Patient Agent

The patient agent is a patient-facing, asynchronous conversational agent that a patient talks to in their own time, over the channel they already use (SMS or WhatsApp via Twilio). It only ever replies to a message the patient sent — it never starts a conversation on its own.

What it can do

Capability	What the patient can do	What the agent may do on its own
Appointments	Ask for a new appointment, move an existing one, cancel one, ask when their next one is	Read the patient’s own appointments; book or move one directly into an open slot on a care team member’s published availability; cancel outright
Health education	Ask questions about their condition, an upcoming procedure, pre-procedure preparation, or medication they were given	Answer from the service’s own content library and the patient’s own context only; hand off anything outside that scope
Symptom triage	Report a new or changed symptom between consults	Classify urgency; on a red flag, immediately redirect to emergency services and escalate to whoever covers the escalation role
Education & counselling	Go through a counselling or education protocol, step by step	Take consent, walk through the protocol from the content library, and record what the patient understood or agreed (with the Education & counselling sessions tool)
Full encounter	Need a longer conversation than SMS allows	Send a link to a full AI encounter on the patient portal (with the Hand over for a full encounter tool)

Do you need one?

The patient portal’s own assistant already answers patients’ questions. Turn on a service’s patient agent only if the service needs its own conversational agent over SMS or WhatsApp. It is off by default.

Setting it up

1. Author the agent

Create a service agent with Audience = Patient on the Agents tab. It gets the patient tool set, and the platform safety preamble is always added on top of your instructions.

2. Bind it in Settings

On the service’s Settings tab, the Patient agent card:

Let patients chat with this service’s patient agent: the on switch.
Patient agent: a dropdown of this service’s active Patient-audience agents. If there are none, it points you to the Agents tab.
Delivery mode:
- Review before sending (recommended): every drafted reply waits on the patient’s page for a staff member to approve, edit or discard.
- Send automatically: replies go straight out. A red-flag emergency reply always sends immediately, in either mode.
Escalation role: which organisation role (Admin, Doctor, Nurse or Receptionist) is notified on a red flag, a failed turn or a hand-off. It is resolved through Coverage to whoever is covering that role right now.
Emergency number: e.g. 995.
Emergency message: a template, seeded with a platform default, that must keep the {emergencyNumber} placeholder. Sent immediately and unedited whenever the red-flag screen fires.
Stay quiet after a staff reply (minutes): when someone on the care team messages the patient from the Messages tab, the agent stays out of the conversation for this long, so it never talks over a person who has taken over. 0 turns the hold off. An emergency reply is never held.
Offer education & counselling sessions proactively: lets a staff service agent message patients to ask whether they would like to go through a counselling protocol from the Content tab, based on where they are in their care. The patient agent still never starts a conversation; it handles the reply.
- Agent that sends the offer: a staff agent with the Twilio message tool.
- Also check right after an encounter note is saved: otherwise offers are only considered on scheduled agent runs.
Hold routine replies overnight: a From/To time and Timezone. A routine reply drafted in this window waits until it closes; an emergency reply is never held.

Click Save.

Safety design

A few things are true regardless of settings, and are not configurable away:

Red-flag screening runs before the model does, on the raw message text — deterministic keyword/pattern matching (chest pain, can’t breathe, stroke symptoms, heavy bleeding, anaphylaxis, self-harm/suicidal ideation, and more). A hit sends the fixed emergency message and opens an escalation without ever calling the LLM. False positives are accepted; false negatives are the thing this is built to avoid.
It never invents an appointment, clinician name, result, or clinical fact. Appointments only ever come from the appointment tools; anything the agent doesn’t know, it hands off to the care team rather than answering.
It never diagnoses, changes a prescription, or interprets an unreleased result.
It discloses that it’s an automated assistant, not a clinician, at the start of every new conversation and on request.
Structured yes/no replies to an existing scheduled message still go through the normal automated flow — the patient agent only picks up free text and anything not already correlated to an open message, so existing automated confirmations stay deterministic.

Reviewing drafts (review mode)

When the service is in review mode, a “Draft(s) awaiting review” card appears at the top of the patient’s page (above the tabs, so it’s never missed) whenever a reply is waiting:

Read the drafted reply, along with its detected intent and triage band badges.
Either:
- Click Approve & send to deliver it exactly as drafted, or
- Edit the text in place, which changes the button to Send edited reply, or
- Click Discard to drop the draft entirely (nothing is sent).

Approving goes through the same delivery path an auto-mode send would — the patient sees exactly what they’d have gotten in auto mode.

Escalations

An Escalations card on the patient’s page (General tab) lists everything the patient agent has flagged: a Red flag, an Agent-flagged triage, a failed turn (Agent turn failed), an Out of scope hand-off, or something Staff flagged manually. Each shows its severity, status, and timestamp.

Click Acknowledge to mark an open escalation as seen.
Click Resolve once it’s been handled.

If no one was actively covering the configured escalation role at the time, the card notes that the service owner was notified instead as a fallback — a visible signal that the rota needs attention.

Preview chat

Before turning the agent loose on real patients (or after editing an already-bound agent), use Preview chat from that agent’s row on the Agents tab:

Pick a real cohort patient to test against. The preview uses their actual context in this service.
Type messages as if you were that patient and see the agent’s replies in a sandbox chat window.

Nothing in a preview is sent to the patient or saved to their message history — appointment bookings and escalations run in a simulated dry-run mode only. This is the recommended way to walk through appointment, education, red-flag, and out-of-scope scenarios end to end before enabling the binding for real, or before switching a service from review to automatic sending.

Recommended rollout

Start a new patient agent in review mode on a single pilot service, with staff approving every reply for a defined window. Only switch to auto mode after you’ve reviewed the draft/edit rate and are comfortable with what it’s producing.