Superintelligence for Clinical Laboratories: What to Prepare Now

Superintelligence, in the sense of systems that would outperform humans at nearly every task, is not a product a laboratory can buy today. What makes a difference now is preparing the operation: well documented sample collection, written protocols, clear limits for any AI agent, and report delivery with an identified responsible professional. Labs that build this base can use agents to inform, schedule and route requests safely, and can later change or expand the technology without redesigning the whole process.

Key takeaways

• Superintelligence is a debate about future systems. The useful decision for a laboratory concerns the agents that exist today and their limits.
• Many errors start at collection, in areas such as patient identification, preparation instructions, tube type and deadlines. An agent cannot fix these errors on its own, so the process must be mapped first.
• Protocols and SOPs must be written, versioned and approved by the technical lead before they become a source of answers for an agent.
• Agents inform, schedule and route. Interpreting results and releasing reports remain the responsibility of a qualified professional.
• Health data requires extra care under LGPD, and any expansion should happen only after a pilot with records and an agreed success metric.

Superintelligence is a future topic, agents are a decision for today

The term appears often in media content and tends to mix two different things. On one side are discussions about general, hypothetical systems that would be far more capable than any current tool. On the other side are language models connected to workflows, with access to systems, rules and communication channels. For a laboratory manager, only the second category has practical effects right now, and it can make mistakes, invent answers and be influenced by text the patient sends.

So the useful question is not whether the laboratory is ready for superintelligence. It is whether the laboratory can describe precisely what the agent may do, with which data and under what supervision. It is worth reading about the difference between AGI, superintelligence and AI agents and about the myths that distort decisions, because inflated expectations often lead to automations with no limits.

Collection: what to organize before any automation

The pre-analytical phase is where an operation depends most on human discipline, and where an agent can help without taking over decisions. Before connecting any digital channel, the team should have the full workflow written down, from the medical request to the start of analysis. A starting checklist:

• Map the complete path: medical request, insurance verification, scheduling, arrival, identification, collection, transport, receipt and start of analysis.
• List the most common questions about preparation, such as fasting, timing, medications and sample type, and name the official source for each answer.
• Define who validates special preparations and who authorizes rescheduling for health reasons.
• Standardize how the patient and the collected material are identified, with a clear rule for registration mismatches.
• Record the time of each step, so any question about the status of a test has a verifiable basis.
• State what the digital channel never handles, such as requests for early results, reports of urgency or clinical complaints.

Once this map exists, the agent has clear boundaries. Without it, the agent only repeats improvised practices faster.

Protocols and SOPs: what the agent can consult and what it must never answer

An agent should respond only from approved documents. In practice, this means sorting content into three tiers, each with explicit rules:

• May answer with approved text: opening hours, address, standard preparation, required documents and deadlines the laboratory publishes itself.
• Must hand off to a human: nonstandard preparation, tests with special collection, rescheduling requests for clinical reasons and any complaint.
• Must never answer: interpretation of values, diagnosis, treatment, comparison with previous exams and early release of a report.

It is also worth defining what happens when a patient reports symptoms. The safest rule is to transfer the conversation to a human immediately and guide the patient according to the laboratory's protocol, without trying to judge severity on its own. Each protocol should have a review date and a named owner for its content, because an agent that answers from an outdated version causes the same harm as an untrained employee.

Report delivery and health data under LGPD

Health data is sensitive personal data under LGPD, which raises the level of care across the entire chain. Before any agent touches orders or results, the laboratory should answer key questions with documents rather than assumptions: which data enters the model's context, who can access the records, how long conversations are stored, and whether processing involves transfer to another country. The topic of international data transfer with AI deserves a specific review with the data protection officer and legal counsel. The national data protection authority, ANPD, is the official reference for guidance on the law.

A practical principle is data minimization. To answer whether a test is ready, the agent does not need access to the clinical content of the report. It only needs the status of the request. For proactive notices, such as a reminder to come in for collection or a notice that a report is ready, the channel requires prior permission from the patient, which is covered in the article on WhatsApp opt-in.

The delivery rule is simple to write and hard to follow if it is not defined in advance: the report is released through the flow the laboratory already uses and signed by the technical lead. The agent may tell the patient that the document is available and explain how to access it through the authorized channel, but by default it should not repeat findings in a WhatsApp message.

A four-stage plan for a pilot

The sequence below works for laboratories of any size and avoids starting with the technology:

• Map: document the current service and collection flow, with timings and failure points, without changing anything yet.
• Limit: define which data the agent can read, which actions it can take and which remain always with people.
• Pilot: choose one real, small flow, such as preparation questions or request status, log every conversation and agree on a success metric before starting.
• Transfer: hand the flow over with a manual, named owners, periodic review and an explicit decision to expand, adjust or stop.

This is the method we follow in projects: map, limit, pilot and transfer. The process makes no promise of results. It only requires that each stage be verified before the next one begins.

Illustrative scenario: a laboratory receives a steady stream of messages about fasting and decides to begin only with that topic. The agent replies only with the preparation text approved by the technical lead. When a patient mentions taking a medication, the conversation goes to the team with the full history. At the end of the pilot, the team notices that a large share of the questions were about timing rather than preparation, and only then decides to extend the scope to scheduling.

How to test and monitor before expanding

Before putting the agent in contact with patients, test it the way you would test any system that makes decisions. Build a set of realistic, anonymized questions with expected answers and with the cases that must be handed off. Include hostile messages, requests for diagnosis and attempts to obtain someone else's report. The OWASP Top 10 for LLM applications lists this kind of manipulation among the risks of language model applications, and the NIST AI Risk Management Framework helps structure the assessment. A guide on testing with mocks and sandboxes shows how to do this without exposing real patients.

In operation, log inputs, responses, actions and handoffs with timestamps. Review these conversations regularly during the first weeks and update the approved texts whenever a recurring question reveals a gap in the protocol. The goal is not to eliminate every failure. It is to make sure each failure is noticed, recorded and corrected by the responsible team.

Frequently asked questions

Can superintelligence be used in a clinical laboratory today?

Not as a tool available for clinical or operational decisions. The realistic use today is language models inside bounded flows, such as answering preparation questions, scheduling and request status. Any discussion of more advanced systems should be kept separate from these day-to-day decisions.

Can an AI agent tell a patient their test result on WhatsApp?

By default, no. Interpreting and releasing results is the responsibility of the qualified professional and the laboratory's delivery process. The agent can let the patient know the report is available and point to the authorized access channel, but it should not repeat values or findings.

Where should a laboratory start organizing collection before using agents?

Start by mapping the full path from the request to the analysis, with the time of each step and the points where errors usually happen. Then list the most common questions and link each one to an approved official source. Without this map, any automation inherits the problems that already exist.

Does LGPD require special care with laboratory data?

Yes. Health data is sensitive personal data and receives enhanced protection. Before connecting an agent, the laboratory needs to know which data enters processing, who has access, how long it is kept and whether there is international transfer. These answers should be validated with the data protection officer and legal counsel.

Sources

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