Customer Support Automation: Deploying a 24/7 AI Chatbot for the Dutch Market

The Challenge: High Ticket Volumes and Language Barriers

High Volume of Repetitive Questions
The client’s support team was overwhelmed by a constant stream of repetitive questions. How do I save points? How do I redeem miles? What’s my balance? Which partners participate? These questions represented the vast majority of inbound support volume, but every single one required a human agent to respond. The repetition drove up wait times, increased the support team’s workload, and pulled agents away from the complex, sensitive cases that actually needed human judgment.

No 24/7 Support and Inconsistent Answers
There was no way for customers to get help outside business hours. And during business hours, the quality of answers varied depending on which agent picked up. Some were thorough, some were brief, and there was no system ensuring every response matched the brand’s tone and programme rules. Customers who couldn’t find answers quickly on the website were forced to wait in line for support, a frustrating experience that eroded trust in the programme.

No Structured Escalation Process
When a customer issue was sensitive, emotional, or required account-specific actions, there was no consistent process for routing it to the right person. Agents made judgment calls on when to escalate, leading to gaps where sensitive cases weren’t flagged and frustrated customers fell through the cracks. The client needed a system that could handle the routine while reliably identifying and escalating what it couldn’t.

The Solution: 24/7 Multilingual AI Chatbot

RAG-Powered Conversational AI in Dutch
AY Automate designed and built a web-based AI chatbot that answers customer questions about the loyalty programme 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 document stored in Google Drive, the programme’s official rules, partner details, saving and redemption instructions, and policies. When a customer asks a question, the AI retrieves the most relevant sections, generates an accurate answer in the brand’s tone, and delivers it instantly through a clean chat interface.

Intelligent Escalation and Confidence Scoring
Every response from the AI is structured with more than just an answer. The system returns a confidence score, an escalation flag, and a question classification for every interaction. When a topic is sensitive, account-specific, or out of scope, the bot automatically sets an escalation flag and shows the customer service contact information, handing off gracefully instead of attempting answers it shouldn’t. This structured output means the system knows its own limits and routes cases accordingly.

Zero-CMS Knowledge Management
One of the most elegant aspects of Airmails is its knowledge management architecture. The entire knowledge base is a single PDF stored in Google Drive. When the content needs updating, the client simply edits the PDF. An n8n workflow automatically detects changes, re-chunks the document, generates new embeddings with OpenAI, and re-indexes everything into Pinecone , no custom CMS, no migration scripts, no technical overhead. The client’s team can maintain the bot’s knowledge base without touching a line of code.

Complete Analytics and Feedback Layer
Every conversation is logged in Supabase with the full context: question, answer, confidence score, escalation flag, session ID, and optional user feedback via thumbs up/down ratings. This gives the client complete visibility into what customers are asking, how well the bot is performing, which topics trigger escalation, and where the knowledge base has gaps. The data is structured for continuous improvement , making it easy to refine the bot and expand the knowledge base over time.

Team Assembly and Delivery
AY Automate embedded a specialist AI engineer with deep experience in RAG systems, conversational AI, and n8n workflow automation directly into the project. The engineer handled end-to-end delivery , from architecture design through to production-ready deployment , ensuring the client received a complete, maintainable system without needing to build internal AI expertise.

The Technology Stack Used

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

The architecture runs entirely on n8n for backend orchestration , no traditional backend server. One workflow handles PDF ingestion from Google Drive, chunking, embedding with OpenAI’s text-embedding-3-large model, and indexing into Pinecone. A second workflow receives chat messages via webhook, queries Pinecone for relevant context, runs an AI agent (GPT-4.1-mini) with the retrieved knowledge and session memory, and returns structured output to the Next.js frontend. A third workflow captures user feedback and updates Supabase.

Results & Impact

24/7 Automated Customer Support
The support chatbot delivers what the client’s support team couldn’t: instant, accurate answers to loyalty programme questions around the clock. Customers no longer need to search the website, wait in line, or call during business hours to get answers about saving points, redeeming rewards, or finding partners. The bot handles the volume autonomously, freeing human agents to focus on the sensitive, complex cases that actually require human judgment.

Brand-Consistent Dutch Language Experience
Every response from the support chatbot is delivered in natural Dutch, consistent with the brand’s tone and programme rules. Unlike generic chatbot solutions, the system was built specifically for the Dutch market, ensuring cultural and linguistic accuracy in every interaction. This consistency builds customer trust , every user gets the same quality answer regardless of when they ask or how they phrase the question.

Built for Continuous Improvement
The platform is instrumented for measurement from day one. Supabase stores every question, answer, confidence score, escalation decision, and user satisfaction rating. Once live, the client can track resolution rates without human intervention, time to first answer, satisfaction trends, and knowledge base gaps , creating a continuous feedback loop that makes the bot smarter over time. The zero-CMS knowledge architecture means improvements to the knowledge base take minutes, not development cycles.

Business Impact and Key Takeaways

* RAG-powered chatbots deliver accurate, brand-consistent answers from existing documentation , no need to build custom knowledge management systems. A single PDF in Google Drive can power an entire customer support AI.
* Structured AI output (confidence scoring + escalation flags) creates reliable automation. The system knows what it can and can’t answer, escalating gracefully instead of guessing , critical for customer-facing applications.
* Dutch-language AI support requires purpose-built solutions. Off-the-shelf English-first tools can’t match the cultural and linguistic accuracy needed for European loyalty programmes. Custom development pays off.
* n8n as a backend eliminates traditional server infrastructure. Running the entire RAG pipeline, chat logic, and feedback collection in n8n workflows reduces complexity, speeds delivery, and makes the system maintainable without dedicated backend engineers.
* Logging every interaction from day one enables data-driven improvement. With questions, confidence scores, escalation decisions, and satisfaction ratings stored in Supabase, the system is ready to optimize from its first conversation.