Triage AI: what it can and cannot do
Apps that ask about your symptoms keep getting better at signposting — and remain limited at diagnosis. Where the help ends and the consultation begins.
12 March 2025 · 7 min
Key points
- A triage tool estimates probabilities and suggests an urgency level; it does not diagnose and carries no clinical responsibility.
- Performance drops where training data is thin: young children, pregnant people, frail older adults, rare disease.
- For chest pain, breathlessness, bleeding or loss of consciousness, call the emergency number — do not open an app.
What a triage tool actually does
An AI-based triage tool receives a set of facts about you — age, symptoms, duration, history, sometimes where you are — and compares them with patterns learned from data. The result is not a diagnosis but an estimate: how likely it is that your situation needs immediate care, a consultation in the coming days, or a conversation with your family doctor.
The genuinely useful part is structure. Many people reach a clinician unable to say when symptoms started or what makes them better. A well-built questionnaire asks the questions a triage nurse would ask and shortens the road to the right decision.
Where it genuinely helps
In signposting: which kind of service to look for, how quickly, what to prepare before an appointment. Tools work here for a simple reason — the problem is well defined and a wrong answer costs little.
For common, well-documented conditions in adults without multiple chronic illnesses, such a questionnaire can order the hypotheses. In education it can explain medical terms and help you formulate questions for the clinician — probably its most underrated use.
The limits that matter
Training data does not cover every population. Young children, pregnant people, frail older adults, people with rare disease or compromised immunity are under-represented in datasets, and a model trained on typical adults performs worst exactly where mistakes cost most.
The tool cannot see you. It does not notice that you are whispering, that your skin colour is unusual, or that you are guarding a part of your body. It cannot listen to your lungs and has no access to test results. All it knows is what you typed.
Then there is error in two directions: too cautious — sending you to emergency care with a cold, which clogs the system — or too relaxed, telling you to wait when you should not. Both happen, and no benchmark eliminates them.
The signs that do not wait for an app
Prolonged chest pain, breathlessness, weakness on one side of the body or slurred speech, bleeding that will not stop, seizures, high fever in an infant, thoughts of harming yourself: all of these are handled by an emergency call or an emergency department, not a questionnaire.
A good tool says this in its first screens, before asking for personal data. If an app has no obvious route to emergency care, close it.
Questions to ask before handing over your data
Who operates the service and what happens to your answers? Is the product registered as a medical device in the country where you use it? Is there a named human behind it with responsibility? How do you report an error? Can you delete your data and your account?
If the answer to any of them is vague, use the tool as a list of questions for your clinician, not as a source of decisions.