Two-Way Appointment Reminders at 500 Calls a Day: Pacing, Voicemail Detection, Caller ID and the Reschedule Flow

Patients rarely forget an appointment. They hit a conflict, lose the prep instructions, or cannot get through the front-desk line fast enough to move the slot. The rigid "reply 1 to confirm, 2 to cancel" message was built for a clinic that only wanted a yes or no. It solves the wrong half of the problem.

Why a one-way reminder leaks appointments

A patient who gets a confirm-or-cancel message has three realistic choices.

1. ▸Confirm. Easy. The appointment holds.
2. ▸Cancel. The slot is gone. The message offers no alternative, so the clinic loses the visit unless the patient calls back.
3. ▸Reply with anything else. "Can I drink coffee before bloodwork?" or "Can I move it to Friday?" The reply lands in a shared inbox or an error. The front desk sees it hours later, if at all.

Options 2 and 3 are where slots leak. The patient did not forget. They had no way to reschedule inside the message.

Staff could follow up by phone, but the arithmetic works against them. Here is an illustrative example. Take 500 calls a day at about 2 minutes each. That is 1,000 minutes, roughly 16 to 17 hours of talk time, before dialing, ringing, voicemail and charting. It is several people on the phone all day, competing with every walk-in. So the calls get skipped, batched badly or left as voicemails.

What a two-way agent does instead

A two-way agent holds a conversation instead of broadcasting a message. It can run over WhatsApp, SMS or voice. When the patient replies "I can't make Tuesday, what about Friday morning?", the agent follows five steps.

1. ▸Parse the intent (reschedule) and the parameters (Friday, morning).
2. ▸Call the practice-management or EHR API for that provider's availability.
3. ▸Offer two or three real slots.
4. ▸On selection, write the booking and release the original slot.
5. ▸Send the prep instructions that match the new appointment type.

The language model handles the conversation. The schedule does not live in the model. Availability, provider rules and slot locking sit in your database and reach the agent as strict API tools.

The model never sees the full calendar. It sees only the slots a tool returns. If the list is empty, it says so.

In the Gulf, put WhatsApp before SMS. Patients already message clinics there, and a WhatsApp reminder is more likely to be read than a text. Confirm the channel mix with a week of your own inbox data. Replies arrive in dialect and mix Arabic with English, so test on real patient replies. During Ramadan, shift the calling window away from the afternoon fast.

The metric to watch is the reschedule-to-cancel ratio. It shows whether patients who would have cancelled are now moving their slot instead. Your own data shows it within the first week.

Latency: one paragraph

If the agent takes longer than roughly half a second to answer after the patient finishes speaking, the patient assumes the line is dead. The usual target is 400 to 500 ms from end of speech to first audio. Reaching it takes streaming transcription, predictive endpointing, a small fast model and streaming speech output. The full latency budget and working streaming code are in the voice latency post.

Writing to the EHR without inventing slots

An agent that talks but cannot update the calendar is an expensive answering machine. The value is in the state change. Which EMRs allow a schedule write at all is the first thing to settle, because older systems may need a gateway or offer read-only access.

The failure to avoid is the model inventing a Thursday 2 pm slot and promising it. The result is a double-booked doctor or a payload the EHR silently rejects. Three guardrails prevent it.

Tool calls, not free text. Slots come from a function call to the practice system. The model picks from the returned list.

Schema validation before the write. Every outgoing payload is checked against a strict schema, such as Pydantic or Zod. Invalid payloads never leave the orchestration layer.

Narrow write scopes. The agent can confirm, cancel and reschedule. It cannot read medical history, change billing or touch anything outside scheduling. If the EHR times out, the agent reads a fallback line and flags the call for a human callback.

Concurrency and pacing

You do not fire 500 calls at 9 am. Carriers flag a number that does this as spam, which collapses answer rates. And patients who miss the call and dial back all land on your front desk at once.

Production pacing looks like this.

• ▸Pull tomorrow's appointment list from the practice system the night before.
• ▸Randomize the order so three family members are not called in the same minute.
• ▸Hold concurrency to a small number of active lines, tuned to your inbound capacity and your carrier's tolerance. 5 to 15 is a starting range, not a rule.
• ▸Spread the calls across a window the clinic wants, usually mid-morning to mid-afternoon.
• ▸Retry no-answers once, later in the window, from the same number so the callback shows a familiar caller ID.

If a reasonable share of called patients dialed back right now, your front desk should still answer within a few rings. Twilio documents a default limit of 1 outbound call per second, raised for accounts with an approved Business Profile.

Caller ID and telecom rules by country

Caller-ID trust and anti-spam rules differ by jurisdiction. Do not apply US guidance to a Gulf deployment or the other way around.

United States. STIR/SHAKEN signs outbound caller ID so carriers can verify the number is not spoofed. For high-volume outbound, register the number through a trust registry so it shows the clinic name and avoids a "Scam Likely" label. Allow weeks of lead time. The FCC ruled on February 8, 2024 that AI-generated voices count as "artificial" under the TCPA, so these calls need prior express consent. What a TCPA-defensible system has to log for every call is covered in its own post. HHS states that appointment reminders count as treatment, so the HIPAA Privacy Rule does not require patient authorization for them.

UAE. The TDRA regulates telecoms and publishes its rules on its site. Carriers there may block or strip unregistered caller IDs on automated outreach. Confirm the registration requirement with your carrier before launch. Local SIP termination through a licensed provider is the clean path, because international SIP into UAE mobiles is often filtered.

Saudi Arabia. The CST regulates numbering and anti-spam. Confirm with your carrier what a licensed local termination requires. International origination can be filtered there too.

Both Gulf markets. Use a local number and keep the greeting short. Patients hang up on long robotic introductions.

Voicemail

Voicemail detection is a separate problem from conversation. Standard speech-to-text cannot reliably tell a live "hello" from a recorded "hello, you have reached."

Two signals combined work well. Answering machine detection at the telephony layer looks at the first second of audio. Semantic analysis of the transcript catches recorded greetings it missed. Twilio documents strict ranges for its detection settings, such as a MachineDetectionTimeout between 3 and 59 seconds.

Once voicemail is confirmed, wait for the beep and play a pre-generated audio file. Do not stream a live model response. The script is short and the same every time.

"Hi, this is an automated reminder from [clinic name] for your appointment on [day] at [time]. Please reply CONFIRM to the text we just sent, or call us back on this number."

What outbound reminders do not solve

Reminders cut some no-shows, not all. Patients who ignore the call still miss, and some who confirm still miss.

The one-way message is the weak link, so the fix is to make every reminder a conversation. A two-way flow gives a cancelling patient an alternative slot in the same thread. It routes a question to an answer instead of a dead inbox. And it keeps the model away from the full calendar.

Clinical questions are the boundary. A reminder agent must never triage. Lock its scope to scheduling, and send any question about results, dosage or symptoms to a human through a handoff.

For cost, see a per-minute cost model with containment rate. It counts the humans you still pay, which a variable-cost figure hides.

Frequently Asked Questions

What should concurrency be set to? Low enough that your front desk can absorb the callback wave. Start small, watch the inbound queue for the first week, and adjust. Your carrier and your inbound capacity set the ceiling, not the dialer.

Will this integrate with our practice management system? If it has an API that reads availability and writes appointment status, yes. Older on-premise systems sometimes need an HL7 or FHIR gateway in front. Without programmatic write access, the agent can still make the calls and deliver a daily report for staff to apply.

What happens if the agent tries to book a slot that is no longer available? It cannot. The booking write is a transaction that fails on conflict, and the agent offers another slot from the tool.

Pull tomorrow's list tonight and count how many patients would land on your desk if 10% called back inside five minutes; that number sets your concurrency.