Build a plate-costing tool with Claude for a restaurant
A Brooklyn restaurant owner, also a fractional CFO, used Claude and Python to turn messy invoices and recipes into plate costs. She reports about four hours saved a week.
Angela Moznerova, a fractional CFO who owns a Brooklyn restaurant, built a costing model with Claude acting as a junior analyst. Claude reads invoices from small vendors and turns them into one price book. It also converts recipe units, including "a pinch", into clean tables. Her chef fills in and checks recipes in a shared Google Sheet. Anything that produces a money figure stays in Python, with a test that must pass and a saved run history so each run can be compared with the last. Any plate cost that moves more than 15% is flagged. She also used the output to raise minimum order values on delivery apps.
How it went
As reported by the interviewer, the first working version took about four hours to build and now saves her about four hours a week. Updating recipe costs takes 20 to 25 minutes a week. The checks caught a pasta recipe sheet listing 3 oz of pepper instead of about 3 grams, which made the dish look like a 33% food-cost problem when the real figure was about 13%. They also caught a 60-liter pot size read as the yield (the real yield was 15 liters). A vendor price gap worth up to a four-figure sum a year was found, but she keeps vendor choice for herself and her chef.
Worth knowing
Self-reported in a single interview. The tool is still in progress and savings have not been measured yet. The owner has finance skills and can write and test code, which most restaurant owners lack. Fuzzy name matching paired "Sage" with "Sausage" and an AI wine price estimate was wrong, so both were replaced with an explicit name map and the invoice price. The article gives US dollar figures and a US setting, and gives no precise software cost beyond an AI subscription plus free tiers.
Try it yourself
Paste into ChatGPT, Claude or another AI assistant and fill in the square brackets.
Act as my junior analyst for restaurant costing. I'll paste [supplier invoices] and [recipe sheet with ingredient quantities and units]. Step 1: turn the invoices into a table of ingredient, unit, price per unit and supplier. Step 2: convert every recipe quantity to a standard unit and list any you had to assume, such as "a pinch". Step 3: list anything that looks odd, such as a batch name that may be a pot size rather than a yield, or a quantity far above what the invoices show. Do not estimate any price; if an ingredient has no invoice price, mark it missing. Do not calculate final plate costs yourself. Output clean tables I can run through my own tested [spreadsheet or script]. Flag any plate whose cost moves more than [15%] from last time. I will check the flagged items with the chef before I change any menu price or vendor.
A third-generation izakaya owner in rural Japan had Claude enter cash receipts into a ledger. Date errors were fixed by adding one line to a written rule sheet.
A New York sandwich shop owner asks Square AI about sales and repeat customers, and says it found that Fridays beat Thursdays and helped him thank top spenders.
A Taiwanese shop owner built a ledger app with Claude and Claude Code, with Notion storing data. AI reads receipts; the owner reports month-end close fell from hours to about 20 minutes.