Orbit Chat AI
The speed of AI, without losing control of the conversation.
Orbit Chat AI is a no-code platform where a business trains an AI assistant on its own website, files and answers, then runs customer conversations across Instagram, Messenger, TikTok and WhatsApp. I designed it from sign-up to daily operations, with three priorities: setup in minutes, an AI the owner can control, and a smooth handover between the AI and human agents.
- Role
- Product Designer
- Platform
- Web & mobile app
- Timeline
- 2.5 months
- Team
- 3 developers, 1 QA
- Tools
- Figma
What I owned: design system, onboarding, bot training, widget setup, shared inbox, products, services, orders, channels, settings and email templates.
Context
Small businesses in Nepal sell and support customers through social media messages. Replies are slow and repetitive, and conversations live on one person’s phone.
Orbit turns a business’s own content into an assistant that answers customers, takes product and service orders, and escalates to a human when needed.
The problem
Businesses want AI’s speed, but they don’t want to write prompts, manage code, or lose control when a customer needs a person.
Owners need a working, branded bot in minutes, without technical skills.
Teams need to see chats, take over from the AI, assign and reuse replies.
Unless products, services and orders connect to the conversation.
Owners need control over what the AI knows, how it speaks and where it replies.
Users & goals
Quick setup, control over tone and sources, usage and billing visibility.
A clear inbox, the ability to take over, canned replies and customer context.
Products, variants and orders captured straight from chats.
A fast, friendly assistant that can still reach a human.
Goals
- Fast time to a first working botTime from sign-up to a published widget
- An accurate, controllable AIChats resolved by AI vs handed over
- Smooth AI-to-human collaborationTime to first human reply; unassigned chats
- Turn chats into ordersOrders created from conversations
Constraints: platform rules like Instagram’s 24-hour reply window; metered usage (credits, storage, channel access); four training sources with different upload and refresh behaviour; and multiple roles with different permissions.
System first, screens second
Before any screens, I defined the design system: colour scales (Primary, Neutral, Success, Warning, Destructive), a type scale from H1 to H5 with paragraph sizes, and reusable components including badges in three sizes, tabs, navigation, cards, tables and dialogs.
Key design decisions
In an AI product, trust comes from visible control: showing who is replying, what the AI knows, and how to take over.
Start from a link, not a form
Onboarding begins with a website link or a file upload. The system fetches the logo, brand and colours automatically, then shows a live widget preview while the owner adjusts settings. They see their bot before they’ve configured anything.
One training pattern, with a staleness warning
URL, File, Text and Q&A share one layout: status, last updated and bulk actions. Learn one, know all four. When a product changes, a banner says “Product changes detected. Please retrain,” so the AI never silently gives outdated answers.
Always show who’s replying
The inbox labels every chat as AI-handled, agent-handled, unassigned or archived. “Take Over AI Response” explains that the AI will stop before an agent steps in, and unassigned chats are view-only until someone claims them, so two people never answer the same customer.
Controls in plain language
Response style, tone, model, language, timezone and custom instructions in one configuration screen, plus AI access per channel. Agents get canned replies by category and “Write with AI” tools to expand, rephrase, formalise or summarise.
Surface the rules, connect the commerce
The Instagram connection explains the 24-hour reply window right where users connect, instead of hiding it in docs. Products with variants, services with pricing models and orders with customer details live next to the conversations, with templated order emails.
System & edge cases
- Empty states for every list: chats, products, orders, members
- Expired password-reset and invitation links
- View-only unassigned chats
- Archived chats with unarchive and export
- Confirmations for every destructive action
- Processing states: fetching brand, crawling links, retraining
More screens
Conversations, AI-only vs handovers, unassigned chats and reply time by channel.
Colours, header, bubble alignment, conversation starters, pre-chat form and embed code.
Reusable replies by category, with group-level access.
Owner, Admin and Agent roles, groups and invitation states.
Credits, storage and per-channel AI access.
Order confirmation and verification code emails.
A full AI support and sales platform, from sign-up to daily operations, on one design system.
What I learned
- With AI products, trust comes from visible control: who is replying, what the AI knows, and how to take over.
- Explain platform limits where people hit them. A one-line note at the Instagram connection prevents a support ticket later.
- I designed the handover first, then built the chatbot around it. Most AI chat tools start with the bot and treat the human take-over as an escape hatch. I worked backwards from the moment an owner steps in: who is replying, what the AI knows, and a clear take-over with a confirmation. Training, tone and widget setup were designed to support that moment, so control is visible rather than hidden in settings.
- I showed the result before asking for effort. The common pattern is an empty dashboard and a long setup checklist. In Orbit, a website link or a file is enough: the brand is fetched and a live widget preview appears while the owner is still setting up. Seeing a working bot early is what keeps a small business owner going.