Data vs intuition: the guide that turns your gut into a system

Step 1: pull 90 days of POS sales and sort by margin, never by popularity

The deliverable of the first step is a sheet with four columns —dish, units sold, recipe cost and contribution margin in dollars— and you verify it by adding the margins up: the total has to match gross sales minus food cost for that same period, within a 2% tolerance. Ninety days is not arbitrary; below 60 days one holiday or a rainy week shifts your median. The case that opened this guide came straight out of this exercise: the sherry tenderloin sold 41 times a month at 3,10 USD of margin while the risotto sold 190 times at 9,40, which is 1.786 USD against 127 a month, fourteen times the cash. Your POS already stores that data and you paid for it: restaurant POS software moves 16.430 M USD in 2025 (SkyQuest Technology 2025). Use it. No data-driven decision is worth anything if the recipes loaded in the system do not match what leaves the kitchen, and that gap is the rule rather than the exception.

Step 2: audit your recipe cards before believing a single number the system shows you

Weigh your ten best-selling dishes during live service, not during a quiet morning test: protein, garnish and sauce, three repetitions each. Acceptable deviation is 5%; above that your theoretical cost is fiction. I got this wrong for years, recommending pretty dashboards built on 2023 recipe cards nobody had touched since, and the outcome never changed: immaculate reports describing a kitchen that no longer existed. A model trained on garbage returns garbage with an authoritative accent, which is worse than having nothing, because you BELIEVE it. The deliverable: ten cards signed and dated by the chef. Write the intuition in a single line shaped like this: «I believe X, and I will know if metric Y moves Z points before day 14». Without that format the hunch is unfalsifiable and you will keep defending it in the same meeting forever. A real example of the format: «I believe the risotto absorbs 2 USD more in price, and I will know if weekly units drop no more than 12% within fourteen days».

Step 3: turn the hunch into a hypothesis with a number, a date and a kill criterion

Fourteen days is enough because it gives you two full weekly cycles, weekends included, and weekends carry 40% or more of sales in most table-service operations. A restaurateur's intuition nails what people will enjoy and fails at what leaves money behind: palate and cash measure different things. A dish can score 4,7 in reviews and run 41% food cost with neither figure contradicting the other. Data without a control proves nothing, and this is the step almost everybody skips. If you raise the risotto price, leave another dish of similar rotation untouched for the same fourteen days and compare both curves; if the two fall together, blame the week, not your change. In multi-unit operations the control is even cleaner: apply the change in two locations and leave two exactly as they were. Availability bias rules the kitchen —you remember the night salmon ran out, not the eleven nights it was left over— and your memory keeps the peaks while erasing the median.

Step 4: measure against a control, because without comparison there is no evidence

Monday's order comes out of that biased memory. A written count carries no such defect. Deliverable: a two-column table, treatment and control, fourteen rows deep, with the difference calculated at the bottom. Five figures are enough to run a restaurant and everything beyond them is noise: period food cost, weighted average contribution margin, average check, waste in dollars and labor hours over sales. Put them in a shared sheet rather than in new software, and open it on Tuesdays at ten with the chef sitting there. What the market sells as decision intelligence for hospitality does not mean the algorithm decides; it means the algorithm BRINGS the question properly framed to the table, and pulling a dish with history remains your call. That is why the weekly ritual matters more than the tool. The money moving around this is enormous —AI in hospitality and tourism goes from 20,39 billion USD in 2025 to 26,53 billion in 2026, a 30,1% CAGR (The Business Research Company 2025)— and almost none of that spending fixes a badly kept sheet.

The four mistakes that wreck the sequence and how to dodge them

The costliest mistake is not picking the wrong metric: it is changing five things in the same week and ending up unable to tell which one worked. Log one change per fourteen-day cycle. The second mistake is measuring in percentages when the decision gets made in dollars; a dish at 28% food cost with 2 USD of margin pays less rent than one at 34% with 11 USD, and I have watched owners pull the second one out of respect for a percentage table. Third: buying automation before auditing inventory, which buys you speed in the wrong direction. Fourth, and most human: asking for the count from the chef who also cooks the service, because that count will happen at eleven at night, exhausted and from memory. Assign counting to someone off the hot line, even if that person is you during the first month. Deciding with data is not a fork in the road against intuition, it is a SEQUENCE: the hunch frames the hypothesis and the number kills or confirms it within fourteen days.

What to expect in one quarter: 2 to 4 food cost points without raising prices

An owner who installs that full sequence trims 2 to 4 food cost points in a quarter without touching menu prices, because the savings come from pulling two dishes that were draining cash, correcting three portion weights that drifted off spec and no longer purchasing from memory. On 60.000 USD of monthly sales, three points are 1.800 USD a month, 21.600 a year, and none of those figures required buying software. At Masterestaurant I carry this same sequence into kitchens across 43 countries and the pattern repeats: whoever measures wins, not whoever spends most on technology. The paradox is that the sector's most profitable tool is still a 40 USD scale used properly. You finished the rollout when you can answer six questions without opening the computer, and that is the exam. One: which are your three highest-margin dishes in dollars and how much do they contribute monthly?

Closing checklist: how to know everything landed right

Two: when was your signature plate last weighed and what was the deviation? Three: which hypothesis is running this fortnight and what day does it expire? Four: which dish did you pull last quarter and what happened to total margin afterwards? Five: what is your food cost for the closed week, to two decimals? Six: who counts inventory and at what hour? Fail two of them and the sequence is not installed, it is decorated. Start tomorrow with the cheapest thing available: export those 90 POS days, weigh everything on Thursday and write one hypothesis with an expiry date. A restaurateur's instinct is EXCELLENT at predicting what guests will enjoy and terrible at predicting what leaves money behind, because the palate and the register measure different things: a dish can hold a 4.7 review rating and a 41% food cost simultaneously, and neither figure contradicts the other.

Where gut-feel decisions actually break?

Availability bias runs the kitchen — you remember the night the salmon ran out, never the eleven nights it sat in the walk-in; memory keeps the peaks and erases the median, and next week's order comes out of that skewed memory.

A written count has no such defect. Decision intelligence applied to hospitality does not mean the algorithm decides, it means the algorithm BRINGS a properly framed question to the table; pulling a dish with history behind it is still your call, with context no model holds. Miscalibration cuts both ways and I will say this against my own trade: I have met owners who ignore the number and consultants who grant a number authority it has not earned, when the sample is eleven services with a national holiday sitting in the middle. Operations automation without prior judgment multiplies error instead of correcting it, because an AI agent reordering stock against a badly loaded recipe card will request 40 kilos too many with admirable punctuality, every Tuesday, unnoticed until closing.