What an AI chatbot in student housing should answer, and what it must hand to a person
An AI assistant on a student residence's WhatsApp gets it right or wrong in the same place: knowing what is not its call. Telling someone when the laundry opens is easy. The hard part is recognising that "my chest really hurts" is not a question, that "my flatmate is threatening me" is not answered with the house rules, and that a complaint is not closed with a thank-you. This guide sorts, category by category, what the assistant can answer alone, what it turns into a ticket and what it must pass to a person, and what that handover should look like.
The short rule: answer information, hand on judgement
Information is whatever is written down and holds for every resident alike: an opening time, a house rule, where parcels are collected. Judgement is whatever depends on the person and the moment: whether a pain is serious, whether noise is a breach, whether a payment can be deferred. An AI can handle the first very well and should not attempt the second, even when it seems to know the answer.
Between the two sits a third thing: what has to be done. A repair is neither answered nor judged; it becomes a ticket with an owner and a deadline. That is where AI earns its place, precisely because it decides nothing: it sorts what comes in and puts it in front of whoever has to fix it.
Category by category
The nine categories that come up most on a residence's WhatsApp, with the three questions to ask of each:
| What comes in | Does it answer alone? | Does it open a ticket? | Does it alert a person? |
|---|---|---|---|
| Opening times (dining hall, reception, laundry) | Yes, with what the building's manual says | No | Only if the manual does not say |
| House rules (visitors, noise, shared kitchen) | Yes, quoting the written rule | No | If someone asks for an exception |
| Dining (cancellations, packed lunches, menu) | Logs cancellations and packed lunches; the menu, if it is written down | No | Allergens and diets: the kitchen confirms |
| Routine repair (bulb, blind, wifi) | Acknowledges it and gives the ticket number | Yes, with its category and deadline | The maintenance person on shift |
| Urgent repair (water, fire, smell of gas, someone locked in) | Only what is essential for safety | Yes, as urgent | The whole on-call post at once, at any hour |
| Health | No: neither diagnosis nor advice | No | Always, straight away |
| Conduct (noise, friction between residents) | Neither mediates nor sanctions | No | Whoever handles conduct in the building |
| Harassment or discrimination | No | No | Yes, and it also points to the confidential channel |
| Complaint | Acknowledges it and does not treat it as closed | No | Whoever answers for that service |
Anything not in the manual goes to the last column by default. A question with no written answer is not an invitation to improvise.
Allergens deserve a line of their own. The assistant can repeat the menu and the allergens in each dish if the dining hall has them in writing, but if a resident asks whether a dish is safe with their allergy, the kitchen answers. The reasons are in the guide to allergens in the dining hall, written for Spain.
And the urgent repair has a catch: the AI should say something useful for safety ("leave the room", "don't use that socket") if the building has it written down as an instruction. What it must not do is stop there. The alert to the person on call goes out at the same time as the reply, not afterwards.
What a good handover looks like
Handing on is not replying "please contact reception". That gives the problem back to the resident, who wrote precisely so as not to have to find anyone. A good handover has four parts:
- To whom. To one named person, chosen by role and shift, not to a staff group where everyone assumes someone else will see it. At night, to whoever is on call. That needs an org chart with shifts written down, not just in the head of management.
- Through what. Through the channel that person actually watches. An email at two in the morning is not a handover.
- With what context. Who is writing, from which room, what they said in their own words and since when. Whoever receives it should not have to ask everything again.
- How fast. Urgent matters, now. Everything else, with a deadline someone watches, with reminders before it runs out and, if it does, a step up to the next person. That is escalation to a human, and it works when it is written in an escalation matrix.
There is a fifth part, for the resident: knowing a person will read the message and, where possible, who and when. Silence after "I've passed this to the team" is what makes them write to the whole-building group next time.
What the rules and the codes ask
That people know it is an AI. From 2 August 2026, Article 50(1) of the EU AI Act, Regulation (EU) 2024/1689, requires that anyone interacting with an AI system is told so, unless it is obvious, at the latest at the first interaction. The duty falls mainly on whoever builds the assistant. And an assistant that serves residents is not high-risk as long as it does not decide admissions or score anyone. Who answers for what, in the guide to the AI Act and a residence's WhatsApp assistant.
That it does not ask too much. What a resident says about their health is special-category data (Article 9 GDPR; in the UK, the same article of UK GDPR). An assistant that asks about symptoms, medication or history in order to "help" is collecting data nobody asked it to collect. The right amount is the minimum needed to know whom to alert. More in the guide to GDPR in student accommodation.
That repairs have a clock. In the UK, the voluntary ANUK/Unipol Code of Standards for larger student developments (the version of 27 February 2026, clause 4.18) sets the standard that the private PBSA operators who join it commit to: emergency repairs completed within 24 hours of the report, urgent ones within 5 days and non-urgent ones within 28. It is a reference framework, not a law, and it binds only those who join. For an AI assistant the lesson is plain: the deadline starts when the resident writes, so the ticket has to exist from that moment, with its time.
That conduct and harassment take their proper route. In Spain, Act 3/2022 on university coexistence changed the disciplinary regime and opened the door to mediation, and each centre needs its own rules to reflect it, as the guide to house rules after Act 3/2022 explains. An AI is not a mediation body and cannot open proceedings: when there is a conflict, its job is to alert someone who can. If what arrives is harassment, the centre's protocol applies, and the one for residents is not the one for staff, as the guide to harassment protocols in student accommodation explains. What cannot be said on the shared WhatsApp also needs a confidential channel, kept apart from the chat; when one is required, in the guide to the whistleblowing channel.
Common mistakes
- Inventing an opening time. The AI cannot find the closing time and gives one that sounds right. The resident goes down and it is shut. The cause is rarely the AI alone: it is a manual with gaps and an instruction that does not forbid filling them.
- Answering about health. "Rest and drink some water" sounds kind and is medical advice to someone nobody has seen. The right thing is to acknowledge and alert.
- Closing complaints. Replying "thanks for your feedback" to a complaint treats it as dealt with when nobody has read it.
- Mediating a dispute. Reminding someone who complains about their flatmate of the quiet-hours rule looks neutral, but it takes sides without knowing anything.
- Handing on to a group. Sending the alert to the staff group spreads responsibility until nobody holds it.
- Not saying it is an AI. Beyond what the rules require, a resident who finds out late that they were talking to a machine stops trusting the good answers too.
Who reviews what it answers
A named person on the team. In the first weeks it pays to read the conversations almost daily: that is where the questions the manual does not cover and the replies that fall short show up. What gets corrected is the manual, not the single conversation, so that the next answer comes out right. And every handover has to leave a record: who wrote, who it went to, when it was accepted and how it was closed. Without that record, at the end of the year nobody can say what took how long.
How Bedel does it
Bedel, which publishes this guide, is an AI assistant on the residence's WhatsApp that follows this table: it answers from the centre's manual in the resident's language, opens every repair with its post, its deadline and its escalation, and puts health, emergencies, conduct and complaints on a person's phone. It says it is an AI from the first message. The detail is in AI for PBSA and student accommodation and on the page about the WhatsApp chatbot for student accommodation.
Frequently asked questions
What should a residence chatbot never answer?
Anything that depends on judging a person or a situation: health, disputes between residents, harassment, money, exceptions to a rule and complaints. Nor anything that is not written in the building's manual. In all of those, the right thing is to acknowledge the message and pass it to a named person.
How does a resident know they are talking to an AI?
Because the assistant itself says so, at the latest at the first interaction, unless it is obvious. Article 50(1) of the EU AI Act requires it from 2 August 2026, and the duty falls mainly on whoever builds the assistant. A clear notice at the start of the conversation is enough.
What happens at night if the repair is urgent?
The alert has to reach the whole on-call post at once, at any hour, rather than wait for reception to open. The AI can give the safety instruction the building has written down, but the alert to the person goes out at the same time. What can wait, on the other hand, can be held for the morning shift.
Who reviews what the AI answers?
A named person on the team with time to do it. Above all at the start, they read the conversations, see which questions the manual does not cover and correct it. The AI should not learn on its own from what residents tell it: what changes its answers is the manual.
Is there a record of every handover?
There should be. Each handover should show who wrote, who it went to, when it was accepted and how it was closed. That is what lets you know whether a deadline was met, answer a complaint with facts and see at the end of the year where the work gets stuck.
Can the AI decide whether something is urgent?
It can classify it with the rules the building gives it: which categories are urgent and at what times. What it must not do is downgrade an urgency on its own. When in doubt, the right call is to treat it as urgent and alert someone.