Can an AI agent triage patients?

Yes, in one direction. An AI agent can sort incoming requests, collect symptoms with an approved question set and move anyone with a red-flag symptom to urgent care at once. It should never lower urgency on its own. Deciding that someone can wait, or assigning a lower clinical priority than the protocol suggests, stays with a clinician.

Why is patient triage asymmetric?

Triage errors come in two kinds. Over-triage sends someone to urgent care who did not need it. That costs a clinician's time and a patient's afternoon, and both are recoverable. Under-triage tells someone with a serious condition to wait. If they deteriorate in the meantime, nothing restores the lost hours.

So the two directions deserve different treatment. Moving a patient up the queue is a protective action that an agent can take on its own. Moving a patient down is a clinical judgement with consequences that cannot be reversed, which is the test set out in why autonomy in regulated industries is a question about reversal.

This also means you should judge a triage agent by its misses, not its average. An agent that sorts most patients correctly but occasionally under-triages a stroke is not acceptable at any volume.

What can the agent do, and what stays with a clinician?

The agent's proposed priority is a draft. A clinician reviewing a list of proposals, each with the answers that produced it, works faster than one reading raw messages. But the clinician sets the priority, and the record shows who did.

How is patient triage different from other triage?

The word is borrowed widely. In security operations, alert triage sorts alerts by likely severity, and in service desks, ticket triage routes requests to the right team. Agents can often take both further, because a misrouted ticket can simply be rerouted. Our playbook on triaging the support inbox with an AI agent shows how much can run unattended there.

Patient triage differs in two ways. The cost of delay is physical, not operational. And the person describing the problem is often frightened, in pain or unsure what matters, so the information is less reliable than a system alert. Both push decisions towards a clinician.

A worked example

Take a primary care practice that receives several hundred online requests each morning. The agent reads each one.

Most are administrative. A repeat prescription request goes to the prescribing queue; a request for a fit note goes to the admin team. The agent routes both at depth 3.

One patient writes that their father, aged 70, has had a "funny feeling" in his left arm since breakfast and his speech seems slow. The agent recognises two red-flag symptoms from the protocol. It replies immediately, telling them to call emergency services now, and alerts the duty clinician with the message. It does not wait for the morning list.

Another patient reports a sore throat for three days with no red flags. The agent proposes a routine priority, shows the answers that led there, and places the request on the duty clinician's list. The clinician agrees and books a routine appointment.

What controls make it safe?

• Clinician-owned protocol. The question set, red flags and priority rules are written and signed off by clinicians. The agent applies them; it does not extend them.
• Escalate freely, never de-escalate. The agent's role scope allows moving a request up. Moving one down is outside its bounds at every depth.
• Tested on past cases. Before go-live, run the agent against the practice's own historical requests and check every under-triage, not just the overall match rate.
• Verbatim capture. The patient's own words are kept alongside any summary, so the clinician sees what was said, not what the agent understood.
• A record per request. Each routing, escalation and proposal is filed with the rule that applied.

Health data is special-category data under GDPR and protected health information under HIPAA in the United States, so the agent sees only what triage needs.