What a language model is and what an agent does
A language model generates text from learned patterns and the context it receives. An application may add search, documents or tools. Two applications based on similar models may therefore deliver very different results.
An AI agent can use tools and carry out steps towards an objective; text generation is only one possible capability. Access to documents does not guarantee correct interpretation. Distinguish between a model-written answer, a retrieved quotation and an action performed on a system.
Potential uses in preventive work
Limited tasks, reviewed materials and checkable outputs support useful application. Examples include turning notes into draft minutes, summarising a report, adapting vocabulary for a specific audience or comparing two versions of a procedure.
It can also suggest investigation questions or flag inconsistencies for human review. Regulatory research requires opening and checking the official source. Asking for “current legislation” is not enough. A risk assessment requires knowledge of actual tasks, equipment and conditions; describing them convincingly does not demonstrate that they have been observed.
Risks that must be controlled
Models may generate nonexistent references, mix obligations from different countries or retain outdated information. They may also omit an important exception, simplify an instruction until it becomes unsafe or assume facts nobody supplied.
NIST’s Generative AI Profile provides a voluntary framework for managing risks throughout system use. In prevention, relevant failures should be recorded and checked for recurrence. No generic accuracy percentage makes every application suitable: the criterion depends on the task and the consequences of error.
Verifying an answer before using it
- Define the task, country, site and expected document type.
- Supply only necessary information and separate facts from hypotheses.
- Ask for sources and uncertainties to be identified.
- Open every reference: check that it exists, supports the claim and is the applicable version.
- Compare measures, figures and conditions with the actual situation.
- Assign a competent review and retain the approved version when it forms part of the preventive documentation.
A second answer from the same model is not independent verification. If the source cannot be found or contradicts site data, the claim must remain unresolved.
Personal data, permissions and responsibility
Before entering information into a tool, review the contract, data uses, access and retention. The GDPR requires a lawful basis and appropriate measures; health data merit particular protection. Pseudonymisation reduces risks but does not automatically make information anonymous.
Use documents prepared for the authorised purpose and exclude unnecessary identifiers. Where the application can act, limit its permissions and define which actions require validation. In Spain, using AI does not transfer the employer’s protective duty to the tool or replace the competence of those performing preventive activities.
Hypothetical example: reviewing a cleaning instruction
A safety professional requests a draft for cleaning a machine using the current manual and site procedure. The response proposes a step absent from the sources and omits an energy isolation condition. The professional flags both issues and returns to the manual and assessment rather than accepting well-written text.
After review, the procedure is validated through the company’s usual process and the draft withdrawn. The failure is recorded to improve the tool’s instructions and testing. Value is measured by preparation and review time as well as errors detected, not by the number of documents generated.
