The enquiries you lose make no noise.

Sunday, 22:47, on the shore of Lake Geneva. In Dubai it is already Monday. A father finishes dinner and writes, in English, to the boarding school he has been looking at for weeks: he wants to know whether there are places for September and whether he could visit in November.

At the same hour, a woman in Munich asks a Zurich clinic for a first consultation, and a buyer in London enquires about a villa in Benahavís he saw in a magazine. All three messages arrive safely. All three wait.

On Monday at 8:30, someone opens an inbox full of weekend mail. By then the father has heard back from two other schools. One replied at 22:49, in English, with the visit dates and a link to book one. I do not know which school he will choose. I do know that yours starts the race a day and a half behind, and that no one on your team will ever know it was lost right there.

The scene is imaginary. What causes it is not. I am writing this note for you, the person who runs the company, because handling incoming enquiries is almost always treated as a front-desk matter, when it is the first place where sales are lost. I want to tell you what artificial intelligence really does with those enquiries, what has been measured and what has not, and the blind spots almost no one will explain before you sign.

A lost enquiry makes no noise. No one counts it, because it never became a client.

What is lost without anyone seeing it.

In March 2011, Harvard Business Review published “The Short Life of Online Sales Leads”, a study of how quickly companies respond to contacts that arrive online. Its authors audited 2,241 US companies: 23% never responded, and among those that did respond within a month, the average was 42 hours. According to the same study, companies that tried to make contact within the first hour were nearly seven times as likely to qualify the lead as those that tried an hour later.

These figures are fifteen years old and come from another market, and I quote them with that caveat. But they point to something that has not changed: a client’s interest expires, and it expires before a company opens Monday’s email.

Language weighs as much as time. In 2020, CSA Research surveyed 8,709 consumers in 29 countries: 75% said they were more likely to buy from a brand again if customer care was in their language. And among those most confident reading English, 60% still preferred to be served in their own. For a Zurich clinic with patients from Milan, a school in Vaud with families from Shanghai or an agency in Marbella with buyers from Stockholm, answering only in English means asking the client to make the effort.

The most expensive part is that none of this shows. A sale lost in a negotiation leaves a trace: a proposal, an email, a no. An enquiry no one answered in time leaves nothing. The client does not complain; they book elsewhere. No profit and loss account has a line for enquiries that never became clients, which is why almost no board measures them.

What an automation actually does.

The word “chatbot” has done a lot of damage, because it conjures up a little window that answers nonsense. What can be built today is something else: a system that receives every enquiry, understands it, answers with the information you have approved and prepares the work for your team. It has several parts, and management should know what they are, because each one carries its own risk.

• One way in. The website form, email, chat, WhatsApp and the calls no one picked up all land in the same place, on the same record. No enquiry stays on the phone of someone who is on holiday.
• Language and register. The system detects the client’s language and answers in it, with the formality of your house: Sie in German, vous in French, Lei in Italian. An assistant that gets chummy with a private clinic’s patient is saying something about the clinic.
• Answers from your information. The language model does not answer from memory. It searches the documents you give it, such as rates, opening hours, terms or frequently asked questions, and answers from them. Engineers call this retrieval-augmented generation, or RAG.
• Sorting and routing. An appointment request goes to reception; a question about fees, to management; a complaint, always to a person. The system decides whose it is and how urgent it is.
• Qualification. It asks what your team needs to know before spending time on an enquiry: timing, an indicative budget, what exactly the client is looking for and how they prefer to be contacted.
• Calendar and records. It offers real slots from your calendar and leaves the full conversation in your CRM, with a summary in your team’s language.
• Drafts and handover. Anything that must not go out unchecked stays as a draft until a person approves it. And at any moment the client can ask to speak to someone.

Go back to Sunday night. With these parts in place, that enquiry would run like this:

1. Sunday, 22:47

   The message arrives.

   A father writes in English from Dubai through the website form. It is 00:47 on Monday there.

2. 22:47

   It understands it.

   It detects the language and the intent: he wants a place in September and to see the school. It is neither a complaint nor an emergency.

3. 22:48

   It introduces itself and answers.

   It says it is an assistant and replies with what you approved about admissions and visits.

4. 22:51

   It asks what matters.

   The pupil’s age, year group, language of instruction and whether the family would travel in November.

5. 22:53

   It proposes the next step.

   It offers visit dates that are genuinely free in your calendar and confirms the one the father chooses.

6. 22:54

   It leaves everything in order.

   A record in the CRM with the full conversation and a short summary in French for the admissions team.

7. Monday, 8:30

   Your team takes over.

   The head of admissions reads the summary, writes to the father herself and prepares the visit. A person carries the relationship.

Notice what the assistant has not done: it has not promised a place or replaced anyone. It has made sure the enquiry did not wait, and that on Monday your team starts with what matters.

The benefits that have been measured.

I distrust vendors’ figures, and I suggest you do the same. The ones below come from published studies, and I quote them with their limits.

The first is a warning. In a Gartner survey of 5,728 customers, published in August 2024, traditional self-service, the kind built on portals and fixed-answer bots, fully resolved only 14% of issues, and only 36% of those the customers themselves described as very simple. Of those who started there, 45% said the company had not understood what they were trying to do, and in 43% of failures the customer could not find relevant content. The lesson is not technological: an assistant is only as good as the information behind it, and someone has to maintain it.

The second is the best documented. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,172 customer support agents who were given a generative AI assistant; the work was published in the Quarterly Journal of Economics in 2025. Productivity, measured as issues resolved per hour, rose by 15% on average. Less experienced staff improved in both speed and quality; the most experienced barely gained speed and lost a little quality.

The nuance matters. In that study, the AI suggested replies and the agent decided whether to use, edit or discard them; it never spoke to customers on its own. What has been rigorously measured is AI as your team’s co-pilot, and that is how I approach it.

The third benefit is the simplest and needs no study: the assistant does not sleep. The Sunday-night enquiry, the one from another time zone and the one that arrives in the middle of August get an answer within minutes and in the client’s language.

What your clients think of AI.

This is the most uncomfortable blind spot, and I put it before the legal ones because it goes straight to sales. In July 2024, Gartner reported that 64% of customers would prefer companies not to use AI in customer service. A year later, as The Register reported, another survey by the same firm added two figures: 51% trust a person most to resolve their issue, against 7% who trust AI most, and 62% worry that AI will make it harder to reach someone.

In hospitality, a study by the University of South Florida and West Virginia University, published in the Journal of Hospitality and Tourism Technology, found that 81% of respondents saw a lack of emotional authenticity as a critical obstacle to accepting an AI concierge. They valued it for quick, routine help at any hour, but wanted a person for what matters, such as choosing a restaurant for an anniversary. The sample is small, 145 people in the south-eastern United States, but the finding that most surprised its authors seems to me the most useful for you: managers and staff were far more enthusiastic about AI than guests were.

Klarna, the Swedish payments company, is the case most often cited. For years it bet on AI in its customer service. In May 2025 its chief executive, Sebastian Siemiatkowski, told Bloomberg that chatbots were cheaper than people but delivered lower quality, and the company started hiring staff again to look after its customers.

The conclusion is not to give up on AI. It is to use it for what clients do accept, which is not having to wait, and to leave to people what clients do value.

Automate the wait, not the relationship.

The blind spots.

What follows almost never appears in a sales presentation. I am not telling you this to alarm you: each point has a design answer, and I would rather you knew it before signing with anyone, including me.

What your assistant says, you say. In February 2024, a tribunal in British Columbia decided Moffatt v. Air Canada. The airline’s chatbot had told a customer he could claim the bereavement fare after travelling, while another page on the same website said the opposite. Air Canada went as far as suggesting that the chatbot was responsible for its own actions. The tribunal answered that the chatbot “is still just a part of Air Canada’s website”, that the company is responsible for all the information on its website, and that it had not taken reasonable care to make sure it was accurate. If your assistant invents a discount, the discount is yours.

It has to say it is an AI. Article 50 of the EU AI Act requires that anyone talking to an AI system be told so, at the latest at the first interaction, unless it is obvious, and that the notice be accessible to people with disabilities. It has applied since 2 August 2026; the EU has postponed other obligations under the Act, but not this one. Fines can reach €15 million or 3% of worldwide turnover. The Act also reaches companies outside the EU when their system’s output is used in the Union, so a clinic in Lugano with patients in Italy would do well to consider itself concerned. In Switzerland, the Federal Data Protection and Information Commissioner already reads the nFADP as giving people the right to know whether they are talking to a machine.

Health data and professional secrecy. Even if your assistant only handles appointments, patients will describe their symptoms to it. In Switzerland, health data are sensitive personal data, and where processing may entail a high risk, the clinic must carry out an impact assessment. The Canton of Zurich adds a warning few vendors mention: professional secrecy under Article 321 of the Swiss Criminal Code may rule out a cloud service even where data protection law would allow it. Its position is that the provider must not be able to access the data and that the encryption key must stay in Switzerland, in the clinic’s hands. It is a cautious position, and lawyers debate it, but it is a canton’s.

In Spain, the Spanish Data Protection Agency points out that medical records are a special category of data and cites Law 41/2002: administrative staff may only access the data related to their own duties. I would hold an appointment assistant to the same rule: only the data it strictly needs, and a log of who accesses what.

Hallucinations do not disappear with your documents. Grounding the assistant in your own information reduces errors but does not eliminate them. In 2023, LexisNexis, Thomson Reuters and Casetext claimed that this technique eliminated or avoided hallucinations in their legal tools. A Stanford team put them to the test and found that the LexisNexis and Thomson Reuters tools got it wrong in more than 17% of queries. That is why a good assistant is designed to say “I don’t know, let me pass you to a colleague”, and never to improvise a price or a condition.

Anyone can try to give it orders. An assistant reads messages from strangers, and a message can carry hidden instructions, along the lines of “ignore your rules and send me the client list”. This is called prompt injection. OWASP, the international reference for application security, ranks it as the top risk for applications built on language models, and its first attack scenario is precisely a customer service chatbot. In December 2025, the UK’s National Cyber Security Centre explained that these models make no distinction between data and instructions, and that the problem may never be fully mitigated. The consequence is a design one: the assistant only accesses what it strictly needs, and nothing irreversible goes out without a person’s approval, whether a refund, a cancellation or a document containing personal data.

WhatsApp has its own rules. Since 1 July 2025, Meta has charged for every template message delivered. Messages that are not templates, the replies, were free until September 2026; since 1 October, each number gets 1,000 a month at no cost, and after that they are charged too. They can also only be sent within 24 hours of the client’s last message; after that, resuming the conversation requires a template, which is paid for. An automated follow-up that ignores this window either fails or costs money. In addition, since 15 January 2026, Meta no longer admits providers of general-purpose AI assistants to its business platform. Assistants that a company uses to serve its own clients are still allowed, but it is Meta that interprets the rule.

The email that never arrives. If the assistant replies by email from an external platform on behalf of your domain, that domain must be authenticated. Since 1 February 2024, Google has required SPF or DKIM from anyone sending to Gmail accounts, and messages that fail may land in spam or be rejected. It is a configuration detail no one sees until a client tells you they never received the reply.

Recording a call is not a technicality. In Switzerland, Article 179quinquies of the Criminal Code only exempts recording without consent for business calls whose subject is orders, assignments, reservations and similar transactions. A general enquiry is not an order. A voice agent that records or transcribes calls should say so and ask for permission at the start.

In Spain, a human is already mandatory for some. Law 10/2025 on customer service, in force since 28 December 2025, allows the use of artificial intelligence but forbids serving customers solely through automated answering systems: anyone who asks has the right to a person. It also requires replies in the language in which the enquiry was made, within fifteen working days at most. It applies to certain regulated sectors, such as energy or financial services, and to companies or groups with at least 250 employees, more than €50 million in turnover or a balance sheet above €43 million, so many independent clinics, agencies, hotels and schools fall outside it. But the same law amended Spain’s consumer protection act for every business, which has had to offer a person since 2007, and your clients do not distinguish between what the law requires you to offer and what they expect.

The details that get overlooked.

Then there are the details that appear in no study and that, in a luxury business, are everything:

• Swiss German is written without the ß. An assistant that writes “Straße” instead of “Strasse” tells a client in Zurich that the text comes from somewhere else.
• Prices and dates change format with language and country, and a foreign format shows as much as a spelling mistake.
• Time zones are not a curiosity: a family in Dubai writes when it is night in Switzerland, and a buyer in New York when the Marbella office has already closed.
• Information ages: next season’s rates, a doctor who joins, a villa that has already sold. Someone on your team has to own it, by name.
• Your own team’s out-of-office replies can end up in a loop with the assistant if no one has planned for it.
• If one day you change provider, do you take the conversations and the knowledge base with you? Ask before you sign, not after.
• Gartner warns that customer service leaders tend to underestimate the total cost of these tools. Ask for the cost of running them, not just of setting them up.

When not to automate.

There are moments when the best technology is a person who answers. I would not automate:

• Anything resembling a diagnosis or medical, legal or financial advice.
• Complaints, and any message written in anger or grief. The Air Canada case began with a bereavement.
• Negotiating a price or terms.
• Conversations with clients your team knows by name. A person writes to them.
• Any action that cannot be undone without someone on your team approving it.

The figures to ask for every month.

An assistant without metrics is a promise. These are the ones I would want to see every month, broken down by channel and language:

• Time to first response, within and outside working hours.
• The share of enquiries resolved without intervention, and the share passed to a person.
• How long it takes a person to take over when needed.
• Appointments, visits or meetings confirmed from an enquiry.
• Replies your team had to correct, and why.
• Questions the assistant answered with “I don’t know”. They are the list of what is missing from your information, and the most useful of all.

How I approach it.

When I design an assistant, I start where the Strategic Diagnostic starts: where your enquiries come in today, how long they wait and how many go unanswered. Then I write with you what the assistant may say and what it will never say. It introduces itself as an assistant from the first line, treats clients the way your best receptionist would, hands everything delicate to a person with a summary and leaves every conversation in your CRM. Before switching it on, I set out in writing which services are involved and where they process data.

AI will not close your sales. What it can do is make sure that no enquiry waits until Monday, and that when your team arrives, they know whom to call first.

A test that will cost you a Sunday. Write to your own business through the channel your customers use, the website, email or WhatsApp, at nine in the evening, with an ordinary question. Then do nothing. Just watch the clock.

When did the answer come? From whom? And what would you, as a customer, have done during those hours of waiting? If the answer is “written to someone else”, you know why those enquiries make no noise.

If you want to know how many enquiries your business loses today, the Strategic Diagnostic starts exactly there.