AI in customer service has gone from promise to daily practice in the past two years. But the picture rarely matches the brochure: some things work better than expected, others stubbornly keep disappointing. An honest overview.

What already works well today

Preparing replies

This is by far the strongest application. AI reads the question, pulls in the relevant context and readies a draft. The employee reads, adjusts where needed and sends. You remove the emptiness of the blank screen, and that saves several minutes on many questions.

What matters is what the AI takes into account. A model with only general knowledge writes general replies. A model that knows your product information, previous conversations and approved example replies writes replies that need hardly any adjustment.

Classifying and routing

Determining what a message is about and who should pick it up is something AI does reliably. It seems small, but it saves the manual distribution of the inbox every morning and ensures urgent messages don’t end up at the bottom.

Summarising

Reducing long conversations to their core is something language models are good at. An employee taking over a file of fourteen messages reads three sentences instead of everything.

Analysing why people get in touch

Reading through thousands of conversations to see which problems recur — nobody does that. AI does. That often delivers the most valuable insights, because you find out which questions you can prevent.

What still disappoints

Handling everything fully autonomously

With simple, well-defined questions it works. With everything beyond that it derails, usually on the exception: the customer who means something slightly different, the order that isn’t standard, the commitment once made by phone.

The sensible setup is therefore tiered: let AI send automatically where confidence is high, and put the rest before a human. Who decides what should be up to you — not the vendor.

Answers without sources

A model that knows nothing about your organisation makes things up with conviction. That’s not a shortcoming of AI, but of the setup. Make sure the system draws on your own documents, orders and previous conversations.

Tone without examples

An instruction like “write friendly and professional” delivers generic text. Approved example replies from your own practice work much better than a description of the desired tone.

The question to answer first

Not which AI tool to buy, but where your time goes. In one team the time sits in looking up order information, in another in phrasing replies, in a third in distributing work. Those are three different solutions.

So measure first. If you don’t know which contact reasons cause the most volume, you also can’t determine where automation pays off. Insight into your contact reasons is therefore a more logical first step than a chatbot.

Where people remain indispensable

With complaints, with borderline cases and everywhere an exception is needed. That’s not a temporary phase. The best result arises where AI does the groundwork and people keep the judgement — see also how we’ve set up human control.

Want to know what’s realistic in your situation? Get in touch — we’d rather look with you than promise something.