resources
Unstructured data masking finds the personal and confidential data inside documents, emails, chats, call transcripts and free text and replaces it consistently, under the same policy you use for your databases, so the files stay usable for testing, AI and sharing. Unlike column masking, the values have to be found first, because they can sit anywhere in the text. When no real file should be used at all, Synthesized generates synthetic documents instead.
PDF, Word and PowerPoint files including scans, ZIP archives, mailboxes with attachments, Slack exports, spreadsheets, HTML, calendar files and the text inside images. Masking database tables and generating synthetic CSV sets are self-serve in the product; document support is delivered engineer-led on the same platform.
Extraction, detection and reconstruction run in the processing environment you configure. When Governor is used, real values are replaced with tokens before the transformation table reaches it, and the token-to-value mapping stays local.
Within a run, yes: each detected value gets one replacement everywhere it appears, and files can be grouped to share replacements across an archive. Nicknames and first-name-only references are worth checking with a labelled sample.
Redaction tools black values out, which breaks documents for testing and AI. Data-security scanners find and report personal data but leave the files unchanged. Synthesized finds the data, replaces it with consistent, realistic values under the same policy as your databases, keeps the documents usable, runs inside your environment and records what it did on every run.
Those tools detect personal data in the text you send them. Synthesized is built for test data: it keeps documents usable, sits alongside the databases you already mask, generates synthetic documents, and checks output against labelled samples. Many teams will use a detector and a test data platform together.
No. Masked values are replaced in the file itself, not covered over. You choose the output: image-based PDFs with no text layer at all, or searchable PDFs whose text layer carries the replacement values.
One document type, inside your environment. A masked or synthetic set, an entity report and a readout, for a fixed fee credited when you go ahead.