What works today
Draft replies
It is a use with a good balance of benefit and risk. The AI reads the thread and the customer’s history and proposes a reply. The agent edits it and sends it. Adjusting text that is almost right takes less time than writing it from scratch.
It works because a person is still in the loop and because mistakes are cheap: if the draft is bad, you throw it away.
Classifying and prioritizing what comes in
Reading an email and deciding its type, priority and queue is a repetitive task AI does well. And mistakes are easy to fix: a misclassified ticket gets reclassified.
Transcribing and summarizing calls
Turning a fourteen-minute call into four useful lines saves the most tedious part of the job. And calls finally become searchable.
Filtering out what is not a request
Spam, newsletters and auto-replies, gone before they create work. Not glamorous, and very worthwhile.
Questions about your own data
“How many hours have we spent on this customer this year?” without waiting for someone to build a report. It works if the system does not let the AI write free-form database queries. The safe approach is for it to propose a plan over approved tables and to show what it looked up.
What does not work well
The chatbot that solves complex technical cases
It handles frequent, documented questions well: opening hours, the status of a ticket or how to do something covered in the knowledge base. Beyond that it gets stuck. The problem is not that it gets stuck, but that the customer has already spent five minutes before reaching a person, and arrives annoyed.
What does work is a chatbot for the frequent questions with a clear, quick way out to a person, not one hidden behind three menus.
AI closing tickets on its own
It is technically possible and commercially dangerous. A ticket closed by mistake leads to a reopen and a complaint. And the complaint is against the company, not the model.
Technical diagnosis without your documentation
A general-purpose model does not know how your specific equipment fails. With your documentation loaded, it helps. Without it, it makes up answers with great confidence, which is the worst thing that can happen.
Predicting failures without data
It gets sold a lot, but it needs data: sensor readings or a long, well-kept history. If your history covers two years and has gaps, no model can predict anything, however good it is.
Four questions before you switch anything on
- How much will it cost me each month? There must be a usage screen and a budget with a cap, not an estimate. Without that, do not turn it on.
- What happens if it gets it wrong? If the answer is an angry customer, a person must review before anything is sent.
- Can I turn it off? Feature by feature, without breaking the rest.
- Where does my data go? Which provider processes each request, and whether your data leaves the region it has to stay in, such as the European Union.
Where to start
With draft replies, for one month, on a small budget, measuring time per ticket. If it saves time, expand it. If not, turn it off; you have lost almost nothing.
What does not work is switching on six things at once. You will not know which one helps and which one gets in the way.
Today, AI in support is a good assistant for agents and a poor replacement for them. Anyone who pitches it the other way around is selling you a problem.
Last updated: