How We Keep the AI Honest
Language models can write confident sentences that are not true. In a medical record that is unacceptable, so Ayah is built to leave the model very little room: a fixed structure, a short list of things it may say, and a physician who verifies the result.
A Fixed Structure
Every note fills a set template for its note type and specialty, with the same sections every time. The AI returns each section as structured data, not free text, so it has no room to invent new ones.
Checked Before You See It
The server checks the AI’s answer against that structure before the note exists. A malformed answer is rejected and asked for again, a failed attempt never changes your note, and an ICD-10 code that is not valid is dropped.
Only What Was Said
Scribe may use only what was said in the visit, in the patient’s intake or in prior notes. It may not add vitals, exam findings, results, diagnoses, medications, doses or plans you did not state, and it writes an exam line only for a system you examined. Audio with no discernible speech produces no text, not a guess.
Suggestions Show Their Basis
A code is a suggestion until you accept it. A related diagnosis is added only when your note documents the finding behind it, and the card quotes that finding. The differential gives a reason for each alternative, and plan suggestions are yours to accept or dismiss. The differential and any working diagnosis come only from the history, exam and results you stated.
Intake Follows a Checklist
The intake chat works through a checklist of questions your clinic can edit, not an open-ended conversation. The AI proposes what a reply answered, and the server accepts it only with the patient’s own words or a tapped choice. The server decides when intake is complete, and the read-back summary is written from recorded answers, not by a model.
Fixed Rules for Emergencies
Emergency wording is read by fixed rules, not left to the AI’s judgment. A patient who describes an emergency happening now sees an immediate call 911 message, self-harm gets the 988 line, and the chart is flagged for your team.
A Physician Verifies Everything
The AI never signs or completes anything. A note stays a draft until you read it, correct it and sign off, and codes it added and exam lines it filled in stay highlighted until you have reviewed them. These safeguards are designed to reduce made-up content, not to replace your judgment. The note is yours, and so is the final check.
Built to Protect Patient Information
Security is part of how every workflow is designed, from the first intake question to the signed note.
- HIPAA-Aligned Design
Built around HIPAA’s safeguards for protected health information.
- BAA Available
A Business Associate Agreement is available for every practice.
- Two-Step Sign-In
Required for every account, with an authenticator app and one-time recovery codes.
- Passkeys
Sign in with Face ID or another passkey instead of typing a password.
- One Account, Every Clinic Kept Separate
Clinicians sign in once with an Ayah Account. Each clinic’s patients and data stay walled off from every other clinic.
- Consent Before Recording
Scribe records only after the clinician confirms everyone agreed.
- AI Processing in the United States
Patient data sent to the AI is processed in the United States.
- Not Used to Train General Models
We do not use patient information to train a general AI model.
Ask Us About the Safeguards and the BAA
We will walk you through the safeguards, the security design and the Business Associate Agreement for your practice.