AI Scenario Simulation for Restaurants: Before and After You Decide With Numbers

Step 1: load your real plate-level costing, not the one your supplier handed you

The first deliverable is a table holding thirteen weeks of sales alongside each dish's raw-material cost calculated from the weighed recipe, and without it the model returns elegant garbage. I got this wrong for years, accepting the theoretical cost printed on the spec sheet instead of the real one coming out of storage, and the gap between them usually runs four or five food-cost points. Weigh what actually leaves the walk-in and compare it against the sheet: that delta is your true waste. House rule says no dish goes above 32% food cost, and with 50% of full-service restaurants already automating inventory according to Restroworks 2025, you have somewhere to pull the number from without typing it by hand. You know the step worked when the summed cost of every dish sold matches your period storeroom issues within 2%.

Step 2: rebuild the sales mix week by week, never the quarterly average

Sales mix is the input that decides whether the simulation is worth anything, and it has to load disaggregated by week because a quarterly average buries the seasonality that actually moves margin. No restaurant runs thirteen flat weeks: there are paydays, there are long weekends, there is rain. Export unit sales per item per week from the POS and bolt them onto the cost from the previous step; what comes out is a thirteen-column matrix where every row is a dish carrying its unit contribution margin. That margin —price minus raw material, with no payroll or rent piled on top— is the only figure the simulation can project honestly. Verify the deliverable by summing contribution margin across all rows and contrasting it against gross profit on the P&L for the same period; a gap above 3% means sales outside the system or comps nobody logged.

Step 3: split out fixed structure and calculate break-even in covers

Put payroll, rent, utilities and insurance for the month into a single column, divide that total by the weighted average contribution margin, and you get how many covers you must sell before you stop losing money. That number, not food cost, governs every scenario you simulate afterward. A restaurant carrying twelve million in monthly fixed structure with a weighted margin of twenty-four thousand pesos per cover needs five hundred covers a month just to reach zero, and everything sold above that drops almost whole to the bottom line. Look hard at the costs eating margin outside the kitchen: delivery apps truly cost between 30% and 40% of order revenue per ActiveMenus 2025, with DoorDash nominal commissions at 15%, 25% or 30% depending on plan (Food On Demand 2026). Verify by running a closed month backwards: the model must reproduce your actual profit.

Step 4: write scenarios as questions with a number, never as intentions

A badly framed scenario returns a useless answer, and the difference lives in the verb: don't simulate "raise prices", simulate "raise the 42,000-peso dish by 7% assuming unit declines of 5%, 10% and 15%". The model then works out implied elasticity and tells you at which point the increase turns negative. Always run three variants per decision —pessimistic, base, optimistic— and leave each one written with its assumption beside it, because a scenario without an explicit assumption is just an opinion dressed as a table. Diego F. Parra pushes this order with the Masterestaurant teams: first the question with a number, then the model, never the reverse. This step delivers a document of five to eight signed scenarios, each carrying its volume assumption, its horizon in weeks, and the threshold at which you reverse the decision.

Step 5: connect the AI layer to your history and validate against a period you lived

Before you believe the model, make it predict a quarter you already survived and compare. An agent wired into sales history builds the full sensitivity in an afternoon, and with 86% of operators reporting they feel at least somewhat comfortable using AI in 2025 according to Toast, the barrier stopped being technical. The barrier is validation. Load January through March, ask for the April-to-June projection, and measure error against what actually happened: revenue deviation above 8% means the model is badly fed or a variable is missing —a closure for renovation, a promotion, a menu change— that nobody mentioned. Automation here is real and cheap next to a full kitchen robotics build, which runs between 150,000 and 250,000 dollars per location (Dataintelo). The deliverable is an error report line by line: revenue, covers, food cost and margin.

Step 6: simulate digital channels with their real commission layered on

Delivery, kiosk and chatbot each change the unit margin of every dish and must be modeled separately, because the same plate carrying a 30% commission can flip from profitable to destructive without food cost moving a single point. Self-ordering kiosks lift average ticket 30% at McDonald's per Restroworks, a guided-ordering chatbot adds between 12% and 18% to order value and moves site conversion from roughly 2% to 6.5% (Zellyfi), and 83% of guests pick another restaurant when their calls hit voicemail more than once (Hostie AI). Load each channel as its own column with its own margin: same dish, four different effective prices. It's done when you can answer, with a number, which dish should NOT sit on your delivery menu.

Four mistakes that wreck a well-built simulation

The mistake repeating across every implementation is feeding the model costings from two years ago, and the second is loading payroll and rent onto the plate, which artificially inflates unit cost and hides the true break-even; payroll and rent belong to fixed structure, always. Third comes staring at consolidated food cost: an aggregate 31% looks calm and usually hides two dishes at 38% subsidized by a starter at 19%, and that changes the menu, not the average. Fourth, and the most expensive, is simulating without defining the reversal threshold beforehand, because then you defend a bad decision out of pride for two quarters. What happens if the pessimistic scenario lands in week three? If you haven't written down what you do that Monday, the model bought you nothing.

Closing checklist: how to know the simulation came out right

Your model is ready when it clears six concrete checks and not before. One: the cost of all dishes sold reconciles with storeroom issues within 2%. Two: total contribution margin reproduces gross profit for the period with under 3% variance. Three: break-even in covers matches what the last closed month's P&L shows. Four: the retroactive projection of a known quarter misses revenue by less than 8%. Five: every digital channel carries its effective margin with the real commission loaded, between 30% and 40% on delivery per ActiveMenus 2025. Six: every scenario has a written assumption, a horizon in weeks, and a reversal threshold. When all six turn green, book a forty-minute monthly review and reload the rolling thirteen weeks: a model nobody feeds expires within a quarter.

The four differences that actually matter

The first difference is horizon. Without simulation you decide looking at last month; with it you decide looking at next quarter, because the model projects sales mix against fixed structure and tells you which week the change turns profitable. Not fortune telling — arithmetic with dependencies nobody holds in their head. Granularity is the second. A conventional dashboard shows a consolidated 31% food cost and you relax; the model cracks that 31% open and finds two dishes at 38% subsidized by a starter at 19%. That detail rewrites the menu, the average does not. At Masterestaurant we measure dish by dish because the consolidated figure hides precisely what needs fixing.

The four differences that actually matter — in practice

Courage is third. Once the pessimistic scenario sits on screen —traffic down 15%, inputs up 11%, twelve weeks of cash— the hard conversations with partners and suppliers happen ahead of the problem. A shared number ends arguments that an opinion would stretch across three months. And the fourth, hardest to explain: simulation does not reduce risk, it makes risk visible. You can still be wrong, but now you are wrong with the size of the error measured in advance, which lets you size the bet instead of praying. I got this wrong for years, selling certainty where there was only a range.