Can you strip client names out before a document goes to an AI tool?
For small digital businesses that handle client documents: agencies, bookkeepers, consultants. Paste text, swap detected names, IBANs, tax IDs, emails and phone numbers for tokens such as [PARTY_01], check what is left, then put the originals back after the AI step. Everything runs in this browser tab; the page makes no network requests.
BEST-EFFORT DETECTION · NEEDS HUMAN REVIEW. This demo uses simple patterns plus the names you list. It will miss things: names you did not list, addresses, record numbers, unusual date or phone formats, and context clues. A person must read every output before it leaves your machine. Even Microsoft Presidio's README says there is no guarantee it finds all sensitive information (source). Nothing on this page makes a document "compliant" with any law.
What this leaves out: names you did not list, street addresses, record numbers, unusual date or phone formats and context clues; whether sending the masked text is allowed by your client contracts and local law; and the security of the machine where the originals stay.
| Vault Token | Category | Local Original |
|---|
Sources & what we checked
- 45 CFR 164.514 (text via Cornell LII): the HIPAA Safe Harbor method in (b)(2) lists 18 identifier types to remove. Dates related to the individual keep only the year; ages over 89 (and date elements, including year, that indicate such an age) may be aggregated into a single "90 or older" category. Under (c), a re-identification code must not be derived from or related to information about the individual, so use random IDs, not hashes of a patient ID. Safe Harbor also requires that the covered entity has no actual knowledge that the remaining data could identify someone.
- Microsoft Presidio README: states there is no guarantee Presidio will find all sensitive information and recommends additional systems and protections.
- All sample names, emails (.example domains), phone numbers (555 ranges), tax IDs and IBANs on this page are fictional test data.