AI for post-treatment follow-up calls: a safer operational model

Post-treatment follow-up is a deceptively difficult workflow. A clinic may need to check whether a patient received the discharge information, confirm a scheduled review, answer an administrative question, collect a photo or document and make sure a coordinator sees anything that needs attention. Each task sounds simple in isolation. Across languages, time zones and a busy patient coordinator team, however, missed calls and incomplete notes can turn a routine follow-up into a preventable service problem.

Voice AI can help when it is designed as an operational layer around the care team. Dring's healthcare tourism workflows and human handoff model are useful starting points: the agent can manage approved questions, remember the case context, continue on WhatsApp or SMS when a file is needed and route a clinical concern to an authorised person. It should not diagnose, prescribe, change treatment or create the impression that a model is a clinician.

Separate follow-up work from clinical work

Before building a script, divide the workflow into three lanes. The first is administrative follow-up: appointment date, arrival instructions, document collection, transport details, payment status and language preference. The second is approved education: information already reviewed and signed off by the clinic, such as how to access a patient portal or where to find a written aftercare instruction. The third is clinical concern: pain, symptoms, medication questions, unexpected changes or anything the patient believes is urgent.

The first two lanes can often be supported by an agent if the clinic approves the knowledge and escalation rules. The third lane should create a clear human route. The distinction must appear in the agent's policy, the test set and the CRM outcome, not only in a training document. A patient should never have to convince the system that their question is important enough to reach the care team.

Start with a consent-aware opening

A follow-up call should explain why it is being made, identify the organisation and make it easy to continue or stop. The opening can be short: “Hello, this is the clinic's follow-up team calling about your recent appointment. Is now a suitable time for two quick administrative questions?” If the workflow involves recording or analysis, use the clinic's approved disclosure and recording policy. If the person is not the intended patient, do not reveal treatment details. A wrong number is an operational stop, not an invitation to ask for more information.

Language choice is also part of consent and comprehension. Dring's 62-language technical capability inventory spans voice, WhatsApp, SMS and email, but coverage is not a reason to assume that every translation is suitable for every clinical context. Ten languages are public launch priorities, and every requested locale/workflow is validated on the actual path before production. Keep patient-facing phrases short, use the clinic's approved terminology and provide a human route when the person says they do not understand. The multilingual terminology governance guide describes how to keep these variants under review.

Use a follow-up decision tree patients can understand

Good follow-up is not a long questionnaire. It is a small number of decisions that lead somewhere. First confirm identity using the clinic's approved method. Then ask whether the patient received the information they were expecting. If a document, image or link is needed, move the interaction to the appropriate channel without losing the case. If the person asks an approved administrative question, answer it and confirm the next action. If the person reports a concern outside scope, say clearly that a trained member of the team needs to review it and set the expectation for the callback or handoff.

Avoid vague reassurance. “That sounds normal” is unsafe when the agent cannot assess the situation. A better phrase is: “I cannot assess medical symptoms on this line. I will record what you have told me and connect you with the clinic team that can review it.” The wording is calm, useful and honest about the boundary.

Turn every call into a care-team-ready record

The value of an automated follow-up is often the quality of the record it leaves behind. A useful CRM note contains the contact result, language, consent state where applicable, appointment or case identifier, question category, information sent, escalation reason, urgency stated by the patient, owner and next promised action. Do not store a free-form transcript as the only record. Structured fields make it possible to audit missed follow-ups and improve the workflow.

Dring's call analytics and Sector Insight can be used to group recurring questions and service friction. For example, if many patients ask where to find the same aftercare document, the clinic may have a communication problem rather than an agent problem. If people repeatedly ask for a human after a particular phrase, that moment belongs in the escalation design and the next simulation set.

Build the exception library before launch

Healthcare follow-up fails in the edge cases: a family member answers, the patient is travelling, the appointment moved, the patient says the instruction was never received, the requested language is unavailable, the line quality is poor or the caller mentions a symptom. Write these cases before the first production call. Each case needs an allowed response, a prohibited response, an outcome and an owner.

Use the Agent Factory to turn the library into simulated conversations. The factory can generate variations in wording, interruptions and language so the team reviews behaviour rather than a single rehearsed sentence. Release only after the clinic has reviewed the difficult paths and agreed what “ready” means.

Measure service quality without hiding the risk

Measure more than answer rate. Useful indicators include completed follow-up rate, correct appointment confirmation, document delivery success, time to human review, handoff completion, repeat contact within a defined window and reviewer-rated safety of the response. Track unanswered calls separately from resolved cases. A call that ends quickly because the patient gave up is not a successful outcome.

Review a sample of calls by language, workflow and escalation type. Look for overconfident language, missed human requests, incorrect identity handling and notes that do not tell the care team what to do next. The quality program should have authority to pause a release when a new risk pattern appears. The goal is not to make the agent autonomous at any cost. It is to let coordinators spend more time on conversations that need their judgement.

A practical post-treatment follow-up checklist

• Define administrative, approved-information and clinical lanes.
• Use the clinic-approved identity, consent and recording language.
• Support the patient's preferred language and offer a human route when comprehension is uncertain.
• Never diagnose, prescribe or reassure beyond the approved scope.
• Pass structured context to the care team when escalation is needed.
• Test wrong numbers, relatives, silence, interruptions, language changes and symptom disclosures.
• Measure completion, handoff quality, repeat contact and reviewer safety, not only call volume.

Post-treatment care is where automation earns trust through restraint. A well-designed agent makes the simple parts easier, keeps the case together across channels and brings a qualified person forward when the question becomes clinical. The patient experiences a clear next step. The clinic gets a better operating record. That is the right ambition for voice AI in follow-up.

Further reading

• WHO on safe, ethical and supervised AI in health
• WHO's principles for the design and use of AI in health
• NIST AI RMF on continuous measurement and management