AI Has Entered the Consult Room. You’re Still Responsible
Artificial intelligence did not arrive in veterinary medicine wearing surgical scrubs and demanding control of the practice.
It arrived quietly.
It offered to transcribe our consults, tidy our clinical notes, draft discharge instructions and help us find the polite wording for “your Labrador is still overweight.”
Then it started reading radiographs.
AI is now appearing in practice-management systems, diagnostic imaging, laboratory analysis, appointment scheduling, client communication, education and clinical decision-support tools. Some of it is genuinely useful. Some of it is impressive right up until it confidently invents something that never happened.
And all of it raises the same important question:
How much should we trust a tool that can sound certain without actually understanding the patient?
The rise has already happened
For many clinics, the first useful encounter with AI has been the clinical scribe.
A microphone records the consultation and software converts the conversation into a structured history, examination and plan. In theory, this means less time staring at a computer and more time looking at the patient.
That is not a small benefit.
Clinical records are essential, but completing them after an already overfilled day is one of those veterinary traditions nobody asked to preserve. If AI can turn a rambling conversation about vomiting, diet changes, three different treats and an incident involving compost into a usable first draft, it may give clinicians some of their evenings back.
Beyond scribing, AI tools are being developed or marketed for:
Interpreting radiographs and other diagnostic images
Recognising patterns in laboratory results
Flagging patients at risk of deterioration
Producing consultation summaries and discharge instructions
Drafting client emails
Automating appointment and administrative tasks
Supporting clinical education and literature searches
Monitoring population-level disease trends
A recent review concluded that AI is beginning to create real value in selected small-animal applications while remaining immature in others. That is probably the least exciting description of AI—and the most accurate. Read the review on PubMed.
Where AI may genuinely help
Veterinary medicine produces an enormous amount of information.
There are histories, examination findings, laboratory results, images, medication records, previous notes and client observations—some of which arrive as a twelve-minute voicemail recorded beside a running washing machine.
Computers are good at processing large quantities of structured information. They do not get tired at 6:45 p.m., become distracted by the spaniel screaming in reception or discover halfway through a report that someone has eaten their emergency biscuit.
Used properly, AI may act as a second set of eyes. It could highlight a subtle abnormality, identify an unexpected pattern or remind a clinician of something worth investigating.
One 2025 study compared commercial radiology software with veterinary radiologists across 50 canine and feline imaging studies. The AI achieved accuracy comparable to the best-performing radiologist in that particular dataset and was good at recognising normal findings. However, it was less sensitive than the human radiologists, was weaker at detecting abnormalities and did not generate differential diagnoses. The authors concluded that its likely role was to complement—not replace—human expertise. Read the radiology study.
That result sounds promising, but it also demonstrates why the details matter.
A tool that is excellent at confirming normality but less reliable at detecting disease may be helpful in some circumstances and dangerous in others. “The computer said it was normal” will provide remarkably little comfort when the computer was wrong.
A published commentary also raised concerns about the study’s small, imbalanced sample and the lack of an independent diagnostic gold standard. Its authors argued that the findings should be treated as preliminary rather than proof that AI can match a specialist. Read the commentary.
In other words: promising technology, interesting evidence and a very good reason to keep thinking.
The confidence problem
Generative AI does not retrieve truth in the same way a clinician retrieves a potassium result from the laboratory system.
It generates a likely response based on patterns in its training and the information supplied to it. When it lacks information, it may say so. It may also fill the gap with something beautifully written and entirely fictional.
This is commonly called a hallucination, although “confident fabrication” is perhaps more clinically useful.
An AI-generated note might insert an examination finding that was never discussed. A literature summary may include a paper that does not exist. A differential list may contain a plausible-sounding condition that makes little sense for the species, age or presentation.
The dangerous outputs are not always obviously ridiculous. They may be mostly correct, professionally phrased and wrong in one small but important way.
The spelling can be perfect while the medicine is unhinged.
You cannot delegate responsibility to the robot
Professional regulators are already making their position clear.
The UK’s Royal College of Veterinary Surgeons says clinical decision-making must not be wholly delegated to AI. Veterinary professionals remain responsible for their patients, and AI-generated records should be manually checked and corrected. Read the RCVS guidance.
Vetboard Victoria has issued similar advice: veterinarians remain responsible for records, diagnoses and other work produced with AI assistance. Errors, omissions and additions cannot simply be blamed on the software. It also advises practices to consider accuracy, privacy, cybersecurity, recording consent and staff training. Read the Vetboard Victoria guidance.
That means an AI-generated medical record is a draft, even when it looks finished.
If it says the heart was auscultated and it was not, that is now your inaccurate record. If it removes an important qualifier from the history, that is your omission. If it gives the wrong dose and you approve it without checking, the software will not be the one explaining the decision.
AI can assist with the work. It cannot assume professional accountability for it.
Privacy is not an optional extra
Consultations contain sensitive information about clients, staff and patients.
Before feeding any of that information into an AI tool, clinics need to know:
Where the information is being stored
Whether it is retained after processing
Whether it may be used to train the system
Who can access it
Whether the system meets local privacy requirements
Whether clients must consent to audio recording
What happens if the provider suffers a data breach
There is an enormous difference between an approved veterinary scribe operating under a suitable privacy agreement and somebody pasting an identifiable clinical history into a free consumer chatbot.
Convenience does not cancel confidentiality.
Bias goes in; bias comes out
AI systems learn from data. If that data is narrow, incomplete or unrepresentative, their performance may be too.
A tool trained mainly on common canine and feline cases from large referral hospitals may perform differently when confronted with an uncommon breed, a poorly positioned general-practice radiograph, an exotic species or a condition rarely represented in its training material.
This does not make AI uniquely flawed. Humans also carry bias, miss patterns and make mistakes.
The difference is scale. One clinician’s error affects the cases in front of them. A systematic error embedded in widely used software can be reproduced across thousands of patients while looking reassuringly consistent.
Before adopting a clinical tool, practices should ask what it was trained on, how it was validated, what its intended use is and whether there is evidence that it performs reliably on patients resembling their own caseload.
If the supplier cannot explain its limitations, that is itself useful information.
Will AI replace veterinary professionals?
Probably not in the dramatic robot-takes-your-stethoscope sense.
Veterinary work depends on context, communication, practical skill and judgement. A computer does not see how an animal moves into the room unless that information is captured. It does not notice the owner hesitating before answering. It does not understand that the theoretically ideal plan is impossible for this particular family.
It cannot restrain a furious cat, place an intravenous catheter in a collapsed patient or explain devastating news with compassion.
However, AI will change parts of the job.
Some administrative tasks will become faster. Clinical workflows will change. People who understand how to use and question these tools may work more efficiently than those who ignore them completely.
The real risk is that every minute AI saves will immediately be filled with another appointment.
If a scribe saves a veterinarian an hour of paperwork but management responds by squeezing in four more consults, the technology has not improved that veterinarian’s life. It has merely made the conveyor belt move faster.
The best use of AI would be to return attention to the patient, restore time for communication and reduce the amount of unpaid work following clinicians home.
A sensible clinic rulebook
Until the evidence and regulation become more mature, a few principles are worth keeping:
Treat AI output as a draft or second opinion—not an authority.
Check every clinical record before approving it.
Independently verify doses, diagnoses, references and treatment recommendations.
Do not enter identifiable client or patient information into an unapproved system.
Understand how consultation recordings and data are stored and used.
Be transparent with clients when AI is recording or materially contributing to their care.
Test tools against your own caseload before trusting them in routine use.
Keep the human in the decision.
Make sure saved time benefits the veterinary team and the patient—not only the appointment schedule.
The tool should make us more human
AI is neither the saviour of veterinary medicine nor the beginning of its inevitable collapse.
It is a collection of tools—some already useful, some overmarketed and some still looking for a problem they can solve.
Used carefully, AI may reduce administrative burden, improve access to information and provide useful diagnostic support. Used carelessly, it can produce inaccurate records, encourage overconfidence and expose confidential data.
The important question is not whether veterinary medicine will use AI. It already does.
The question is whether we will use it with enough curiosity to benefit from it, enough scepticism to challenge it and enough backbone to prevent “efficiency” from becoming another word for doing more work with fewer people.
Let the computer help carry the paperwork.
The judgement, responsibility and humanity still belong to us.
Comments