A problem everyone knows
A customer messages your website at 11:40 pm. They ask about prices, delivery or opening hours. Nobody answers until morning — and by then they have already bought from a competitor. Classic scripted chatbots break on the first unusual question, while “just plugging in ChatGPT” means handing your customer conversations to a model nobody controls: it can invent a discount, reply in the wrong language or “forget” that the question should have gone to a manager long ago.
We wanted an assistant a business can trust. That is why in xik the model is only one of the executors. Rules, permissions, conversation history and money stay under the platform’s control.
What the assistant can do
Answers from your own knowledge. The assistant relies on the organisation’s knowledge base: documents, service descriptions, FAQs. For frequent questions you can define standard answers, so the customer gets verified text instead of improvisation.
Speaks the customer’s language. The assistant detects the language of the person it is talking to and replies in it. 15 languages are supported today, including English, Ukrainian, Polish, German, Spanish, Japanese and Arabic. The default language can be set in the agent’s settings.
Calls a human when needed. An operator can join any conversation at any moment, take it over and later hand it back to the assistant. The customer sees that a manager has joined. A language bridge is built in: the customer writes in their language, the operator reads and replies in theirs, and the platform translates both ways automatically.
Knows your schedule. Opening hours and the business time zone are part of the settings, so the assistant can correctly say “we are closed right now, but…” instead of promising a call at 3 am.
Uses your words. System messages shown to customers (greeting, “an operator has joined”, “please wait” and so on) are editable in the admin panel for every language and support variables: phone, email, agent name and opening hours.
Launch without a developer
Creating a consultant does not require writing prompts or understanding models. In Agent Studio (in the web dashboard and the mobile app) an organisation admin:
1. picks the “Text consultant” template;
2. describes the business and its rules in plain language;
3. checks the assistant in a test chat;
4. activates it and configures the widget and opening hours.
The website widget is embedded with a single line:
<script src="https://…/public/v1/widget.js" data-site="my-shop" defer></script>Besides the website, the assistant works in Telegram. The channel architecture is open, so WhatsApp, Slack and others plug in as separate adapters without changing the core. For sales there are demo links: you can show a prospect the assistant before it is installed on their site.
Security: the model is an untrusted party
The most important architectural decision in xik is simple: we do not trust the model. User messages, knowledge base content and model output can all contain malicious instructions (prompt injection), and this cannot be solved 100% at the model level. So the goal is different: even a “compromised” model cannot exceed the permissions that the platform core checks deterministically.
- Tenant isolation at the database level (PostgreSQL Row-Level Security): one company’s data is unreachable from another company’s queries.
- Minimal widget permissions. A website visitor gets a short-lived, read-only identity and cannot trigger other agents or tools through the widget.
- Secret protection. Keys and tokens are automatically redacted from text before it reaches the model or the history.
- Provider keys never reach the browser. The web dashboard works through a server-side BFF with server sessions and mTLS to the core.
Predictable costs
An AI assistant should never be a surprise on your invoice. That is why xik has two levels of control:
- Technical budget: a cost and token limit per reply, set by the platform.
- Commercial quota: how many customer messages an agent handles per month. When it runs out, a “quiet mode” kicks in: the model is not called, and the customer never sees words like “limit” or “quota”. The quota can be topped up with packages that never expire.
Not locked into one AI provider
The xik core knows nothing about any specific model. The assistant can run on OpenAI, Anthropic Claude or Groq, and the super admin chooses the allowed models and executors for each organisation. If a cheaper or better model appears tomorrow, the business will not have to migrate conversations, settings or its knowledge base.
Who it is for
- Small businesses and freelancers who need to answer customers 24/7 without a support team.
- Online shops and services whose customers write in different languages.
- Support teams that need a “first line” able to hand a conversation to a person at the right time.
- Agencies and integrators running assistants for several clients: every organisation is isolated and has its own knowledge, quotas and settings.
Summary
The xik text assistant is not “yet another GPT chatbot”. It is a consultant that answers from your knowledge, speaks the customer’s language, knows your schedule and hands the conversation to a person in time. And it works within the limits you set, not the ones the model chooses.
Share
