"People tell me I need a chatbot." For two years now, that sentence has come up in almost every first meeting. Sometimes it's an excellent idea: the business answers the same question thirty times a week and wins those hours back. Sometimes it's budget spent on a widget nobody opens. The difference isn't the technology — it's the same in both cases — but questions that are rarely asked before signing.
Three very different things share one name
When two people talk about a "chatbot", they almost never mean the same thing. And because all three versions look alike on screen — a bubble in the bottom right — people compare quotes that aren't remotely comparable.
The useful distinction isn't about which AI model is used. It's about what the chatbot is allowed to read, and what it's allowed to do. That's what determines its usefulness, its cost and its risk.
The answering machine
It knows a list of pre-written questions and answers. Outside that list it knows nothing. Useful for a first pass, frustrating as soon as you step outside the script.
The assistant that reads your documents
It answers from your real content: terms, prices, product sheets, procedures. It does nothing but answer — but it answers correctly.
The agent that acts
It checks a calendar, creates a request, opens a file, sends a quote. At that point it's not a chatbot: it's software with a conversation in front of it.
In practice, the vast majority of SMEs need level 2, and believe they need level 3. An assistant that answers correctly from your real documents solves the heart of the problem, for a fraction of the budget and without touching your existing systems.
The three situations where a chatbot genuinely pays off
A chatbot doesn't create demand: it absorbs a load that already exists. So it pays off exactly when that load is real and measurable.
The same questions, on a loop
Opening hours, lead times, areas covered, documents to provide, guarantees. If your team answers the same ten questions every week, that time can be quantified — and won back.
The hours when nobody answers
A significant share of enquiries arrive in the evening and at weekends. A visitor who gets their answer at 9pm doesn't go looking elsewhere at 9am the next day.
Multilingual, in Switzerland
Answering properly in French, German, Italian and English is expensive in staff. That's the case where a well-built assistant delivers the most value, and the fastest.
The three situations where it's an expensive gimmick
It has to be said just as plainly: in a fair number of cases, the honest answer is no.
When traffic is too low. A chatbot on a site with a few visits a day will save nothing, because there's nothing to absorb. The problem to solve comes earlier: being found. That's visibility work, not automation — the subject of our offers for digital marketing.
When the conversation is precisely what sells. On a high-value project, the human exchange isn't a cost to remove: it's the moment trust is built. Putting a robot in front of a qualified prospect means placing an automated switchboard at the entrance to your best deal.
When there's nothing reliable to give it to read. This is the most frequent case, and the least visible. If your prices live in the owner's head, your terms in a PDF from 2019 and your procedures nowhere, the chatbot has no reliable source. It can only improvise — and that brings us to the real risk.
The real risk: an AI that invents commits you
This is the point least talked about, and the only one that can cost you money directly.
A language model doesn't "know" anything: it produces the most plausible sequence of words. When it doesn't know an answer, it has no natural reflex to say "I don't know" — it fills the gap with something that resembles a good answer. On your website, that becomes an invented delivery time, a guarantee that doesn't exist, a slightly wrong price.
And this isn't a quality detail: a chatbot placed on your site speaks in your name. A client who picked up a promise on screen has grounds to hold you to it. The problem is no longer technical, it's commercial and contractual.
That's the question to ask any provider: "what does your chatbot answer when it doesn't know?" If there's no clear, demonstrable answer, there's no guardrail.
Talk it through with us →The good news is that the problem is manageable. Not by changing model, but by keeping it on a leash: forbidding it to answer outside your documents, closing the scope to the subjects you master, providing a handover to a human as soon as the question leaves the frame, and keeping the conversation history so you can see what's actually being said. A chatbot that answers "I'm passing your question to a colleague" is infinitely more professional than one that improvises.
Your data, the nFADP and the GDPR
Second question people forget to ask: where do the conversations go? A visitor typing into a chat bubble sometimes puts in their name, their address, the description of a dispute, even information about their health or finances. That's personal data, and it falls under the Swiss nFADP and the European GDPR.
Three points deserve an answer in writing before anything goes live: which provider processes the messages and in which country, how long conversations are kept, and whether they're used to train a model. It's also a matter of commercial common sense: nobody wants their clients' enquiries feeding somebody else's model.
None of this is a blocker — these are architecture choices made at the start, not fixes bodged on afterwards. We set out this logic in our article on cyber risks for SMEs.
What it really costs
As always with us: nobody can give a firm price without knowing the scope, and a price quoted blind is either padded to cover the unknown, or too low to win the signature. What we can be perfectly clear about is what moves the invoice.
The state of your content
This is the first item, and it surprises everyone. If your information is up to date and gathered together, the assistant is built quickly. If everything is scattered or out of date, the real work is putting things in order — and that work serves you far beyond the chatbot.
Connections to what exists
Answering from documents is one thing. Checking a calendar, stock or a client file is another: every system to connect is a small project within the project.
The level of autonomy
An assistant that informs needs far fewer guardrails than an agent allowed to book, modify or commit. The more it can act, the more rules, tests and traceability are required.
An honest order of magnitude: an assistant answering from your documents is counted in weeks, an agent connected to your business tools in months — the same rule as for the brochure site and the web app. And you have to budget a line quotes often forget: upkeep. Your prices change, your terms evolve; an assistant nobody updates becomes wrong within a few months, which is worse than no assistant at all.
The mistake that costs the most
It's almost always the same one: putting the chatbot first.
A chatbot is a layer laid on top of reliable information. Installed on a shaky base, it corrects nothing: it just spreads approximate answers faster and to more people. Many businesses buy the assistant hoping it will make up for the lack of clear documentation — exactly the opposite happens.
The order that works is unspectacular, but it works: gather and update the information you already repeat, publish it where your clients can read it, then plug an assistant on top. Often, by step two, part of the problem has already gone.
Our real work: telling you when a chatbot is pointless
At Renova Softwork, the most useful part of our job comes before the first line of code: the scoping. For a chatbot it fits into one simple question — how many hours a week does your team spend answering repetitive questions? When the answer is "an hour", there's no project. When it's "a day and a half", there is one, and we already know what it's worth.
We regularly turn down chatbot projects. Not for lack of appetite — it's an object we enjoy building — but because the need described can be solved with a well-written FAQ page, for a fraction of the budget. Selling the rocket to someone who needs a van means winning a project and losing a client.
And when the assistant is justified, we build it on a leash: plugged into your real content, with a closed scope, a handover to a human planned, and a readable history. That's the point of our offers for AI chatbots and custom software — start useful, and grow when the need really grows.
Frequently asked questions
Can a chatbot replace my customer service?
No, and that's not its job. A well-built assistant absorbs the repetitive questions — opening hours, lead times, terms, documents to provide — and leaves your teams the exchanges that call for judgement. The right measure of its usefulness isn't the number of conversations, but the number of hours given back to your team.
What happens if the chatbot gives a wrong answer?
That's the real risk, and it has to be handled at design time. A language model that doesn't know an answer tends to invent a plausible one, and that answer commits you because it appears in your name. The guardrails are to force it to answer only from your documents, close the scope, provide a handover to a human as soon as the question leaves the frame, and keep the conversation history.
Does my customer data go to a foreign artificial intelligence?
It depends entirely on the architecture chosen, and it's a question to ask before anything goes live. Three points must have an answer in writing: which provider processes the messages and in which country, how long conversations are kept, and whether they're used to train a model. Conversations contain personal data and therefore fall under the nFADP and the GDPR.
How long does it take to set up a chatbot?
An assistant answering from your existing documents is counted in weeks. An agent connected to your business tools — calendar, stock, client files — is counted in months, because each system has to be connected and its permissions framed. The most unpredictable factor isn't the technology: it's the state of your content at the start.
Is a chatbot useful if I have few visitors?
Rarely. A chatbot attracts nobody: it absorbs an existing load. On a site with a few visits a day there's nothing to absorb, and the budget is better invested in visibility. The first question to ask is: how many hours a week does my team spend answering the same questions?