
Few topics are discussed as loudly right now as AI agents. For small and medium-sized businesses, the real question is more sober: which task in my company gets done faster, cheaper or more reliably?
1. Triage incoming requests
Reading emails, contact forms and tickets, classifying them and routing them to the right person is something language models do very well today. What matters is that decisions stay traceable. I always have the model return a short reason and a confidence score; unclear cases go to a human.
2. Extract data from documents
Delivery notes, invoices, contracts: instead of retyping data, a model extracts the relevant fields into a fixed schema. Results are checked against rules (does the total add up? is the date plausible?) before they enter your system.
3. Drafts instead of final answers
An agent that replies to customers on its own is risky. An agent that prepares a good draft which an employee approves with one click saves time immediately, without giving up control.
Why projects fail
- No clear goal: “We want to do something with AI” is not a project. “Quotes in 10 instead of 40 minutes” is.
- No real test data: A prototype with three examples shines in the demo and fails in daily use.
- Costs out of sight: Without logging, nobody knows after three months what the automation costs.
My advice: start small, with one process that is annoying today and measurable. Once that works, move on to the next.