AI is quietly reshaping the front door of telehealth. Before you reach a human provider on many platforms, an AI system has already assessed your symptoms, estimated urgency, and routed your case. Some patients find this efficient; others find it unsettling. Here's what's actually happening, what the research shows, and what you should know.
What AI triage is
AI triage systems use algorithms — increasingly powered by large language models — to evaluate your symptoms and determine the appropriate level of care. At the simplest level, this means structured symptom checkers that ask branching questions and suggest whether you need emergency care, urgent care, or a routine appointment. At more advanced levels, AI can analyze unstructured text (your description of symptoms in your own words), pull in relevant medical history, and generate a preliminary clinical assessment for the human provider who ultimately reviews your case.
Where it's being used
AI triage is deployed across multiple telehealth settings. Pre-visit intake systems use AI to collect and organize patient information before a human provider reviews it, reducing the time a clinician spends gathering basic data. Symptom routing engines direct patients to the right care level — emergency, urgent, or routine — based on described symptoms and risk factors. Clinical decision support tools present providers with AI-generated differential diagnoses and recommended workups, which the provider can accept, modify, or reject.
Companies like Infermedica, Buoy Health, and Babylon (now part of eMed) have deployed AI triage systems used by millions of patients. More recently, general-purpose AI systems like OpenAI's ChatGPT Health (launched January 2026) have entered the consumer health space as symptom guidance tools.
What the research shows
AI triage performance is mixed — and the most recent independent research highlights real concerns. A February 2026 study published in Nature Medicine tested ChatGPT Health's triage recommendations across 60 clinical scenarios and found an inverted U-shaped accuracy pattern: the system performed well for "textbook" emergencies but made dangerous errors at clinical extremes. It failed to identify the urgency of 48% of true emergency presentations and over-triaged 35% of clearly non-urgent conditions.
The researchers noted a particularly concerning pattern: the AI would correctly identify dangerous findings in its explanations but then still reassure the patient — recognizing the red flag intellectually without translating it into appropriate urgency in its recommendation.
What patients should know
If a telehealth platform uses AI in its intake or triage process, you should understand that AI assessment is not a diagnosis — it's a routing suggestion. A human provider should always review your case before treatment decisions are made. If an AI tool tells you your symptoms aren't urgent but your intuition says otherwise, trust your instincts and seek care. AI systems can't detect non-verbal cues, assess your appearance, or factor in the subtle clinical context that experienced providers pick up on. The best AI triage systems are transparent about their limitations and consistently route ambiguous cases toward human review.
The regulatory landscape
AI triage tools that make clinical recommendations may be subject to FDA regulation as medical devices. Infermedica's symptom checker, for example, is certified as a Class IIb medical device under EU regulations. In the U.S., the FDA is still developing its regulatory framework for AI in healthcare, and many consumer-facing tools operate in a gray area between wellness app and medical device.
Our Assessment
AI triage is a useful efficiency tool when it serves as a front door to human clinical care — not a replacement for it. The technology is good at the middle of the severity spectrum and unreliable at the edges. Patients should treat AI triage recommendations as suggestions, not diagnoses, and should always escalate concerns that don't feel right. The telehealth platforms using AI most responsibly are those that route edge cases to human providers rather than auto-resolving them.