What characterizes an agent
An AI agent combines a model with tools and rules to work toward a goal. It can query information, organize steps, and perform permitted actions in a digital environment. Its scope varies: some prepare a written response, while others interact with records or workflows. The word “agent” alone does not imply a level of autonomy or a guarantee of reliability.
In occupational health and safety (OSH), it’s important to describe the specific function: what it receives, what it consults, what it can modify, and what it delivers. This description allows for an evaluation of its usefulness and risks. A convincing demonstration is not a substitute for verification with the actual data, procedures, and situations of the organization that will use the system.
Preventive uses and limits of the assignment
An agent can support inspection preparation by gathering background information, helping to locate documentation, or drafting a report from available records. They can also facilitate task tracking when connected to the appropriate systems. These uses require defining the expected outcome and who validates it before taking action.
Preparing information should not be confused with verifying a working condition. A summary of previous inspections does not prove that a protective measure is still in place. Nor does a list of training courses alone certify practical competence. Prevention requires maintaining the connection between digital records and the physical verifications that give them meaning.
Data, context and traceability
The outcome depends on the quality of the data and the context. It must be clear which company, center, period, and document version are being used. If records are missing or discrepancies exist, the agent should acknowledge this limitation rather than supplementing the information with seemingly certain assertions.
Digital traceability helps to verify the origin of a figure or conclusion. It is useful to be able to access the original record, distinguish retrieved information from generated text, and maintain the relationship with the version used. A reference included in a response must be verified: its visual presence does not guarantee that it actually supports the claim.
Permits and actions with consequences
The configuration should limit access to what is necessary for each task. Viewing, modifying, approving, and distributing an assessment are different actions and may require different controls. Agent permissions must respect the organization’s responsibilities and restrictions on personal data, especially health-related data.
It’s also important to define how a task is stopped, how changes are reviewed, and how the process is recovered from a failure. An error in a query can be corrected before it’s used; an error that propagates to multiple records requires a different response. The evaluation of AI use should consider these consequences and the workflow steps that could amplify them.
Human supervision and responsibility
Human oversight requires competent individuals with the knowledge and capacity to intervene. It must be determined which outputs need review, what checks are performed, and when specialized support should be requested. Routine approvals without sufficient time or access to the relevant sources can become a mere formality lacking effective control.
The addition of an agent does not replace the company’s preventive obligations, nor does it transform the system into a responsible healthcare professional or technician. Drafts must be incorporated into the workflow with the corresponding validations. The goal is to free up time for prevention and improve the consistency of information, while maintaining clarity regarding the final decision.
Magnus and Ethos in Sabentis
Sabentis offers the Magnus and Ethos tools. Its public presentation describes Ethos as supporting preventive documentation, with drafting, summarizing, and response functions linked to sources. Magnus is presented as connected to the organization’s data, with queries on records and monitoring of preventive information.
The same public documentation outlines applications related to reporting, inspections, and training, and specifies human review before results are distributed. These descriptions should be understood within the documented scope and available configuration of each environment. They do not imply that the agent can independently verify any physical risk or replace professional action that must be performed in the workplace.
Practical example
In a hypothetical example, a technician requests to prepare an inspection of a facility. The agent gathers previous findings and open measures and drafts a proposal of points to review. Upon checking the sources, the technician discovers that a production line has changed and that the background information no longer covers part of the facility.
The proposal is supplemented with current information, and the site visit verifies the physical conditions. The findings are recorded, and the necessary actions are assigned. The agent’s value lies in facilitating preparation and reducing searches, while the inspection and professional judgment provide the evidence needed for decision-making.
How to check its usefulness
Implementation should measure relevant results: accuracy of references, omissions, necessary corrections, review time, and effective task tracking. The number of documents generated does not demonstrate preventive improvement. It is also important that people can identify limitations and report errors without simply accepting incorrect answers out of habit.
Testing should be repeated when sources, permissions, models, or processes change. It’s advisable to start with clearly defined tasks and expand use based on evidence of successful operation. Sustainable improvement occurs when the agent is integrated into clearly defined responsibilities and helps make better-informed decisions about occupational safety and health.
