No automated clinical advice for patients
Patient-facing screens never give automated clinical advice. Symptom questions go to the clinic's care team, and emergencies to 911.
AI boundaries
AI reads and drafts. People decide.
AI in OvaFlow is designed to help with reading and drafting, always grounded in a source and always reviewed by a person. “Uncertain” is an acceptable answer.
Grounded, or empty
Every field points back to its source
Every AI result can come back empty or uncertain, with a reason. A missing field is shown as missing.
Signed source · trigger instruction, version 3
Ovidrel 250 micrograms subcutaneous, once, at 20:00 PDT on Oct 30. Signed by Dr. Mira Lee at 17:10 PDT.
- Medication and dose
- Ovidrel 250 microgramsFrom the signed source
- Route and frequency
- Subcutaneous, onceFrom the signed source
- Planned time
- 20:00 PDT, Oct 30From the signed source
- Signed by
- Dr. Mira Lee, 17:10 PDTFrom the signed source
- Retrieval time
- Not in this documentShown as missing, not guessed
Scope
What AI may do, and what it never does
Permitted
- Extracting fields from a source document, with a link back to where each came from
- Comparing two versions to show what changed
- Summaries grounded in the source
- Administrative drafts for a person to review
- Suggesting a task category
Excluded
- Diagnosis or triage
- Choosing or changing doses
- Ranking embryos
- Predicting the chance of success
- Inventing a dose, unit, time, signature, identity, acceptance or confirmation
Design prototype · Fictional patient data
Capture with consent
Speech capture is opt-in
Capture follows the clinic's policy, starts only after consent, is always visible and can be paused at any moment. A draft from speech is reviewed by a clinician before it goes anywhere.
- Ready on the wrist. The microphone stays off until a visit is chosen.
- The visit and consent are confirmed on the paired phone before capture starts.
- A visible recording state with elapsed time and a deliberate stop.
- Saved on Watch means saved locally. It is not yet a clinical note.
Patients and model changes
Model changes are governed
A model update waits for the clinic's medical director and stays within approved operational scope. It is not clinical decision support.
How AI output is checked
Gated, grounded, and human-reviewed
OvaFlow's AI boundaries are designed around NIST AI RMF 1.0, WHO AI for Health guidance, US federal clinical decision support transparency principles, and ASRM laboratory guidance [1, 2, 3, 4].
Four checks before use
AI safety checksFrom signed source document to human review in the EMR
- Check 01Signed source inputStarts from a clinician-signed order, note, or lab PDF—never an unverified prompt.
- Check 02Source-linked spanEvery extracted field highlights the exact words in the source document it came from.
- Check 03Abstain when missingIf a value is absent or unclear, it is shown as missing rather than guessed.
- Check 04Human review in EMRA qualified staff member edits and commits the record inside the clinic's own system.
Hard exclusions at every step: No diagnosis · No dosing · No symptom triage · No embryo ranking · No outcome prediction
Published health-AI guidance
How OvaFlow's limits align with clinical AI guidance
NIST AI Risk Management Framework (AI 100-1) · Govern, Map, Measure, Manage
Standard: Define explicit context-of-use boundaries, measure failure modes, and keep human governance over model updates [1].
OvaFlow design: Restricts context to reading/drafting beside the chart, logs abstentions, and gates model changes behind clinic medical-director approval.
Never deploys autonomous background model changes into clinical workflows.
WHO Guidance on Ethics & Governance of AI for Health (2021 / 2024 LMM)
Standard: Protect human autonomy, ensure transparency and explainability, and preserve clinician accountability [2].
OvaFlow design: Every extracted value highlights its exact source span; visit capture is strictly opt-in with visible consent status.
Patient-facing screens never deliver automated clinical advice or symptom triage.
US Federal Clinical Decision Support (CDS) Transparency Principles (2022)
Standard: Software supporting clinicians should enable independent review of the underlying basis so clinicians do not rely blindly on opaque outputs [3].
OvaFlow design: Displays the signed source document side-by-side with any extracted field or version comparison.
Never generates ungrounded recommendations or hidden risk scores.
ASRM Practice Committee — AI in the IVF Laboratory (2026)
Standard: Highlights validation, generalizability and oversight requirements for embryo evaluation and laboratory AI [4].
OvaFlow design: Deliberately excludes embryo morphology ranking, ploidy prediction and live-birth probability calculators.
Embryology grading and selection stay 100% with embryologists and lab systems.
See it
Watch it hold in a walkthrough
Recordings of the interactive design prototype, with fictional patients and staff.
Cryo custody, donor updates, outages and model governance
Shipping clearance isn't clinical clearance, a bounced email isn't abandonment, and a model update waits for the medical director.
nAbleIVF: consult capture, pause, save check and a nursing task
Consult notes captured with consent, paused for private moments, reviewed, then checked as saved rather than assumed.
Care conversations: audience change, duplicate send, clinical routing
When someone new joins the thread just before a clinical update, the send pauses. Duplicate reminders are caught, and dose questions go to a nurse instead of an automated reply.
AI boundaries
Ask us what the AI will not do.
We will show every place AI appears in the prototype, and where it is deliberately absent.