Ambient AI scribe or dictation: what changes for a fertility practice

2026-08-31|By Sergei Gorlovetsky, Founder & CEO

Clinics rarely choose between documentation methods in the abstract. The decision arrives attached to something else — a physician threatening to cut a session, a nurse manager pointing out that nobody has written up last week's calls, a new location opening with no coordinator. By the time anyone is comparing options, there is usually already a preferred answer and the comparison is being done to justify it.

This post is an attempt at the comparison anyway, from the clinic's side rather than the vendor's. We build an AI medical scribe for fertility clinics, so the interest is declared; what follows is written so that a practice could reasonably read it and choose dictation.

The three options

A human scribe, in the room or on a video link, listening and typing into the EHR while the clinician works. The oldest of the three and still the most capable: a scribe understands context, asks the clinician a question when something is ambiguous, and can navigate a chart.

Dictation with transcription, where the clinician speaks a note — during the visit, between visits, or after hours — and speech recognition, a transcriptionist, or both turn it into text. This is what most practices already do, often without describing it as a system.

An ambient AI scribe, which listens to the encounter itself and drafts the note from the conversation, for a clinician to review and sign.

They are not three points on one axis. They differ most in who does the work of turning speech into a record, and when.

What each one costs, structurally

Specific prices depend on region, volume and contract, and anyone quoting you a number without knowing yours is guessing. The shapes of the costs are stable, though, and the shape is what determines whether an option survives contact with a fertility practice.

A human scribe is a salary or an hourly rate, and it scales linearly with clinic hours. Two physicians running parallel sessions need two scribes. The cost is fully visible on a budget line, which is why it is the first thing cut and the first thing measured. It also carries a recruiting and training cost each time the role turns over, which rarely appears in the comparison.

Dictation looks nearly free, and that is its most misleading property. The licence is cheap. The cost is the clinician's time, moved rather than removed — from typing at the keyboard to speaking into a microphone and then correcting the output. That cost lands after hours, where it is invisible to every operational report the practice runs, and where the literature consistently locates burnout.

An ambient scribe is a per-clinician subscription, so it scales with headcount rather than clinic hours — a physician running a long day costs the same as one running a short one. What it does not remove is review: someone still reads and signs every note, and if that takes as long as typing would have, the subscription has bought nothing. The independent work here is worth reading before signing anything: the Peterson Health Technology Institute's 2025 assessment found early adopters reporting improved burnout but an unclear financial return, and attributed the highest adoption rates to note customisation and hands-on training rather than to any property of the model[1].

Where accuracy actually comes from

This is the part of the comparison that gets flattened into a percentage, and the percentage is nearly always the wrong number to look at.

The most useful evidence is a 2018 study of 217 dictated clinical notes across two health systems, which measured errors at three stages. In the raw speech-recognition output the error rate was 7.4%. After a medical transcriptionist reviewed it, 0.4%. In the note the physician actually signed, 0.3%[2]. Of the errors in the raw output, 5.7% were clinically significant.

Read that carefully, because it is the whole argument. Dictation is not accurate; dictation plus review is accurate. The accuracy of every option in this comparison is a property of the review step, not of the capture step. Which means the real question is not "how good is the transcription" but who reviews, how long it takes them, and whether they actually do it.

That reframes each option:

With a human scribe, review is fast, because the scribe has already resolved most ambiguity in the room and can be asked. The evidence on scribes is largely about clinician experience rather than accuracy — the first randomised trial found large improvements in physician satisfaction with chart quality and accuracy, and more charts closed within 48 hours, with no effect on patient satisfaction[3].
With dictation, review is the clinician re-reading their own words, which is the hardest text in the world to proofread. The 7.4% → 0.3% path in that study went through a professional transcriptionist. A practice that has dropped the transcriptionist to save money has quietly dropped the step the accuracy came from.
With an ambient scribe, review is reading someone else's draft of a conversation you were in. That is easier to proofread than your own dictation and harder than a scribe's note, and it depends entirely on whether the draft makes its uncertainty visible. A draft that reads uniformly confident forces a full re-read; a draft that flags what it could not support lets a clinician check the flags.

The ambient evidence is genuinely good on burden and genuinely modest on magnitude: a multicentre study reported outpatient burnout falling from 51.9% to 38.8% with ambient scribes and around an hour less after-hours documentation[4]; an Ontario evaluation measured a 69.5% reduction in note-writing time in a controlled setting and roughly three hours a week less after-hours admin in practice[5]; a UCLA trial found about 10% less time per note and around a 7% improvement in burnout scores[6]. The range across those is not noise — it is what fit looks like when it is measured in different settings.

What each one covers

For a fertility practice this is where the comparison stops being close.

All three options above document the consultation. That is the encounter they were designed around, and for a specialty whose work is mostly consultations, choosing between them is a reasonable way to spend an afternoon.

A fertility practice's documentation load is not mostly consultations. It is:

Incoming referrals — faxed, scanned, uploaded, handwritten — which somebody retypes into the record before a patient exists in it at all.
Monitoring visits, four minutes each, several per patient per cycle, mostly numbers.
Nurse calls, carrying dose changes and trigger times, documented from memory if at all.
Patient messages, which are clinical communications and are treated as correspondence.

A human scribe does not read your faxes. Dictation does not capture the call a nurse made at four o'clock. An ambient scribe that only listens in the consultation room covers the encounter type you have fewest of.

We are obviously not neutral about this — it is the argument the rest of our product rests on. But a practice can test it without us: count the documentation events in one week by type, and see what fraction of them happen in a consultation room. In most fertility clinics we have looked at, that fraction is small.

The three options are not equivalent under health privacy law, and the difference is mostly about recordings.

A human scribe hears the encounter and writes it down. Nothing is retained but the note. Dictation produces an audio file, usually short-lived, usually of the clinician rather than the patient. An ambient scribe records the patient's voice, for a period, and sends it somewhere to be processed.

Ontario's Information and Privacy Commissioner published AI Scribes: Key Considerations for the Health Sector on 28 January 2026, and it is the most concrete guidance available in Canada. Three points from it that a clinic should raise with any vendor:

Consent has to be real, and declining has to be free. The guidance asks custodians to "ensure patients who withhold or withdraw consent to AI scribe use receive the same level of care as consenting patients"[7]. In practice that means a documented path for the visit where the patient says no — which is a workflow question for the clinic, not a feature question for the vendor.
Retention is a decision, not a default. The guidance asks custodians to apply data-minimisation "throughout the AI scribe's lifecycle", including scrutinising "whether it is necessary to retain full audio recordings or transcripts"[7]. Ask what is kept, where, and for how long, and ask to be able to change the answer.
Model training is a separate question from service delivery. The guidance addresses what happens when information collected during AI scribe use contributes to model development[7]. "We don't train on your data" is a sentence worth getting in the contract rather than on the website.

Similar guidance now exists in British Columbia and Alberta, and PHIPA, PIPEDA and HIPAA all bear on this differently. None of it makes ambient scribing inadvisable. All of it makes an unexamined procurement inadvisable.

A checklist for deciding

Not a scoring matrix — those tend to produce the answer whoever built the matrix wanted. Six questions whose answers usually decide it:

Count your documentation events for one week, by type. Consults, monitoring visits, referrals, calls, messages. If consults are a minority, an option that only covers consults is solving a minority of your problem.
Time the review step, honestly. Have a clinician review five real drafts with a stopwatch. Compare it to what they do now. If review is slower, the option fails regardless of what it promises.
Ask who carries the risk of an unreviewed note. Every option here produces a draft that a clinician signs. The signature is the control. Any option that makes signing feel automatic has weakened it.
Ask what happens on a patient's "no". Not whether consent is captured — every vendor will say yes — but what the visit looks like afterwards.
Ask what is retained, and get it in the contract. Audio, transcripts, derived data, training use, deletion on termination.
Ask what it does with a referral. If the answer is nothing, you have costed a consultation tool against a clinic-wide problem.

The honest summary: for a practice whose documentation is genuinely consultation-shaped, all three options work and the choice is mostly about budget shape and appetite for change. For a fertility practice, the consultation is the part that is already handled least badly, and the choice is better made on what happens outside the room.

See it on your own week: if you want to test the counting exercise above against real output, book a demo of the AI medical scribe for fertility clinics and bring your own de-identified referrals and call recordings.

References

Related:

Sergei Gorlovetsky, CEO, Fertiligent