AI in the kitchen: what actually works, what's theater
Five years of 'AI for restaurants' pitches, and a shortlist of the things that are genuinely useful on the line.
There's been a wave of "AI for the kitchen" products over the last three years. Most of them are theater. A few are genuinely useful. The line between the two has almost nothing to do with the underlying models and almost everything to do with whether the builder has ever stood on a line during a Saturday rush.
What's actually working
Prep forecasting. Not "here's how many covers you'll do Saturday" — that's the easy part. What matters is "here's how many ounces of the short rib reduction you'll go through Saturday, given the weather, the event at the venue two blocks over, and the fact that your scallop special is going to steal share from the steak". That level of decomposition actually lets your chef prep to a number instead of to a feeling.
Waste tracking by station. Cameras pointed at the prep table and the garbage can, tagged to station and shift. Two weeks in, you know your Sunday brunch sauté wastes 18% of its mise every shift, and you can do something about it. Not sexy. Enormous.
Recipe cost drift alerts. Your béchamel costs 42 cents per ounce on Monday and 51 cents on Friday. Why? Cream went up at the supplier. The system catches it same-day and flags the dishes that use it. Menu pricing stays current; you stop finding out at month-end.
What's theater
Robot pizza / robot fry / robot burger. Every few months, a venture-funded concept shows up with a mechanical arm that can make one dish in one format. The unit economics never pencil. The dishes look like the photos on a printer manual. The companies that push hardest on this are uniformly the ones farthest from actually running a restaurant.
"AI-generated recipes". The models are fine at generating recipes. Your chef is better. The problem isn't recipe ideation — the problem is execution, consistency, and margin. An AI that produces a recipe for "spiced lamb with mint chimichurri" has not solved any problem your kitchen actually has.
Voice AI that takes drive-thru orders. Works in controlled demos. Falls apart in the real world — regional accents, ambient noise, menu substitutions, the guest who pauses mid-order to discipline their kid. The operators we know who've deployed these have either pulled them out or kept a human ready to take over 30% of orders.
The common thread
The useful AI tools share one property: they do a narrow, mechanical task that a human was doing poorly because it was boring. Counting, measuring, watching, reconciling. The theater tools try to do creative work that a human was doing well because it was satisfying.
That's how we decide what belongs in Restaurant OS. If AI can do it and your chef doesn't want to, it's a candidate. If a human would rather do it than not, we stay out of the way.