What clinical questions should a voice agent escalate?
A pharmacy voice agent should escalate anything clinical, anything it is not confident about, and anything the caller asks a human for. In practice the boundary is simpler than it sounds: the agent handles questions about the state of a prescription, and a pharmacist handles questions about the medication itself.
Status versus substance. "Is it ready" is status. "Should I take it with food" is substance, and it belongs to a pharmacist.
Where is the boundary?
The workable dividing line is between the *state of a prescription* and the *medication itself*. Whether an order is ready, when it can be collected, whether a delivery went out, whether the pharmacy is open — these are facts about a record and a shop. They are not clinical, no matter how they are phrased.
Anything about the drug — what it does, how to take it, what it interacts with, why the patient feels strange since starting it, whether they should continue — is substance. It requires clinical judgement, it is a pharmacist’s job, and it does not become the agent’s job because the question sounded simple.
That last point is where these systems fail. "Can I take this with my other tablets" sounds like a lookup, and answering it correctly requires knowing the whole medication list, the patient, and the context. An agent that answers it because the retrieval was easy has made a clinical decision it had no business making.
What should always escalate?
These are not edge cases to be tuned later. They are the design, and they should be conservative by default — the cost of an unnecessary escalation is a pharmacist’s minute, and the cost of a missed one is unbounded.
Anything about interactions, dosing or side-effects
The core clinical territory. No exceptions, no confidence threshold high enough, no "but this one is well documented".
Any symptom report
A patient describing how they feel has moved out of scope entirely. The agent captures it and routes it — it does not assess whether the symptom sounds serious, because assessing that is the clinical act.
Should I keep taking this?
A question about whether to continue a therapy is a clinical decision with real consequences either way. It is also frequently the moment a patient is about to stop, which makes it the most valuable call of the month to get to a pharmacist.
Anything the agent is not confident about
Low confidence is an escalation trigger, not an invitation to improvise. A system without a well-calibrated notion of its own uncertainty should not be on the phone with patients.
Any request for a human
Immediately, without a retention attempt. An agent that makes a patient ask twice has destroyed the trust the channel depends on to save one call.
Distress, confusion, or vulnerability
A caller who is upset, disoriented or clearly struggling needs a person. The agent’s job is to recognise it early and get out of the way.
Anything suggesting an emergency
Routed immediately, with the caller directed to emergency services rather than held in a queue. This path should be the most tested one in the system and the one nobody optimises for cost.
Why draw it conservatively?
The costs are wildly asymmetric. An over-escalation costs a pharmacist a minute and mildly annoys them. An under-escalation is a patient who received a clinical answer from software that was not qualified to give one — and the consequences of that range from a bad outcome to a regulatory event to a person being harmed.
There is also a regulatory dimension that is easy to miss. Software providing information used for a therapeutic decision is treated as a medical device in a growing number of jurisdictions — in the EU, MDR Rule 11 is written about exactly this, and the EU AI Act adds obligations on top. An agent that answers "should I keep taking this" may have quietly reclassified the product it lives in. Escalating is not just safer for the patient; it is what keeps the system on the right side of that line.
And the commercial argument runs the same direction, which is convenient. The routine calls are the volume. Conceding all of the clinical ones costs little of the value and removes most of the risk — refusing to concede them buys a small amount of extra deflection at the price of the whole deployment.
What does a good escalation hand over?
The transcript, not a summary of it. A pharmacist should be able to read what the patient actually said, in their words. Summarisation is lossy in exactly the places that matter.
The trigger. Which rule fired, or what the agent was unsure about. A pharmacist picking up a call should know why it reached them in one line.
The context. Who the patient is, what they are on, what the agent had already established. The entire point of escalating well is that the patient does not start again from zero — an escalation that makes them repeat themselves has spent the goodwill the automation earned.
And it should be a live handover where the patient is still on the line. A "we will call you back" is not an escalation, it is a queue with better manners.
Frequently asked
- What clinical questions should a voice agent escalate?
- All of them. The workable line is between the state of a prescription — is it ready, when can I collect it, are you open — which an agent can handle, and the medication itself — dosing, interactions, side-effects, whether to continue — which belongs to a pharmacist. Also escalate any symptom report, any request for a human, any sign of distress, anything suggesting an emergency, and anything the agent is not confident about.
- Why should the escalation boundary be conservative?
- Because the costs are asymmetric. An unnecessary escalation costs a pharmacist a minute. A missed one means a patient got a clinical answer from software not qualified to give it. There is also a regulatory dimension: software providing information used for a therapeutic decision is treated as a medical device in a growing number of jurisdictions — EU MDR Rule 11 is written about precisely this.
- What should a voice agent hand over when it escalates?
- The transcript in the patient’s own words rather than a summary, the specific trigger that caused the escalation, and the context the agent had already established — so the patient does not start again from zero. It should be a live handover while the caller is still on the line; "we will call you back" is a queue with better manners, not an escalation.
- Should a voice agent answer "can I take this with my other tablets"?
- No. It sounds like a lookup, which is exactly what makes it dangerous — answering it correctly requires the full medication list, the patient and the context. An agent that answers because the retrieval was easy has made a clinical decision it had no business making.