Superintelligence in Clinics: How to Prepare Without Risk

Superintelligence in Clinics: How to Prepare Without Risk

Superintelligence in clinics is not something you can buy today. It is a hypothesis about systems far more capable than current ones. What a clinic can do now is prepare for more competent AI agents: keep medical records and clinical decisions outside the agent, and treat health data as sensitive data under Brazil's LGPD. A clinic that sets these boundaries first can swap models later without rebuilding the process.

Key takeaways

• Superintelligence is a hypothesis. What a clinic runs today are agents with a limited scope, and that is what you govern.
• Medical records and clinical decisions stay outside the agent: it does not write, edit or decide diagnosis, treatment or prescription.
• Health data is sensitive personal data under LGPD. That calls for a clear purpose, the minimum necessary data, access logs and care with vendors and with transfers outside Brazil.
• Model capability may rise, but permissions only rise through a human decision that is documented and tested.
• The safe path is to map, limit, pilot one administrative flow, and only then decide whether to expand, adjust or stop.

What does superintelligence mean for a clinic today?

The term comes from the debate about systems that would outperform people at most intellectual tasks. No such product exists, and no clinic should plan its operation around a promise. The difference between the concepts is covered in Difference Between AGI, ASI and AI Agents. The practical point is that a clinic operates agents that read messages, query systems and carry out bounded tasks.

It is reasonable to expect these agents to get better at understanding ambiguous requests, summarizing histories, drafting replies and using tools. That is a hypothesis, not a fact. So preparation does not depend on getting the forecast right. It depends on an architecture where the model's capability can grow without its power over data and decisions growing with it.

Think of it this way: a very capable professional hired for the front desk still has no access to the medical record and no authority to decide treatment. Their competence improves the service, but it does not change who is clinically accountable. The same applies to the agent.

What changes in patient service when the agent becomes more capable?

In practice, the most visible change is at the front desk and in patient relationships, not in the consultation room. A more capable agent tends to handle loose WhatsApp messages better, such as "I need to move that follow-up up, but I can only do mornings," and turn them into an organized action: check the schedule, propose times, confirm and log.

What can improve, with care:

• Scheduling, rescheduling and confirmation, with clear rules for notice and overbooking.
• Administrative guidance: address, logistical preparation, documents to bring, accepted payment methods.
• Topic triage: separating what is administrative, what is financial and what needs a professional.
• Conversation summaries for the team, so nobody asks for the same information twice.
• Reminders and follow-up messages, as long as the patient has given permission for outbound contact.

What does not change, however capable the model becomes: urgency and symptoms go to a person. The agent should recognize signs that the conversation has left the administrative realm and hand it to the team with a short summary, without trying to give clinical guidance. For schedule and reminder design in a sensitive practice, read AI for Psychology Practice Scheduling: Privacy-First Setup, which covers care with what appears in the message.

What should stay outside the agent: medical records and clinical decisions?

The design rule is simple: the medical record belongs to the professional, and the clinical decision belongs to the professional. The agent is not the source of truth for either.

In practice, this means:

• No writing to the record. The agent does not create, edit or delete clinical entries.
• No broad reading. If the agent needs a piece of information, such as the time of a follow-up, it calls a tool that returns only that field, not the patient's history.
• No treatment guidance. The agent does not interpret tests, suggest diagnoses, advise on doses or say whether a symptom is serious.
• No decisions about access to information. Requests for reports, medical certificates and copies of the record follow the clinic's own process and validation.
• No clinical memory of its own. If the system stores conversation history, it stores only what is needed for administrative service and has a defined disposal period.

When the clinic wants to use AI to support the professional, for example summarizing a consultation or suggesting a draft, that is a different project with a different level of risk. It should be treated as a professional's tool, reviewed and signed by the professional, in an environment separate from the patient service agent. Check the rules of your professional council and get legal advice before starting.

Keep in mind that an incoming message can carry malicious or simply confusing instructions. The less the agent can see and do, the smaller the damage. The OWASP guidance on risks in applications built on language models is a good reference for the technical team, and our article on prompt injection protection for AI agents goes into the problem in detail.

How does LGPD apply when the data is health data?

Under LGPD, health data is sensitive personal data, with stricter rules for processing. This applies to what the patient writes on WhatsApp, to what sits in the schedule and to any record the agent touches. The starting point is understanding the purpose of each piece of data and the applicable legal basis, with legal support, because the right basis depends on the case.

The safeguards that tend to appear in any project:

• Purpose and necessity: collect and use only what the flow requires. To confirm an appointment, you do not need the reason for the visit.
• Transparency: the patient should know they are talking to an automated service and how to ask for a person.
• Security: access control, a log of who accessed what, and protection of integrations.
• Vendors: know where data is processed, under what conditions, and whether it is retained or used for training. If processing happens outside Brazil, read International Data Transfer Under LGPD When Using AI Providers.
• Data subject rights: access, correction, deletion where applicable and review of automated decisions. The article on LGPD automated decision-making explains the right to review.
• Documentation: records of processing operations and, when the case calls for it, an impact assessment. The ANPD (Brazil's data protection authority) publishes official guidance that the clinic should follow.

One design consequence: the less health data the agent receives, the less the clinic has to protect and justify. Minimizing is not a technical detail. It is the main compliance decision.

How do you prepare the clinic in five steps?

The method we use in this kind of project is to map, limit, pilot and hand over. Applied to a clinic, it looks like this:

• Step 1, map: describe the current flow at the front desk and after the visit. Who receives the message, which system holds the schedule, how long it takes, where errors happen. Set a baseline before any AI, so you can tell later whether anything improved.
• Step 2, limit data and actions: list exactly which fields the agent can read and which actions it can perform. Everything not on the list is denied by default. Medical records and treatment decisions stay out.
• Step 3, define the handoff to a human: triggers such as symptoms, urgency, complaints, requests for reports, billing questions and any uncertainty from the agent. Define who receives the handoff, how fast and with what summary.
• Step 4, pilot one real flow: for example, only appointment confirmation and rescheduling. Keep a decision log, agree on a success metric beforehand, and have someone on the team review conversations regularly.
• Step 5, hand over and decide: deliver documentation, permissions and the monitoring routine to the clinic, and decide whether to expand, adjust or stop. Stopping is also a valid outcome.

To structure the project's risk assessment, the NIST AI Risk Management Framework offers a useful vocabulary for governing, mapping, measuring and managing risk, without being specific to healthcare. The question it helps keep alive is: what happens when the agent gets it wrong?

When the model is replaced by a more capable one, repeat step 4 on a small scale. Run the same test conversations, compare the answers and only then release it. Permissions do not go up just because they are possible.

Illustrative scenario: the front desk of a specialty clinic

This is an illustrative scenario, unrelated to any real case. A clinic with several professionals receives WhatsApp messages all day. The front desk spends much of its time confirming appointments and rescheduling, and sometimes messages mix administrative requests with descriptions of symptoms.

The clinic maps the flow and decides to start with confirmation and rescheduling only. The agent calls a tool that returns open slots per professional and creates or changes a booking, nothing more. It cannot see the medical record, cannot see the reason for the visit and does not answer clinical questions.

A patient writes: "I need to reschedule, I've had a lot of pain since yesterday and wanted to know if I can wait until next week." The agent reschedules if the patient asks, but it recognizes the report of pain, does not offer an opinion on waiting, and passes the conversation to the front desk with a summary and a priority flag. A person calls or replies. The decision about what to do is clinical and stays with the team.

A quick decision guide for any message in this flow:

• Purely administrative request with data already in the schedule: the agent resolves it and logs it.
• Administrative request that needs an exception, such as an overbooking outside the rules: the agent prepares it and the front desk approves.
• Mention of a symptom, medication, test or urgency: the agent gives no guidance and transfers immediately.
• Request for a report, certificate or medical record: transfer to the clinic's formal process.
• Out-of-pattern message, with odd instructions or an attempt to get another person's data: the agent declines and flags it for review.

If the clinic wants to send outbound reminders, check permission and channel rules first, as described in WhatsApp Outbound Messages: Templates, Opt-In and Best Practices.

Frequently asked questions

Can superintelligence already be used in clinics? There is no evidence that superintelligent systems exist for commercial use. What exists are AI agents with growing capabilities, which a clinic can use for well-bounded administrative tasks. The right preparation holds for any level of future capability.

Can the AI agent access the patient's medical record? The safest design is that it does not. If a lookup is truly indispensable, it should return only one specific piece of data, with minimum permission, access logging and approval from the person responsible at the clinic. Writing to or changing the record stays outside the agent.

What does LGPD change for those who serve patients on WhatsApp with AI? Health data is sensitive, so the clinic needs a clear purpose, minimal collection, security, transparency with the patient and control over vendors. It also needs to know where data is processed and offer a way to reach a person. Get legal advice to define the legal basis and documentation for your case.

Where do you start without putting operations at risk? Start with a low-risk administrative flow, such as confirmation and rescheduling, with a measured baseline, an agreed metric and a log of the agent's decisions. After a pilot period, review the conversations and decide whether to expand, adjust or stop. If the clinic wants help structuring this, the work follows the stages of map, limit, pilot and hand over.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework
• OWASP Top 10 for LLM Applications
• Anthropic: Building effective agents