How to choose an AI assistant for a student residence: ten checks and a one-week test

What an AI assistant for a residence is, and what it should do

In two sentences. It is the front line of the centre's channel: it reads every message that comes in, settles the ones that are plain information and turns the others into a job for someone on the team. And it should do five jobs, not one: answer from the house manual, take in faults, handle dining-hall notices, send out the team's announcements and step aside for a person when the matter calls for one.

What sets it apart from a menu chatbot is that it understands a sentence written any which way, without making anyone press "1" for maintenance. What sets it apart from a language model open to the public is that it holds no opinions: it only talks about your house. What AI is good for in this sector, with the messages of an ordinary day, is in AI for student residences. This page is about how to buy it.

Key takeaways

• Test it on what it does not know. The question that marks out a dependable assistant is one your manual does not answer.
• Follow one report to the end. A fault is only well handled if it reaches the phone of whoever fixes it and the resident learns the outcome.
• Ask who sees what, inside the team and outside it, and under which contract.
• Count your own work, not just the provider's. You write the manual, and it has to be kept alive.
• Decide on a week of real messages, not on a slide deck.

Ten things to check before you sign

They come in the order in which they tend to fail. Each has a test you can run yourselves, with no technical knowledge at all.

1. It answers from your manual and says when it does not know

Give it the house rules and the timetables, then ask it something that is not in them: whether a bicycle can be taken up to a room, say. The right answer is "I have nothing on that, I am passing it to the team". If it makes up a rule that sounds reasonable, rule it out. A resident cannot tell an invented rule from a real one, and you will find out the day somebody quotes it back to you. That failure has a name, a hallucination.

2. It steps aside for a person, and you know which one

Ask for the list of matters it never settles on its own. It has to include, at the very least, health, a fee or a deposit, an exception to the rules and friction between housemates. Then ask which phone each one lands on, and what happens if that person is off that day. "It is handed to a human" is not an answer. A post, a shift and a stand-in are. The guide what a residence's AI should answer and what it must hand on has the full list.

3. Faults arrive with a photo on the phone of whoever is on shift

Send the photo of a scorched socket without typing a word. A numbered report should open, and it should ring on the phone of the technician working at that moment, not in a shared inbox. Do it again on a Sunday. And watch the return journey: when the technician marks it as fixed, does the person who reported it know, or do they have to ask again?

4. The dining hall, if you have one

In a catered centre, half the inbox is about food. Check that it tells apart skipping dinner, asking for a packed lunch and asking for a plate to be kept, because for the kitchen they do not add up the same way. And try giving notice late: the sensible behaviour is to say the cut-off has gone and ask what to do, not to log it quietly or turn it down flat.

5. Notices from the team to the whole house

The channel also runs the other way: a power cut on Tuesday, the exam timetable. Ask who a notice can be addressed to (everyone, one floor, families), whether the number of recipients shows before sending and what happens with someone who has not written for a month. WhatsApp only allows open text during the 24-hour window; once that window has closed it requires an approved template.

6. Languages

Write to it in English halfway through a chat in Spanish. It has to switch with you, and breakfast has to be at the identical time in both languages. Ask as well what it leaves untranslated: general notices often go out exactly as the team words them, and that is worth knowing before September.

7. Who sees what: data protection

Three questions. Inside the team: can the kitchen read what a resident tells management? Outside the team: which country holds the database, where does the model run, and is it trained on your conversations? And on paper: the residence is the data controller and the provider its processor, so there has to be a data processing agreement you can read before the demo. What you need to have in order on your side is in the guide to GDPR in a student residence.

8. It says that it is an AI

The resident has to know there is a machine at the other end, and the EU AI Act asks for it. Look for it in the first message a new number receives, and see whether that notice explains how to ask for a person. How the duties are split between whoever builds the assistant and the centre that uses it is in the guide to the AI Act and a residence's WhatsApp.

9. It works on the channel residents already use

Open the reception phone and count where the last seven days came in. If nearly all of it arrived on WhatsApp, an assistant that lives on a website or in a separate application forces you to move the whole house before it is any use. The two routes, compared task by task, are in residence app or WhatsApp.

10. What you have to do to set it up and keep it current

No assistant arrives knowing what time your front desk closes. Ask how many sessions it takes, who has to sit in them, whether you keep your number and, above all, how a wrong answer gets put right: whether changing one line of the manual is enough or a request has to be raised with the provider. A manual nobody updates turns a good assistant into one that repeats, very politely, last year's timetable.

Questions for the demo, and the answer that should worry you

How to test it in a week

Seven days of testing teach more than three meetings. There is no need to open it to the whole house: the team and a handful of trusted residents are enough.

• Day 1. Hand over the manual as it stands, unpolished, and the list of who looks after what. That shows how much really has to be prepared.
• Day 2. Twenty real questions, copied from the reception phone. Note how many it answers well, how many it passes to the team and how many it makes up. The third figure has to be zero.
• Day 3. Three faults: one with a photo and no text, an urgent one in the small hours and a minor one at the weekend. See whose phone rings for each, and at what time.
• Day 4. Dining and notices: a cancellation in good time, a late one and an announcement to a single floor.
• Day 5. The delicate ones: a message about health, one about money and one about a housemate. None should get a substantive reply, and all three should end up in front of someone.
• Day 6. Another language, and a voice note.
• Day 7. Put three answers right by changing the manual and repeat the questions. Then sit down with the team: has it taken work away, or handed you a different kind?

If a provider will not agree to a test like this, on your own cases, that is information too.

Where Bedel fits, and what it does not do

Bedel, which publishes this guide, is an assistant of this kind for university residences and colegios mayores. It is only fair that it goes through its own ten checks:

• It answers from the centre's manual and, when the fact is not there, says so and passes it to the team. It does not learn on its own from what residents write: an answer is put right by changing the manual.
• Money, health, exceptions and conduct always end up with a person, at the post the centre's org chart sets. Anyone who writes HUMANO is asking to be attended by a member of staff. The detail is in escalation to people.
• A fault comes in as text, a photo or a voice note, gets a number and goes to the WhatsApp of the technician on shift, who accepts and closes it with buttons. Whoever reported it learns of each step. It is in faults and incidents.
• It handles dining cancellations, packed lunches and kept portions, with a cut-off time per meal. Announcements go out from the dashboard or from the team's WhatsApp, to the whole house or to a set of rooms.
• It replies in the language of the last message. The ones that are tested are Spanish, English, French, Italian, German and Portuguese, and the manual is written once, in Spanish.
• In the dashboard, each user has each section hidden, read-only or with permission to manage. The database is in the European Union, in Ireland. The model runs in the United States under standard contractual clauses, and its provider does not train on the centres' data. The data processing agreement is published, and the summary is in GDPR.
• It introduces itself as an AI assistant from the first message. Residents install nothing: they write to the centre's number, which can be the one they already had.

What it does not do. It is not a property management system: it does not allocate rooms, does not hold contracts and does not invoice. It takes no payments by any route, and sends no payment links through the chat. It does not control access: no turnstiles, no cards, no QR codes. It has no group view, because one account is one centre. It does not translate announcements or what the team types by hand, and it does not tell you who has read a notice. And it does not count anyone who gives no notice: if nobody says they are skipping dinner, the kitchen expects that person. The other tools that sit alongside an assistant like this are in software for student residences.