AI for Occupational Health Clinics: ASO Scheduling With Clear Limits

AI for Occupational Health Clinics: ASO Scheduling With Clear Limits

AI for ASO scheduling and client reports in Brazilian occupational health clinics, with restricted health data and human medical decisions.

•

Want to apply this in your company?

Book a call with Juan →

AI in occupational health, in practice, means using it to organize the schedule of workplace medical exams, track expiration dates, and produce status reports, without the software interpreting results or deciding anyone's fitness for work. For companies that contract occupational health services, the gain is a more predictable flow: the company knows who needs an ASO (Atestado de Saúde Ocupacional, the occupational health certificate), by what deadline, at which clinic unit, and what is still missing. The central discipline is treating every health data point as sensitive, with access restricted by role and medical responsibility kept intact.

Key takeaways

  • AI can organize the ASO schedule, reminders, and status, but the fitness opinion remains the act of the responsible physician.
  • Health data is sensitive personal data under the LGPD, so the agent should have minimum access defined by role.
  • Reports for HR and management should show status, deadlines, and pending items, never the clinical content of exams.
  • Before any expansion, run a pilot with one unit, log every action, and agree on one metric in advance.
  • Every message or report that touches health goes through human review and leaves an auditable record.

What AI does and does not do in ASO routines

The ASO is the document that records the outcome of an occupational exam, and its issuance is part of the Occupational Health Medical Control Program, the PCMSO, required by NR-7. The most common types are the pre-employment, periodic, return-to-work, job-change, and dismissal exams. Each has its own trigger and deadline, and this is where operations usually break down: the information is scattered across spreadsheets, emails, and the clinic's systems.

AI is useful for administrative and repetitive tasks. It can cross-check the employee list against due dates, suggest appointment times based on unit availability, send reminders, confirm attendance, check whether mandatory documents were received, and consolidate status by department. It can also draft a report that a professional reviews before it is sent.

What should not be delegated is clinical interpretation. Classifying fitness, commenting on a test result, answering questions about a diagnosis, or suggesting a course of treatment are acts of health professionals. An agent that does any of these, even with good intentions, creates medical, legal, and trust risks. That is why the first design decision is an explicit list of prohibited actions.

Which data the agent can see

The principle is minimization: the agent sees only what it needs to perform the task. To schedule and track due dates, it usually needs an internal identifier or name, role, department, unit, exam type, expected date, and status (scheduled, completed, pending document, overdue). It does not need, and should not have, access to results, anamnesis history, or any information that reveals a health condition.

When the workflow passes through a model hosted outside Brazil, the question gains another layer. Sending personal data abroad requires attention to LGPD rules, and the topic is covered in detail in international data transfer with AI. In many cases, the simplest solution is to send no clinical data to the model at all, keeping AI only on administrative data that has already been pseudonymized.

How to separate access by role

The access architecture should reflect who does what. OWASP publishes authorization guidance that works well as a foundation for this design: least privilege, verification on every request, and no shortcut to full access. In practical terms, each role gets a defined scope:

  • The scheduling agent sees only administrative data and status, with no clinical fields.
  • HR at the client company sees status, pending items, and due dates by department, with no medical reports.
  • The responsible physician and the clinic team access the medical record only inside their own systems, with authentication and logging.
  • The team operating the agent should not have access to clinical content, only technical logs and metadata.

This design also prevents a common problem: the AI system becoming a shortcut to data the person could not see in the original system. The agent's access should never exceed the access of the user it represents.

Implementation plan in stages

A safe rollout usually follows this sequence:

  • Map the current flow: who receives the request, where the schedule lives, who issues the ASO, how status reaches the company, and where delays happen.
  • List the data each step uses and mark, one by one, what is clinical and what is administrative.
  • Define in writing the permitted and prohibited actions, including what the agent should do when it is unsure or receives an out-of-scope request.
  • Choose the employee contact channel. WhatsApp reminders require care with permission and message templates, which is covered in templates and opt-in for WhatsApp outbound messages.
  • Test in a simulated environment with fictitious data before any real contact.
  • Run a pilot with one unit or one client company, keep a log of each action, and measure an agreed metric at the start, such as the number of ASOs that expire without prior notice.
  • Review the records with the clinic and the company, decide whether the flow continues, is adjusted, or stops, and only then expand.

Reports for client companies without exposing medical records

The report is where many operations leak data without noticing. The rule is simple: the report answers management questions, not health questions. Questions such as how many ASOs expire in the next thirty days by department, which document gaps block scheduling, and which units have a queue are legitimate. Questions about the reason for a work absence or the result of an exam are not.

A good monthly report shows totals by exam type, pending items with an owner, overdue deadlines, and a field for administrative notes completed by the reviewer. For audit purposes, it can include a list of issued documents with dates, without their content.

Illustrative scenario: a manufacturer with several departments asks its occupational health provider for a monthly report for HR. The agent consolidates due dates by department and sends a draft to the clinic coordinator, who reviews, adjusts, and only then releases the document to the company. No exam result leaves the clinic's system, and HR receives only what it needs to decide on schedules and deadlines.

Risks, records, and human review

The risk of a scheduling agent is not only technical. Incoming messages can contain malicious instructions or requests for third-party data, and OWASP lists this kind of risk in its Top 10 for LLM applications. The agent must treat any received text as untrusted and never perform a data access action without checking identity and scope.

Records are the other half of control. Every query, message sent, and report generated should have an author, date, purpose, and result. The NIST AI Risk Management Framework offers a structure for mapping these risks and defining responsibilities, and the ANPD is the national authority for organizations handling personal data that want to understand their obligations.

Automated decisions also require care. If any step in the flow affects a worker's situation, such as blocking a function or changing a record, there must be human review and a channel to contest the outcome. The topic is detailed in automated decisions and Article 20 of the LGPD.

How we apply the method to an ASO operation

In our work, the sequence is map, limit, pilot, and hand over. Mapping means understanding the real flow before proposing any automation. Limiting means defining data, roles, and prohibited actions in writing. Piloting means testing a real flow with logs and an agreed metric. Handing over means delivering a documented operation and deciding with the company whether it should be expanded, adjusted, or stopped. We do not promise productivity or compliance outcomes. What we measure is what was agreed at the start of the pilot.

Sectors with strong confidentiality, such as psychology, face a similar challenge, and the reasoning behind scheduling with clear limits applies well there. See superintelligence in psychology practices with confidentiality and scheduling for the parallel.

Frequently asked questions

Can AI decide whether a worker is fit for a role? No. Fitness is a medical act, recorded on the ASO by the responsible professional. The agent can organize the schedule and status, but it should not classify, interpret, or suggest conclusions about an exam.

Does the client company need to see exam results to track deadlines? No. For deadline management, the document status, expected date, and administrative gaps are enough. Results and clinical content should stay in the clinic's system, under the access of the health professional.

Where should a project like this start safely? Start by mapping the current flow and listing the data at each step, marking what is clinical. Then define permitted and prohibited actions, test with fictitious data, and run a small pilot with logs and an agreed metric before going live.

Do workers need to be told that AI is used in this process? Transparency is recommended. Employees should know that scheduling contacts may be supported by automation, which data is used, and how to reach a person. This communication should be reviewed by the legal team of both the company and the clinic, since every case has its own specifics.

Sources

occupational healthAI agentsLGPDhealth data privacyclinic operations