Here's a truth that should make you choke on your breakfast: an AI agent that scans a cereal box can't reliably tell you which one has less sugar. A study accepted at IEEE EMBC 2026 — the flagship biomedical engineering conference — put AI vision agents to the test. The results? On global products, they hit 88.9% accuracy. On local Swedish products? A dismal 59.5% — basically a coin flip.
That's not a tech hiccup. That's a health-safety warning. The fact that this question showed up at a biomedical engineering conference, surrounded by cancer research, tells you everything: the academic world now treats AI nutrition advice as a patient-safety issue. Not a marketing gimmick. A health risk.
Picture Maria in Stockholm, a diabetic, scanning two muesli boxes with her phone. The AI says the second one has less sugar. It's wrong — as likely as not. She trusts it. Her blood sugar spikes. Who's liable? The AI vendor? The grocery store? Or Maria, for believing a machine that was trained mostly on American and global foods?
Here's the twist: this isn't just a supermarket problem. It's a symptom of a bigger trust crisis. Data exec Carly Taylor put it bluntly on LinkedIn: most teams can't vouch for their own data, yet they're shipping AI agents anyway, hoping nothing goes wrong. They're building houses on swamps and selling them as mansions.
And the stakes are climbing. Agentic commerce is turning product images into machine-readable evidence. When your AI shopping assistant decides what to buy, a misread sugar label isn't just a bad snack choice — it's a wrong decision multiplied across thousands of purchases.
The winners here are the teams that do the boring work: data governance, testing, transparency. Trust isn't a buzzword; it's a market differentiator. The losers are the rest of us — and the vendors who ship untested agents into our kitchens.
So, can we trust AI agents in the supermarket? Not yet. Not with local products. Not without knowing what they've been trained on. The question for the industry isn't whether AI can be useful — it's whether we'll have the guts to admit when it can't. As for the future, trends suggest we'll see a push for certifications or labels on AI nutrition tools — but that's a might, not a certainty.