
Superintelligence for Safety Training: Certificates Under Human Control
How to organize classes, reminders and training records with an AI agent, and why certificates should only be issued after human review.
Want to apply this in your company?
Book a call with Juan →An AI agent can organize Brazilian workplace safety training (NR courses), confirm attendance, send reminders and keep a record of what happened at each step. What it should not do is issue a certificate on its own. Issuance has to pass through review by a person responsible for the training, who checks the attendance list, the assessment and the requirements of the regulation before any document leaves the company. In practice, this is a narrow agent with well-defined tasks, not a superintelligence, as we explain in narrow AI, general AI and superintelligence: a scale for managers.
Key takeaways
- Class scheduling, reminders and certificates are three steps with different owners, and they should not be treated as one automation.
- The agent prepares the paperwork and the draft, but issuance stays with a person authorized by the company.
- WhatsApp reminders require the recipient's permission and must follow the platform's sending rules.
- Every certificate needs a trail: who reviewed it, when they reviewed it, and which attendance list they relied on.
- Before the agent goes live, test the workflow with fictitious data and an agreed success criterion.
Why the certificate cannot be issued automatically
A training certificate states that a person completed a course on a specific subject. An error in that document affects the worker, the company, and how the training will be read during an inspection or after an incident. Language models produce fluent, confident text, which makes it harder to notice a misspelled name, a wrong regulation or an inconsistent date. The NIST AI Risk Management Framework treats human oversight and governance as part of managing the risk of AI systems, and that logic applies directly to this case.
It is also necessary to be clear that the content, workload, validity and refresher requirements for each NR are set in the regulation itself. They must be confirmed by the company's technical lead, usually the occupational safety professional or the head of the occupational health and safety department (SESMT), where one exists. The agent should not interpret these rules on its own. They enter the system as parameters reviewed by that person, and any change to them goes through the same review.
How to separate class, reminder and certificate
These three tasks look similar, but their risks differ. A class booked with an error creates a scheduling conflict. A wrong reminder creates annoyance and, if it is an outbound message sent without proper permission, can create a problem with the contact. A wrong certificate produces a document that circulates under the company's name. For that reason, map each step with three questions: who decides, who executes, and who checks.
In most sensible designs, the agent handles the repetitive work, the training team decides what goes out, and the record remains available for audit. This division may feel bureaucratic at first, but it is what prevents a convenient automation from becoming an improper issuance.
How to organize classes with the agent
A workflow that tends to work follows this order:
- Registration: a list of employees with role, department and pending trainings, loaded by HR or the safety team, with a defined purpose.
- Planning: the agent suggests dates, room and instructor based on available seats and the queue of pending trainings, without confirming anything on its own.
- Invitation: the approved class is communicated to participants, and each response is recorded.
- Attendance: the list is signed or validated by the instructor on the day. The agent receives this list, it does not produce it.
- Assessment: the result is entered by the instructor or assessor, using the criteria defined by the program.
- Closing: the agent opens the queue of draft certificates and sends it for review.
Also define what the agent cannot change: attendance, grades, issue date and class cancellation. Any request in these areas must be routed to a person, with the reason recorded.
Reminders without becoming a nuisance
Reminders are the most natural use of WhatsApp in this context, and also where mistakes happen most often. Before a class, participants can receive a confirmation and a short reminder with the location, time and what they need to bring. For training that is about to expire, the notice should go to the employee and, if company policy provides for it, to the manager of their area, with enough lead time to schedule the course.
For contacts who have not started a conversation, use approved message templates and confirm that a registered sending list exists. See how the 24 hour WhatsApp API window limits what the agent can answer freely, and how outbound messages with templates and opt-in must be prepared before the first send. Reminders should also never disclose assessment results to third parties, and the agent should not answer questions about another person's grades.
The review gate before the certificate
This is the central point of the process. The agent can generate the draft and flag discrepancies, but the release is a recorded action taken by a person. Before releasing a certificate, the reviewer should check:
- Full name matched against the document presented at registration, not only against the version typed into a conversation.
- Course, regulation and workload matching the class that was actually held.
- Attendance validated by the instructor, with the date and class identified.
- Assessment recorded according to the criteria set by the technical lead.
- Issue date and validity following the guidance of the person responsible for the program.
- A release record with the name of the approver and the time of approval.
If any item fails, the certificate stays blocked and the case moves to the correction queue. The agent should not try to work around the failure, nor fill in missing information by guessing.
Permissions and personal data
Certificates contain personal data and, depending on the case, information tied to occupational health or assessment results. Brazil's LGPD (the general data protection law) requires a clear purpose, data minimization and security in processing. Official guidance on the subject is gathered on the ANPD page on gov.br, which is worth reading together with the company's data protection officer.
In the technical setup, the agent should have read access to the employee list, write access only to drafts, and no permission to issue. The principle of least privilege, described in the OWASP authorization recommendations, exists precisely so that a workflow error does not turn into an issued document. Also avoid sending certificate PDFs over WhatsApp to numbers that have not been confirmed.
Illustrative scenario
Imagine, purely as an example, an industrial maintenance company organizing a refresher class. The agent builds the list from the HR spreadsheet and finds an employee whose name is spelled differently in the safety department's records. Instead of correcting it on its own, the agent flags the discrepancy, keeps the certificate in draft, and sends the question to HR. The HR team checks the name on the identity document, corrects the record and releases the issuance. The log shows the discrepancy, who made the decision and when. The agent issued nothing, and the error was caught before anything left with the company's name on it.
How to apply the method in a pilot
The method we use has four moves: map the process and its baseline, limit data and actions, pilot a real workflow with logs and an agreed metric, and then transfer the operation and decide whether to expand, adjust or stop. Applied to safety training, this means mapping current classes and how long each step takes today, limiting the agent to registration, suggestion, invitation and drafting, and piloting a single NR with one real class.
A useful pilot metric is simple: zero certificates released without a recorded approval, and the number of discrepancies caught before issuance, tracked by the team. Before the real class, test the workflow with fictitious data, as described in how to test AI agents with mocks and a sandbox. A successful pilot does not prove that the agent should issue certificates. It shows whether human review works and whether the trail is complete.
Frequently asked questions
Can an AI agent issue a safety training certificate?
Technically, a system can generate the document, but this article recommends that issuance require recorded human approval. The person responsible for the training decides whether to release it. The agent can prepare everything so that this decision is quick and safe.
Do workers need to authorize expiry reminders for their safety training?
It depends on the channel, the relationship with the contact and the type of message. For outbound WhatsApp messages, use approved templates and a sending list with recorded permission. Confirm this point with legal counsel and the company's data protection officer before the first send.
What should we do if the agent finds an error in a certificate that has already been delivered?
Do not edit the document on your own. Record the case, notify the person responsible for the program, and correct it through the company's formal process, which usually includes reissuance with a new approval. Then review the log to understand at which step the error entered.
Where should we start without building a complete system?
Start with the registration spreadsheet and a simple invitation and confirmation flow, always with a record of every action. Measure it over a few classes, identify where discrepancies appear, and only then move on to automatic reminders and certificate drafts.