Almost every contact centre knows chatbots by now. But in 2026 the focus is shifting: away from rigid, rule-based bots towards agentic AI — agents that understand a request on their own, plan the steps and see a task through from start to finish. What lies behind the term, where do companies really stand, and how does a sensible start look?
What separates agentic AI from a classic chatbot
A classic chatbot follows fixed rules or decision trees: it recognises keywords and plays back predefined answers. An AI agent goes considerably further. It captures intent, context and mood in real time, breaks a request into steps, reaches into connected systems and acts — rather than merely replying.
The difference shows in the awkward cases. A rule-based bot fails as soon as a question is phrased unexpectedly. An agent works out what is meant and carries on.
Where companies actually stand
A current trend study of the German-speaking contact centre market paints a sober picture — neither crisis nor revolution. Around 49 per cent of the companies surveyed use chatbots, voicebots with open speech recognition reach roughly 40 per cent for the first time, and generative AI is in use at about a third. Agentic AI is running productively at some 12 per cent — clearly an early-stage figure with plenty of room ahead.
Worth noting: many of the bots in use are still rule-based rather than genuinely AI-driven. Alongside them, supporting technologies such as AI-based routing, speech analytics and next-best-action tools are gaining ground because they underpin decisions with data.
Where autonomous agents have the greatest leverage
AI agents work hardest where large numbers of similar, standard requests arrive. For such tier-one enquiries, well-implemented agentic solutions reach containment rates above 80 per cent — the request is settled without human involvement. Typical cases: delivery status, appointment changes, address updates, straightforward billing questions.
Where it gets complicated, or where an exception is needed, the agent hands over. That handover is the decisive point: it has to carry the full context, or the customer starts from the beginning and the gain turns into annoyance.
An example
A customer writes in the evening: “My delivery has not arrived and I need it urgently.” A rule-based bot would link to a FAQ page. An AI agent recognises the urgency, retrieves the shipment status through the connected systems, identifies the delay, offers a replacement delivery or a credit note without being asked, and documents the case. Only if the customer remains dissatisfied, or an exception is required, does it hand over to a colleague — with the whole conversation attached. That saves time on both sides and avoids frustration.
The real bottlenecks: data, skills, infrastructure
Technology alone does not make a good AI agent. According to the study, data quality, staff skills and suitable IT infrastructure are simultaneously the greatest levers and the greatest bottlenecks. An agent is only as good as the knowledge base and the systems it can reach.
Anyone introducing agentic AI should therefore start with the knowledge base and the interfaces, not with the model. A perfectly tuned agent in front of an outdated FAQ produces confident wrong answers — which is worse than no answer at all.
How to start
- Pick one clearly bounded use case with high volume, not the whole service at once
- Clean up the knowledge base for that case before anything else
- Define what the agent may decide on its own and where it must hand over
- Measure containment and follow-up contacts — a case closed by the bot that returns the next day was not settled
The transparency obligations of the EU AI Act apply from August 2026: an AI agent has to identify itself as such. That is worth building into the concept from the outset rather than retrofitting.
