Superintelligence for Speech Therapy Clinics: What to Automate Safely

Superintelligence, in the sense the market usually gives the term, is not a product a clinic buys to replace the speech therapist. In practice, the part that can be automated safely is the scheduling layer: appointment confirmation, reminders, rescheduling and waitlists, always under rules the professional has defined. Session notes, assessments and any therapeutic decision remain the professional's responsibility and should stay in a system with controlled access.

Key takeaways

• Scheduling and reminders are the best starting point, because mistakes are cheap and the rules are easy to test.
• Health data is sensitive personal data under the LGPD, so the agent should never treat a clinical record like an ordinary conversation.
• Session notes are clinical documentation: the professional writes, reviews and signs them, and AI can, at most, organize a draft under supervision.
• Every agent needs explicit limits, a handoff to a human and a log of the decisions it made.
• A pilot with one real workflow, a success criterion agreed in advance and regular log reviews is more useful than a broad platform from day one.

What can be automated in scheduling

A speech therapy clinic has a repetitive routine: patients ask for times, request changes, forget sessions or wait for an opening. These tasks follow rules that can be written clearly, such as each professional's working hours, the length of each session type, the insurance plans accepted, the gaps between appointments and the deadline for cancellation.

Within those rules, it makes sense to automate:

• Confirming, rescheduling and cancelling slots that already exist in the calendar, respecting what the professional has set.
• Sending reminders at fixed times, with a simple instruction on how to cancel.
• Offering waitlist spots when a slot opens, with human confirmation if the rule requires it.
• Collecting only the minimum administrative data, such as name, phone number and type of service, without asking for clinical information.
• Passing to reception or to the professional any message that falls outside the script.

If the channel is WhatsApp, the design also has to cover what the agent may say and when it may start a conversation. The guide on automatic appointment scheduling on WhatsApp with AI explains how this kind of flow works at a basic level.

Why session records are a different problem

A session record brings together the complaint, observations about speech, language, voice, swallowing or hearing, the approach used and the patient's progress. This is health data, which the LGPD treats as sensitive personal data, with stricter requirements for purpose, security and access control. The rules of Brazil's Federal Council of Speech-Language Pathology on records and technical responsibility should also be read directly from the source, not summarized by an assistant.

There is a difference between organizing and deciding. An agent may, for example, turn notes the professional already wrote into a draft with standard fields. It should not interpret assessment results, suggest a course of treatment, change a signed record or answer a patient about what happened in a session. Any generated draft must be reviewed, edited and approved by the person who treated the patient.

Audio recording of sessions deserves special care. It multiplies the volume of sensitive data and requires a justification, consent and clearly defined storage rules. When in doubt, the practical recommendation is not to record during the first implementation cycle.

How to separate scheduling from records in five steps

Before choosing any tool, the clinic needs to draw the boundary between scheduling and records. A simple way to do this:

• List every task that touches the calendar or the record, and note who performs it today.
• Classify each data item as administrative or clinical, and decide which system it may live in.
• Write the scheduling rules in plain language: length, gaps, deadlines, exceptions and topics the agent never answers.
• Define the handoff point to a person and the hours when the team can respond.
• Choose one workflow for the pilot, with a simple criterion such as the share of reschedules completed without a phone call, and log every decision the agent makes.

This list often shows that part of the problem is not technological. If the calendar has no written rules, no agent will follow them reliably.

Which decisions stay with the professional

Use the list below to decide when a case leaves the agent and goes to the team:

• Any question about symptoms, progress, diagnosis or the effectiveness of a treatment goes to the professional.
• A request for discharge, a change of therapist, a complaint or any sign of urgency goes to a person immediately.
• A change to the therapy plan, frequency or referral happens only through a recorded decision by the speech therapist.
• Access to session records is limited to the clinician treating the patient and authorized staff, with a log of who accessed what.
• Clinical content sent by mistake to the service channel must not be answered by the agent and should be handled as an internal incident.

Illustrative scenario

Illustrative scenario: a clinic with a team of speech therapists wants to reduce missed appointments without overloading reception. The agent sends a reminder the day before, offers rescheduling within each professional's rules and logs every reply. When a patient writes that she can no longer attend because she is feeling unwell and asks whether she should stop therapy, the agent does not give an opinion. It tells her the team will reply, marks the message as priority and forwards the case to the professional. The therapy content stays in the clinic's system, outside the agent's reach.

How the method applies to a clinic

The work follows a four-part method: map, limit, pilot and hand over. Mapping means surveying the current calendar and the points of friction, without assuming the problem is technical. Limiting means giving the agent only the administrative data it needs and no permission over clinical records. Piloting means testing one real workflow, with logs and an agreed criterion set before starting. Handing over means returning the operation to the clinic and deciding, based on the records, whether the flow expands, changes or stops. Before serving real patients, it is worth testing the agent's behavior with fictitious data, as described in the article on mock tools for testing AI agents in a sandbox.

The limits here are similar to those in other areas of mental health and rehabilitation. The article on superintelligence in psychology practice offers a complementary view of that boundary, and the piece on superintelligence for physiotherapy clinics applies the same logic of setting up scheduling before the agent books anything.

Frequently asked questions

Can AI book appointments for a speech therapy clinic? Yes, for scheduling and reminders, as long as each professional's rules are written down and there is a handoff to reception when a conversation goes beyond the script. The agent should check the real calendar and never create slots that do not exist. Before turning on the flow, the clinic should validate these rules with the people who actually book sessions.

Can the agent record session progress? Only as support for organizing a draft, reviewed and signed by the professional who treated the patient. The agent should not interpret results, suggest treatments or have access to clinical conversations through the service channel. Decisions and interpretations stay with the clinician.

Is scheduling data also health data? It depends on the content, but the simple fact that someone is a speech therapy patient, with a type of service attached, can already reveal health information. That is why the LGPD should be considered from the design stage of the flow, with minimal collection and a clear purpose. The National Data Protection Authority (ANPD) publishes guidance on the LGPD that is a sound starting point for this analysis.

Where should a clinic start without hiring a large platform? Start by mapping the current scheduling process and choose one workflow, such as confirmation and rescheduling. Test with fictitious data before using real patients, define the success criterion and name who reviews the logs each week. The decision to expand, adjust or stop comes after the pilot, based on what was recorded.

Sources

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