How should an AI patient assistant route calls across clinics and departments?
A practical design for multi-site patient call routing that finds the right administrative destination without making clinical decisions.
The short answer
An AI patient assistant should route a call by confirmed administrative intent, location, service, appointment context and opening status. It should use an approved directory owned by the healthcare organisation, repeat the destination to the patient and finish with a visible result such as transfer completed, callback case created or staff review required.
It should not infer a diagnosis, decide clinical urgency, choose a specialty from symptoms or rank patients for access. If the caller introduces symptoms, deterioration, medication risk or another clinical concern, the administrative flow stops and follows the organisation's approved human or emergency route.
Start with the administrative job, not a department name
Patients often describe what they need rather than naming the right team. They may want to change an appointment, ask where to arrive, send a referral, request a copy of a document, discuss an invoice or leave a message for a clinician. The routing design should first identify that job, then apply location and service rules.
A department name alone is unreliable. Two sites may both have cardiology, while only one handles new referrals or a particular diagnostic test. The assistant should confirm the requested action and the known appointment or referral context before it offers a destination.
Build a versioned directory that the clinic owns
The source of truth should list each site, service, team, supported request, opening hours, transfer number, case queue, fallback owner and exceptions. Give every record an owner and review date. A language model should not invent a route from a public web page or a familiar sounding department name.
Temporary changes need an expiry time. If a reception desk closes early, a clinic moves building or a team changes its callback queue, the old rule should not remain active indefinitely. The case record should show which directory version produced the route so staff can investigate a wrong transfer.
Collect the minimum needed to disambiguate the destination
Useful routing fields may include the caller's preferred language, site, service, appointment date, whether the call concerns a new or existing referral and the requested administrative action. Identity verification should match the action. General opening information needs less proof than changing an appointment or discussing a patient record.
A patient should not have to give a detailed symptom history to reach administration. GDPR data minimisation and protection by design support a narrower approach: collect what the route needs, restrict who can see it and keep clinical details out of an administrative case unless the approved handoff requires them.
Keep routing separate from clinical triage
Administrative routing answers questions such as which team owns the referral, where the appointment takes place and whether a callback queue is open. Clinical triage assesses health information and urgency. The same conversation can cross that boundary, so the workflow needs an explicit stop condition rather than a vague instruction to use judgement.
When a caller asks whether symptoms are serious, whether treatment can wait or which specialty is medically appropriate, the assistant should not improvise. It follows the approved message and transfer path. Extending the product into diagnostic or therapeutic decision support changes its intended purpose and needs a separate clinical, legal and regulatory assessment.
Design transfers and callback cases as equal outcomes
A live transfer is useful only when the destination can answer. Outside opening hours, during overload or when the line fails, repeated blind transfers create another dead end. Each route should define whether to transfer, create a callback case, send approved information or place the request in a named review queue.
If a calendar, HIS, directory or ticket integration is unavailable, the assistant should preserve the confirmed facts and mark the case as pending. It should not claim that a transfer, booking or message delivery succeeded. The patient receives a realistic update time and an approved way to call back.
Confirm the route in the patient's language
Multilingual routing needs more than translated department labels. Test local clinic names, abbreviations, street names and the difference between similar services. Critical details such as site, date and destination should be repeated in the language used during the call.
End with a concrete status: connected now, callback requested from a named team, information sent, staff review opened or urgent human route started. The patient should know what will happen next and when, without hearing internal codes or unnecessary clinical data.
Measure wrong routes, corrections and unresolved calls
Track routing accuracy by destination, transfers that return to reception, manual corrections, duplicate cases, abandoned calls, callback completion and time to the first useful response. Review results by site, language, service and request type. A high answer rate can still hide a poor patient journey if the call reaches the wrong queue.
AI may identify an administrative request, consult approved records, create a case, transfer a call and communicate a status. Authorised people decide clinical urgency, diagnosis, treatment, care priority, eligibility, exceptions and any change to the system's intended medical purpose.
FAQ
Can an AI patient assistant choose the right specialty from a caller's symptoms?
Not as part of an administrative routing flow. Questions about diagnosis, urgency or the medically appropriate specialty require an approved clinical process and authorised people. The assistant can route an existing referral or appointment using confirmed records.
What happens when two locations have the same department?
The assistant should confirm the site and administrative action, then use the organisation's current directory and appointment or referral context. If the records conflict, it creates a review case instead of guessing.
Should every call be transferred live?
No. Each route should define when a live transfer is available and when a callback case, approved information or staff review is the better outcome. The patient must hear the actual status and expected next update.
Sources and further reading
- WHO: Ethics and governance of artificial intelligence for health
- WHO Europe: Knowledge Community on responsible artificial intelligence in health
- EUR-Lex: General Data Protection Regulation
- EDPB: Guidelines 4/2019 on Data Protection by Design and by Default
- European Commission: Guidelines on AI Act Article 50 transparency obligations
- European Commission: MDCG 2019-11 rev.1 on software qualification and classification
- NIST: Artificial Intelligence Risk Management Framework 1.0