The chatbot that actually knows your business

There's a version of AI chatbot that every SMB has seen. It sits in the corner of a website, opens with "Hi! How can I help you today?", and answers every product question with "I'd be happy to help! You can find more information on our website." That's not a customer service tool. That's a polite dead end dressed up as one.

The problem isn't the chatbot format. The problem is that the AI behind it knows nothing about the business. It's a general-purpose language model that has been told to sound helpful. It hasn't read your product catalogue, your return policy, your measurement guides, or the FAQ your support team has updated forty times. It improvises. And improvisation in customer service is where trust goes to die.

I spent the past year building a different kind of system for a European client. They make specialized medical equipment, sold through clinicians and certified professionals, with a product line complex enough that even their own sales team occasionally needs to look things up. Three languages. Regulatory documentation that has to be cited precisely. Technical specifications where a wrong answer isn't just unhelpful — it's a clinical problem.

The system I built doesn't know everything. It knows what the client has written down. Every product FAQ, every instruction document, every measurement guide, every policy — ingested, chunked, embedded, and stored in a vector database on EU-hosted infrastructure. When a user asks a question, the agent retrieves the relevant pieces of that documentation, synthesizes a response, and tells you exactly which source it drew from. If the answer isn't in the knowledge base, it says so and offers to escalate.

That distinction matters more than any benchmark. A system that says "I don't know, let me get a human" is more trustworthy than one that confidently invents an answer. The confidence is the problem. SMBs have been burned by AI tools that hallucinate with complete authority.

The other thing I built was a workflow layer. The chatbot isn't just answering questions — it's capturing structured business events. A user wants a quote: the system accumulates the line items across the conversation and submits a formal record. Someone has a complaint: it opens a complaint case with a reference number and the right regulatory fields. A prosthetist wants sizing guidance: the agent prompts for a photo upload, runs it through a quality check, and delivers the measurement guide only when the image passes. These aren't chat transcripts. They're business data.

For SMBs in the EU, there's a third layer that matters: compliance. Most off-the-shelf chatbot products send your conversations to US servers. GDPR doesn't prohibit that, but it does require you to know it's happening, tell your users, and have a lawful basis for the transfer. For medical devices, financial services, professional services — anything regulated — that's a real legal and reputational exposure. The system I built runs entirely on EU-hosted infrastructure, uses EU-region Azure services for inference, and has a full deletion and anonymization workflow. No American processors for PII. Clean data residency. Audit logs for everything.

The businesses that get the most out of a system like this are the ones with genuine knowledge problems. A company that spends half its support hours answering product fit questions. A clinic that needs consistent answers across three languages without hiring three more staff. A B2B supplier where every customer conversation could become a sales record if someone was paying attention. The AI isn't replacing customer service — it's making the knowledge those businesses already have actually accessible at scale.

It took about twelve weeks to build the full system. The core agent, the retrieval pipeline, the admin dashboard, the knowledge base ingestion, the GDPR controls, the CI/CD, the multilingual handling. That's not a product you install from a marketplace. But it runs forever with minimal maintenance, the client's team updates the knowledge base through a browser interface without touching code, and the cost per conversation is a fraction of what a support agent costs per hour.

Most businesses don't need this level of complexity. A lot of them just need something that stops making their customers feel ignored. But for the ones that have built real knowledge and want it to actually work for them — the ones for whom a wrong answer has real consequences — the generic tool isn't going to cut it.

If you're building something like this, or thinking about it, I'm happy to talk through what's actually involved.