AI assistants already advise shoppers, and in the US they now complete the purchase too. What agentic commerce is, how far it has got in Germany, and what a retailer should do now without chasing the hype.
Lesomat
AI
One question ran through the week: how much should an AI agent be allowed to do? Anthropic ships an agent that asks before anything delicate, 1Password lets agents sign in without seeing the password, and the banks and Microsoft make distrust the default.
Three releases from the AI week, sorted for the working business: Anthropic builds its own source control, Google and Meta field new frontier models, and the Claude Code team shows how to review AI-written code without reading every line.
Three releases from the AI week, sorted for the working business rather than the hype cycle: an inexpensive Claude model with small print, a free embedding model from Google, and AI agents that live on your own machine.
An AI that talks nonsense in a chat window is embarrassing. An AI that acts on nonsense inside a workflow is expensive. How to catch the errors before they reach a customer: validation rules, a human in the loop, and tests that measure rather than hope.
A friend who runs a small joinery outside Rosenheim told me recently, with the particular pride of a convert, that he no longer googles anything. He asks. The other week he wanted someone to wire his warehouse software into his accounts, so he put the request to ChatGPT and let it name the firms worth […]
Agentic AI does not merely answer, it acts: given a goal, it works through the steps on its own. Where that genuinely pays off for a mid-sized firm, and where it does not.