AI chatbot for customer service: what it should handle, what it shouldn’t, and how to launch one

An AI chatbot for customer service is one of the most requested AI projects I get — and one of the easiest to get wrong. The technology is no longer the hard part. Modern language models can answer questions fluently in dozens of languages. What decides whether a chatbot saves your team hours or quietly drives customers away is scope, data and the handoff to a human.

This guide covers what a customer service chatbot should handle, what it should not, what it costs, the new EU transparency rules, and a practical plan to launch one.

Chatbot, assistant or agent: what are you actually building?

The word “chatbot” now covers three very different things:

Most businesses should start with a knowledge assistant and add actions one at a time. Jumping straight to a fully autonomous agent is where projects stall.

Why companies are investing now — and the cost reality

Gartner predicts that agentic AI will autonomously resolve 80% of common customer service issues without human intervention by 2029, leading to a 30% reduction in operational costs. At the same time, Gartner warns that the cost per resolution for generative AI in customer service may exceed $3 by 2030, as complex use cases consume more tokens and vendors push for profitability.

Both predictions point to the same lesson: an AI chatbot for customer service pays off on high-volume, well-defined questions. It does not pay off when it is used for everything.

What an AI chatbot should handle

• Repetitive questions with clear answers — prices, delivery times, opening hours, return policy, how-to steps.
• Status lookups — “Where is my order?”, “Is my appointment confirmed?” — with read access to your system.
• Simple, reversible actions — booking or rescheduling an appointment, updating a contact detail, sending a document again.
• Lead qualification — asking the right questions, collecting contact details and passing a clean summary to sales.
• After-hours and multilingual coverage — answering at 11pm, in English, Albanian, Italian or German.

What it should hand over to a person

• Complaints, refunds and anything involving money beyond a set limit.
• Angry or upset customers — detect it and escalate early.
• Legal, medical or safety questions.
• Any question the assistant cannot answer from your approved sources. “I don’t know, let me get a colleague” is a feature, not a failure.

The handoff must be smooth: the human should receive the full conversation and a short summary, so the customer never has to repeat themselves.

The EU AI Act transparency rule you cannot skip

If you serve customers in the EU, Article 50 of the EU AI Act applies from 2 August 2026: people must be informed when they are interacting with an AI system, unless it is obvious from context. In practice, that means a clear label such as “You are chatting with our AI assistant” and an easy way to reach a person. It costs nothing to do and it builds trust. For the full timeline, read our guide to EU AI Act compliance.

What makes customer service chatbots fail

1. Bad source data. If your FAQ, policies and price lists are outdated or contradictory, the chatbot will be too. Clean the knowledge base first.
2. No grounding. A chatbot that answers from the model’s general knowledge will invent policies. It must answer from your documents, using retrieval (RAG), and say when it does not know.
3. No escape route. Customers trapped in a loop with a bot remember it. Always offer a human option.
4. No measurement. Without tracking resolution rate, handoff rate and customer ratings, you cannot tell if it works.
5. Trying to automate everything on day one. Start narrow and expand based on real conversations.

How to launch an AI chatbot for customer service: a 5-step plan

1. Pull the last 3–6 months of questions from email, WhatsApp, chat and phone notes. Group them. The top 15–20 topics usually cover most of the volume.
2. Write or fix the source answers for those topics. This is the most valuable hour you will spend.
3. Build a grounded assistant on your website and, where your customers are, on WhatsApp or Messenger — with citations and a human handoff.
4. Test with real questions before launch, including hard and hostile ones. See how to evaluate a RAG system.
5. Launch, review conversations weekly, then add actions — bookings, order lookups — one at a time.

How much does it cost?

Costs split into a one-time build (knowledge base, integrations, interface, testing) and running costs (model usage, hosting, and someone reviewing conversations). A knowledge-only assistant is the smallest project; every system it connects to adds work. We break down the arithmetic in how much an AI agent costs to build and run.

Frequently asked questions

Can an AI chatbot answer in Albanian?

Yes. Current models handle Albanian well, and the same assistant can answer each customer in their own language — as long as your source content is accurate.

Will it replace my support team?

Not realistically. It removes repetitive work so your team can spend time on the conversations that need a human. The best results come from AI and people working together.

Can it work on WhatsApp?

Yes. Through the official WhatsApp Business Platform, the same assistant can answer on WhatsApp and on your website, with one knowledge base behind both.

How do I stop it from making things up?

Ground every answer in your approved documents, show sources, instruct it to say “I don’t know” when the answer is not there, and test it before launch.

Plan your customer service assistant with us

Shkulaku builds AI agents and RAG systems that answer from your real data and connect to your booking, order and CRM systems. If you are considering an AI chatbot for customer service, tell us how many questions your team handles each month and where they come from. We will suggest a realistic first scope and an honest estimate of what it would save.