AI chatbot on your site: when it actually makes sense

Two years ago, putting an AI chatbot on your site was a fun project to look modern. In 2026 the discourse has shifted: the question is no longer «when do I install it?», it's «what does it solve and what does it cost to operate?». That's good news — it separates chatbots that contribute from those that just decorate.

The real state in Spain

Only 12% of Spanish small businesses use AI operationally today. That means two things: first, having a well-built chatbot is still a competitive advantage, not commodity. Second, most of those who jumped in did it badly — that's why you'll hear stories of chatbots causing more problems than they solve.

When it works

A well-designed chatbot resolves between 60% and 80% of queries hitting a website. Not all of them — that's exactly what separates a good chatbot from a lying one. Cases where it actually adds value:

• 24/7 reception for FAQs: hours, public pricing, location, what a service includes. People hit your site at 22:30 and leave if they don't find the answer. A bot that replies in 3 seconds retains that visit.
• Qualified lead capture: the chatbot converses, understands what the customer needs, and only then asks for email/whatsapp. Better than a cold contact form because the visitor is already engaged.
• Filtering unfit traffic: the bot detects when someone is looking for something you don't offer and tells them clearly, instead of leaving you a stale email you have to manually reply to say no.
• Pre-qualification for bookings: clinics, garages, advisory services. The chatbot asks what type of consultation it is, saving you the first email-tour of "what exactly do you need?".

There's a documented case of a dental clinic that saved €1,200/month (part-time receptionist) by replacing them with a €28/month chatbot. And a mechanic shop that automated €3,800/month in basic quotes. The data is there.

When it does NOT work

• Expert attention in regulated sectors: legal, medical, tax. A bot hallucination («yes, that's covered by social security», when it isn't) has real cost. If your profession demands legal, medical or financial precision, don't put generative AI customer-facing.
• Complex decisions: a coach selling €200 sessions shouldn't let the chatbot close. The closing conversation is done by a human, always. The bot captures the lead and stops there.
• Businesses without volume: if you get 5 visits a week, the chatbot solves a problem you don't have. Before automating, validate that there's traffic to automate.

How much it really costs

There are two very different brackets depending on the model:

• Pre-built SaaS (Tidio, Drift, Intercom): €15-50/month. Plug and play, standard templates. Pros: fast. Cons: comes out generic, lock-in, doesn't get the fine context of your business.
• Custom implementation (what I build): €1,500-4,000 setup + €0-15/month operation (Groq model, Supabase, no fixed fee). Personalized, no lock-in, with your own prompt and integrations. Typical payback: 3-6 months if the use case is well identified.

And if your case is very standard (FAQ + lead capture, nothing weird), an in-between solution can work: good SaaS + €200-500 of my setup so it doesn't come out cheesy.

The 3 rules that separate a useful chatbot from a depressing one

1. Human escape route always visible. "Talk to a person" or "WhatsApp me" button on every turn. Without it, the visitor feels trapped and leaves.
2. It must know how to say it doesn't know. A chatbot that makes up answers erodes your credibility. The system prompt has to be explicit: «if you're not sure, say you don't know and defer to the human».
3. Clear metrics. Not "interactions", but: how many qualified leads did it capture this month? What percentage of visits get resolved without intervention? What questions couldn't it answer? Without metrics, you don't know if it's helping or just there.

Try it right here

This site has a chatbot. It's at the bottom right. I built it. It runs on Groq (Llama 3.3 70B with fallback chain), conversations are stored so I can improve the bot with real data (with notice to visitors), and the logic is aligned with my own service: if you ask me about something I don't offer, it tells you. If you ask a price, it gives it. If you want to talk to me directly, it gives you the email.

Try it. Then decide if you want a similar one on your site.