AI scribe accuracy in practice: what six months of daily use actually showed

When I first started using an AI scribe, I expected it to be more accurate than my memory. That was not a high bar. After a long day of back-to-back consults, a vet's recall of what was said in appointment six is already compromised by appointments seven, eight, and nine. Anything that captured the conversation while it was happening rather than after the fact had an obvious advantage.

What I did not expect was that my standard would rise. The better the tool got, the more precisely I noticed where it fell short. That is probably the most honest thing I can say about accuracy after sustained daily use: the goalposts move as you get used to it, and they move in the right direction.

The accuracy question worth asking

Many AI scribe vendors will give you an accuracy percentage if you ask. Most of those numbers describe transcription quality, meaning how faithfully the tool heard what was said. That is the lowest bar. A SOAP note is not a transcript. It is a clinical record of what happened, what was found, what was decided, and what happens next. Getting from the conversation to that record involves three distinct skills, and a tool can perform well on the first of them while failing quietly on the other two.

The first is hearing correctly. The second is translating informal speech into clinical meaning: recognising that "she has been a bit off her food" belongs in the history, not as a verbatim quote from the client. The third is structural judgement: putting information in the right section, separating what the client reported from what was found on examination, and filtering out the logistics questions and small talk that fill a real consultation. A 2025 instrument validation study testing AI scribe technology found that the errors that mattered clinically were not simple mishearing. They were information added, omitted, or quietly inconsistent with what was actually said.

When a practice asks a vendor "what is your accuracy rate?", they are asking the right concern in the wrong terms. A more useful question, which I came to understand only after extended use, is about speed. A scribe can be made very accurate by taking longer to process the consultation. The practical question is whether the result comes back fast enough to fit into the consulting flow. A tool that takes two minutes to return a note while the next client is waiting in reception is not useful, regardless of the percentage.

What a two-week trial cannot tell you

A short trial period almost always produces results that are more favourable than the technology deserves, not because the tool gets worse over time, but because the cases get harder. In the first two weeks, most vets use the scribe on the consults they are confident about: vaccination boosters, simple rechecks, weight checks, medication reviews with one client speaking clearly about one issue. Those consults are exactly what an AI scribe handles best.

In my experience, you need at least a month before you understand the real accuracy pattern. By then you have encountered the consults that expose where the tool struggles: multiple animals in the same appointment, multiple owners talking over each other, clients whose thinking is scattered rather than structured, noisy rooms, nervous animals, or consultations where the important information is implied rather than stated directly. A two-week trial will not capture enough of those to tell you what you actually need to know.

A retrospective cohort study published in JAMA Network Open in October 2025 found that clinicians using an AI scribe spent less time in the EHR and in notes over a three-month pilot. The broader lesson for practices is that the benefit depends on how consistently the tool is adopted and how well it fits the clinical workflow. That matches what I observed: the vets who got most out of the tool were the ones who used it on every consult rather than selecting only the simpler ones.

A practice evaluating an AI scribe should deliberately put it through the difficult cases during the trial period, not just the straightforward ones. That does not mean using it unsupervised on high-risk consults. It means comparing the AI note against the clinician's own judgement on the consult types most likely to expose weaknesses. That comparison will tell you more than any vendor accuracy figure.

Where it earned trust

Single-issue consults with one client speaking clearly produced notes I could review in five to fifteen seconds. Not five minutes. Seconds. Vaccination boosters, simple rechecks, uncomplicated medication reviews: the scribe handled these reliably enough that over time the review became a focused scan rather than a line-by-line reconstruction of the consult. A Lupa case study with Thompson's Pet Care reported an estimated saving of around one hour per user per day across the practice workflow. That figure reflects the broader Lupa platform rather than Lupa Notes in isolation, but it illustrates why getting repeatable, high-volume consults right consistently is where practice-wide time savings are actually built.

The second strength is one that is easy to undervalue until you experience it: the scribe is present for the whole consultation. When I am working through a complex case, my attention is on the animal and the client. Details that belong in the history sometimes get compressed or lost by the time I am writing the note. The AI has been listening throughout. On the busiest days, that continuity of attention catches things I would have trimmed or forgotten.

The clearest example of this came from an out-of-hours emergency involving complex interactions with a client and the police. "My interactions with the client and the police were long and complex," Dr Nick recalls. "I would never previously have captured the detail and nuance that the scribe allowed me to. I was able to produce very full and complete notes, which gave me peace of mind in that I had done the very best for the animal in the situation, and were available for the ensuing legal process." A high-pressure situation involving multiple parties, competing accounts, and potential legal scrutiny is exactly the context in which a vet's own post-consult notes are most likely to be compressed by stress and time. The scribe was there for all of it.

Where it needed a proper read

The errors worth flagging are not the dramatic kind. A scribe that produced obviously wrong or nonsensical notes would be abandoned within a week. The real risk is subtler: a note that is plausible, well formatted, and almost entirely correct, with one detail quietly wrong in a way that is easy to miss on a fast review. That is exactly what I found over time: picking out the one incorrect thing from an otherwise very good summary.

Multi-pet and multi-speaker consults were the most consistent source of this kind of error. When two owners are discussing two animals in the same appointment, the scribe occasionally attributed a symptom or a plan to the wrong patient. I learned to compensate for this by using verbal cues during the consultation itself, making explicit statements that named the animal before discussing its history or plan. That helped, but it also required changing how I consulted, which is worth being honest about.

The same adaptation applied to multi-speaker consults. When two owners talked over each other or moved quickly between topics, the notes were weaker. I started making clarifying statements mid-consult, partly for the scribe's benefit, partly because it is good clinical practice anyway. The two are not in conflict.

Brand names, diet products, supplement names, and some medication terminology were a second recurring weak point. A specific product name was sometimes transcribed correctly but interpreted strangely, treated as a literal description of something rather than a proper noun. None of those errors were clinically dangerous alone. All of them needed a human to catch.

The sections that were weakest across the longest period of use were the History and Plan sections of the SOAP note. Those are the two sections that require the most clinical judgement to construct well. History depends on interpreting what a client said and placing it in the right clinical frame. Plan requires reflecting decisions that were made but sometimes not stated explicitly during the consultation itself. Those are hard things for any non-clinical system to do reliably, and they remained the areas where I paid the most careful attention throughout.

How review effort should actually be distributed

Every AI-generated note is still a professional clinical document and should be reviewed before it enters the record. That is not in question. The more useful question is how that review effort should be distributed across different note types, because treating every note identically either wastes time on the straightforward ones or creates false confidence about the complex ones.

As a personal working rule, I found note review often took around five per cent of the consultation length. For a ten-minute vaccination appointment, that means around thirty seconds. For a thirty-minute complex case, closer to ninety seconds at minimum, sometimes more. That sounds granular, but the instinct for it builds naturally once you have been using the tool consistently for a month or more.

What changes over time is not that the tool gets easier to trust wholesale. It is that your own consulting becomes cleaner, which makes the summaries more reliable. The scribe reflects what was said in the room. If the consultation was well structured, with clear verbal transitions between topics and explicit statements about which animal and which issue you are addressing, the note reflects that structure. If the consultation was compressed, rushed, or chaotic, the note reflects that too. The quality of the input shapes the quality of the output in ways that a vendor accuracy figure cannot capture.

The practical summary

AI scribe accuracy is not a percentage to believe or dismiss. It is a working pattern that develops over time, shaped by how consistently you use the tool and how well your consulting style fits what the tool can do.

The practices getting most out of Lupa Notes have learned, through real use on real consults, which notes can be scanned in fifteen seconds and which need a proper clinical read before they go near the record. That knowledge does not come from a vendor trial or a percentage. It comes from a month of daily use, across the full range of consult types, including the difficult ones.

If you are exploring what an AI scribe is and what it can do for your practice, or deciding between a standalone scribe and a PMS with AI built in, ask the accuracy question in those terms: what happens on the difficult consults, not just the straightforward ones, and how long does the result take to come back?