Every restaurant operator I talk to is thinking about AI right now. Most of them are thinking about it in the wrong place. They are looking at the front of the house — the chatbot, the drive-thru voice bot, the thing customers can see. Meanwhile the money is sitting in the back, in the walk-in, on the prep table, in the case of chicken you bought and never sold. That is where food waste lives, and that is where your margin is going.
We have two prime costs in this business: labor and food. In this post I want to focus on food, because it is the one where the data already exists and almost nobody is using it. And the numbers are not small. For some brands a four percent swing in food cost is hundreds of thousands of dollars.
Take a restaurant doing a million dollars at a 30% food cost. That is $300,000 walking out the door in product every year. Trim six percentage points off that — 30% down to 24% — and you have kept $60,000. Even shaving 6% off the spend itself is $18,000 straight to the bottom line. On larger units, or across a portfolio, you are into six figures. That is not a technology project. That is profit.
We have all watched Gordon Ramsay or Robert Irvine walk into a struggling restaurant and get genuinely appalled at what is rotting on the shelf — product sitting for days, sometimes weeks, never used. The uncomfortable truth is that most kitchens have a smaller version of that happening right now, and nobody can see it because nobody is measuring it.
Start with demand forecasting
Demand forecasting means using your own history to figure out what you should have used, then comparing that against what you actually bought. Your POS already knows what sold. The trick is rolling menu items back into ingredients.
Think about chicken breast. Some of it went out as tenders. Some as a sandwich. Some in the chicken pasta. Individually none of that tells you much. Aggregated, it tells you exactly how much chicken you should have consumed. Now put that next to your purchasing:
- I bought 14 cases of chicken breast over the last two weeks.
- My sales say I should have used 13 cases worth.
- I ended the period with X on hand.
- So where did the difference go?
Historically, answering that required somebody with very intimate knowledge of the menu, the recipes, and the vendor invoices — and the patience to reconcile all three. That is exactly the kind of work these tools are good at. You can hand the system your purchases, your theoretical usage, and your ending inventory, and then simply ask it what happened.
Then ask why
This is the part I think gets underused. Once the variance is in front of you, start interrogating it:
- What was the weather doing that week?
- What was happening in the local area, or the wider region?
- Was there construction affecting access to my parking lot?
- What events were going on that would have moved traffic up or down?
Any one of those can explain a bad week. The point is not to excuse the number — it is to learn the pattern so the next forecast is better.
Stop ordering on autopilot
Here is the habit I see almost everywhere: this is what we ordered last week, so this is what we will order this week. It is understandable and it is expensive.
I have worked with restaurants in high-traffic urban locations, in heavily seasonal markets, and literally on the beach. In every one of those, the forecast is not a straight line. Weather changes it. Events change it. So point the system forward instead of backward and ask what is coming:
- Kids going back to school — and note that the school calendar shifts year to year.
- College move-in weekends.
- High school and college sports starting back up.
- Pro games that fill a stadium near you.
- Major events. We are just coming off the World Cup, which moved behavior significantly in a lot of markets.
We are recording this in the back half of summer, which means in a few weeks a whole lot of kids head back to school and off to college. Does that drive your volume up or down? If you have to guess, that is the gap.
The institutional knowledge problem
The old answer to all of this was a veteran manager who just knew. That answer is disappearing. Our teams have less tenure than they used to, and we are having a hard enough time keeping the people we have.
When a manager with 12 years of pattern recognition walks out the door, that knowledge leaves with them unless it lives somewhere else. Feeding this data into a system that keeps learning is how you stop paying for the same lesson twice. Think of it as building a second brain for the operation — one that remembers the sold-out game three miles away, and what it did to your covers.
Track your food waste, or none of this works
The gap in almost every one of these programs is waste. How often are you prepping too much ranch? Do you actually know how much you made versus how much you threw out?
If you are not capturing food waste and over-portioning, your forecast will keep telling you a story that does not match your P&L. The more you track and feed in, the more you get back out. That is the whole deal.
One serious caution
Everybody likes to joke about AI being slop, and honestly, untended it will be. This is not a system you install and walk away from. You have to train it, enhance it, feed it, and nurture it — continuously.
It is no different from your loyalty program, your point of sale, your online ordering, or your labor management system. None of those got better on their own. Behaviors are going to change inside your brand and outside your four walls, and the model only knows what you keep telling it.
Treat it like a team member. It is going to make mistakes. Train it, educate it, talk to it, give it the data. The more data you give it — and the more you let it go look at what is happening in the world around your restaurant — the better it gets. The better it gets, the better you get. And the better you get, the more money you have to invest in the things you actually want to do.
The bottom line
- Roll your POS sales back into ingredient-level theoretical usage, then compare it to what you purchased.
- Interrogate the variance — weather, events, construction, region — so the next forecast learns from it.
- Point the forecast forward at the calendar instead of copying last week’s order.
- Capture waste and over-portioning, or the whole model is working from bad inputs.
- Keep training it. This is a system you tend, not one you install.
As I say on a lot of these episodes: if you are not doing this, your competitors are. You are already investing part of your time by listening to the show — now go take the next step and actually learn the systems.
If you have not already subscribed, please do. My name is Jeremy Julian, I run the Restaurant Technology Guys podcast, and I have been in the restaurant technology space for over 30 years helping restaurants succeed. I am also Chief Revenue Officer at CBS NorthStar, where we built the NorthStar point of sale solution for multi-unit operators — you can find us at cbsnorthstar.com. Make it a great day.
