An AI assistant that answers from your documents, not from the open web.

An ordinary chatbot works with predefined menu trees or with general world knowledge. A XICBOT assistant answers from your own content: sourced statements carry the excerpt they rest on, and an answer without a used citation is recorded as a gap. On top of that come tools that actually do something — show product cards from your catalogue, assemble a cart in the chat, take appointments, record enquiries and, in the Custom package, query a connected system. Every tool has a kind and an approval mode: reading may run without asking, anything that writes shows the visitor in plain text beforehand what will be sent. What the assistant cannot support with a source it is meant to say openly, and it can hand over to a human. Which functions are included is shown in the overview of functions.

From your own material: website and subpages, documents and PDFs, price lists, frequent questions and, on request, the shop catalogue with prices and availability. This becomes the knowledge base every answer rests on: each source is split into sections that keep their heading and origin, and these are combined into one state. A new state is built to completion first and then switched over in one go, so that answers are not half old and half new; older states stay alongside it for a while. If a run yields not a single section, it is discarded rather than switched in. What is not in there, the assistant is instructed not to answer on its own initiative.

It usually says so and offers a handover to a person — carrying the conversation so far, so that nobody has to ask twice. The model phrases that sentence under a fixed instruction; at the same time the question is recorded as a knowledge gap, and from the Pro package the list of knowledge gaps shows you which source is missing. For complaints and for legal or medical topics it hands over instead of advising: four protective rules stand above role, topic scope and handover rule and cannot be deselected. A language model can be wrong, which is why the handover is built in — and the detection behind it runs without a language model, so it still works when something is currently stuck at the model.

It leads into the shop; the selling happens there. The product_search tool type knows four forms — search, compare, bundle and also bought — returns at most six hits and only from the catalogue of your account; a stated article number beats the search term. The cart tool type knows six kinds of change, coupon among them, with at most 99 units per position. The cart is drawn as a card with line items, total and a shipping line. Price and stock come from the shop: the assistant assembles the cart in the chat, and “Continue” leads to the shop address you have stored, with a one-time pickup key attached; for the items to sit in the cart there, your shop needs an integration of its own. More under assistant for online shops and product cards.

Operation is in Germany; the answers are generated by a specialised model provider, also outside the EU on the basis of standard contractual clauses. There is a data processing agreement, a deletion concept and full data sovereignty on your side. Content and conversation data are neither passed on nor sold. The subprocessor register lists the self-operated language model gateway in Germany; what is passed to it is the conversation history and the excerpts found for it in your knowledge base, and behind it the answers are generated by a specialised model provider, also outside the EU on the basis of standard contractual clauses. A visitor's address is never stored in clear text but as a salted fingerprint per account. How long which data stays is up to you through retention periods per kind of object. Details under privacy and hosting.

With one line in your source code. The loader is the only file that runs on your page, and it never calls the network itself: everything that fetches data runs inside the isolated frame. A page with a strict content policy therefore does not have to weaken anything for the embed. The frame is built only once your page has finished loading, and it carries exactly one permission, namely to execute scripts, without access to your page's origin. Its two parts carry version names derived from the content together with a checksum. This works with common content management systems and with Shopware Community Edition. More under integration.

In the language it is written to in. That language is detected in our own code through function words and characteristic characters, with no external service and no extra model call; word lists exist for nine languages: German, English, French, Spanish, Italian, Dutch, Polish, Portuguese and Turkish. If a sentence is not clear-cut, the field stays empty instead of guessing German. Detection happens per sentence, which is why a switch mid-conversation is carried too. You do not have to maintain your content several times over: the knowledge base is not filtered by language, language acts as a ranking bonus. More under multilingual assistant.

Yes. Every project begins with a free initial consultation, and on request we set up a free demo with a test assistant for your own website. The demo is its own, leaner package: one assistant, one knowledge source and a small monthly allowance — enough to see how the assistant reads your content and answers questions before you choose a plan. An AI assistant can make mistakes, which is why we bind its answers to your own sources: sourced answers name their source, an answer without a used citation is recorded as a gap, and handover to a human is available on request.

The website assistant starts at 49 € per month net, the shop assistant at 149 € net, plus a one-off setup from 990 € and from 2,900 € net respectively. Included are hosting in Germany, maintenance of the knowledge base, security and model updates, and the data processing agreement. Operation comes with a monthly allowance: 500,000 tokens in the Start package, 3,000,000 in Pro. Consumption is shown in the administration under Billing next to the allowance, so a rise becomes visible before it is noticed elsewhere. The Custom package bills operation by scope, because there the connected systems determine the effort. All services can be cancelled monthly; there is no minimum term.

You do, if you want to: the interface shows sources, documents and conversations, and a changed source is read in again. Every source has its own refresh interval — preset to 7 days, adjustable between 1 and 365, and the value 0 explicitly means: not automatically. A run in the early morning works through the sources that are due; a source only counts as stale once more than twice its interval has passed since the last check, and at least 14 days. Unchanged pages are skipped, and a run that wrote nothing deletes nothing — otherwise a maintenance page would clear out your knowledge base. The analytics show which questions come up often and where an answer is missing. If you would rather not do it yourself, you hand it to us.