An AI assistant on a boarding school website has to know how to answer. Above all, it has to know when not to.

Picture this: a Sunday at 10:50 p.m., and the assistant on a boarding school website in the Bernese Oberland receives a message in Arabic. It is not about fees. A mother writes from Dubai that her son told her on the phone he does not want to go back to his boarding house, that he does not know who to tell, and that she does not know who to write to either. The assistant replies in flawless Arabic: it thanks her for her interest, summarises the wellbeing programme and offers to book a visit.

The assistant did what it was asked to do: answer. The problem is that at a boarding school, not everything that arrives is an enquiry. I build assistants like this, which is why I start with what they must not do. I think the first rule for AI on a school website is not what it can answer, but what it can recognise.

What is already the rule in England.

Families who know English schools arrive with different reference points, because there child protection comes with written rules. The UK government’s statutory guidance for schools in England, Keeping children safe in education, in its 2026 edition in force since 1 September, asks every school to have a child protection policy that is reviewed at least once a year and available publicly, via the school website or by other means. It also asks that the systems for pupils to report any form of abuse be well promoted, easily understood and easily accessible.

The same guidance says staff should never promise a child that they will not tell anyone about a report of abuse. And the standards for independent schools in England require the school to publish on its website how it meets its safeguarding duty. A family coming from that system knows what to look for, and notices when it is missing.

What Swiss law already asks of staff.

In Switzerland, since 1 January 2019, Article 314d of the Civil Code has required professionals in education, care, teaching and other fields with regular contact with children to report to the child protection authority when there are concrete indications that a child’s physical, psychological or sexual integrity is at risk and they cannot remedy it within their own work. Those bound by professional secrecy under the Criminal Code are excluded; they have a right to report. Reporting to a superior also fulfils the duty. And the cantons can add further obligations.

In other words, the school already has people who, by law, know what to do with a concern. What it often lacks is a path for a message written on the website, at night and in another language, to reach one of them.

The standard the UK government sets for AI.

Since January 2025, England’s Department for Education has published safety standards for generative AI products in education, updated on 19 January 2026. For tools used by pupils, it expects the product to identify disclosures that may indicate a safeguarding issue and alert the people responsible. It also expects the school to enter its safeguarding lead’s contact details during setup, for those details to be confirmed before activation, and for high-risk alerts to reach that person within an agreed timescale.

What I do not know.

I do not know what obligations your canton has added. I have found no public data on how many messages like this a boarding school receives, and I suspect few schools count them. And the English standard is written for tools used by pupils in England: an assistant that talks to families in Switzerland is not bound by it. I use it as a yardstick because it is the most concrete one I know.

The blind spot: language and time.

A concern like this almost never arrives with the word “abuse”. It arrives in the family’s language, at odd hours, wrapped in a practical question: whether a room can be changed, whether a child can come home early, whether someone can call their son tonight. An assistant set up for fees and visits reads it as one more enquiry. And the alert, if there is one, lands in an inbox that is opened on Monday.

Sometimes the best answer an assistant can give is to hand the conversation to a person.

Where I stand.

I believe no boarding school should switch on an AI assistant on its website without first deciding, in writing, which messages it does not answer, whom it alerts and how quickly. The assistant does not assess and does not promise secrecy. It gives the name and phone number of the person on duty and alerts that person there and then, with the original message and its translation. An adult does the rest. And the child protection policy, published in every language of the website, tells the family the same thing the assistant does. An assistant that does not know when to stay quiet has no place on a boarding school website.

What I would check.

• The published policy. Whether your child protection policy is on the website, in every language of the website, with the date of its last review.
• The person on duty. Whether the website says whom to call at night and at weekends, with a phone number, not just an email address.
• The assistant’s rules. Which messages it does not answer, in which languages it recognises them and what it does with them.
• The alert route. Whom it reaches, through which channel, how quickly, and who covers it during the holidays.
• The tests. Whether anyone has tried the assistant with ambiguous messages, in Chinese, Russian or Arabic, before switching it on.

Where to start.

1. Today

   Read your website as a worried family

   Look, in each language, for whom to write to if something is wrong, and count the clicks.

2. This month

   Write the rules

   With your safeguarding lead: which messages always go to a person, to whom and how quickly.

3. Before switching on

   Test with difficult messages

   Ambiguous sentences, in several languages and at odd hours, checking where each alert ends up.

4. Every school year

   Review alongside the policy

   When the child protection policy is reviewed, review the assistant’s rules and the phone numbers too.

Write, yourself, the hardest message your website could receive on a Sunday at two in the morning. Not the one from a family with questions about fees, but the one from a pupil who does not know who to tell.

Now answer honestly: who would read it, when, and what would the website say in the meantime? If the answer depends on an assistant, do you know what it would reply? And if there is none, what does that pupil understand when nobody answers?

If you would like me to review how your website, and your assistant if you already have one, responds today when a message is not an enquiry, start with the Strategic Diagnostic. What your canton requires and how it applies to your school is for your legal advisers and your safeguarding lead to decide, not a designer.