How a New York restaurant can build next week's prep list with AI

Sunday night in the walk-in, clipboard in hand, guessing at par levels from memory while a storm rolls in and a street fair sits on the calendar for Saturday. Every kitchen manager has stood there, doing arithmetic on feel. This piece shows how to let AI turn last week's sales, local events and the weather forecast into next week's prep list in a few minutes, and how to turn that into a routine that runs every week.

What AI does here

AI tools such as ChatGPT and Claude are chat programs you type questions into and get written answers back from. Given last week's sales numbers, a list of nearby events and a forecast, one of them can draft prep quantities per item, flag which days look heavy, and write it all out as a list your prep cook can work from. It is doing the pattern-reading you already do in your head, except it holds all three inputs at once and shows its reasoning.

The output is a draft, not an order. A language model can produce wrong answers, so anything important gets checked before it drives purchasing. The routine below builds that check in as a step, not an afterthought.

Who this is for

This fits an owner-led restaurant or a kitchen manager who already tracks last week's sales in some form, whether that is a POS report, a spreadsheet or a nightly count sheet. It works in New York City, on Long Island, upstate, anywhere the sales data exists. It does not fit a kitchen with no record of what sold last week, and it does not replace the judgment of someone who knows their walk-in. It gives that person a faster first draft.

Build the list

The whole setup can be built in one sitting with ChatGPT or Claude, and it needs no new software beyond that subscription, which the restaurant pays for directly. No booking platform, no integration, no API connection. Just the numbers you already have.

Here is a shape that works. Say what the restaurant is, paste last week's counts per item by day, paste the seven-day forecast and the events list, then ask for a prep par for each item for each day next week, in the same units as the counts, with one line of reasoning per item. End by telling it to flag anything it is unsure about and to ask a question if a number looks off. That last line catches typos before they turn into purchases.

The saved prompt is the reusable part. After two or three weeks it carries the corrections the model needed, the items that run hot on event days, the sauces that hold. Weather and events are where the model earns its keep, but they are also where it guesses, so the prompt should state your kitchen's quirks plainly: the item that dies in a heat wave, the cut that always moves with a home game. That context turns a generic answer into one that sounds like your kitchen wrote it.

Keeping the list honest

A weekly routine needs a weekly check. Each Sunday, before running the prompt, compare last week's suggested pars with what actually sold. Where the model over- or under-called an item, say so in plain words and it will carry that into the next draft. Over a month the list stops looking like a stranger's guess and starts looking like yours.

Where a hand helps

Building a routine like this on a working call is the kind of task New York AI Lab's AI Concierge is made for. One of the team's people learns the business, works alongside the owner on real tasks, and helps with small setups like this one, with two 1-hour calls a month and voice messages and email between calls. Everything built lives in the restaurant's own accounts and is theirs to keep. A full walkthrough of the service sits on the site.