Superintelligence for Elderly Care Homes: Family, Medication and Visits

A care home can use an AI agent to answer families about visits, schedules and routine administrative matters, as long as the agent has a narrow scope: it shares what is public or routine, routes health topics to the responsible staff, and does not disclose medication, medical records or clinical progress without a recorded authorization. In practice, the gain comes less from generic intelligence and more from written rules, an organized channel and records that show who asked what and who responded. The useful question for leadership is not how much the AI knows, but what it may do on its own and what must pass to a person.

Key takeaways

• The agent should treat communication with relatives as a flow with permissions: who is asking, about what, and under which authorization.
• Health data is sensitive under Brazil's LGPD, so the default rule is not to disclose, and any exception must be documented.
• The visit schedule is often the biggest gain, because it reduces repeated calls and scheduling conflicts.
• Any clinical, medication or complaint topic should leave the agent and go to a responsible person with a committed response time.
• Before any agent contacts families, the institution must test, log and review its answers.

What superintelligence means in a care home

The term superintelligence appears often in AI content, but it rarely describes a tool that a manager can buy and operate. For a care home, what exists in concrete terms is an AI assistant connected to channels such as WhatsApp, to a spreadsheet or scheduling system, and to a set of instructions. If you want to separate concept from product, read the superintelligence glossary for executives before speaking with vendors.

The difference matters because the promise of a system that "knows everything" leads to poor decisions, such as giving the agent broad access to resident records so it "can help better." A safer approach works in the opposite direction: define first what the family needs to know, then what the staff must decide, and only then design the agent to cover the repetitive part.

Which information can be answered for families

Not every question from a relative needs clinical attention. A large share of the volume is administrative: visiting hours, documents needed for a companion to enter, activity days, instructions on what to bring in terms of clothing or personal items, already formalized fee questions, and procedures for reporting absences. These answers can be standardized and approved by management.

The resident's health status, test results, medication changes, falls, agitation episodes and anything in the medical record should stay out of the agent's reach. Even when a relative has the right to know, the information should come from the person with professional responsibility, through a channel that records the conversation. The agent can say that the nursing team will reply within an agreed timeframe, without trying to summarize the resident's condition.

How to organize the visit schedule without exposing routines

The visit schedule has two functions: organizing who enters the building and protecting residents' routines. It should be easy to check and should not show the person booking anything about other residents or about the clinical condition of the person being visited. A good design separates the list of available time slots from the list of who is authorized to visit each resident.

In practice, this means the agent checks open slots by unit and time, confirms the visitor's name and their relationship to the resident according to the institution's registry, and records the visit. If the relationship is not on file, the request goes to reception or the coordination team. Cancellations, rescheduling and temporary restrictions, such as an isolation period, need their own rules and a named person who decides.

Step by step: rolling out family communication

The rollout should start small and with records. The sequence below works as a decision roadmap for management or the operations lead:

• Step 1: map the types of questions that arrive today by phone, WhatsApp and reception, separating administrative, scheduling and clinical questions.
• Step 2: define, for each type, whether the agent answers on its own, answers with approved information, or hands off to a person with a committed response time.
• Step 3: list the data the agent may access and the data that stays blocked, applying the minimum amount of data needed for each answer.
• Step 4: write the agent's instructions with examples of refusal, including what it says when a question is clinical.
• Step 5: test with real questions, anonymized, and with attempts to obtain another resident's data, before any real contact with families.
• Step 6: run a pilot on a single channel and a single unit, logging every handoff and every answer the staff corrected.
• Step 7: review the pilot with management, decide whether to expand, adjust or stop, and only then widen the scope.

In step 3, record the legal basis for handling each data type, because using health information requires more care than using ordinary registry data. The superintelligence guide for psychology practices on confidentiality and scheduling addresses a similar problem with different limits and can be a useful reference for designing permissions.

Illustrative scenario: a question about medication

Illustrative scenario: a relative writes on WhatsApp asking whether their mother has already taken her afternoon medication and whether the time can be changed. The agent recognizes that the question involves medication. It does not confirm or deny administration, does not name the drug, and does not suggest a schedule change. It replies that the nursing team has been notified, that a response will come through the same channel within the timeframe the institution has set, and that in an urgent case the family should call the coordination number. The conversation is logged with the time, the reason for the handoff and the staff member who took it over.

This answer is less satisfying in the short term than a technical answer, but it protects the institution, the resident and the family. The most common mistake in projects like this is trying to make the agent answer everything to reduce waiting time, which usually increases the risk of incorrect information.

How we apply this method in practice

In our work with companies, the method follows four stages: map the current process and its baseline, limit data and actions, pilot a real flow with records and a success metric agreed in advance, and then hand over the operation and decide whether to expand, adjust or stop. In care homes, the limiting stage is the most sensitive, because it defines which data the agent will never see. The pilot stage should include human review of every sensitive response, and the metric must be agreed with management before launch, for example time to correct handoff, not just the volume of messages answered. When the operation is managed by an outside partner, that partner should handle maintenance and monitoring of the agents under the same written rules.

Channel rules also matter. Messages started by the institution have their own restrictions on WhatsApp, and the WhatsApp 24-hour customer service window defines how the agent can reply or resume a conversation. Visit reminders are a common case and usually require the relative's prior opt-in.

Which risks management needs to monitor

The main risks are not only technical. A poorly written instruction can lead the agent to give health advice. An outdated registry can allow someone without a relationship on file to book a visit. A malicious person may try to extract another resident's data through indirect questions. For that last case, apply the prompt injection prevention principles published by OWASP, and clearly separate what the agent is allowed to look up.

The NIST AI Risk Management Framework offers a useful structure for mapping risks and assigning responsibilities, and Brazil's national data protection authority, the ANPD, is the Brazilian reference for guidance on personal data protection. None of these documents replaces a legal assessment of the institution, but they help organize the internal conversation.

Frequently asked questions

Can the agent tell a family whether the resident has taken their medication?

No. Information about medication administration is health data and must be given by the responsible staff, through a channel that records the reply. The agent should recognize the question, log the handoff and state the response time, without confirming or denying anything.

Who can receive updates about a resident?

Only people registered with a relationship and an authorization formalized by the institution, following the procedure management defines. The registry should be reviewed regularly, because relationships change. If there is doubt about who may receive the information, the case goes to a person before any reply.

Do families need to be told they are talking to an AI?

Yes. The communication should make clear that the first contact is handled by an automated system and explain how to reach a person. This notice avoids confusion and builds trust. Management should approve the wording, which can appear in the channel's first message.

Where should a care home start if it has never used AI?

Start by mapping the questions that arrive today and defining which ones are handed to people. Then choose a single channel and a single question type, such as visit scheduling, for a pilot with records. Expand only after the staff has reviewed the answers and agreed on what the tool should do and what it should not.

Sources

• NIST AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados
• OWASP: LLM Prompt Injection Prevention Cheat Sheet