AI support agent

The agent answers questions from your own sources: help articles, product documentation, terms, past tickets. For each question the system retrieves the matching passages and gives them to the language model as its basis. If nothing fits, the agent says so and hands over — it does not fill the gap with guesses.

Account data comes from your database: order status, plan, open invoice. That way every user gets an answer about their own case rather than a general hint.

What separates it from a chatbot is actions. The agent can move an appointment, change an address, resend an invoice or start a refund. Each action is a fixed function with its own rules: who may trigger it, up to what amount, with which confirmation.

Critical steps — cancellations, refunds above a threshold — are only prepared by the agent. A person approves them. Responsibility stays where it belongs.

Before launch we build a test set with expected answers from real tickets. Every change to the agent runs against that set before it goes live. This shows when an improvement in one place makes something else worse.

In operation three numbers matter: how many requests the agent resolves itself, how many it hands over, and how often the same user returns with the same topic. The last one is the most honest.

The state of your knowledge base is decisive. If the help texts are current and consistent, setup is quick. If they are not, cleaning them up is the first project step — and a gain even without AI.

• Channels: in-app chat, website widget, WhatsApp or email
• The number and risk of the actions the agent may perform
• Connection to ticketing and database — for us often Supabase and n8n
• Languages of your users