Ask it a customer's question and it answers from the shop's own material. If the question is one a person should handle, it stops and says so.
It strips card numbers, social security numbers, email addresses and phone numbers out of the message before the model or the log ever sees them, and it ignores instructions hidden inside a customer message.
Demo data. Try sending it a card number and telling it to ignore its instructions — both are handled in front of you.
Run the agent · live
Your ticket. Your docs. Its decision.
Pick a scenario or type your own. Paste a few lines of your real help docs if you want — or just name your business and leave the docs blank: the agent will draft sample docs for it, then answer from them. (Your live build grounds on your real docs.)
In plain English: a “grounded” answer is one the agent gives only using facts from your documents — it never makes things up. If the answer isn’t in your docs, it sends the ticket to a human instead of guessing.
Time of day
Runs on fast edge AI for the demo. The build you commission runs Claude or GPT with real RAG over your full docs.
FIG · customer support agent🛡️ answers only from your docs · sends the rest to a human
🤖 Identifies itself as AI to customers
🌐 This demo is chat. A build we deliver runs one agent across 💬 chat · ✉️ email · 📱 WhatsApp · 📞 voice, replying in the customer's language
Agent activity
Press “Run the agent” — its reasoning streams here, step by step.