Illustrative scenario. A garage chain rolls out an agent that assigns work orders to mechanics based on urgency and load. It works: cars leave sooner. Two months in, a mechanic asks why he always gets Friday's long repairs, the staff representative asks how the agent decides, and nobody in management can answer with anything beyond "the AI does it." The project was an operational win and a legal gap.
The thesis in one line: in Spain, an agent that assigns tasks, shifts or evaluations is no longer just a product decision. It is information the workforce has a right to receive, and the company should be able to write it on one page before anyone asks.
Do you have to tell employees about AI use at work?
It depends on what the system decides, not on how much AI is inside it. If an algorithm or AI system influences working conditions, the company has two different duties to inform, and it pays not to mix them up: a collective one toward the workforce's legal representatives, and an individual one toward each worker.
- Collective. Article 64.4.d of the Spanish Workers' Statute, added in 2021 by the so-called "rider law" (Ley 12/2021), gives the works council or staff delegates the right to know the parameters, rules and instructions behind algorithms or AI systems that affect decisions on working conditions, access to and retention of employment, including profiling.
- Individual. Royal Decree 723/2026 of September 9, published in the Spanish official gazette (BOE) on September 15 and in force since October 5, 2026, requires informing each worker in writing that algorithmic or automated decision-making systems exist and, when they are used to decide working conditions, of their guidelines, criteria and operating rules.
This describes Spanish law. The EU directive the decree transposes is common, but each member state has written it into national law in its own way, so if your team works in another country, this list does not carry over as is.
What changed on October 5, 2026 with Royal Decree 723/2026?
The decree partially transposes Directive (EU) 2019/1152 on transparent and predictable working conditions and repeals Royal Decree 1659/1998. For anyone deploying agents, one letter matters: point k) of article 3.2, which adds to the information given when an employment relationship starts the existence of algorithmic or automated decision-making systems.
The text lists, as examples, the decisions that count: the length and distribution of working time, task assignment, pay, career progression, place of work and termination of the contract. Three practical details:
- When. The information is given before the employment relationship starts (art. 7.1); if the system later changes, no later than the day the change takes effect (art. 7.3). This chapter only applies to relationships lasting more than four weeks (art. 2.2).
- Existing staff. The transitional provision does not require handing out paperwork on October 5, but it does require providing the information within thirty working days to anyone who asks, and reporting any later change.
- Format. On paper or in an accessible electronic format (art. 6.2). The public employment service (SEPE) is to publish a model document, but until it exists the text says the company must inform anyway.
Does your agent count as an "algorithmic decision-making system"?
The useful question is not whether it is generative AI, but whether it decides, sets or changes a working condition. Some law firms go further and read the rule as also covering a tool that only recommends and leaves the sign-off to a person; if the agent proposes and the manager approves almost every time, the difference is more theoretical than practical. It is the same problem as the levels of autonomy of an agent: what counts is what happens in practice, not what the diagram says, and a human in the loop who approves without reading does not take you off the radar.
| Agent | Does it touch working conditions? | What you must be able to explain |
|---|---|---|
| Assigns work orders or tickets among team members | Yes: task assignment | Allocation criteria (urgency, load, specialty) and who can change them |
| Proposes shifts or schedules | Yes: working time and its distribution | What data it uses, which constraints it respects and how to ask for a review |
| Scores performance or calculates incentives | Yes: pay and progression | What it measures, with what weight, and what it does not measure |
| Screens job applications | Yes: access to employment, and likely high-risk under the AI Act | Screening criteria and real human review |
| Summarizes emails or drafts replies | No, unless someone uses it to evaluate people | Nothing specific; a usage policy, yes |
What has to be put in writing, and do you have to show the code?
The text talks about guidelines, criteria and operating rules. It does not mention source code or model weights. That changes the job: the point is not to open the box but to describe it honestly on a page that a non-engineer can follow. For each system, that page answers six questions:
- What it decides or recommends, and about which working condition.
- What personal data it uses and where that data comes from.
- Which criteria or rules it applies, in plain language and with exceptions.
- Who reviews the output and in which cases they can change it.
- How a person can ask for a decision affecting them to be reviewed.
- When and how changes will be announced.
Writing it forces you to know how your agent works, and that is the hidden advantage: if you cannot explain why it distributes tasks the way it does, you cannot defend that distribution to a customer, an auditor or the team itself.
How does a company prepare without slowing the project down?
You do not need a program with a steering committee; you need order. A sequence that fits in a month:
- Inventory. List the agents and automations that use staff data or affect people's work. If you still lack the agent inventory your company does not have, this is a good reason to build it.
- Classification. Flag which ones touch working time, tasks, pay, progression, place of work or termination, and which only assist.
- One-page sheet for each flagged system, with the six questions above.
- Delivery. The sheet goes to the workforce's legal representatives and through whatever individual channel HR uses: contract annex, supplementary document or employee portal.
- Changes. A change to rules, data or model that alters how the agent decides is a modification and must be announced in time. Hang it on the change process you already use to govern automation, as the guide on governance and control of AI automation sets out.
- Trail. Keep what the system decided and under which rule; without that, the sheet is a promise. That is what AI decision traceability covers.
Agents that handle job applications or staff data also have their own path under the EU AI Act, on a different timeline: complying with the AI Act while operating AI covers it. And for anything HR-related, first read where automating HR with AI helps and where it gets close to illegal.
When don't you need to inform?
When the agent does not shape people's work: customer service, invoice extraction, document classification. Even so, watch for drift. An agent that assigns tickets and is one day asked to also prioritize by performance has changed category without anyone writing it down. Review the classification every time its function changes.
Note: this article is not legal advice and uses no third-party figures. The regulatory data comes from the text of Royal Decree 723/2026 as published in the BOE; the exact wording of article 64.4.d of the Workers' Statute and the specific scope for your case should be confirmed with your labor counsel. The scenario is illustrative.