Superintelligence in Clinical Laboratories: What an AI Agent Can Do

Superintelligence in clinical laboratories, in practical terms, means a narrow AI agent that handles administrative and informational work: booking sample collection, repeating the preparation instructions the lab has approved, and telling a patient which stage a result is in. It should not interpret values, suggest a diagnosis or comment on a report, and it must treat health data as sensitive personal data from the first sketch of the workflow. The gain is relief for the front desk and the phone line, not a replacement for the judgment of the biomedical scientist, the clinical biochemist or the physician.

Key takeaways

• The AI agent belongs in the lab for scheduling, test preparation and result status. Clinical interpretation stays out, always.
• Preparation instructions should come from a text approved by the technical manager, and the agent only repeats it, without improvising.
• Health data is sensitive personal data under LGPD (Brazil's data protection law). That calls for minimal collection, restricted access, usage logs and authentication before any individual information is shared.
• Every case outside the script (clinical question, abnormal result, complaint, urgency) should be handed to a person, with the context already summarized.
• The safest path is to map, limit, pilot a single workflow with logs, and only then decide whether to expand.

Superintelligence in the lab: what do you actually run?

When people talk about superintelligence, they usually mean a hypothetical system far more capable than any person at almost every task. That is not what a laboratory operates today. What it can put into production is an AI agent with a narrow scope: a language model connected to a few tools, such as the appointment calendar, the library of preparation instructions and a status lookup. To understand how the concepts differ, read the difference between AGI, ASI and AI agents.

That clarity changes decisions. If the expectation is a system that understands the patient and answers questions about health, the project starts with too much risk. If the expectation is a digital receptionist that handles three repetitive tasks within clear limits, the project is viable, auditable and easy to explain to the technical manager and the data protection officer.

What the agent can do: scheduling, preparation and status

A large share of laboratory customer contact is made of predictable questions. That is where the agent helps.

For scheduling, the agent collects what is needed to book: name, preferred unit, type of visit (collection at the unit, or home collection if the lab offers it), date and time of day. It checks the real calendar through a tool with limited permissions and confirms the slot. Rescheduling and cancelling follow the same logic. Technical details of this kind of workflow appear in automatic appointment scheduling on WhatsApp with AI.

For test preparation, the agent shares what the lab has already defined: fasting, medications to pause, hydration, documents to bring, care on the day of collection. The central point is that the text comes from a library approved by the technical manager, and the agent delivers it without rewording its meaning. If a patient asks whether they can take a specific medication, the correct answer is to refer them to a qualified person, not to guess.

For result status, the agent tells the patient which stage the order is in: collection registered, sample being analyzed, result released, available on the portal or ready for pickup. It does not describe the content of the report. When the result is released, it points to the official access channel, with authentication.

These three workflows have something in common: the correct answer already exists in a lab system or document. The agent only finds it and delivers it in the conversation.

What the agent should never do

The boundary has to be written down, tested and visible to the team. It is not a configuration detail.

• Do not interpret reference values, abnormal or normal results, or compare them with earlier tests.
• Do not suggest a diagnosis, a course of action, a medication or the need to repeat a test.
• Do not reassure or alarm the patient about a result.
• Do not send the content of a report in an open message without a policy defined by the lab and without authentication.
• Do not change a record, a result or a medical order on its own.
• Do not answer questions about third parties (relatives, employers, insurers) without the authorization foreseen in the lab's process.

The reason is practical. A language model can produce a fluent and wrong sentence, and in healthcare the cost of that error is of a different kind. Also, interpreting a result depends on the patient's clinical context, which the lab often does not have. The same boundary principle applies in other sensitive fields, as shown in the article on AI limits in psychology practice.

Three layers help enforce the limit: a clear instruction to the model about what to decline, tools that simply give no access to report content, and a handoff rule for when the subject leaves the script. The second layer is the strongest, because what the agent cannot query, it cannot leak or comment on.

How to protect health data in this workflow

LGPD treats health-related data as sensitive personal data, which calls for a higher level of care. In design terms, that turns into concrete decisions. The ANPD, Brazil's data protection authority, is the official starting point for guidance, and the lab's legal counsel or data protection officer should validate the legal basis and the purpose of each processing activity.

A few practices that make a difference:

• Collect the minimum. To book an appointment, name and contact are usually enough. Do not ask for a diagnosis, history or a photo of the medical order if the workflow does not require it.
• Authenticate before sharing individual status, using details only the patient knows or a link from the lab's portal.
• Separate permissions. The calendar tool does not need to see reports, and the status tool does not need to see the text of the result. Authorization principles are well described in the OWASP Authorization Cheat Sheet.
• Log what the agent queried and answered, restrict access to those logs, and define how long they are kept.
• Check where the model and the data are processed. If anything is sent outside Brazil, read what LGPD requires for international data transfer when using AI providers.
• Protect the agent against malicious instructions. A patient, or an attached document, may try to manipulate the model. The OWASP Top 10 for LLM applications lists these risks and works as a review checklist.

Another point is the right to human review. Even if the agent decides nothing clinical, any answer that affects a patient's access to a service deserves a clear path to a human attendant. The topic is covered in LGPD automated decision-making and Article 20.

Illustrative scenario: the question about the result

This is an illustrative scenario, unrelated to any real laboratory. A patient writes to the lab's WhatsApp asking whether her blood test is ready and then whether the value of one item is good.

The agent recognizes the status intent and asks for identity confirmation through a method the lab has defined. After verification, it calls the status tool, which returns only the stage of the order, for example: released on the portal. The agent replies that the result is available and explains how to reach the official channel.

When the patient asks whether the value is good, the agent does not comment. It explains, in simple terms, that it cannot interpret results, that the assessment belongs to the physician who requested the test, and offers, if she wishes, to forward the question to the lab team. If she writes that she has severe symptoms, the agent drops the script, advises her to seek urgent care through the appropriate means and hands the conversation to a person with a summary.

In this scenario, the agent never saw the content of the report, and the refusal does not depend on the model getting a hard judgment right. It follows from the design.

A step-by-step plan to start

The method has four moves: map, limit, pilot and hand over (meaning hand the operation to the team and decide the next step).

Map: survey how customer contact works today. Which questions arrive by phone, WhatsApp and at the counter, who answers them, which system holds the calendar and how status is looked up. Set a simple baseline, such as the time the team spends on these tasks and how many calls are only status requests, without inventing numbers: measure what the lab can already measure.

Limit: write the list of allowed and forbidden tasks, the approved sources for preparation instructions, the data that may be collected and the tools with their permissions. Define the triggers for handoff to a human and the notice telling the patient they are talking to an automated assistant.

Pilot: choose one real workflow, usually result status or appointment confirmation, and run it with logs and a success metric agreed with the lab's manager. Test beforehand with hard cases: clinical questions, attempts to bypass authentication, requests from third parties. Simulated environments are useful here, as described in testing AI agents with mocks and a sandbox. The NIST AI Risk Management Framework helps structure the risk review.

Hand over: at the end of the pilot, give the team the documentation, the logs and the adjustment procedure, and decide from the data whether to expand, adjust or end the workflow. Stopping is a legitimate pilot outcome.

Anthropic's article Building effective agents recommends starting with the simplest design that solves the problem, which suits a laboratory well: deterministic flows for scheduling and status, and the language model only to understand the question and draft the reply.

Frequently asked questions

Can an AI agent deliver test results over WhatsApp? It depends on the lab's policy, on patient authentication and on a security and legal review. A more prudent approach is to use WhatsApp to notify the patient that the result is available and direct them to the authenticated portal, without sending the report content in the chat.

Can the agent explain what an abnormal value means? No. That is clinical interpretation and should stay with the qualified professional, usually the physician who requested the test. The agent should recognize the limit, say so clearly and offer a referral to the team.

Do I need the patient's consent to use AI in customer service? The answer depends on the legal basis chosen for each purpose, and health data calls for special attention under LGPD. The data protection officer or the lab's legal counsel should define the basis, the notice to the patient and the record. In any case, telling patients that an automated assistant is handling the conversation is good practice.

Where do I start without putting the lab at risk? Start with a single administrative workflow, such as appointment confirmation or a stage status lookup, with authentication, logs and handoff to a human. Run it as a pilot, evaluate it against an agreed metric and only then consider widening the scope.

Sources

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