AI Agents for Ophthalmology Clinics: Scheduling, Exams, Follow-Ups

AI agents for ophthalmology clinics work best when they take on repetitive, well-defined tasks: offering appointment times, sending the right preparation instructions for each exam and reminding patients about follow-up visits. They use as little health data as possible and hand the conversation to the front desk quickly when a case leaves the script. They do not diagnose, interpret results or decide clinical care. The gain is that reception stops repeating the same messages and has time for the cases that need a person.

Key takeaways

• Start with three flows: scheduling, exam preparation and return reminders. They are repetitive, predictable and low in clinical risk.
• Collect only what is needed to book: name, contact, type of visit, payment method and preferred time. Complaints, reports and images stay out of the agent.
• Exam preparation texts come from the clinic and are reviewed by the medical team. The agent sends them and never improvises.
• Define in writing when the agent must hand off to reception, and make the handoff carry the context already collected.
• Pilot one real flow with conversation logs and an agreed goal, and only then decide whether to expand, adjust or stop.

What can AI do, and not do, in an ophthalmology clinic's daily routine?

People often reach for the word superintelligence on this topic. What a clinic actually operates today is more modest and more useful: an agent with a narrow scope, connected to the schedule and to a set of approved messages. Thinking this way avoids the expectation that AI will solve everything and helps you draw clear limits.

What belongs to the agent:

• Explaining which types of visits the clinic offers and how to book them.
• Checking open slots and proposing options to the patient.
• Recording the patient's choice and confirming the appointment.
• Sending the correct exam preparation, using the text approved by the clinic.
• Reminding patients of booked visits and inviting them to the return visit the doctor has already indicated.
• Alerting reception whenever the conversation leaves the script.

What does not belong to the agent:

• Assessing symptoms or saying whether something is serious.
• Explaining or commenting on exam results.
• Advising on eye drops, medication or changes in treatment.
• Deciding when a patient needs to come back.
• Negotiating financial terms outside a price table that is already defined.

The line is simple: the agent handles logistics and the team handles health. The doctor decides the return. The agent only helps book it.

How do you handle scheduling with minimal health data?

The design question is: what is the smallest set of information that lets you book well? For most clinics the answer is short:

• Patient name and phone number (which already comes with the WhatsApp conversation).
• Type of visit: consultation, exam, return visit or procedure.
• Private pay or health insurance, and which plan, when relevant.
• Preferred day and time of day.

All of this is needed for logistics. Details such as the clinical reason, medical history, photos of prescriptions and reports are not needed to book, and every extra piece of data adds risk and responsibility. Health data is sensitive personal data under LGPD (Brazil's data protection law), so the practical rule is not to ask for what you will not use.

If a patient writes something on their own, such as "I've had eye pain for days", the agent does not respond about the subject. It thanks the patient, says it has passed the message to the team and transfers to reception, which decides the path with clinical support. This behavior must be in the agent's instructions and tested before going live.

On the schedule integration, the agent should only be allowed to read open slots and create a reservation. Editing patient records, deleting appointments or viewing medical charts is not part of its job. The reservation is only saved after the patient confirms, and the system should reject a slot that is already taken. When the insurer requires prior authorization, the agent marks the booking as pending and reception completes it. To understand the mechanics in detail, read automatic appointment scheduling on WhatsApp with AI.

How do you prepare patients for exams without giving clinical advice?

Exam preparation is one of the best uses of AI in ophthalmology, because patients tend to arrive with similar questions and reception answers the same ones every day. The risk is letting the model generate the text on the spot. The solution is to reverse it: the clinic writes and approves the texts, and the agent only picks the right one and delivers it.

A preparation library works like this:

• Each type of exam has its own text, written or reviewed by the medical team.
• Each text has an owner, a review date and a version, so nobody sends an outdated instruction.
• The agent sends the text for the exam listed in the booking, without summarizing or rewriting it.
• Questions outside the text go to reception or to the clinical team.

Examples of items a clinic might include, according to its own guidance: warning that the exam may involve pupil dilation and that vision may be uncomfortable afterward, suggesting the patient bring someone along in those cases, asking the patient to bring glasses and documents, reminding them of the insurance authorization form and advising on contact lenses according to the clinic's rule. These items are illustrative. What applies is defined by the service, not by the agent or this article.

A common question is "Can I take my medication before the exam?" This is a clinical question, and the agent should only say it will pass it to the team and then transfer. This case belongs on the handoff list, with testing beforehand.

How do you remind patients of a return visit without exposing health information?

The doctor indicates the return, usually at the end of the visit, and it is recorded in the schedule or in the clinic's system. The agent does not decide when the patient should come back. It reads that information and invites the patient to book. This avoids a serious mistake: the AI suggesting a return, or dismissing one, that it was never in a position to judge.

In the message text, less is more. A reminder does not need to mention a condition, exam or procedure. Something like "Hello, this is the clinic's reception. We can now arrange your return visit. Would you like to see the available times?" is enough. If someone other than the patient glances at the phone screen, nothing sensitive appears.

For messages the clinic starts, WhatsApp has its own rules. You need the patient's permission (opt-in), approved message templates and a simple way to stop receiving messages. We cover this in WhatsApp outbound messages: templates, opt-in and best practices. It is also worth setting a frequency rule: if the patient does not reply, the agent tries a few times and then leaves the decision to reception instead of pushing.

The same applies to appointment confirmations, rescheduling notices and document reminders. All of them use short messages with no clinical data and a clear option to speak with a person.

When should the agent hand the conversation to reception?

The handoff is what makes the system safe. Without it, the agent tries to answer what it should not. With a badly built one, the patient repeats everything to reception and loses patience. That is why we define the triggers before switching anything on.

Checklist of handoff triggers:

• Any description of symptoms, pain, sudden change in vision, trauma or worsening. The clinic, with the medical team, defines the list of terms and situations that signal urgency.
• Questions about exam results, reports or prescriptions.
• Questions about medication, eye drops or contact lens use.
• Requests for a price outside a defined table, or any negotiation.
• Complaints, dissatisfaction or an irritated tone.
• A patient who explicitly asks to speak with a person.
• A minor, or a third party speaking on the patient's behalf, when the clinic's rule requires confirmation.
• The agent lacking the information to answer, or being uncertain itself.

At handoff, the agent gives reception a short summary: who the patient is, what they asked, what has already been collected and why the conversation was transferred. Outside business hours, it says when the team will reply instead of promising something it cannot deliver. If the conversation involves possible urgency, the message should direct the patient to seek urgent care according to the protocol the clinic has defined, without trying to assess the case.

All conversations and agent decisions should be logged, so the clinic can review errors, adjust texts and account for what happened when needed. This connects to the right of review for automated decisions, covered in LGPD automated decision-making and the right to review.

Illustrative scenario: a day of messages at a clinic

The scenario below is illustrative. It does not describe any real clinic or measured result.

A mid-sized clinic receives messages in the evening. A patient asks for a slot for a routine exam and says she has health insurance. The agent asks which plan, offers three times, records her choice, marks the insurance authorization as pending and sends the approved preparation for that exam. The next morning, reception finds the booking ready, with only the insurance check left to do.

The same day, another patient replies to a return-visit invitation saying that since yesterday he has been seeing points of light in his eye. The agent does not comment on what that might mean. It replies that it is alerting the team now, transfers the conversation to reception with priority and includes a summary of what was said. Reception, following the clinic's protocol, calls in the clinical team.

The value of the flow is in the two extremes: the simple case leaves the queue without occupying anyone, and the sensitive case reaches the right person fast.

How to start: map, limit, pilot, hand over

For clinics, the method we apply in AI projects has four steps, and none should be skipped.

Map: understand how the schedule works today, which messages reception repeats most, where no-shows, reschedules and forgotten returns happen, and which systems are involved. Record a baseline before changing anything, so you have something to compare against.

Limit: decide which data the agent can see, which actions it can take and which questions it never answers. If the model provider processes data outside Brazil, you need to assess the legal basis and safeguards, a topic covered in international data transfer under LGPD when using AI providers. The ANPD (Brazil's data protection authority) publishes guidance on data protection that the clinic should consult, and the data protection officer, where one exists, should be part of the design from the start.

Pilot: choose one flow, for example return reminders or preparation for one type of exam, and run it with conversation logs and an agreed goal, such as fewer repeated messages at reception. Test beforehand with simulated conversations, including ones that try to pull the agent out of scope.

Hand over: give the operation to the team, with clear owners for who reviews the logs, who updates the preparation texts and who answers the handoffs. After the pilot, the clinic decides whether to expand, adjust or stop. Stopping is a legitimate decision.

Summary plan for the first month of a pilot:

• Choose one flow and one type of visit.
• Draft the texts the agent may send and have the medical team approve them.
• Define the list of handoff triggers and who receives each one.
• Set minimal permissions on the schedule and test with simulated conversations.
• Release to a small group of patients who have already agreed to receive messages.
• Review the logs every week and fix texts and rules.
• Compare with the baseline and decide the next step.

To structure the agent and its limits, Anthropic's guide to building effective agents and the NIST AI Risk Management Framework are useful references for thinking about risk and oversight.

Frequently asked questions

Can AI answer patients' questions about an exam result? It should not. Results and reports are clinical content and require the medical team. The agent should recognize the question, say it will pass it on and transfer to reception or to the channel the clinic has defined.

Do I need consent to send return reminders on WhatsApp? For messages the clinic starts, it is best to have the patient's permission on record, in addition to following WhatsApp's rules for message templates. The legal basis for processing the data should be defined with legal support and with the data protection officer, where one exists. In every case, the patient must be able to stop receiving messages easily.

Is it safe to let the agent read the medical chart? At the start, it is not necessary and we do not recommend it. Scheduling, preparation and reminders work with schedule data and minimal registration details. If a future use case requires more access, it should go through a new risk assessment, restricted permissions and review with the clinical team and the person responsible for data.

How long does it take to put a flow like this into operation? It depends more on internal organization than on technology: approved texts, an accessible schedule and clear handoff rules speed everything up. That is why we start with a single flow, with a pilot and log review, and only then expand. Every clinic moves at its own pace, and promising a timeline before mapping the process would not be honest.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework
• Anthropic: Building effective agents