What an AI scribe has to get right at an IVF monitoring visit

Most ambient scribes were designed around a particular kind of encounter: a patient arrives with a problem, describes it, a clinician examines and reasons out loud, and a plan is agreed. That encounter is mostly language, and language is what these systems are good at. The monitoring visit is not that encounter. It is four minutes long, it is mostly numbers, and almost nothing that matters is said in a full sentence.

That gap is the reason we built an AI medical scribe for fertility clinics rather than adapting a general one. What follows is not a product tour — it is the list of things we found a scribe has to get right before it is safe to put anywhere near a stimulation cycle.

The monitoring visit is a different shape of encounter

A patient in stimulation may be seen every day or two for a week or more. Each visit produces a small, dense set of observations: follicle counts and sizes for each ovary, sometimes an endometrial measurement, sometimes bloodwork, a decision about whether the medication changes, and a decision about when she comes back. The consultation that started the cycle was a conversation. The monitoring visit is a measurement, a judgement and a date.

Three things follow from that.

The record is the output, not the narrative. In a primary-care note the prose is the documentation. Here the prose is nearly worthless: what the clinic needs downstream — for the trigger decision, for the cycle summary, for its registry reporting — is discrete, structured values in named fields. A beautifully written paragraph that says "several follicles were seen bilaterally, the largest around fifteen millimetres" has destroyed the data.

Context comes from outside the room. Almost every number recorded today is only interpretable against yesterday's. "Fourteen, thirteen, eleven" means one thing on cycle day 6 and something else on cycle day 10. A scribe that begins each encounter with a blank page is being asked to document a sentence while being shown one word of it.

Protocol varies, legitimately. It is tempting to build one canonical monitoring template. Clinics differ, and the evidence supports their differing: the 2026 ESHRE update on ovarian stimulation carries 121 recommendations, 17 of them on monitoring, and several are more restrictive than common practice — it reports that adding oestradiol measurements to ultrasound monitoring is "probably not recommended", that a panel of oestradiol, progesterone and LH alongside ultrasound is "probably not recommended", and that routine monitoring of endometrial thickness during stimulation is "probably not recommended"[1]. Plenty of good clinics do all three anyway, for reasons of their own. A scribe's job is to document the protocol the clinic actually runs, not to impose the one it was trained on — and certainly not to leave a field blank because it expected a value the clinic deliberately does not collect.

What a generalist gets wrong

These are the failure modes worth testing for. None of them is exotic; all of them are quiet.

Laterality. "Right ovary: 12, 11, 9. Left: 14, 13." Spoken quickly, with the sonographer half-turned to the screen, this is one of the easiest things in medicine to transpose. A generalist scribe has no reason to treat left and right as structurally different fields, so it writes a sentence, and the sentence is plausible whichever way round it is. Nothing downstream can catch it.

Units and magnitude. Follicles in millimetres, endometrium in millimetres, oestradiol in pmol/L or pg/mL depending on the lab and the country. A model that has learned "1.4" and "14" are both reasonable numbers near the word "follicle" will occasionally produce the wrong one, and a decimal point in the wrong place is the kind of error that reads perfectly well.

Cycle day. Almost never said aloud, because everyone in the room already knows it. It is also the single value that makes the rest of the note interpretable. If the scribe cannot read it from the cycle record, it will either omit it or infer it, and the inference is a guess dressed as an observation.

Trigger timing. "We'll go tomorrow night" is a complete instruction to everyone present and an incomplete one to the record. Which drug, what dose, what time, and therefore what time the retrieval is. This is the sentence in the whole visit that is most costly to get wrong and most casually spoken.

Counts versus measurements. A clinician who measures the four largest follicles and counts the rest has said two different kinds of thing. Flattened into prose they become one list, and the count is lost.

What "cycle-aware" has to mean

The phrase is doing real work, so it is worth saying what it has to mean in practice rather than in a headline.

It knows what cycle this is and where in it we are. The note is written against the cycle record — protocol, start date, cycle day, medication and dose as currently prescribed — not against an empty template. That is a data-integration problem before it is an AI problem.

It knows what this visit type must carry. A monitoring visit has a required shape. If the clinic's protocol says an endometrial measurement is recorded and none was spoken, the right behaviour is to leave the field empty and say so — not to omit the field, and certainly not to fill it from context. The distinction between "not measured" and "not documented" is the whole ballgame, and it is the distinction a fluent language model is worst at preserving.

It refuses to interpolate. If the largest follicle was not stated, there is no largest follicle in the note. Our own scribe traces every sentence in the draft back to the line of transcript it came from, and flags anything the encounter did not support rather than filling the gap. That constraint costs completeness and buys the only thing that matters here, which is that a number in the record was a number someone said.

It carries the decision, not just the observation. Dose change, no change, coast, cancel, trigger. And the next appointment, because a monitoring visit that does not produce a next date has not finished.

The note has to be reviewable in under a minute

This is the requirement that shapes everything else, and it is the one most easily lost.

A clinician doing eight monitoring scans before nine o'clock has, at most, a few seconds per note. If reviewing an AI draft takes longer than typing the numbers would have, the scribe is a tax. Worse, it becomes a rubber stamp — clinicians click sign without reading, and now the clinic has automated the production of unverified records, which is strictly worse than the paperwork it replaced.

What makes a monitoring note reviewable at that speed:

• Numbers presented as fields, not sentences. The eye checks a column of values far faster than it parses prose.
• Yesterday beside today. The reviewer's actual question is "does this progression make sense", and that question cannot be answered from one visit's note alone.
• Uncertainty marked, not smoothed. Anything the scribe is unsure of should look different from everything it is sure of. A draft that reads uniformly confident forces a full re-read, which nobody has time for, so nobody does it.
• Nothing to clean up. If the reviewer's job includes deleting a paragraph of narrative filler, the scribe has added work.

The independent evidence on ambient scribes reports documentation-time reductions from roughly 10% to 70% depending on setting, alongside a measurable fall in clinician burnout[2][3][4] — and the health systems that have deployed them at scale attribute their adoption rates to note customisation and hands-on training rather than to the model[5]. That is a finding about fit, not about intelligence. In a monitoring clinic, fit means the note matches the four minutes the visit actually took.

Nurse-led monitoring and the calls afterwards

In most fertility clinics the monitoring visit is only half the encounter. A nurse calls in the afternoon with the plan: your dose is changing, come back Thursday, trigger at ten. That call is a clinical communication, it changes the patient's treatment, and in a great many clinics it is documented — if at all — as a line typed from memory an hour later.

This is the largest undocumented surface in fertility care, and it is the reason we do not treat the scribe as a consultation-room product. The same reasoning that applies to the scan applies to the call: it carries a dose and a date, it is spoken quickly to someone who is anxious, and the record of it is currently whatever the nurse remembers. Capturing patient calls and messages the same way as encounters is the part of this that clinics tend to be most surprised by and, in our experience, ask about first.

What to test, if you are evaluating one

Take five recorded monitoring visits from your own clinic — de-identified — and check the drafts for exactly these:

1. Is laterality right in all five?
2. Is every follicle measurement present, in the right units, with counts distinguished from measurements?
3. Is cycle day in the note, and is it correct?
4. Where something was not measured, does the note say "not measured" rather than nothing?
5. Is the medication change and the next visit date captured exactly as spoken?
6. How long did it take a clinician to review each one?

A scribe that passes the first five and fails the sixth is not usable. A scribe that passes the sixth by skipping the first five is worse than none.

See it on your own visits: we would rather you measured this than took our word for it. Book a demo of the AI medical scribe for fertility clinics and bring your own de-identified recordings, or watch the scribe work first.

References

• [1] ESHRE guideline: ovarian stimulation for IVF/ICSI: an update in 2025 — Human Reproduction (2026)
• [2] Clinical Evaluation of Artificial Intelligence and Automation Technology to Reduce Administrative Burden in Primary Care — Alliance for Healthier Communities
• [3] Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout — JAMA Network Open
• [4] UCLA study finds AI scribes may reduce documentation time and improve physician well-being — UCLA Health
• [5] Adoption of Artificial Intelligence in Healthcare Delivery Systems: Early Applications and Impacts — Peterson Health Technology Institute (2025)

Related:

• Measured Impact: Clinical Evaluation Proves AI Reduces Documentation Burden
• AI in Fertility Clinics: Why Every Practice Needs an AI Platform

Sergei Gorlovetsky, CEO, Fertiligent