Ask KJ · Healthcare cybersecurity
Using AI Without Exposing Patient Information
Evaluate AI scribes, transcription, chat, and other tools before using them with PHI.

By Kimm C. King Jr. · Reviewed
Start with the exact data flow
Ask what the product receives, including audio, prompts, attachments, drafts, error logs, screen capture, and support cases. Review permitted uses, model training defaults, storage, retention, subcontractors, and ability to export or delete. Determine whether a BAA is needed for the use case and verify scope before allowing PHI.
Pilot in a safe environment
Use synthetic records or data properly de-identified under an approved process. Give staff individual access and a review step before generated content enters a patient chart or communication. Compare outputs against the approved source, record inaccuracies, and define what to do when the system is unavailable.
Example involving ambient documentation
Have the privacy and clinical leads review when audio capture starts, how patients are informed under applicable requirements, which provider sees the draft, and how corrections are saved. Verify that turning off an option actually changes the configured workflow across staff accounts.
Acceptance evidence
Keep the vendor review, approved use case, configuration evidence, reviewer training, and pilot results. Agree on an offboarding process so the practice can remove access, retrieve authorized records, and manage downstream copies.
My answers to related practice questions
Official sources
Apply this to your practice
Confirm requirements for your organization, services, and jurisdiction with your privacy or legal lead. Use fictional or approved de-identified examples when testing. Keep PHI out of public forms.
Find healthcare resources by state or city · All healthcare cybersecurity guides
Explore cybersecurity consulting · Read the website’s healthcare data notice