Superintelligence in Diagnostic Labs

Superintelligence in diagnostic labs does not mean an agent that interprets test results. At most, it means a more capable AI setting in which agents handle scheduling, confirmation and preparation instructions under clear rules, and leave every interpretation of a result to a qualified health professional. The model may get more capable over time, but the limit on what it does remains a company decision, not the model's.

Key takeaways

• Scheduling, rescheduling, confirmation and sending approved preparation instructions are suitable tasks for an agent, as long as it reads only content the lab has validated.
• Interpreting a result, saying whether it is normal or abnormal, suggesting next steps or comparing it with earlier tests stays outside the agent, even if the model can technically do it.
• A more capable AI increases the need for explicit limits, because more fluent answers feel more trustworthy than they really are.
• Test data is sensitive personal data under Brazil's LGPD, so the design needs minimal access, authentication and a record of every action.
• The safe path is to map the process, limit data and actions, pilot a single flow with an agreed metric, and only then decide whether to expand.

What changes when AI gets more capable in a lab?

When people talk about superintelligence, the conversation tends to jump to what AI might do someday. For someone running a lab, the useful question is different: what does a more capable system change in day-to-day risk? To get the terms straight, it helps to read about the difference between AGI, ASI and AI agents. In practice, what a company operates are agents with defined tools and permissions, not a general intelligence.

A more capable model writes better, understands messy messages and keeps the context of a long conversation. That helps customer service. But the same fluency creates a new risk: the patient gets an answer so natural that they take it as clinical guidance. The more convincing the agent, the more important it is that it has a structural boundary, and not just a polite instruction in the prompt.

That is why the design question is not whether the model could interpret a blood count. It is whether the lab wants its service channel to do that, with which responsible person, with what record and with what right of review. In most cases the answer is no, and the technical design should reflect that decision.

What the agent can do in scheduling

Scheduling is a process with objective rules: location, type of test, collection window, availability, health insurance or private payment, required documents. It is good ground for automation, and the principle is the one described in how automatic appointment scheduling on WhatsApp with AI works: the agent checks the real calendar, proposes time slots and writes the outcome to the system.

A well-bounded agent can generally:

• Identify which test or package the patient wants to book, confirming with them when something is ambiguous.
• Check open slots in the lab's calendar and offer options.
• Book, reschedule and cancel appointments according to the company's notice rules.
• Provide the address, opening hours and documents to bring.
• Send reminders and ask for attendance confirmation.
• Hand off to a staff member when the request falls outside the pattern, such as a test without a legible medical order or a coverage question.

What the agent should not do on its own: decide whether a test is appropriate for the patient, swap one test for another by itself, or promise insurance coverage without checking the system. When the information is not in the approved source, the right answer is to say so and pass the patient to a person.

If the lab uses outbound reminders on WhatsApp, the rules on templates, opt-in and best practices for outbound messages apply here too. A collection reminder is useful, but it should only go to people who gave permission and in the format the channel allows.

Preparation instructions: where the agent helps and where it needs care

Test preparation (fasting, pausing something, type of collection, use of a sample container) is where service generates the most questions and where a mistake has a real cost, because the collection may have to be redone or the result may be compromised. The agent is useful here, but only as a courier of instructions the lab has already written and approved.

The recommended design is simple: preparation content lives in a document maintained by a technical lead, with a version and a review date. The agent retrieves the section matching the scheduled test and passes it along without rewriting its meaning. This is an implementation recommendation, in line with what Anthropic's guidance on building effective agents suggests: start with simple flows, clear tools and controlled sources before granting autonomy.

The limit shows up in questions the document does not cover. Examples:

• Can I take my regular medication before the collection?
• I am pregnant, does the preparation change?
• I felt unwell while fasting, what should I do?
• My child is small, how do I adapt the instructions?

These questions depend on the individual case and should not be answered by the model improvising. The agent recognizes the type of question, says the guidance needs a lab professional, and opens a handoff with the context already summarized, so the patient does not have to repeat everything.

What the agent never interprets

The most important line in the project is this: the agent does not interpret test results. That includes saying a result is normal, abnormal or worrying, comparing it with reference values in its own words, explaining what a number means for the patient's health, suggesting a repeat test or indicating which specialist to see. Even a reassuring phrase, like it doesn't look serious, is already interpretation.

There are three practical reasons to keep this rule, no matter how capable the model is:

• Interpretation depends on clinical history, medications, age and context that the lab often does not have.
• As a rule, analyzing the result is up to the qualified health professional, and the lab should confirm with its legal counsel and technical lead how this applies to its case.
• An error in tone or content reaches the patient as if it were the lab's own word.

The agent can handle status: say that the result is available, guide the patient to the authenticated portal, explain turnaround times the lab has already published and indicate how to request a conversation with the responsible person. The difference is between saying where the result is and saying what it means.

For delivery over WhatsApp, the prudent choice (a design decision, not a specific legal requirement) is to send only an availability notice with a link to an authenticated area, without attaching or summarizing the content in the chat. This reduces exposure of sensitive data on a channel that other people may see on the patient's device.

How should the agent react when a patient asks about a result?

In practice, patients will ask. They receive the PDF, open the chat and write: what does this value mean? The agent's behavior at that moment has to be tested, not just written into the prompt.

Illustrative scenario: a patient logs into the portal, sees an item flagged in the report and writes on WhatsApp asking whether it is serious. The agent does not comment on the value or reassure her. It replies that it cannot interpret results, explains that the analysis should be done by the doctor who ordered the test or by a lab professional, offers to open a contact request with the responsible person and records the request with time and channel. If the patient reports intense symptoms, the agent follows the urgency protocol previously defined by the lab's clinical lead, which usually means advising her to seek emergency care, and it alerts a human immediately.

The key point of this scenario is that the refusal does not end the conversation. An agent that only says I can't help frustrates the patient. An agent that explains the reason, hands off and confirms the next step preserves trust. This behavior should be part of the pilot's test set, with messages from worried patients, ambiguous messages and attempts to get around the rule, such as asking the agent to just say whether the value is within what is expected.

Health data, LGPD and action logs

Health data is sensitive personal data under the LGPD, which calls for an appropriate legal basis, a clear purpose and more care with access and sharing. ANPD, Brazil's data protection authority, publishes guidance and is the primary source for following how the law is interpreted. Final validation of the design should go through the lab's data protection officer and legal counsel.

Some architecture decisions reduce risk from the start:

• Authenticate the patient before any lookup of individual data, instead of trusting the phone number alone.
• Give the agent access only to what the task requires: calendar and status, with no access to the content of result reports.
• Apply the principle of least privilege to the tools, as the OWASP authorization best practices describe.
• Log every agent action, the source consulted and the reason for each handoff.
• Define retention and deletion for conversations, and assess carefully whether any model provider processes data outside Brazil, a topic covered in international data transfer under LGPD when using AI providers.

There is also the matter of automated decisions. If a flow affects the patient in a significant way, such as denying an appointment or coverage, the right of review under the LGPD needs an operational path. The article on LGPD automated decision-making and the right to review (Art. 20) details what to watch. To organize the risk of the project as a whole, the NIST AI Risk Management Framework offers a useful vocabulary for mapping, measuring and managing risk.

How to start: map, limit, pilot, hand over

The method we apply in projects like this follows four moves, with no promise of results: map the current process and measure the baseline, limit data and actions, pilot a real flow with logs and an agreed metric, and hand over the operation with a decision to expand, adjust or stop.

A step-by-step plan for a lab:

• Map: find out how scheduling requests arrive, which preparation questions repeat, how many go to staff and which ones require human judgment.
• Classify each type of message into three groups: the agent resolves it, the agent prepares and a human decides, or the agent never touches it.
• Limit: write the approved preparation document, define the permitted tools (check calendar, create appointment, open handoff) and explicitly forbid access to result reports.
• Define the interpretation rule: any question about what a result means goes to a human, with an approved response text.
• Pilot: pick one flow, for example scheduling a simple collection at one location, with an agreed metric before starting, such as correct handoff rate and response time.
• Test with simulated conversations, including pressure cases and attempts to bypass the rule, before any real patient is involved. The approach of mock tools and sandbox testing for AI agents helps do this without risk.
• Hand over: deliver logs, rules and owners to the lab's team, and decide in a meeting whether the pilot expands, adjusts or stops.

It is worth remembering that stopping is a legitimate outcome. If the pilot shows the flow creates more wrong handoffs than time saved, the report with logs supports the decision not to move forward.

Frequently asked questions

Can an AI agent explain a test result to the patient? It should not. Interpretation depends on clinical context and, as a rule, belongs to the qualified health professional. The agent can say that the result is available, guide access and pass on the request for a conversation with the responsible person.

Can the agent answer questions about fasting and preparation? Yes, as long as it answers from instructions the lab has approved and versioned. Questions about medication, pregnancy, children or feeling unwell should go to a professional, because they depend on the individual case.

Does a more capable AI change these rules? It changes the quality of the conversation, not the boundary of responsibility. More capable models tend to sound more certain, which makes structural limits, such as restricted tools and mandatory handoff, even more necessary.

Can I send the test result over WhatsApp? Technically it is possible, but it is a risk decision. Because health data is sensitive, the most prudent practice is to send only an availability notice with a link to an authenticated environment, and to validate the choice with the data protection officer and legal counsel.

Sources

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