96% on Mercedes' line: the robot story is the other 4%

Robot news usually reads like science fiction. This doesn't. A robot stands at a line and moves metal brackets from big bins into small trays, in the order the next car needs them. That's it. Which is exactly why it's worth a look.

What happened

• Who: Google DeepMind with Gemini Robotics 2.5, a model that links vision, language and action. The hardware is Apptronik's Apollo 2 humanoid.
• Where: active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany.
• What: kitting, sorting brackets from bulk bins into trays per vehicle. Success rate per DeepMind: 96%, explicitly a proof of concept.
• How: two models split the work. Gemini Robotics ER tracks the workflow, Gemini Robotics 2.5 controls arms and hands.
• Context: across 200+ tasks, DeepMind reports 30% to 100% higher success rates than the previous version and 20% to 30% faster completion. For now, the model goes to robotics partners only.

The best first AI job is boring

What I find interesting: nobody let the robot build a car. The job is tightly scoped, follows a clear sequence and ends in a result anyone can check instantly. Is the right bracket in the tray or not?

That's exactly what a first agent project in marketing should look like. My rule of thumb: one clearly scoped job that ends at a concrete decision and delivers a result someone can judge. Say: prepare a reactivation campaign, hand over the package for review. Not: "do our marketing."

The catch: who catches the 4%?

96% sounds strong. Translate it to your world. An agent that personalizes newsletter variants correctly 96% of the time sends roughly 2,000 wrong emails out of 50,000. In a plant, a wrong bracket will likely get caught on the line. In marketing, the error lands straight with the customer.

The announcement also leaves out key details: how long the pilot ran and how success was measured. It's a vendor number, not an audit. And Gemini Robotics 2.5 isn't something a mid-sized company can buy tomorrow.

So the real work isn't in the model, it's in the process around it: where's the check, who approves, where do errors go? Human in the loop doesn't mean a person standing next to the machine. It means someone knows where the 4% end up.

What I'd do now

• Find your office kitting job. Tasks with clear sorting logic: routing leads by criteria, filing assets by channel, triaging inbound requests.
• Set the acceptable error rate up front. What rate is fine, and where is the check before anything goes out?
• Build the error path. One place where wrong results land, and one person who clears it every day.

Bottom line: 96% is the headline. Whoever catches the other 4% decides whether AI works in real life.