Over the past couple of years, nearly every business has been offered some version of an "AI chatbot" for its website. Most of these are generic widgets wired up to an off-the-shelf AI model, with no real connection to the business they're sitting on. They can respond fluently — but often incorrectly, or too generically to be genuinely useful.
We didn't want to simply resell a version of that to our own clients. So we built our own.
Irozon AI Assistant is a reusable AI assistant platform we designed, built, and production-tested — and it's currently running live on irozon.se. This article explains what the platform actually does, how it's built, and why that hands-on experience directly informs our work in AI Automation & Custom Systems.
Grounded in real business knowledge
An AI assistant that doesn't know a business's actual details is unreliable at best and misleading at worst. Ours answers from structured, approved business content — services, pricing, past projects, locations, and contact information — rather than guessing or inventing an answer. If something isn't in the approved knowledge base, the assistant honestly says it doesn't know, instead of filling the gap with a plausible-sounding but ungrounded guess.
Secure by design
AI provider API keys (OpenAI and others) never leave the server, and are never exposed in the browser, in any API response, or in any error message. Every call to the AI provider happens server-side, exclusively. What actually reaches the website is a public project identifier — comparable to a Stripe publishable key — that tells the backend which project's configuration to apply, not a credential in itself.
Isolated configuration for every client
The platform is built to run many independent client projects from a single deployment, without any project ever able to see or affect another. Every project has its own configuration, its own knowledge base, its own allowed website domains, and its own request limits — separated at the storage level, not just in application logic. That kind of isolation is exactly what's required to run the assistant for multiple clients at once without any information ever crossing between them.
A reusable React integration
The assistant is a self-contained React component that drops into a modern website with a handful of lines of code. It supports two modes:
- Hosted — the website connects to Irozon's own already-running backend using a public project key. Irozon handles the AI provider credentials, hosting, and capacity.
- Self-hosted — a business runs its own backend on its own infrastructure, with its own AI provider key, entirely independent of Irozon's hosting.
Both modes share the exact same component and the same streaming protocol, so a project can start self-hosted and move to hosted later (or the reverse) without rebuilding the integration.
Real-time responses, not waiting
Replies stream in token by token as the AI model generates them, rather than the visitor waiting for a complete answer to appear all at once. It makes longer responses feel noticeably faster and more like a real conversation.
A multilingual interface
Every part of the interface — buttons, placeholder text, error messages, and screen-reader text — ships built-in in both Swedish and English, and can be extended with additional languages without switching frameworks. That includes accessibility text for screen readers, not just the visible copy, so the experience is consistently translated end to end.
Built for mobile, not just desktop
The chat panel is built to work as well on a phone as on a desktop — including details that are easy to miss, like sizing the message input's font so mobile browsers don't unexpectedly zoom in when a visitor starts typing.
Conversations that persist, and control when you need it
An in-progress conversation is saved locally in the visitor's browser tab, so it isn't lost if the page accidentally reloads. A visitor can stop a response mid-stream with a Stop button, or start over completely with New Chat — both ordinary, expected behaviors that simpler chat widgets often skip.
Limits that protect against abuse and runaway cost
Every project has configurable limits on how many requests a single visitor, and the project as a whole, can generate within a time window. That protects against both abuse and a single traffic spike driving costs out of control — a basic requirement for offering this responsibly to more than one client at a time.
Built for both React and WordPress
Irozon AI Assistant exists in two independent forms, not one shared codebase forced to fit two different platforms. The reusable React assistant described above is built for modern web applications. In parallel, we've built a separate, dedicated WordPress plugin for WordPress and WooCommerce sites, supporting the same underlying principles — business-specific knowledge and secure, server-side AI provider configuration — implemented natively for the WordPress ecosystem. The WordPress plugin has been submitted to the WordPress.org plugin directory and is currently awaiting review.
The reason for two separate implementations, rather than one universal solution, is straightforward: different businesses use different platforms. Building separately for each ecosystem lets us adapt the integration to the platform a business already uses, instead of forcing companies to rebuild around a specific technical stack.
Why practical AI should solve a real problem, not just add a chat bubble
Adding a chat bubble to a website is easy. Building something that actually knows the business, stays within safe boundaries, works across every device and language, and can be run responsibly for multiple clients at once is considerably harder — and considerably more valuable. Our starting point is that AI on a business website should solve a concrete problem for the visitor or the business — faster answers to common questions, less manual work, a clearer path to contact — rather than exist as a technical demonstration.
How Irozon can help
Building and running this platform ourselves, in production on our own website, is exactly the kind of hands-on experience behind our work in AI Automation & Custom Systems — practical internal tools, integrations, and AI-supported workflows built around your actual processes and needs, not a generic solution bolted on afterward. Get in touch and we'll talk through what would be genuinely useful for your business.