A while back I took my dog Bowie to the vet because her back legs were bothering her. Partway through the appointment the vet mentioned she was carrying extra weight. I said she seemed fine to me. The vet scored her at 5.5 on the nine point body condition scale, said she should lose half a pound, worked out in her head that she needed around 235 calories a day, and we moved on.
She was right. Bowie was over a pound heavy, which on a dog that size is a serious proportion of her, and it was almost certainly why she was struggling on the stairs. The vet had eight minutes, her own eyes, a scoring scale and a standard formula. She used all of it well.
What interests me now is the iPad she was holding.
It had Bowie's name, age and vaccination history. It had what I owed at the end, to the penny. It also had every weight Bowie had ever been recorded at, sitting in a list, going quietly up, and it had nothing whatsoever to say about them. It did not flag the trend. It did not know that her breed mix is prone to weight gain. It did not check the calorie figure, and it had no mechanism to revisit that figure later, when it stopped being right.
None of that is a fault in the product. It is the product working exactly as designed.
Veterinary practice management software belongs to a category we have been building for forty years, usually called a system of record. Booking, billing, inventory, reminders, notes, compliance, the audit trail. These systems are genuinely good at that work and clinics could not operate without them. A record is reliable precisely because it is inert. It holds what you put in it, does not editorialise, and can be produced in court a decade later unchanged.
The design assumption underneath all of it is that the knowledge lives in the clinician and the software handles the paperwork. For most of the history of this software, that assumption was correct, because the clinician's head was the only place clinical knowledge could live. The division of labour was sensible. Software took the half of the job it could actually do.
What has changed is that the other half is now tractable, and that changes what it is reasonable to expect the software to do.
The specific failures in that consulting room are worth separating out, because they are not the same failure. The first was a series that nobody interpreted: years of weights that formed a shape, sitting in a database that stores numbers and has no concept of a trend. The second was missing context: the fact that this individual, of this breed mix, at this age, carries a different risk profile from the dog seen before her. The third is the one I find most interesting, because it is not about the moment of the appointment at all. The vet's 235 calories was a reasonable figure derived from a standard formula. It was also going to stop being correct as soon as Bowie's weight moved, and there was nothing anywhere in the loop designed to notice that or revise it. The vet would not see her for another year. The record was a record. And I had no idea it needed revising.
I want to be precise about this, because it looks like a data problem and it is not one. Nothing was missing. The weights existed. The breed information existed. The formulae exist in the veterinary literature and have for years. What was missing was anything in the loop capable of noticing.
That is the distinction I would draw between a system of record and the thing that is starting to replace it. A record cannot be wrong on its own, which is its great virtue and the reason we built everything this way. But the same property means it cannot correct itself, cannot tell you when it has stopped being useful, and has no stake in the outcome. It remembers. It does not notice. Those are different jobs, and we have spent four decades building software that does the first one extremely well and then left the second one to whichever human happens to be in the room, on the clock, with an owner who is politely disagreeing with them.
Human medicine is worth watching here, because it digitised first and largely digitised without getting smarter. A time and motion study published in the Annals of Internal Medicine in 2016 observed physicians spending 27% of the office day in direct contact with patients and 49% on the record and desk work. That is what you get when you computerise the filing cabinet and stop there. You do not get a system that knows anything. You get a clinician who has been quietly reassigned to data entry.
Veterinary software is standing roughly where human medicine stood fifteen years ago, and it has the advantage of being able to watch what happened.
What software that actually knows something looks like, in practice, is less dramatic than the phrase suggests. It means the clinical calculations that currently run in a vet's head, or not at all, are held in the system and applied per animal. It means those calculations update when new data arrives rather than being frozen at the moment they were made. It means the system carries the relevant literature on breed, age and build, so the number it produces is specific to the dog in front of you rather than to dogs in general. None of this is speculative technology. Most of it is arithmetic that has been sitting in textbooks for decades, waiting for somewhere to live.
The practical consequence for a clinic is continuity. A vet sees eight minutes of a dog's year. The owner sees all of it and often cannot perceive what is happening, because you do not notice gradual change in something you look at daily. Between those two failures of attention there is a gap of about fifty-one weeks, and that gap is where most weight problems actually happen.
It also changes the conversation in the room. A vet raising weight with an owner is currently offering a professional opinion against that owner's feelings about their own dog, in a few minutes, with nothing to show. Put the trend on the screen, with the target and the reasoning behind it, and the exchange stops being a matter of competing impressions. The judgement remains the vet's. She simply gets to make it with something in her hand.
That is the shift, and it is a smaller and more practical thing than most of the current conversation about AI in clinical settings suggests. The record was never the problem. Decades of dutiful note-taking by people who thought they were doing paperwork are the reason any of this is possible now. But a record that only remembers is half a system, and we have been treating it as a whole one for a long time.