Deep Seek and Customer Service in 2025: How to build an AI Chatbot customers actually use

Crazy weeks in the world of AI lately. Deep Seek changed the world of AI power. There was a lot of discussion about if it is truly is cheaper and how much GPU's they used (probably more like 50'000).

Fact is, it is truly innovative what they built. Imagine it a bit like your brain. If you are asked how to build an AI Chatbot, currently AI models activated all their parameters. This takes a lot of compute power and energy. Deep Seek managed to build an architecture similar to our brains, where different parts are used for different tasks. The question for example only requires a specific area to answer, so instead of activating the whole brain or in LLM terms all 600 billion parameters it only activates around 37 billion. This method is called Mixture of Experts and mimics the brain's architecture. Amazing Podcast if you want to learn more.

Now that you got updated on the latest AI talks let's deep dive into a topic that I spent the last 18 months to perfect. How to build Chatbots that customers actually use, don't suck or just waste your time. Together with our customers Aldi Mobile, Digital Republic and GGA we put together the 5 steps that deliver true customer value!

The Messy Library Problem

If you want to read it in German you are in luck!

Imagine walking into a library where books are scattered across the floor—no labels, no system, just chaos. Would you bother searching for what you need? Probably not.

This is precisely how a poorly trained AI assistant feels to customers—disorganized and frustrating.

Despite the appeal of quick AI deployments, rushing to build an AI assistant often leads to failure.

Why?

Let’s explore the key reasons and learn how to build a smart AI assistant the right way.

1. Data is the foundation: Curating a knowledge library

Retrieval Augmentation Generation (RAG) —the technology behind generative AI that accesses your knowledge library—works best with well-organized, accurate data. Without it, an AI assistant is like that messy library—filled with irrelevant or outdated information.

Take the time to

• Curate the library think Quality over quantity – With GenAI, the focus shifts from sheer volume to precision. A well-curated library ensures the model finds accurate, relevant answers, making quality?far more important than quantity.
• Remove outdated data to avoid misinformation.
• Organize the knowledge base to ensure GenAI achieves the highest performance.

Air Canada was one of the early adopters of GenAI Chatbots. Unfortunately they did not do a clean job here, and left outdated or wrong information in the databank. Which let to very favorable answers when it came to flight cancelation. - A strong foundation of reliable data could have prevented this.

Read more here (https://www.theguardian.com/world/2024/feb/16/air-canada-chatbot-lawsuit).

2. The human touch: Guiding AI’s growth

Even the most advanced AI can’t replace the nuance of human understanding. Think of AI as a talented intern—it’s eager to help but needs mentorship.

Here’s how human involvement enhances AI assistants:

• Real-world testing
• Define boundaries
• Build it for humans

We experienced this last issue when building an AI assistant for the Longines CSIO, an international horse tournament. Spectators frequently asked, ‘Who won the tournament?’ The assistant searched the library and responded with last year’s winner. Technically, it wasn’t wrong—the question was open-ended, and the trophy even had a different name this year. But clearly, it wasn’t the answer the spectators were looking for.

Read more about the case

3. The customer experience: Quality over speed

Your AI assistant is often your brand's first impression, and a bad experience can drive customers away for good.

Picture this: I lost my car keys on a deserted beach in Brazil and desperately needed help. Localiza car rental's chatbot trapped me in endless loops, offering no solution or human contact option. The result? I’ll never use their services again.

4. The hybrid model: Humans and AI working together

The last two steps on our Blog and subscribe to our newsletter :)