Oct 2025 – Present
AI support agent
A Claude-based agent that answers customers on Messenger and Instagram and knows when to hand over
- AI
- AI agent
- Duration
- 1 year
- Company
- Boxette Group
- Role
- Designed, built and maintains the service
Overview
An AI agent that answers Boxette’s customers on Messenger, Instagram and the other Respond.io channels from a curated knowledge base. It reads their photos, voice messages and files, and passes the chat to a human operator when it should not answer itself.
Tools
- Claude
- Claude Agent SDK
- Next.js
- Gemini
- SQLite
- Docker
The same questions, slow replies
Most customer messages repeat the same questions about parcel status, customs declarations and tariffs, and replies were slow, especially outside working hours. The first bots lived inside n8n with prompts that were hard to change, and support staff had no safe way to update what the bot knew.
From an n8n bot to an agent service
It began in 2025 as an n8n Messenger agent with Respond.io, Manychat, a Postgres message log and an attachment analyzer. In April 2026 it moved to a dedicated Next.js 16 service that calls Claude with the rules and knowledge base as the system prompt and one tool to reply or escalate. The n8n side debounces incoming messages, analyses images, audio and files, and builds the conversation history. It then either sends the answer or adds an escalation note and switches the chat to a human.
Knowledge the team can edit
Staff change the knowledge base and rules in a password-protected editor by chatting with a Claude Agent SDK agent. Every change shows as a diff to approve and is versioned in SQLite with restore. When the bot got branch opening hours wrong at weekends, the fix was to compute each branch’s open or closed status on the server and put it in the prompt, so the model no longer does the time arithmetic.
Result
The August 2026 conversation analysis found that the bot wrote 44.5% of outgoing replies (31% in July 2025), actually helped the customer in about 67% of the cases it handled, and needed a human in only 10.9%. Median first response time fell from about 35–130 minutes to about 3 minutes from February 2026, a change the report credits to the bot together with process changes. The branch-hours fix passed 5 weekend scenarios and 2 regression checks in production.