Customer Support AI Chatbot Deployment

The challenge

The support team was drowning in repetitive questions. How do I save points? How do I redeem miles? What is my balance? Which partners participate? These were the majority of inbound volume, but each one still needed a human agent to answer. The repetition pushed wait times up, increased agent workload, and pulled people away from the cases that actually needed human judgment.

There was no help outside business hours. During business hours, answer quality varied depending on who picked up the ticket. Some agents were thorough, some were brief, none of it ran against a checked source of truth on programme rules. Customers who couldn't find what they needed on the website were stuck waiting in line, which slowly cost the programme trust.

There was no consistent escalation process. When a case was sensitive, emotional, or required an account action, agents made the call themselves on when to flag it. Some sensitive cases got missed. Frustrated customers fell through the cracks. The team needed a system that handled the routine and reliably caught the cases it couldn't.

What we built

AY Automate built a web-based chatbot that answers loyalty programme questions in natural Dutch directly from the brand's official knowledge base. The system uses Retrieval-Augmented Generation (RAG) to pull answers from a single PDF stored in Google Drive, the programme rules, partner details, redemption instructions, and policies. When a customer asks a question, the AI finds the relevant sections, generates an answer in the brand's tone, and returns it through a clean chat UI.

Escalation and confidence

Every response from the AI carries more than the answer. It returns a confidence score, an escalation flag, and a question classification. When a topic is sensitive, account-specific, or out of scope, the bot sets the escalation flag and shows the customer service contact, handing off cleanly instead of guessing. The system knows what it can and cannot answer.

Knowledge base in a PDF

The full knowledge base is one PDF in Google Drive. When content changes, the client edits the PDF. An n8n workflow watches the file, re-chunks it, generates new embeddings with OpenAI, and re-indexes into Pinecone. No CMS, no migration scripts, no developer involved. The team owns the knowledge base without touching code.

Analytics

Every conversation is logged in Supabase with full context: question, answer, confidence score, escalation flag, session ID, and optional thumbs-up/thumbs-down feedback. The client sees what customers ask, how the bot performs, which topics trigger escalation, and where the knowledge base has gaps. The data structure is set up for continuous improvement.

Team

AY Automate placed a specialist AI engineer with RAG and n8n experience inside the project. The engineer handled architecture, build, and production deployment. The client got a complete, maintainable system without building internal AI capacity.

The stack

Next.js, n8n, OpenAI (GPT-4.1-mini, text-embedding-3-large), Pinecone, Supabase (PostgreSQL), Google Drive.

The whole backend runs in n8n. There is no traditional backend server. One workflow handles PDF ingestion from Google Drive, chunking, embedding with OpenAI's text-embedding-3-large, and indexing into Pinecone. A second workflow takes chat messages over webhook, queries Pinecone for context, runs the GPT-4.1-mini agent with that context and session memory, and returns structured output to the Next.js frontend. A third workflow captures user feedback into Supabase.

Results

The bot delivers what the support team could not: instant, accurate answers about the loyalty programme around the clock. Customers no longer wait, call during business hours, or hunt through the website to find out how to save points, redeem rewards, or find a partner. The bot handles the volume on its own. Human agents are free to work on the complex cases that need judgment.

Every response is in natural Dutch, written in the brand's tone. The system was built for the Dutch market, not adapted from an English default. Customers get the same quality answer no matter when they ask or how they word it.

The platform is set up for measurement from day one. Supabase stores every question, answer, confidence score, escalation decision, and satisfaction rating. The client tracks resolution rates without human help, time to first answer, satisfaction trends, and knowledge gaps. The PDF-based knowledge architecture means updates take minutes, not development cycles.

What we learned

• A RAG chatbot pulls accurate, brand-consistent answers from existing documentation. A single PDF in Google Drive can run the whole customer support AI without a custom CMS.

• Structured AI output, confidence score and escalation flag, makes the automation reliable. The system knows what it cannot answer and hands off cleanly. That matters when customers are watching.

• Dutch-language support needs a purpose-built solution. English-first tools cannot match the cultural and linguistic accuracy a European loyalty programme needs.

• n8n as the backend removes a layer of infrastructure. The full RAG pipeline, chat logic, and feedback collection run in workflows, which keeps complexity down and shortens delivery.

• Logging every interaction from day one is what makes the bot improve. With questions, scores, escalations, and ratings stored in Supabase, the system can be tuned from the first conversation.