Question 1

What data does the AI chatbot learn from?

Accepted Answer

Your chatbot is trained exclusively on your business data, documents, FAQs, knowledge base articles, CMS content, and product catalogues. We use RAG (Retrieval Augmented Generation) with Pinecone vector storage and LangChain orchestration to ensure the AI retrieves verified information before generating any response. This means the chatbot never fabricates answers, it responds only with content grounded in your approved data sources, with citations traceable to the original document.

Question 2

Does the chatbot support Croatian and other languages?

Accepted Answer

Yes, we deploy multilingual LLM models (GPT-4o, Claude Sonnet) that handle Croatian, English, German, and 50+ additional languages simultaneously without extra configuration or cost. The chatbot detects the visitor's language automatically from their first message and responds in kind. This is particularly valuable for tourism, real estate, and e-commerce businesses serving international clients across multiple EU markets from a single deployment.

Question 3

Where can the chatbot be deployed?

Accepted Answer

On your website, mobile app, WhatsApp, Viber, Facebook Messenger, or as an internal Slack or Microsoft Teams bot. We build one AI core powered by LangChain and Pinecone, then deploy it across multiple channels through n8n workflow automation. This means every channel shares the same knowledge base and conversation quality. Adding a new channel takes days, not weeks, because the AI logic is decoupled from the delivery interface.

Question 4

How do you measure chatbot performance?

Accepted Answer

We track four core metrics through a real-time analytics dashboard: automated resolution rate (percentage of queries resolved without human handoff), average conversation length, user satisfaction ratings collected post-interaction, and conversions directly attributed to chatbot interactions. The resolution rate is tracked from the first week, so you see what the agent handles on its own before relying on it. The dashboard also surfaces unanswered questions, so your knowledge base improves continuously based on real visitor behaviour.

Question 5

Will the chatbot make answers up?

Accepted Answer

It can, if you let it, which is why we do not build it that way. A language model naturally produces text that sounds convincing, so without constraints it will answer things it does not know. The fix is to have it answer only from your own documents and pages, retrieving the source before generating a reply, and to explicitly allow it to say it does not have the information. Before launch we test it against real customer questions, including ones with no answer on the site.

Question 6

Does a chatbot replace our support team?

Accepted Answer

No, and it should not. A chatbot handles repeat questions well, and those are most of the volume: opening hours, availability, pricing, order status, how something works. Anything requiring judgement, an exception, or a person taking responsibility stays with people. The realistic benefit is that your support team stops spending the day on the same three questions, not that the team disappears.

Question 7

What happens when the chatbot does not know the answer?

Accepted Answer

It says so and hands the conversation on. That behaviour is deliberate: a clear handover to a contact form, email or a real person beats a guess that sends the customer in the wrong direction. Every handover is logged, so you can see which questions keep coming back unanswered. That list is usually the most useful by-product of the whole project, because it shows what your pages are missing.

Question 8

What does a chatbot cost to run once it is built?

Accepted Answer

The running cost is not our fee but model usage and hosting, and it scales with the number of conversations, the length of answers and the model chosen. Using a smaller model for simple questions and a larger one only for complex queries keeps that low, which is how we build it. Before launch we estimate a range from your expected traffic, and after launch you see actual usage instead of an estimate. What we do not do is a flat monthly figure that stays the same whether the bot gets used or not.

Question 9

Is a chatbot GDPR compliant?

Accepted Answer

It is when it is built that way, and that is part of the work rather than an afterthought. In practice: visitors are told they are talking to a bot, conversation content is not used to train models, only data with a purpose is retained and it has a deletion period, and processing runs with a provider covered by a data processing agreement. If the bot touches customer records in your systems, the scope of that access is defined before anything is connected.

Question 10

How much content do we need for the bot to be useful?

Accepted Answer

Less than most people expect, but it has to be accurate. The bot does not invent knowledge, it searches what you give it, so a handful of well-written documents on services, pricing, process and common questions usually beats a hundred pages of general marketing copy. If your answers contradict each other from document to document, the bot will expose that. Uncomfortable, but useful: the contradictions it surfaces were there before the bot existed.

Question 11

Can the bot complete an action, not just answer?

Accepted Answer

Yes, when it is connected to the system that performs the action. It can look up an order status, check an open slot, send a quote or open a ticket in your system, provided that system exposes an interface for it. Without that connection the bot stays at the level of information. That is why, before estimating, we ask what exactly you want it to do and which system it has to talk to.

Question 12

How do you prevent the chatbot from being abused?

Accepted Answer

By limiting scope, capping messages and monitoring usage. The bot answers only within your subject and refuses to be steered into unrelated requests, message counts per visitor are capped so cost and abuse cannot grow unchecked, and conversations are logged so unusual patterns are visible. None of this is perfect protection, but it covers the realistic cases: automated resource burning and attempts to push the bot into a different role.

AI Agents That Qualify Leads and Support Clients 24/7 in San Francisco