Two scenes, the same kind of system, opposite outcomes.
First scene. A private clinic patient, sixty-two years old, realises on Sunday evening that his blood test is scheduled for Wednesday. He has a question: does he have to fast in the morning, or can he have a light breakfast? He has always called on Monday morning for such things, but this time he picks up his phone, opens WhatsApp, looks up the clinic among his contacts, writes the question. He receives a reply after twenty-six seconds: "Good evening, Mr Rossi, for the exam booked on Wednesday at 8 a full eight-hour fast is required, only water is allowed. If you have any doubts about the preparation, I am here." Signed "Such-and-Such Clinic, Patient Line". The patient is satisfied. The message came from an automatic system that knows his appointment, knows the clinic's procedures, and replied while the receptionist was at home.
Second scene. A potential client visits the website of an independent financial advisory firm. He found the name among those recommended by a friend. On the "services" page a chat appears at the bottom right with a smiling avatar. He writes: "I'd like to understand if you take clients with assets of 200,000 euros or less". After seven seconds the reply begins to appear: "Hi! Thanks for your message! Our company serves clients with various asset profiles. Please fill in the contact form to be contacted by one of our advisors!". The potential client closes the site. He does not fill in the form. The next day he forgets the company's name.
The system underneath the two scenes is practically the same. A language model that replies. A small knowledge base. An integration to send. But the results are diametrically opposite. The first scene creates a more loyal client. The second loses an opportunity. The difference is not technical. It is the environment.
The ignored variable: latency tolerance
In any conversation, there is a waiting time that the user considers normal before the other side replies. This time varies enormously depending on the channel.
On a website chat, with avatar and animated "typing...", latency tolerance is less than three seconds. The user expects an instant reply because the environment suggests it: the interface is built like a live conversation, the avatar smiles in real time, everything communicates presence. As soon as the reply is delayed, the user interprets the delay as something anomalous and often closes the window.
On WhatsApp, latency tolerance is radically different. A user writes to a company and expects a reply in an hour or even more. WhatsApp is, by culture, an asynchronous channel. Even when the reply arrives in thirty seconds, the user perceives it as quick, not as owed. This creates an enormous operational margin that website chat does not offer.
This difference is the variable nobody discusses when deciding to install a chatbot. People discuss features. They discuss the quality of the replies. They discuss the price. But the latency tolerance of the specific channel decides, in advance, whether the system can win or whether it is doomed to fail.
An AI system that answers well takes time. If it has to consult the company's knowledge base, verify a client history, decide the right tone, formulate a reply consistent with the company style, and in some cases request confirmation from a human operator for sensitive steps, the time needed is never under ten to fifteen seconds for quality replies.
On a site widget, where the user expects a reply in three seconds, that time is not there. To stay within three seconds, the chatbot must reply in superficial mode: keyword pattern matching, pre-written phrases, invitation to fill in a form. This mode produces the embarrassing replies we have all seen. "Thanks for your message!", "Please get in touch again!", "Our company offers quality solutions". The user notices immediately and leaves.
On WhatsApp, where tolerance runs in the order of minutes, the system has the time to do what it needs to do well. It can consult documents, decide whether a question requires a human or an automatism, generate a personalised reply, include client-specific information. The same technological complexity that on a website would be a sentence, on WhatsApp is an advantage.
This explains why website chatbots, even those built with advanced language models, tend to have a poor reputation, while AI systems operating via WhatsApp are starting to gather approval among the companies adopting them. The difference is not in the underlying tool; it is in the congruence between channel expectations and the response time that tool requires to produce quality.
The Italian professional audience
To this logic a second factor is added, often ignored in chatbot discussions. The Italian audience over forty-five, which in many professional sectors (law firms, tax consultancies, clinics, financial advisory, estate agencies) represents the majority of clients, does not use website chats. They find them vaguely annoying. When they look for contact with a company, they prefer two channels: telephone and WhatsApp.
Installing a chatbot on the website, even a good one, for this audience means speaking in a channel that for them does not exist. It is the equivalent of placing an intercom in front of a shop where clients enter through the side door.
WhatsApp, by contrast, is already open on the phone of anyone over forty-five. It requires learning nothing. It has no installation friction. The company number sits among already saved contacts or is added with two clicks. For the target audience of many Italian and Ticino SMEs, WhatsApp is not a messaging channel: it is the de facto interface with the world of services.
Operational implications
A company that recognises these two dynamics (different latency tolerance by channel, WhatsApp as default interface of its audience) makes different choices compared to the average.
It does not install the chat on the website. Or if it keeps it, it treats it as a secondary channel, not as the spearhead of automatic reply. The spearhead runs on WhatsApp.
It dedicates time to building a WhatsApp automatic reply system consistent with the perceived quality of the company. The tone, the style, the ability to distinguish between questions that can be answered automatically and questions that must pass to a human. A system that makes these distinctions well is accepted by clients as a normal extension of the practice, not as an obstacle.
It accepts that, to work well, the system requires a structured business knowledge base. It cannot reply correctly without knowing who that client is, what he has already received, which procedures apply to his case. This base is preliminary work, not an extension of the WhatsApp Business subscription.
Most companies that decide to "do conversational AI" start from the wrong side. They look for the best widget for the website, compare plugins, evaluate prices. This search, even if well done, inevitably arrives at a product that in the chosen channel produces the effect of the second opening scene: hasty replies, lost opportunities, erosion of perceived quality.
Work done well starts instead from the question: on which channel does my client prefer to converse? For most Italian SMEs, the answer is WhatsApp. From there, everything changes. The choice of system, the investment in the knowledge base, the type of integrations with internal systems, the escalation path to human operators.
A tax consultancy that discovers the pattern of clients writing on Sunday evening about weekly deadlines, and builds a WhatsApp system that replies competently on those questions, gains hours of secretarial time on Monday and improves the perception of responsiveness. The same practice, had it installed a chatbot on the website, would probably have lost the chance because no client of that audience opens the website on Sunday evening.
A practical criterion: look at the first source of contact for your last ten clients. If they are predominantly WhatsApp, telephone, direct email, recommendations from acquaintances, the channel on which to concentrate conversational AI is WhatsApp. If they are predominantly site forms filled in by users who found the company on Google, then site chat may have a role, but almost always a secondary role compared with optimising the form itself.
This distinction is the most important one to make before investing in a conversational system. It is a choice that is not resolved with a demo. It is resolved by observing how your real clients behave today.
This observation on who uses what is the base on which a conversational system that actually solves a problem is built, instead of adding another encumbrance. Without it, you end up investing in the wrong channel for your audience.
If you recognise the scene of the website losing an opportunity, and you notice that your clients are already using WhatsApp even though you have never formally proposed it, that is the moment to treat WhatsApp as a work surface, not as a courtesy channel. It is the kind of reorganisation worth mapping in a dedicated conversation, before deciding what to install.