"Can't the AI just do that?" I hear this question regularly now. My answer: a large part of the groundwork – yes. The decision at the end – no.
Where AI really helps in support
I have worked in 2nd/3rd-level support for years and build my own AI tools on the side. In daily work, three tasks have turned out to be especially useful:
Summarising long logs. A ticket arrives with several thousand lines of log. A language model can say within seconds which errors appear first and when, which repeat and which components are involved. It doesn't replace reading the log, but it shows where to look first.
Finding similar past cases. Many issues have happened before, just described in different words. A semantic search across knowledge articles and resolved tickets – for example with RAG, a search that retrieves matching documents and gives them to the model as its basis – finds them even when the terms don't match.
Preparing a first draft of the reply. Structure, polite wording, the steps in a sensible order. I check the content, but I don't start from a blank page.
Each of these tasks saves only a few minutes. With many tickets per week, that adds up.
Where AI reaches its limits
Language models sound convincing even when they are wrong. A command with an option that doesn't exist, a configuration key from another version, a plausible but invented cause – all of this happens.
More importantly: AI doesn't take responsibility. Whether a workaround is enough or the cause lies deeper, whether a change to a production system is acceptable right now, whether it's better to call a customer than to write – that is still decided by a human who knows the system and the situation.
A model only sees what you give it. It doesn't know the context outside the ticket – last week's maintenance, a known quirk of a setup, a customer's mood after the third contact.
My rules for AI in support
- Sources, not claims. Answers from a knowledge base should state what they are based on. Then you can check within seconds whether they are right.
- Commands are reviewed before they run. No model suggestion goes onto a customer system unread.
- Data stays where it belongs. Logs and tickets often contain personal data. What an external service may see is a question of data protection and company policy, not convenience. Locally hosted models are one option here.
- A human closes the ticket. The AI prepares; a human is responsible for the decision and the reply to the customer.
A fast colleague for the groundwork
For me, AI in support is a very fast colleague who reads logs, gathers notes and puts a draft on the table. Tireless and often surprisingly good. But not the one who closes the ticket.
Root cause analysis shows this clearly: a model can spot anomalies and suggest hypotheses. Whether a hypothesis is right only becomes clear when someone tests it on the system.
How to get started
- Pick one recurring piece of groundwork, for example log summaries.
- Clarify which data may be processed, and where.
- Deliberately double-check the results for a while.
- Only then hand over further tasks.
That way trust in the tool grows where it is justified – along with a clear picture of where it isn't.