Patient access agents on Agentforce: what hospitals should build first

12 October 2026
Mostafa Sabbah

A hospital group’s patient access team handles the same requests all day. Book me with a dermatologist next week. Move my appointment. Has my referral arrived? What do I bring, and do I need to fast? They arrive by phone, web chat and WhatsApp, in Arabic and English, often after the contact centre has closed, and the answers sit in the booking system.

AI agents can carry much of that work. The hard part in healthcare is the boundary: a conversation that starts with a booking can turn to a symptom in the next message, and the agent has to recognize that moment and hand over. This guide sets out the four agents we would build first on Agentforce, what each needs, where clinical staff stay in charge and how to test them. It builds on our guide to Health Cloud for Gulf hospitals and our guide to hospital websites on SitecoreAI.

Four agents worth building first

Salesforce’s action reference lists standard Health actions for agents: slot search, booking, waitlists, scheduling eligibility, provider matching and identity verification. Health Cloud’s invocable actions add appointment updates, cancellations and referral creation.

Appointment requests and rescheduling, for patients on web chat and WhatsApp. The agent verifies the patient, finds slots for the right clinic and doctor, books, moves or cancels, and offers the waitlist when nothing fits. It works against whichever system owns each clinic’s schedule.
Referral status and pre-visit instructions, for patients and referring clinics. It tells a verified patient whether a referral has arrived, is under review or has been booked, and sends the department’s instructions: documents, arrival time, preparation and parking. The Health Cloud page describes AI that prioritizes referrals, and a clinician confirms each priority.
Follow-up and reminders after a visit or discharge, for patients on WhatsApp, strictly non-clinical: the next appointment, a lab booking, how to collect a report. Salesforce says its service agent only supports inbound conversations and closes a session 72 hours after it starts, so reminders go out from flows, the agent answers the replies, and longer follow-up lives in a case or a care plan task.
Contact centre assist, for staff in the console. Before the call is answered, an employee-facing agent summarizes the patient’s history of requests: open cases, upcoming and missed appointments, referrals, complaints and preferred language. The Health Cloud page says Agentforce gives contact centre teams real-time electronic health record (EHR) data and AI summaries. We start with the request history and keep clinical summaries for a clinically governed project.

Agent	Channel	Data and systems it needs	Actions it may take	Where a person takes over
Appointment requests and rescheduling	Web chat, WhatsApp	Patient identifier, clinics, doctors and visit types, open slots from the scheduling engine	Verify identity, find slots, book, move or cancel, offer the waitlist	Any symptom or urgency, a booking the rules do not allow, a request for a person
Referral status and pre-visit instructions	Web chat, WhatsApp	Referral status, approved instructions by department and language	Report status, send instructions, book once the referral is accepted	Priority, rejected referrals, questions about the reason for referral
Follow-up and reminders	WhatsApp, with reminders sent from flows	Appointments, discharge and no-show events from the hospital information system (HIS), consent by channel and purpose	Confirm, move or cancel the next appointment, book a lab slot, record feedback	Symptoms, results, medication, distress and complaints
Contact centre assist	Console, for staff	Cases, appointments, referrals and complaints	Summarize the request history, draft a callback note	Always: staff decide and send everything

What the agents need

Records and identifiers. Patients, appointments and referrals in Health Cloud, linked to the HIS by one identifier. Salesforce lists the Data Cloud User permission among the requirements for the Health booking action, so plan Data 360 (formerly Data Cloud) into the design.
Identity before personal data. Health Cloud’s identity verification actions give an agent the questions to ask and check the answers. Until a patient passes, the agent answers general questions only: locations, visiting hours, parking and how to book. After that, bind every action to that patient’s record.
The HIS and the schedule. Slots, bookings, registrations and status events come from the HIS or the EHR’s scheduling engine through MuleSoft or your integration layer. The agent calls the same interfaces as the contact centre and gets no other access. Callbacks and complaints become Service Cloud cases with the transcript.
Knowledge with owners. Pre-visit instructions and visiting rules live as knowledge articles in Arabic and English, each with a clinical owner and a review date, matching the website. Salesforce’s multi-language guide says the retriever filters out articles whose language does not match the conversation, so every Arabic answer needs an Arabic article.

Guardrails and handover

The Dubai Health Authority’s policy on AI in healthcare, issued in 2021 for DHA-licensed facilities, asks that AI be supervised or monitored by its end users, that designated users can override or reverse significant decisions, and that it be clear which functions are AI-enabled. So tell patients at the start that they are talking to an AI agent and how to reach a person, and write these rules into every patient-facing agent:

Rules that stay with people. The agent never diagnoses, interprets symptoms or results, or advises on medication. It gives no triage advice, including whether to come in, wait or go to an emergency department. When a message suggests an emergency, its first reply gives the emergency number, 998 for an ambulance in the UAE or the local number elsewhere, and it passes the conversation to a person at once. It verifies identity and checks consent before it discloses or changes anything about a patient, and it sends reminders only on the channels and for the purposes the patient agreed to.

Write the triggers into an escalation subagent (subagents are what Salesforce used to call topics): symptoms, results, medication, distress, complaints, a request for a person, and anything the agent cannot resolve. Salesforce’s deployment unit shows the escalation flow passing the session and conversation history to a queue, and says escalation is attempted only once per session: if nobody is free, the agent resumes. So the emergency reply comes first, from the instructions, and never waits for a free nurse. Give clinical handovers a queue staffed by nurses, and route Arabic conversations to people who read Arabic.

Testing before launch, measuring after

Salesforce lists Arabic as Beta for Agentforce and warns that beta languages may retrieve information inconsistently. Its reference page for the Health booking action adds that the product’s user interface is available only in English and may not be fully supported in other languages. Build the test set from real requests: Gulf and Egyptian dialects, Modern Standard Arabic, Arabic typed in Latin letters, and English clinic and drug names inside Arabic sentences. Add boundary cases on purpose, such as chest pain mentioned in a rescheduling request, and check that each one reaches a person.

Salesforce’s testing tools cover the build. The Agentforce Builder preview simulates a conversation with mock data or runs a live test on real data. Session tracing records each step and needs Data 360. Test Suites, labelled Beta, run batches of cases against the expected subagent and action, in a sandbox. Set the end user language in preview to test each language, as the multi-language guide shows for the legacy builder.

After launch, measure against a baseline taken before it: containment (requests finished without a person), handover quality (whether staff could act without asking again), the routing of clinical and emergency messages, booking accuracy against the HIS, and complaints. A clinician and an Arabic reviewer read a sample of conversations every week.

Our work in healthcare

Arcsen is a Salesforce Summit partner with 170+ consultants and 400+ certifications and offices in Dubai, Riyadh, Jeddah, Doha and Cairo, and we deliver Arabic-first. Our healthcare clients include Almoosa Health, SSMC, AlJazira Healthcare and Kaizmed. See how we deliver Agentforce in Arabic and English.

Where to start

Sort a sample of real requests into the four jobs above, in both languages, and mark the ones that touch clinical questions.
Write the clinical boundary with your medical director: the words and situations that trigger the emergency reply or a clinical handover, in Arabic and English.
Pick one agent and one clinic, usually appointment changes, where slots and identifiers are already clean.
Build the test set before the agent, with routine requests, boundary cases and emergencies in both languages, and rerun it after every change.
Agree the measures and the baseline, and name who reviews conversations each week.

Planning a patient access agent? The free Agentforce Readiness Assessment picks your first three use cases, checks the data and permissions they need and gives you a costed pilot plan in ten working days. See our healthcare work, talk to our healthcare team or send us your RFP.

Frequently asked questions

Can an Agentforce agent book hospital appointments?

Yes. Health Cloud has standard agent actions to find slots, book and manage waitlists, and invocable actions to update and cancel appointments. The agent verifies the patient first.

Should a patient access agent give triage advice?

No. Diagnosis, triage and medication stay with clinical staff. The agent recognizes symptoms and emergencies, gives the emergency number at once and hands the conversation to a person.

Can the agent send reminders on WhatsApp?

Not the service agent itself, which handles conversations the patient starts. Reminders go out from flows, and the agent answers the replies. Send them only where the patient agreed to them.

Does Agentforce work in Arabic for patients?

Salesforce lists Arabic as Beta for Agentforce. Test with real Arabic and mixed-language messages, write Arabic knowledge articles, and route Arabic handovers to Arabic-speaking staff.

Sources

Mostafa Sabbah

Mostafa leads technical architecture at Arcsen: integration, data, identity and security design across Salesforce clouds. 19x Salesforce certified, including System and Application Architect.