Clinics redact all the time: records requests, referrals, insurance appeals, case studies, a chart excerpt for a lawyer. Most of the tools that would speed this up are cloud services, which means sending protected health information to another company and signing a business associate agreement first. Veil sidesteps that by running on the computer. The question is whether its detection is good enough to rely on.
What HIPAA mode changes
HIPAA’s Safe Harbor method lists 18 identifiers that have to come out for health information to count as de-identified: names, small geographic areas, most dates, phone and fax numbers, email addresses, record and account numbers and so on. Veil’s HIPAA mode tunes its detector to that list. It switches off categories that aren’t on it, like occupations and companies, so a chart doesn’t end up with “nurse” and “Medicare” blacked out.
The most useful detail is what it leaves alone. In every mode, Veil never redacts clinical codes and values: ICD-10 and HCPCS codes, dental codes, form numbers, doses and vitals. A redacted record stays clinically readable.
Medical documents, RedactionBench (F1, higher is better)
Source: Veil developer's published benchmarks. Only 16 medical documents, so treat small differences as noise.
Those are good numbers, and switching on HIPAA mode clearly helps. But an F1 in the mid-0.80s means some identifiers will get through on any given batch of records. With only 16 test documents, nobody should read much into the gap between Veil and OpenAI’s filter.
Gaps to watch for
Dates. Safe Harbor requires removing every date tied to a patient except the year, including admission and discharge dates. HIPAA mode keeps birth dates strict but only redacts other plain dates when it’s confident. Check every date by hand, or add a rule for your date format.
The “actual knowledge” test. Safe Harbor also requires that you don’t know the remaining information could identify the patient. A rare diagnosis in a small town can do that with every name removed. No software can make that judgment for you.
Scans in other scripts. Veil’s text recognition reads Latin script only, so a name written in Cyrillic, Chinese or Arabic on a scanned form may be missed.
The AI assistant. By default it sends redacted text to your AI provider, which is the safe choice. But if someone allows the full text, patient information goes to that provider, and most consumer AI plans don’t come with a business associate agreement. A clinic should make that rule explicit.
What makes it a good fit anyway
Because Veil processes everything on the computer, using it doesn’t put patient data in a vendor’s hands, which removes a whole category of paperwork and risk. Batch redaction handles a stack of records at once. The redaction certificate gives you a record of what was removed, by category and page, without storing the removed details. And the way it exports, rebuilding redacted pages from images, means nothing is left hiding under a black box.
Verdict
For a small practice, Veil with HIPAA mode and the Accurate engine is a sensible way to cut redaction time while keeping records in-house. Treat it as the first pass, not the last word: have a trained person check every page, especially dates, before anything is released. For formal de-identification of datasets, talk to your privacy officer about Safe Harbor or an expert determination.