Skip to content
USE CASE: HEALTHCARE & MEDTECH

PHI-aware AI protection

Redact PHI before it reaches your model, and judge every reply against the clinical-advice rules your counsel sets.

Start for FreeView documentation
WHAT IS REDACTED, AND WHAT IS KEPTschematic
detected and redactednhs_numberMOD-11 CHECKSUMmedicareFORMAT + SUFFIXau_medicareCHECKSUMmedical_licenseFORMATdate_of_birthPATTERNamber is the redaction verdict, not a warningthenthe event that is writtenDECISIONredactedREASONpii_leak · nhs_numberENTITIES3TIMESTAMP2026-08-09T11:04:22ZPROMPT BODYnot writtenzero retention, per project
Two separate questions, and a hospital asks the second one first. Left: which identifiers the engine can actually name and validate. Right: what is left on disk afterwards, with zero retention on, the prompt body is never written at all.

Where PHI is exposed

  1. 01

    Controls for Your HIPAA Program

    HIPAA is not offered: there is no Business Associate Agreement. For PHI, run PromptGuard Air-gapped so data never leaves your environment, and use its redaction and audit trail as controls in your own HIPAA program.

  2. 02

    PHI Detection & Redaction

    Detects identifiers with a fixed shape, SSNs, dates of birth, phone numbers, email addresses, Medicare and medical licence numbers, IP addresses, and redacts or tokenizes them. Names and free-text details are not caught by pattern.

  3. 03

    Disclaimer Policies You Write

    An output policy can block or flag a reply that gives clinical direction without the qualification your counsel requires. You define the rule; PromptGuard judges the answer against it rather than editing the answer.

  4. 04

    Dangerous Advice Prevention

    An llm_guard output policy can block replies that give dosages, diagnoses or treatment plans. You write the rule in plain English; an LLM judges each reply against it.

  5. 05

    Agent Tool-Call Checks

    Validate an AI agent's tool calls before they run. Calls that are dangerous are refused, and calls that need a person come back marked for human review.

  6. 06

    Audit Trail

    Every scan is recorded with its decision and threat type, and a masked preview unless zero retention is on. Export it as evidence for your own audits.

Why PromptGuard for healthcare?

  • PHI redaction before the model sees it
  • Output policies for clinical advice
  • Masked previews, or zero retention
  • Agent tool-call checks
  • Air-gapped: PHI never leaves your environment

How it works for healthcare

  1. 01

    Deploy

    Integrate PromptGuard with your healthcare AI applications. Works with patient portals, clinical assistants, and admin tools.

  2. 02

    Protect

    Choose the PII types to detect and redact on the project, and add output policies for clinical advice.

  3. 03

    Review

    Use the audit trail and compliance reports as evidence in your own HIPAA program.

Securing healthcare AI applications

python
from promptguard import GuardClient

guard = GuardClient()  # reads PROMPTGUARD_API_KEY

# What is detected, and whether it is redacted or blocked, is set on the
# project: the PII types, the strictness level, and an llm_guard output
# policy for clinical advice. There is no healthcare switch in code.
messages = [{"role": "user", "content": patient_question}]
decision = guard.scan(messages, direction="input")

if decision.blocked:
    raise PermissionError(f"Blocked: {decision.threat_type}")

# A redacted request comes back with the values replaced
messages = decision.redacted_messages or messages

reply = call_your_model(messages)

# Judge the model's answer against your output policies before showing it
verdict = guard.scan([{"role": "assistant", "content": reply}], direction="output")
if verdict.blocked:
    reply = "Please speak to your clinician about this."

The controls this sector runs on

Secure your healthcare AI

Deploy AI in healthcare and medtech with PHI redaction and an audit trail.

Get started freeContact sales