aceArt · AI Agency

AI Chatbot Development

A chatbot that answers from general world knowledge is about as useful to a company as a search engine in the lobby. It gets interesting when it knows your manuals, your product data, your price list and the internal instruction from last autumn. That is what we build, and we build it so that it says where an answer came from.

Get expert advice

Two use cases, one technique

The bot on your website

It answers what customers ask for the twentieth time, around the clock, without anyone in support getting up at night for it. What matters is less the range than the boundary. A bot that honestly hands over to a person when a question goes beyond its knowledge will be accepted. One that invents something costs trust that nobody gets back later.

The internal assistant

It sits behind the login and knows what lies scattered around the company: the wiki, manuals, work instructions, old quotes, technical data sheets. The benefit usually does not show up in hours saved, but in new colleagues working independently much sooner and experienced ones being interrupted less.

RAG, or why the bot does not make things up

A language model knows nothing about your company. Ask it anyway and it will produce an answer that sounds plausible, because it is trained to sound plausible, not to be right. That is the core of the hallucination problem, and no prompt gets rid of it.

Retrieval-Augmented Generation turns the order around. Before the model answers, your documents are searched for matching passages. Those passages are handed to it along with the instruction to answer only from them. That shifts the task from knowing to summarising, and a language model is good at summarising. In detail in the glossary.

What that means in practice

  • Every answer can name its source. Anyone who wants to check clicks through to the document.
  • New documents are available immediately. Nothing has to be retrained.
  • If the search finds nothing suitable, the bot says so instead of guessing. That rule is the most important one of all.
  • Access rights stay intact. Anyone who is not allowed to see a document does not get it from the bot either.

Incidentally, the quality of a RAG system is not decided by the model but by the search in front of it. If the wrong passages are found, even the best model cannot help. That is where the craft of the work sits: how documents are split up, how they are searched and how the hits are ranked.

Where your data sits

Whatever is sent to an external model leaves the building. That is not a legal nicety but the central decision of the project, and it belongs at the start rather than in the acceptance test. Depending on the requirements we work with providers that process inside the EU, with a properly regulated data processing agreement, or with models that run entirely in your own environment.

The last route costs more in operations and often gives slightly weaker answers. In return, not a single sentence leaves your network. Which route is the right one depends on what is in the documents. A bot on public product information is measured differently from one on personnel files.

From chatbot to agent

A chatbot answers. An agent acts: it calls systems, opens a case, pulls data out of the CRM and works through several steps in a row. The transition is fluid, and many projects start as a chatbot and grow in exactly that direction as soon as the question comes up, "can it not just enter that straight away?".

You can read that in search behaviour too, where the term agent is currently overtaking the chatbot. Technically it is the same foundation, but the demands on boundaries and approvals rise sharply. More on this on the page about AI agents.

How we go about it

  • We first clarify which questions the bot should answer, and above all which it should not.
  • We go through the documents. Often that is the moment when someone notices that three versions of the same instruction are in circulation.
  • An early version goes to a small group of real users. The questions that come in there teach you more than any amount of planning.
  • We measure how often the bot finds nothing, and work on exactly those gaps.
  • After going live the bot stays under observation. Documents change, models get replaced, and the costs run along with usage.

Frequently asked questions about AI chatbots

What does an AI chatbot cost?
Two items belong apart here. Development depends on how many documents there are, how clean they are, and how deeply the bot reaches into your systems. On top of that come running costs, because every request to a model is billed. We work that second item out in advance, because it does not exist like this in classic software and is therefore regularly forgotten.
What is RAG, and why does the bot not hallucinate with it?
With Retrieval-Augmented Generation your documents are searched before the answer is given, and the model may only answer from the passages that were found. So it does not have to know anything, it only has to summarise. A wrong answer cannot be ruled out entirely, but it becomes rare and it can be checked, because the source is delivered with it.
Can the chatbot access our documents?
Yes, that is the actual point. PDF, Word, wiki pages, intranet content, product databases and ticket systems are all common. The format matters less than the question of whether the content is looked after. A bot on outdated documents reliably gives outdated answers.
Where does our data sit?
You decide that and we follow. Possible are providers that process inside the EU, data processing with clear rules, or a model that runs entirely on your side. The last variant costs more in operations and often gives slightly weaker answers, but not a single sentence leaves your network.
How long does the development take?
A first usable version on a manageable set of documents is realistic within a few weeks. It takes longer when rights are finely graded, when many formats come together or when the documents have to be tidied up first. That is why we prefer to start with one area rather than with the whole company.
What happens if the bot says something wrong?
There are two safeguards for that. First, it answers only from sources it has found and it names them, so a mistake can be traced. Second, we limit what it answers about at all. On legally or financially delicate topics it hands over to a person instead of trying its luck.
Can we look after the bot ourselves?
Yes. Normally looking after it means keeping documents up to date, because the bot draws its knowledge from them. For adjustments to behaviour and tone we set up an access with which your team manages without us.
Does this work on a WordPress or TYPO3 site as well?
Yes, embedding it into an existing website is the normal case. We build websites in both systems ourselves, so the connection is routine for us and not the actual work. More on this under AI integration.

Related topics

AI agents

For when the bot should not only answer but act. To the AI agents.

AI integration

The connection to your website, CRM or ERP. To AI integration.

AI consulting

For when it is still open whether a chatbot is the right way in at all. To AI consulting.

Ready for your KI project?

Tell us about your project. We'll get back to you promptly with an honest assessment.

Contact us
HOME › AI-CHATBOT-DEVELOPMENT