Almost everyone has met a bad website chatbot. Some are rigid decision trees that can’t understand a simple question. Others are general-purpose AI that answers confidently and wrongly, quoting a price that doesn’t exist or a policy that was changed last year.
Neither is acceptable when a customer asks “are there seats left on Saturday?”, “what’s your cancellation policy?” or “can I check in late?”. The fix isn’t a smarter model. It’s a grounded assistant.
What “grounded” means
A grounded assistant answers only from sources you control:
- approved knowledge: FAQs, policies, product and venue information that your team has reviewed;
- live lookups in your own systems, such as availability, bookings or order status, through tools you allow;
- context: which page the visitor is on, the current date and time, and (when they sign in) who they are.
When none of those sources has the answer, a grounded assistant says so honestly and offers a way to reach a person. It never invents prices, availability, opening hours or policies.
Five things a production assistant needs
1. A review process for what it knows
Everything the assistant can say should pass through review before visitors see it, with automated checks for conflicting statements, gaps and duplicates. Every approved version should be kept, so you can compare changes or roll back.
2. Clear rules about actions
Looking things up is low risk. Changing things, like adding seats to a cart, booking an appointment or updating a record, needs rules. A good assistant only uses the tools you’ve switched on, and asks the customer to confirm before anything changes. For payments, it hands over to your own checkout rather than taking card details in chat.
3. Honest handover to people
Some questions need a human. The assistant should recognize them, say so plainly, and route the conversation to the right team, by email, Teams, Slack or your help desk, with the context attached so the customer doesn’t have to repeat themselves.
4. Access control
Not every assistant should be open to everyone. Hotel concierges should only work for current guests; member services only for members. Look for open, recognition and invitation-only modes, and access that expires with a stay, term or season.
5. Privacy by default
Conversations should be kept only as long as you decide, data should be separated per organization, and the AI services underneath should be configured so your conversations aren’t used to train their models.
Beyond the website chat bubble
The same grounded assistant can work across channels: a website widget, a full-page chat, a QR code in a hotel room or on a theater seat, a home-screen app, and voice, with hands-free calls that can switch language mid-conversation. Internal versions live in Microsoft Teams or Slack and answer staff questions from your own documentation.
Two real examples
- The AI Chatbot for Tessitura gives performing arts patrons a box office that never closes. It answers from live Tessitura data, adds seats to the patron’s cart and sends them to the venue’s own TNEW checkout.
- The Sectorly.ai platform runs grounded assistants for hotels, clinics, appointments and more on one governed core.
Platform or custom build?
For most customer-facing assistants, a configurable platform is the fastest route: you shape the assistant’s knowledge and behavior without writing code. When the workflow is unusual, the systems are internal, or hosting rules require everything to run in your own cloud, a custom build makes more sense. Bryant Technologies does both. Read more on our AI Assistants & Chatbots page, or tell us who your assistant should help.