AI chatbots in patient support: what actually works vs. what gets sold
Every health tech conference in 2026 has a panel on AI chatbots in patient care. The pitch is always the same: reduce the burden on clinical teams, improve patient access, scale support without scaling headcount.
The pitch is not wrong. It is incomplete. The gap between what gets sold and what actually works in a clinical setting is where patient safety lives.
What works: routing, triage, and intake. A patient messages at 9pm about their next appointment. The chatbot confirms the booking, provides preparation instructions, offers to reschedule. No clinician needed. No clinical judgment required. Administrative, and automation handles it well.
What also works: structured symptom collection before a consult. The patient answers a guided intake form. History, current medications, reason for visit. The responses are available to the clinician before the appointment starts. The consult is more efficient because the preparation happened asynchronously.
What does not work, and keeps getting sold: clinical advice at the point of patient uncertainty.
A patient finishes a consult. They have questions about dosage, side effects, what happens if they miss a dose. They ask the chatbot. The chatbot responds with information drawn from literature, guidelines, maybe even the patient's own record.
The information may be accurate. The problem is context. The chatbot does not know that this patient's anxiety means they will interpret "common side effect" as a reason to stop taking the medication. It does not know that the prescriber chose a lower dose because of an interaction the patient forgot to mention. It does not know what the clinician would have said, because the chatbot is not the clinician.
The design constraint in clinical AI is not accuracy. It is the boundary between preparation and decision.
A chatbot that prepares information for a clinician to review is infrastructure. A chatbot that delivers clinical guidance directly to a patient is practising medicine without a licence, regardless of how good the underlying model is.
Most implementations get this wrong because they optimise for the wrong metric. The goal becomes deflection: how many patient queries can be resolved without a clinician? That metric rewards the chatbot for answering questions it should be escalating.
The better metric is safety-weighted resolution. Not "did the patient stop asking?" but "did the patient get the right response from the right source?" Sometimes the right response is a chatbot confirmation. Sometimes it is a message forwarded to the clinical team with context. Sometimes it is a flag that triggers a follow-up call.
Building AI into patient support is not the same as building a chatbot that answers patient questions. The first is infrastructure. The second is a liability waiting for its first adverse event.