You set the contact policy: how many messages, at what hours, and when the agent stops

When an operations director hears that an agent is going to message their patients, the question that follows is operational: how many times, at what hour, and what happens the day a patient gets annoyed. Here is the answer, detailed enough to forward to your compliance officer.

We arrive with a default: a policy distilled from the protocols contact teams already have written down, plus what the channel imposes. Every number and every stop rule on this page is a parameter your organization sets.

Who sets the contact policy?

You do. Implementation consists of walking every dial with you until the written policy is your organization's. We publish it openly because the objection that kills these projects is that a machine will talk to patients, and that objection gets answered with caps and a record.

This is the same principle running through the whole product, the one we explain in how we work: your protocols and your rules, on top of Surface, followed with 99.9999% accuracy. The agent executes your criteria the same way on a Tuesday at 9 in the morning as on a Saturday at 7 in the evening.

How many times does the agent write, and over what span?

By default, 3 proactive messages per patient over 14 days: the first within 24 business hours of the order entering the system, the second at 72 hours if there was no reply, the third around day 10. Once the last attempt is spent, the case closes with its disposition and returns to the team's queue.

The cap is counted per patient and does not multiply with each order. Someone with three open referrals gets a single conversation covering all of them and a proposal that packs them into the fewest trips. Asking a patient for three separate leaves from work is the fastest way to make them attend none.

At what hours and on which days, and why those windows?

The 20:00 cutoff comes from contact teams themselves: it is the rule they already have written for their own people. An agent should not hold a permission a human scheduler does not hold. The 9:00 floor follows the same logic.

The window applies to messages the agent starts. If the patient writes at 1 in the morning, the agent answers and books the appointment. The asymmetry is deliberate: a patient who writes and gets no answer within minutes abandons the chat, and that abandonment is the largest source of bookings that fall through.

What makes the agent stop?

• The patient books, and only the reminder for that appointment remains.
• The patient asks not to be contacted, in any wording.
• The patient asks something the agent cannot resolve.
• The appointment was already booked elsewhere, or the order is clinically cancelled.
• The last attempt is spent with no reply.
• The patient blocks or reports the number.

All six are evaluated before each message is composed, against the state of the system. Opt-out is read by intent: “stop writing to me”, “I already had it done elsewhere”, “leave me alone” and a thumbs-down are all an immediate stop.

Meta's messaging policy requires the same thing: a business must respect all requests, on and off WhatsApp, to block, stop or otherwise unsubscribe. Complying is a condition for keeping the channel running.

What does the agent never say?

• Diagnoses, or any interpretation of test results.
• Clinical advice. If the patient asks whether their exam is urgent, the case goes to a person.
• Invented availability: it offers slots that exist at the moment of offering and confirms before closing.
• Pressure. No countdowns, no manufactured scarcity, no repeating the same question until the patient gives in.
• Prices beyond what your organization publishes, or promises of coverage.
• Data it does not need in order to book.

An open chat invites any question, and that is where a generic assistant gets into trouble. The agent's scope ends at handling a medical order already written by a professional at your organization. Anything outside that goes to a person.

When does a human take over, and how does the handoff work?

1. The agent handles the normal path

   It identifies the order, maps it to the bookable service, offers times compatible with preparation rules, and closes the booking.

2. It detects the edge of its scope

   A clinical question, a complaint, irritation, data that does not fit, a situation outside protocol. Any of these flags the case.

3. It stops writing

   From the moment the case is flagged, the agent sends nothing further. Two voices never work the same chat.

4. It hands over the full thread

   The scheduler receives the entire conversation, the original order, the times offered and what has been tried. The patient repeats nothing.

5. The case closes with a disposition

   Without a call disposition the interaction stays open and behaves like an unmanaged referral.

What can the patient always do to end the conversation?

Three things, and all three work without depending on anyone: reply that they do not want to be contacted, block the number from WhatsApp, or ask a human scheduler on another channel, because the opt-out is applied in the same system that governs the agent.

The exit instruction goes in the first message of every sequence. Meta asks for it explicitly: a business must provide clear instructions for how people can opt out and honor those requests, and clearly state the name of the business doing the writing.

There is also a mechanical punishment for ignoring this: people can block or report businesses, and Meta limits how many messages a business with a low quality rating can send over a sustained period. Harassing patients degrades the very channel you later need to confirm appointments.

Where is the line between a reminder and harassment?

Start from the legal definition that does exist. Chile's Ley 19.496 defines advertising as «la comunicación que el proveedor dirige al público por cualquier medio idóneo al efecto, para informarlo y motivarlo a adquirir o contratar un bien o servicio», and its article 28 B grants the right to request suspension of promotional or advertising communications.

SERNAC runs that right through the No Molestar registry, whose page states the scope: it suspends advertising and promotion, and leaves out debt collection and informational calls. A message about an order the patient's own doctor wrote two weeks ago lives in that second group. The final call belongs to your legal team.

Even so, the default behaves as if every message fell under the stricter rule. That is what makes an appointment reminder defensible in front of a regulator or a complaint: a single verifiable purpose, an attempt cap, an hours window, a visible exit, opt-out by intent, and a record showing it patient by patient.

Three flows travel the same channel and deserve different caps: handling an open order, reminding about an appointment already booked, and the day-before appointment confirmation. Mixing them is the shortest route to patient harassment, because the patient counts the total messages received, adding up all three.

A patient who feels hounded files a complaint, and that complaint arrives with your organization's logo at the top. That is why the default is conservative and why the ceiling is yours to set.

How do you change any of these parameters?

You ask in writing and it runs the same day. Moving the window from 20:00 to 19:00, dropping from 3 attempts to 2, excluding an entire contract, or splitting cadence by unit are configuration changes that apply and revert within hours.

Every change is versioned with a date and an author, and every message sent is tied to the version in force. When someone asks why this patient was messaged on a Saturday, the answer comes out of a timestamped record.

Watch the effect before rolling it out across the network. Raising attempts raises booking and, past a point, also raises no-show, because part of those appointments get taken under pressure. So the dials move against measurement, with the denominator in plain sight, and implementation starts small: one unit, one order type, two weeks.

The default is what we hand you on day one. From there, you write the policy.