Healthcare
Use AI on Clinical Text Without Sending PHI to the Model
- 01
Identify sensitive values
- 02
Replace with stable tokens
- 03
Restore only inside your boundary
574
Built-in detectors for personal data and secrets
47
Country packs, checksum-validated where the format has one
9
AI providers behind one gateway, plus the desktop agent and proxy
Where PHI meets AI in a care setting
Clinicians summarize discharge notes and draft referral letters with AI. Patient-communication teams draft replies to portal messages. Developers build retrieval over records and guidelines. Each of those requests can carry a name, a date of birth, a record number, and a diagnosis.
Health data is special-category data under the GDPR and PHI under HIPAA. Sending it to a general-purpose AI provider is the exposure most care organizations try to prevent with a ban, which mostly moves the work to personal accounts.
Maskflare puts one rule set in front of every AI provider: identifiers become tokens on the way out and real values on the way back, with nothing kept of the clinical text.
Where healthcare is actually exposed
5 failure modes specific to this environment, not a restatement of general security advice.
01
Clinical notes pasted into a chatbot
A discharge note pasted for summarization carries the patient's name, date of birth, and record number. The desktop agent and browser protection mask them before the prompt reaches the provider, and put them back where the summary repeats them.
02
Clinical assistants built on third-party models
Assistants built on OpenAI, Anthropic, or Gemini send every note to the provider. Through the gateway, identifiers leave as tokens and the answer comes back readable.
03
Identifiers with no universal format
Medical record, member, and license numbers vary by organization. They are detected after their label, and your own patient identifiers can be matched exactly with Exact Data Match.
04
Retrieval over patient records
Embedding records for retrieval puts identifiers into a vector store. Embeddings requests get stable stand-ins, so retrieval still works on masked data.
05
Documents and images
Referral letters and lab reports arrive as PDFs and Office files. Document scanning reads their text; metadata stripping removes author and location data. Text in images is read only when image scanning is on.
Who this page is for
- CISO / Head of Security
- Needs AI use on clinical text to stop being an uncontrolled route for PHI.
- Data protection officer
- Has to show what health data reaches each AI provider and on what legal basis.
- Chief medical information officer
- Wants clinicians to get AI assistance without a privacy incident.
- Clinical AI and engineering
- Needs one integration for every model, with identifiers handled the same way everywhere.
- Compliance
- Has to map the control to HIPAA Safe Harbor identifiers and GDPR obligations.
How a team would actually run this
Summarizing clinical notes with AI
- 01 Turn on the healthcare pack, person names, and dates of birth, plus the country packs for your patients' national identifiers. The HIPAA Safe Harbor profile does this in one click.
- 02 Route the summarization tool's requests through the Maskflare gateway, or roll out the desktop agent to clinicians' machines with Intune, Jamf, or Group Policy.
- 03 Names, dates of birth, and record numbers become tokens; the summary comes back with the real values.
- 04 The traffic log shows how many identifiers each rule masked, with no clinical text stored.
Building retrieval over records
- 01 Send embeddings requests through the gateway: identifiers are replaced with stable stand-ins before the text is embedded.
- 02 Queries are masked the same way, so they still match the stored vectors.
- 03 If masking can't complete, the embeddings request fails closed instead of going out unmasked.
- 04 The vector store never holds a patient identifier in clear text.
Which modules apply, and why
Each links to the module page, where the boundaries and development status are set out in full. Or start at the Maskflare platform overview.
Detection engine
Maskflare Detection Engine
Detects and masks patient names, dates of birth, record and member numbers, and health identifiers before a prompt reaches an AI provider, and can restore them in the response.
For your apps
Maskflare AI Gateway
Puts clinical assistants and retrieval pipelines behind one set of rules, with embeddings masked deterministically and failing closed.
For your people
Maskflare Workforce
The desktop agent covers clinicians using AI chat on the web, deployed through device management and fail-closed by default.
Detection engine
Maskflare Exact Data Match
Matches your own patient and record numbers from keyed hashes, without sending the clear values to Maskflare.
Detection engine
Maskflare File Security
Reads the text of attached referral letters and lab reports and strips EXIF and GPS metadata from images.
Regulatory context
Readiness and relevance, not certification. Nothing here is a claim to hold an audit or authorization we do not have.
- HIPAA (US)
- Maskflare's healthcare and core detectors map to the HIPAA Safe Harbor identifiers it can detect in text, and the compliance coverage report exports that mapping as evidence. Masking is a technical safeguard, not de-identification on its own. Raise BAA requirements during your demo.
- GDPR Art. 9 (EU)
- Health data is special-category data. Masking limits what reaches an AI provider; reversible tokens remain pseudonymized personal data for you.
- SOC 2 / ISO 27001
- SOC 2 Type II and ISO 27001 audits in progress. The Trust Center states current status.
What Maskflare does not do
- No product makes a workflow HIPAA or GDPR compliant on its own. Maskflare is a technical safeguard you would cite.
- Person-name detection covers English, German, French, Spanish, Italian, and Albanian names. Recall on names is lower than on checksum-validated identifiers.
- Text inside images is read only when image scanning is on (classic OCR, best effort, API requests only); scanned PDFs aren't read.
- Maskflare is not a medical device and doesn't interpret clinical content.
4 questions to ask any vendor here
Including us. If our answer is worse than someone else's, you should know that before you buy, not after.
- 01Which HIPAA Safe Harbor identifiers can the tool detect in free text, and which can't it?
- 02Does the AI provider ever receive a patient name or record number, or only tokens?
- 03Can it match our own patient identifiers without receiving them in clear text?
- 04Is any clinical text stored, and where is the service hosted?
Workflow guides for this industry
How a team applies each of these, including the contrast with the tool it replaces.
Healthcare questions we get asked
Which health identifiers does Maskflare detect?
NPI and DEA numbers, Medicare MBI, NHS numbers, the German KVNR, and medical record, health plan member, license, and EHIC numbers after their label, plus names, dates of birth, phone numbers, emails, addresses, and national IDs from the core and country packs.
Is masking the same as de-identification under HIPAA?
Not on its own. Masking removes the identifiers Maskflare detects from what the AI provider receives, which supports a Safe Harbor approach, but de-identification is a determination about the whole dataset and its use.
Can clinicians keep using ChatGPT?
Yes. With the desktop agent installed, text typed into ChatGPT, Claude, Gemini, DeepSeek, and Perplexity on the web is masked with the same rules, and the real values come back where the answer repeats them.
Where is Maskflare hosted, and does it keep clinical text?
Maskflare is hosted in the EU by default, with on-premises deployment available for enterprise. Prompts and responses are never stored; masked values are kept encrypted for 30 days so responses can be restored.
See it on your own data
Start a free trial or talk to us with the identifiers and AI workflows your healthcare team actually handles. Bring the case that makes it hard, because that is the part worth testing.
See it on your own data
Try Maskflare free.
Bring your hardest prompts.
28-day free trial or 10,000 requests, whichever comes first. Point one app at the gateway or enroll a few laptops,
and watch what gets masked. Or talk to us first.
- SOC 2 Type II: In progress
- ISO 27001: In progress
- EU hosting by default · on-premises available
- Prompts and files never stored
- Never used to train AI models
- GDPR · nDSG · DPA available
- Trust center →