Does the triage result replace phoning the clinic?

No, and it is not designed to. It helps you understand how fast to move and helps you organise what you will say on the phone. If the picture worsens, if a new sign appears or if you simply have a doubt, phoning stays the right move even with a low level. Time spent on the phone with the clinic is almost always shorter than time lost by waiting.

Why do I get a high urgency level even when the animal seems normal?

Because the costs of the two errors are not symmetrical. Sending a well animal to the clinic costs one consultation; failing to send an unwell one can cost far more. Every serious system is tuned around that asymmetry and therefore over flags. The practical consequence is that a high level should be read as an invitation to check quickly, not as a prediction of severity.

Does a system trained on more cases automatically become more reliable?

Not necessarily. What matters is where the cases came from and how much they resemble the animal in front of you. Data collected in a population full of rare diseases produce skewed results when applied to a general population, and the reverse holds too. Missing information matters as well: a huge collection with no weight, no age and no current treatments stays blind on exactly the variables that move the decision.

What should I bring to the clinic if I have used triage?

The sequence of events with dates and times, the values you measured at home, an up to date list of medicines and supplements given, and a video if the phenomenon is episodic. The urgency level returned by the tool is the least useful part of the conversation: the clinician redoes the sorting anyway with the animal present, and what they lack are the facts from the interval between visits.

Technology and data10 min readUpdated on August 4, 2026

Triage is not diagnosis: the real limits of artificial intelligence

What an automated system can do about an animal's health, what it cannot do, and how to read an urgency level without making the wrong call.

Audience: Pet owners, Veterinarians
Species: All species
Scope: Valid everywhere, European Union

An automated system that receives a list of symptoms and returns an urgency level does exactly one thing: it sorts. It takes what was written, compares it with rules and examples, and indicates how quickly you should move. It does not look at the animal, does not palpate it, does not listen to the chest and measures nothing that someone else has not measured first. Calling that output a diagnosis is a category error, and it produces two opposite harms: people reassured by a low level postpone a visit that was needed, while people alarmed by a high level take an emergency slot that another animal needed. This article sets out what such a system can do, on which data, where the boundary is sharp, and how to read the result without turning it into a decision it has no business making.

What a triage system actually does

Triage was born in emergency medicine as a sorting tool: with several patients and one consulting room, someone has to decide who goes first. It does not establish what the patient has, it establishes how long that patient can wait before the risk grows. Automated triage applies the same logic to a list of signs reported by the person living with the animal. It is useful because it makes a threshold explicit, and it is limited because it works only on the information handed to it.

Separate jobs for an automated system and for a clinician with the animal present
Job	Automated system	Clinician in consultation
Sorting by urgency what has been reported	Yes, consistently and repeatably	Yes, with the animal present
Detecting a sign nobody reported	No, it does not exist for the system	Yes, that is why physical examination exists
Measuring heart rate, temperature, mucous membranes	No, it can only receive values already measured	Yes
Recalling the whole recorded history at once	Yes, if the history was recorded	Only if it is consulted and readable
Reaching a diagnosis	No	Yes, with a stated level of certainty
Prescribing a treatment	No	Yes

The second row matters most. An animal in abdominal pain that has been hiding for two days can look calm to the person watching it. If nobody writes down that it is hiding, that sign does not exist for the system. The clinician, on the other hand, sees it in the posture, finds it on palpation, infers it from the fact that the animal will not let one specific area be touched.

Why a probability is not a diagnosis

The second misunderstanding is about numbers. A system that claims to recognise nine cases out of ten sounds reliable, but the percentage that matters to the reader is a different one: given a positive result, how likely is it to be true. That percentage depends on how common the condition is in the population being screened, and it changes the meaning of the identical output completely.

Picture 1000 animals going through the same automated check and a condition affecting 10 of them, that is one per cent.
A system that recognises 9 affected animals out of 10 finds 9, and misses 1 that gets no flag at all.
The same system, if it is wrong on five per cent of healthy animals, produces roughly 50 flags on animals that are well.
Total flags therefore come to 59, of which only 9 correspond to a genuinely affected animal: fewer than one in six.
Change how common the condition is and everything changes: identical stated accuracy produces opposite results in a selected population and in a general one.

This is not a flaw in the system, it is arithmetic. And it is why an automated result should be read as a well posed question rather than as a closed answer. Clinicians run the same reasoning daily under a precise name: the probability that a disease is present before the test conditions how far the test moves the conclusion.

Where the data come from and what is missing

A system is only as trustworthy as the sources it declares. Animiyo uses open sources, listed one by one in the integrations register together with the way each is queried. None requires a paid key, and the distinction between a call from the server and a call from the browser is not a technical footnote: in the first case the provider sees the server address, in the second it sees the address of the reader's device.

Open sources declared by the catalogue and how each is queried
Source	How it is queried	Key required
openFDA	From the server, at api.fda.gov	No
PubChem	From the server, at pubchem.ncbi.nlm.nih.gov	No
Open Pet Food Facts	From the server, at world.openpetfoodfacts.org	No
Open-Meteo	From the server, at api.open-meteo.com	No
OpenStreetMap	From the server and from the browser, depending on the feature	No
WSAVA, AAHA, ACVIM, EMA, WALTHAM	References transcribed in the code, no network call at all	No

The typical failures and what they cost

The ways automated triage causes harm are few and repetitive. Knowing them in advance costs one reading and is worth more than any accuracy claim, because nearly all of them arise in the handover between the person and the tool rather than inside the tool.

False reassurance: the description leaves out the decisive sign, the level comes back low and the visit slips by two days. The most expensive failure and the quietest.
Constant alarm: a system tuned to miss nothing flags too much, and after a few weeks the person stops reading it. An ignored alert is worth zero.
Rewriting until the desired answer appears: answers get changed until the hoped for level shows up. This happens to the worried owner and to the one who does not want to leave the house.
Vague language: off colour, strange, not himself. True sentences and unusable ones. A system works on signs described with onset, frequency and trend.
Misplaced authority: the output is reported at the clinic as a diagnosis already received, and the conversation starts from a conclusion instead of from the facts.

One simple counterweight exists and it is recording. An episode noted with date, time and duration stays comparable; a sensation recalled days later distorts. The difference between arriving at the clinic with three dates and arriving with a general impression is the difference between data and anecdote.

What Animiyo does and what it does not

It is worth being precise about the perimeter, because a vague promise is the first step towards misuse. The features listed below exist in the web application; the ones marked as outside the free plan need an active subscription.

Symptom triage: available on the web version, included in the free plan. It orients on urgency, does not diagnose and does not prescribe.
First aid: operational cards written for the minutes before the clinic.
PetAI assistant: available on the web version, not included in the free plan. It answers on what has been recorded and remains a tool for orientation.
Plain language report and unified clinical timeline: they turn scattered documents into a readable story before the appointment.
Reception queue with clinical triage: this belongs to the clinic workspace and serves to order who goes in first.
Second opinion between professionals: a clinic workspace feature, not a service aimed at owners.

On the other side, and this deserves to be said in full, there is no automated image reading, no direct integration with analysers, no connected wearable devices, no video consultation and no automated diagnosis of any kind. If any of these ever exist, they will appear in the feature catalogue like everything else.

How to use a result without getting hurt

The operating rule is single: an automated result helps decide when to phone, not whether to phone. In practice four steps make the use solid and remove almost all of the failures described above.

Describe facts before interpreting
Write what you saw, since when, how many times and on what trajectory. Avoid labels: vomiting four times in six hours is data, upset stomach is a hypothesis.
Treat a high level as a time limit
A high level means the useful window is short, not that the diagnosis is severe. Phone the clinic and report the facts, not the level.
Do not use a low level to close the matter
A low level holds while the picture stays the same. Set a precise moment to reassess, for example the same evening, and write down what would have to change to trigger the call.
Bring the record to the visit, not the screen
The clinician needs the sequence of events with dates and times, not a screenshot of the result. The recorded timeline is the part that adds information to the consultation.

Triage separates those who can wait from those who cannot. It says nothing about what either of them has.
Founding principle of emergency medicine

Frequently asked questions

Does the triage result replace phoning the clinic?: No, and it is not designed to. It helps you understand how fast to move and helps you organise what you will say on the phone. If the picture worsens, if a new sign appears or if you simply have a doubt, phoning stays the right move even with a low level. Time spent on the phone with the clinic is almost always shorter than time lost by waiting.
Why do I get a high urgency level even when the animal seems normal?: Because the costs of the two errors are not symmetrical. Sending a well animal to the clinic costs one consultation; failing to send an unwell one can cost far more. Every serious system is tuned around that asymmetry and therefore over flags. The practical consequence is that a high level should be read as an invitation to check quickly, not as a prediction of severity.
Does a system trained on more cases automatically become more reliable?: Not necessarily. What matters is where the cases came from and how much they resemble the animal in front of you. Data collected in a population full of rare diseases produce skewed results when applied to a general population, and the reverse holds too. Missing information matters as well: a huge collection with no weight, no age and no current treatments stays blind on exactly the variables that move the decision.
What should I bring to the clinic if I have used triage?: The sequence of events with dates and times, the values you measured at home, an up to date list of medicines and supplements given, and a video if the phenomenon is episodic. The urgency level returned by the tool is the least useful part of the conversation: the clinician redoes the sorting anyway with the animal present, and what they lack are the facts from the interval between visits.

What to do next

Run a test right now, calmly and with no urgency: describe an episode that already happened using only facts, onset, frequency and trend, and notice how hard it is to write without slipping an interpretation in. Then set the rule for the real occasion: the automated result decides how fast you phone, never whether you phone, and the record of events is what you take to the visit.