Illustrative scenario. A veterinary clinic chain rolls out an agent that coordinates shift cover over WhatsApp. At 11:10 p.m. the agent texts a nursing assistant: “Can you cover tomorrow at 8?” She doesn’t answer, because she’s asleep. At 7:15 a.m. the agent repeats the message, marks the shift “uncovered” and alerts the manager. Nobody told it to do any of that, but nobody told it not to. The agent did exactly what it was built to do, and the design never asked what time it’s allowed to speak.
The thesis in one line: the right to digital disconnection doesn’t care whether the message comes from a boss or from a system acting for the boss. If your agent can message staff, your hours policy has to say when it can and when it can’t, and the agent has to comply by construction, not by good manners.
Does an AI agent that messages employees have to respect digital disconnection?
In practice, yes. The right is exercised against the employer, and an agent that messages staff does so as the employer. Whether a person types the message or a model generates it doesn’t change who answers for it. That is a reasonable reading of the rules, not settled case law on agents: the law speaks of employers and their managers and never mentions AI agents.
What the rules do say, in Spain:
- Article 88 of Organic Law 3/2018 (data protection and digital rights): workers have a right to digital disconnection outside legal or contractual working time, so that rest, leave, holidays and personal and family privacy are respected. The details are set by collective bargaining or, failing that, by the company through an internal policy, after hearing the workers’ representatives.
- Article 20 bis of the Workers’ Statute: recognizes that right under the data protection and digital rights legislation.
- Article 18 of Law 10/2021 on remote work: repeats the right for remote workers and requires the internal policy to define training and awareness actions on the reasonable use of technology tools.
This describes Spanish law. If your staff works in another country, the rules on rest and out-of-hours contact are different and this list won’t carry over as is.
What makes an AI agent different from a boss who writes late?
A boss who writes at 11 p.m. usually hesitates before sending and knows the message carries weight. An agent doesn’t hesitate. And it has three properties that make the problem worse:
- It has no working hours. Unless someone restricts it, it works at 3 a.m. exactly like at 11 a.m., because for it there is no difference.
- It scales. It can message the whole workforce in a minute. A pushy manager bothers one person; a misconfigured agent bothers two hundred.
- It persists. Retries and automated follow-ups are new messages. Every reminder for not answering at 11 p.m. is another interruption, and when the agent flags it as “no reply,” rest starts to look like non-compliance.
There is also a perception effect: a company message that arrives in the tone of an instruction reads as a demand, even if a model drafted it. A person who sees the alert on their phone can’t tell whether “the agent” expects an immediate reply, and that doubt is exactly what the policy has to remove.
What should the hours policy say for an AI agent?
It has to classify the messages the agent can send and say, for each class, what happens inside and outside working time. A one-page table is enough:
| Message type | During working time | Outside working time |
|---|---|---|
| Informational notice that asks for nothing | Sent | Held and sent in the next working window |
| Request that expects a reply (cover a shift, confirm details) | Sent | Not sent; waits for the working window unless someone with authority has declared an emergency |
| Reminder about an unanswered request | Sent, with a retry cap | Not sent. Silence outside hours doesn’t count as “no reply” |
| Security or critical-continuity alert | Sent | Only to the designated on-call person, never to the workforce |
Two rules are worth writing down explicitly. First: silence outside working time is not a breach, and no agent process may read it that way or escalate because of it. Second: the exception has an owner. Who can declare an emergency and who is on call are company decisions, with names attached, not something the agent decides on its own.
How do you make the agent actually comply?
A paragraph in a policy doesn’t bind an agent; its design does. A good part of this is the same problem as the levels of autonomy of an agent: “sending a message to a person” is an action with an effect in the world, and you have to decide in advance how much freedom the agent has. What works in production:
- A calendar per person. Working hours, shifts, holidays and sick leave live in the HR system, and the agent checks them before sending. They aren’t copied by hand into the prompt.
- An outbound queue with a window. The agent doesn’t write directly: it drops the message into a queue that only releases inside each recipient’s working window. Compliance then doesn’t depend on the model “remembering.”
- On-call, not broadcast. Outside hours, the only possible recipient is the on-call person, and only for the defined categories.
- Capped retries. A maximum number of attempts and no escalation to a superior because of silence.
- A log. Every message sent records the time, the recipient, the rule that allowed it and who can be asked to account for it. Without that, the policy is a promise; with it, it is AI decision traceability.
- A kill switch. A way to stop all of the agent’s outbound messages within a minute, for the day something fails.
All of this hooks into what you should already have to govern and control AI automation: rule changes with an owner, periodic review and a channel for staff to report that something isn’t working. And if you have read what the law asks about telling employees about AI use, the one-page sheet described there is the natural place to write when each agent may speak.
When don’t you need this much policy?
When the agent only talks to customers or to channels nobody receives as a phone notification. There the risk is different: customers have their own communication window and their own rights. But as soon as one of your staff is the recipient of an agent message, disconnection comes into play, and you want it in writing before someone asks why the system messaged them at 11 p.m.
Note: this article is not legal advice and uses no third-party figures. The legal references are Article 88 of Organic Law 3/2018, Article 20 bis of the Workers’ Statute and Article 18 of Law 10/2021; the exact scope for your case, including what your collective agreement says, should be confirmed with your labor counsel. The scenario is illustrative.