Disclosure: the author is CEO of MyBabyBridge, an IVF company. The most useful AI in a fertility clinic may not touch the embryo at all. Tools that guide ovarian stimulation dosing and trigger timing, watch lab equipment and performance, and keep patients informed between appointments act on bigger, more fixable losses. Their evidence is still early, so clinics should demand prospective data before paying.
Every conversation about AI in IVF seems to end up in the same place: a screen showing a blastocyst with a score beside it. It is an irresistible image. It is also, on the evidence so far, one of the less convincing places to spend money. The first large randomized trial of a deep-learning embryo selector, published in Nature Medicine in 2024, could not show it was as good as embryologists at producing pregnancies, although it was roughly ten times faster.
Meanwhile, the parts of a fertility clinic that decide how many good embryos exist in the first place, whether they survive storage, and whether patients stay in treatment long enough to use them, get far less attention. That is where the argument of this piece sits. A clinic choosing where to put AI should look first at stimulation, lab quality control and patient communication, because the losses there are larger, more measurable and more often preventable than the marginal gains from reordering a queue of embryos.
That is a claim about where the upside is. It is not a claim that the tools in these areas are proven. Most are not, yet.
Where does a fertility clinic actually lose outcomes?
Think of an IVF cycle as a funnel. Stimulation produces follicles. Retrieval yields eggs, some mature. Fertilization and culture produce embryos, some of which reach the blastocyst stage and can be frozen. Transfers turn some of those into pregnancies. And across the whole sequence, patients decide whether to continue.
Embryo selection works at one narrow point near the bottom: choosing the order in which already-existing embryos are transferred. If a clinic freezes surplus embryos and transfers them singly, better ranking mainly shortens the time to pregnancy. It does not create viable embryos that were not there.
The steps higher up are different. A stimulation that recovers two more mature eggs, a lab that catches a drifting incubator before it harms a week of cultures, a patient who comes back for a second transfer instead of quietly dropping out: each of these can change whether a baby is born at all. They are less photogenic than an embryo score. They are also where the money is lost.
Can AI improve ovarian stimulation protocols?
Stimulation is full of repeated, data-rich decisions that doctors currently make on experience and rules of thumb: what starting dose of follicle-stimulating hormone to give, when to adjust it, and which day to give the trigger shot that matures the eggs before retrieval.
Two published studies show what machine learning is being asked to do here. In 2021, Hariton and colleagues reported in Fertility and Sterility a causal-inference model trained on 7,866 ICSI cycles from 2008 to 2019. It estimated, for each patient, whether triggering on a given day or a day later would yield more fertilized eggs and usable blastocysts, and it could separate patients who would benefit from waiting from those who would not. In 2022, Fanton and colleagues described in Reproductive BioMedicine Online a nearest-neighbour model that uses age, BMI, anti-Müllerian hormone and antral follicle count to match a patient to similar patients from more than 18,000 cycles at three US clinics, and so to build an individual dose-response curve for her starting FSH dose.
The appeal is obvious. Gonadotrophins are expensive, and many patients may be given more than they need. A model that suggests a lower starting dose with the same predicted number of mature eggs saves drug cost directly. One that nudges trigger timing to add even one usable blastocyst on average changes cumulative success, which embryo ranking cannot.
The caution is equally clear. Both studies are retrospective. Causal inference from historical data is only as good as its assumptions about why doctors made the choices they did; a clinician who delayed a trigger may have seen something in the scan that the model never recorded. Prospective studies and randomized trials of stimulation decision tools are under way or registered, and they are the evidence that should decide adoption. A clinic that buys now should treat the tool as a second opinion inside its existing protocol, audit the decisions where it disagreed with the doctor, and compare outcomes over time rather than trusting the vendor’s retrospective curve.
How can AI help IVF lab quality control?
In March 2018, on the same weekend, two US fertility clinics suffered storage failures. At University Hospitals in Cleveland, more than 4,000 frozen eggs and embryos were affected after a cryogenic tank malfunctioned on 4 March; the problem went unnoticed for a period because a remote alarm meant to warn staff about temperature swings had been turned off. Pacific Fertility Center in San Francisco had its own tank failure the same weekend. Both led to dozens of lawsuits.
No clever algorithm was needed to prevent that loss in Cleveland. A working alarm was. But the episode shows the category of risk that monitoring systems address: slow, silent failures in equipment that holds irreplaceable material, noticed too late because humans are not watching every sensor all night. Modern monitoring that tracks liquid nitrogen levels, tank temperature, incubator gas and temperature, and flags unusual patterns rather than only hard thresholds, is a natural place for anomaly detection. It is also the kind of system that must be designed so it cannot be quietly switched off.
The second form of lab QC is performance monitoring. Labs track key performance indicators such as fertilization, blastocyst development and pregnancy rates, broken down by embryologist, incubator, culture media batch and procedure. The problem is statistical: in a small or mid-sized lab, a genuine drop in one embryologist’s ICSI outcomes, or a subtle media problem, can hide inside normal variation for months. Software that continuously compares expected with observed outcomes, adjusting for patient age and case mix, can surface drift sooner. Some of that is classical statistical process control; some of it benefits from machine learning. Either way, the payoff is catching a problem that affects every patient in the lab, not just the order of one patient’s transfers.
Can AI chatbots support fertility patients between appointments?
Fertility treatment is administratively brutal. Patients juggle injection schedules, dose changes phoned through at short notice, blood tests, scans and bills, often while working and often without telling their employer. Much of the anxiety comes in the gaps between appointments, when a question arises at 10pm and the clinic is closed.
That matters for outcomes because patients leave treatment. A 2012 systematic review in Human Reproduction Update by Gameiro and colleagues looked at why people discontinue fertility treatment despite a reasonable prognosis. Postponement of treatment was the most common reason, at about 39%. Physical and psychological burden accounted for about 19%, and relational and personal problems for about 17%. Organizational problems (about 12%) and clinic problems (about 8%) also featured.
Some of those reasons are beyond any software. But organizational problems, clinic problems and a good part of the psychological burden are about information and friction: not knowing what happens next, not understanding a result, not getting through on the phone. Tools that answer routine questions accurately, send clear medication reminders, explain what a result means in plain language and route anything clinical to a nurse quickly target exactly that.
The risks are real and familiar from elsewhere in medicine. A chatbot that confidently gives wrong dosing information, or a model that drafts replies a busy nurse approves without reading closely, can do harm. We looked at that failure mode in the real risk of AI-drafted patient messages. The safer design keeps the AI in the role of scheduler, explainer and triage layer, with clinical content drawn from the clinic’s own approved protocols and any patient-specific medical question handed to a human.
A worked example: where would a clinic get its return?
Hypothetical, with round numbers for illustration only. A mid-sized clinic runs 1,500 stimulation cycles a year and is choosing between three AI investments of similar cost.
Option one is embryo ranking. Suppose it reliably shortens time to pregnancy for some patients. Because most surplus embryos are frozen and eventually used, the number of babies per egg collection barely moves; the gain is fewer transfers for some patients, which is valuable but largely accrues as saved time and cost for them rather than extra births.
Option two is stimulation decision support. Suppose, optimistically, it trims the average gonadotrophin dose modestly without reducing egg yield, and adds a fraction of a usable blastocyst per cycle through better trigger timing. Across 1,500 cycles, lower drug costs are immediate and measurable, and more usable blastocysts raise cumulative live birth per collection.
Option three is communication and retention. Suppose it persuades even a small share of the patients who would otherwise drop out after one failed transfer to return for a frozen transfer. Each returning patient has embryos already in storage; the marginal cost is low, and the marginal chance of a baby is substantial.
None of these numbers is real, and every assumption needs testing. The point is structural. Options two and three act on quantities, eggs, embryos and patients who continue, that set how many births are possible. Option one acts on sequence. A clinic that can only afford one should think hard before choosing the most visible.
What is the case for starting with embryo AI anyway?
The strongest counterargument deserves a hearing. Embryo selection tools are the most mature commercial AI products in the field; they plug into existing time-lapse incubators; they deliver a measurable, immediate benefit in embryologist time and grading consistency; and they have been tested in at least one large randomized trial, which is more than most stimulation or communication tools can say. Operational AI, by contrast, often means integrating with messy clinic software and changing how doctors and nurses work, which is slower and riskier.
That is a fair point about implementation. It is weaker as a point about value. A proven time saving is worth paying for as a time saving. It should not be mistaken for a proven gain in live births, and it should not crowd out investments that act on the steps where more births are actually won or lost. The operational tools need their own trials, and clinics should ask for them. For a buyer’s checklist that applies well beyond fertility, see buying digital health: the questions hospitals ask too late.
What this means for clinic leaders, clinicians and patients
For clinic leaders, map the funnel before buying anything. Know your stimulation outcomes, your lab KPIs by operator and incubator, and your dropout rate after the first failed transfer. The biggest leak is usually the best first target. Insist that any vendor show prospective data, not only retrospective curves, and build an audit into the contract.
For clinicians and embryologists, the useful role for AI is as a second opinion that is logged and reviewed. Record when you override the tool and why. Over a year, that record tells you more about its value in your clinic than any brochure.
For patients, the question to ask is not only “do you use AI?” but “where?” A clinic that uses software to watch its storage tanks around the clock, monitor its own lab performance and answer your questions promptly may be using AI more wisely than one that advertises an embryo score. Any question about your own medication or treatment plan belongs with your clinical team.
Frequently asked questions
How is AI used in fertility clinics besides embryo selection?
Beyond ranking embryos, AI is being applied to ovarian stimulation (choosing starting hormone doses and trigger timing), lab quality control (monitoring storage tanks, incubators and embryologist performance), and patient communication (reminders, explanations and triage of questions). Most of these tools have promising retrospective data but limited prospective or randomized evidence so far.
Can AI choose the right IVF medication dose?
Machine learning models have been built to suggest starting FSH doses and trigger days using data from thousands of past cycles, including a 2022 model drawing on more than 18,000 cycles. The published evidence is mainly retrospective, so these tools should support, not replace, a fertility specialist’s judgment until prospective trials report. Patients should discuss dosing only with their own clinician.
Why do patients stop IVF treatment?
A 2012 systematic review in Human Reproduction Update found postponement was the most common reason for discontinuation, at about 39%, followed by physical and psychological burden (about 19%) and relational and personal problems (about 17%). Organizational and clinic problems also featured. Better information and support may address part of that burden.
Can monitoring technology prevent embryo storage failures?
It can reduce the risk. In 2018 a cryogenic tank failure at a Cleveland clinic affected more than 4,000 eggs and embryos, and a remote alarm meant to warn staff had been turned off. Continuous monitoring with alarms that cannot be silently disabled, plus clear escalation procedures, addresses exactly that kind of slow, unnoticed failure.
The fertility industry tends to buy the AI it can photograph. The better return may lie in the tools nobody puts in a brochure: the ones that save a few eggs, catch a failing tank at 3am, and bring a discouraged patient back for the transfer that works.
If you run or work in a fertility clinic, where in your funnel do you lose the most, and would you trust software to help fix it?
