How should an AI patient assistant handle pre-visit instructions?
A practical workflow for giving approved appointment preparation information, confirming receipt and sending clinical questions to staff.
The short answer
An AI patient assistant can give a patient approved information for the correct appointment, repeat it at a comfortable pace, send the agreed written version and record whether the message was delivered. The workflow needs one controlled source, a clear identity check for appointment-specific details and a route to staff when the patient asks a clinical question.
The assistant should not create preparation instructions, decide whether a missed step is safe, change advice about medicines or fasting, interpret symptoms or certify that a patient is ready for a procedure. It may convey an authorised message. A clinician or another authorised staff member must handle interpretation, exceptions and decisions that affect care.
Define one instruction set for each appointment type
Start with the appointment types that create repeat questions. For each one, the clinic should approve the location, arrival time, documents to bring, contact route and any preparation message that the assistant may deliver. Link the set to the actual appointment type and site. A generic script is unsafe when two departments use different preparation rules.
Give every set an owner, version, approval date and review or expiry point. Separate public logistics, such as parking, from information that reveals the patient's appointment or treatment. If the system cannot find a current approved set, it should offer staff contact instead of assembling an answer from similar records.
Verify the patient before revealing appointment details
Opening hours and a public address may not require an identity check. The time, specialty, procedure or preparation attached to a named patient can reveal health information. The clinic should set a proportionate check before the assistant reads those details or sends them to a stored contact channel.
Outbound calls need particular care because another person may answer. Before verification, use a neutral introduction and disclose that the caller is interacting with AI, without naming a specialty or procedure. If the check fails, stop the disclosure and provide an official route to staff. Do not confirm that a record exists or treat the failed check as evidence of fraud.
Read the current instruction instead of improvising
The assistant should retrieve the approved version tied to the appointment and deliver that content without adding plausible details. A content record should show who owns it, which appointments it covers, when it became valid and whether it has been withdrawn. This makes a correction traceable when a department changes a location, arrival time or preparation step.
Conflicting sources need a stop rule. If the scheduling system points to one clinic while the instruction set names another, or if two active versions appear, the assistant should not choose the answer that seems most likely. It should create a review task, tell the patient that staff need to confirm the details and avoid sending either version as final.
Confirm receipt without declaring clinical readiness
Break a long message into short parts and allow the patient to repeat or replay them. The record can distinguish between delivered, acknowledged, written copy sent, question raised and handed to staff. A yes to 'Did you receive this?' is not proof that the patient understood every clinical implication or can safely proceed.
A patient may say that they ate, missed a preparation step, took a medicine, feel unwell or are unsure whether the instruction applies to them. Those answers change the nature of the call. The assistant should preserve the question and follow the clinic's human route. It should not decide that the appointment can continue, be cancelled or be moved.
Prepare each language and an accessible alternative
Approve the meaning and terminology in every language used in production. Clinical preparation text should not be translated freely during the live call. Test dates, clock times, units, medicine names where they appear in an authorised message, local place names and the way the assistant handles a correction from the patient.
Offer a slower pace, repetition and an approved written copy through a channel already held by the clinic. Some patients will need an interpreter, accessible format or staff member. The assistant should offer that route without guessing a disability, capacity or language ability from the person's voice. A caller acting for the patient also needs a separate authority check.
Send the question to staff with enough context
A useful handoff includes the appointment reference, verification status, instruction version delivered, exact point that caused the stop and the approved contact route. It should not copy an entire call or extra health details into a broad queue when a short, restricted case is enough. Access and retention should match the purpose of the follow-up.
Name the team that owns each exception and what happens when it is unavailable. Questions about medicines, fasting, symptoms, pregnancy, consent, capacity, suitability for a procedure or the effect of missed preparation stay with authorised people. The assistant may measure whether the handoff was accepted. It cannot make the underlying clinical decision.
Test the whole journey, including a withdrawn message
Acceptance tests should cover the wrong person answering an outbound call, two appointments on one day, a changed location, a replaced instruction, an unsupported language, a dropped call, a clinical question and an unavailable source system. Withdraw one test message and confirm that the assistant stops using it across voice and written follow-up.
Measure completed delivery, written-copy success, staff handoff, repeat contacts, corrections and any use of the wrong version. A change in missed appointments can be reviewed, but it should not be attributed to the assistant without a sound comparison. Pause the workflow if testing finds disclosure to the wrong person, outdated instructions or clinical advice outside the approved scope.
FAQ
Can an AI patient assistant answer questions about fasting or medicines?
It may repeat the clinic's current approved wording, but it should not adapt or interpret that wording for an individual patient. Questions about missed fasting, medicine changes, symptoms or whether a procedure can continue need an authorised clinical route.
What if the clinic changes an instruction after the patient has been called?
Withdraw the old version, identify affected appointments and use an approved correction process. The record should show which version was delivered and whether the corrected message reached the patient. A silent content update is not enough.
Does every pre-visit call need identity verification?
No. Public information can often be given without matching a patient. Appointment-specific details and messages that may reveal health information need a proportionate check before disclosure or delivery to a contact channel.
Sources and further reading
- European Commission: principles of the GDPR
- European Data Protection Board: Guidelines 02/2021 on virtual voice assistants
- World Health Organization: ethics and governance of artificial intelligence for health
- World Health Organization: ethics and governance guidance for large multi-modal models
- European Commission: guidelines on AI Act Article 50 transparency obligations