Operations automation: the before and after, measured at the till

Step 1 — Time one real week before you buy anything

Start by measuring, because without a baseline you cannot prove automation worked, and the deliverable of this first step is a sheet listing seven days of admin tasks, each with its actual minutes and the name of whoever performed it. A manager at a 120-cover restaurant spends close to 11 hours a week retyping supplier invoices, recounting bottles that failed to match, and copying the POS report into an email, and that block is your recovery budget. Do it with a stopwatch rather than from memory, since memory always underestimates repetitive work. Verification is simple: if by Friday you still cannot say how many minutes it takes to close the register on a Tuesday, step one is unfinished, and any license you sign now is a blind bet. With that sheet in hand, tag each task across three columns —high or low frequency, high or low judgment, data already digital or still on paper— and automate this month only the ones that clear all three.

Step 2 — Sort every task by frequency, judgment and data format

Diego F. Parra has applied this same filter in Masterestaurant audits for years, and the order matters more than the tool: inventory counts, register reconciliation, review replies, shift building and supplier orders almost always qualify; menu design and the conversation with a guest never do. Toast closed 2025 with 164,000 locations on its platform against 134,000 in 2024 (Toast 2025), so finding software is no longer the problem. The deliverable is a short list, four to six tasks, prioritized and signed off by you. Costing comes first because marketing amplifies whatever unit economics you already have, and if a dish loses 1.40 USD per unit, extra traffic merely speeds up the bleeding. Food and labor costs each climbed roughly 35% since 2019 in the United States (National Restaurant Association 2024), while Colombian operators raised menu prices 9.8% from February 2025 to sustain 98,000 jobs (ACODRES 2025), which means the recipe you costed a year ago is lying to you today.

Step 3 — Automate costing before marketing, no exceptions

Agents that read a PDF invoice and push it straight into costing are already cheap. The measurable deliverable: your twenty highest-volume dishes with food cost refreshed automatically every time an invoice lands, and not one dish above 32%. Before signing, ask the vendor three concrete questions and keep the answers in writing: whether a read API exists, whether the full history downloads to CSV whenever you want it, and what happens to your data the day you cancel. A stack where the POS will not export and inventory lives inside a closed cloud turns every future improvement into a commercial negotiation, and at that point the lock sets the price, not the market. Count the real integrations too, not the promised ones. Digital adoption did help the trade —contactless payment grew 260% between 2020 and 2023, per the 2024 restaurant POS market report— yet that convenience arrived wrapped in contracts almost nobody reads to the end.

Step 4 — Demand data ownership before signing any license

The deliverable: an email from the vendor confirming those three answers. A dashboard nobody opens is an expense, so step five ends not when the screen lights up but when a thirty-minute meeting exists, the same weekday each week, where three numbers trigger an action. Pick few and hard ones: food cost by family, prime cost over sales, margin by category. Alcohol is named a top-margin category by 46% of surveyed U.S. operators (Technomic / Nation's Restaurant News 2024), and that single line often justifies the whole exercise once you discover your wine list turns over every nine weeks. Verification is deliberately uncomfortable: ask for the minutes of the last four meetings. No minutes means no decision was made, and the dashboard has quietly become expensive decoration. Here sits the tension almost nobody resolves properly, and the way out is not balance but SEGREGATION: everything from the kitchen door inward and across the back office gets automated guilt-free, and everything touching the guest gets reinforced with the time you freed.

Step 6 — Segregate the floor: machine in the back office, person at the pass

Hospitality lives on the server who remembers you do not eat cilantro, and a bot answering reviews with one recycled template destroys in two weeks what took two years to build. Write the boundary on a sheet, with a named owner on each side. The deliverable is a one-page document, pinned in the office, where any new team member understands which decision a system makes and which one stays human. That separation decides whether technology adds or subtracts. The costliest error is not picking the wrong tool but automating a broken process, because an inventory miscounted by hand becomes an inventory miscounted at machine speed while looking like a serious report. Three others follow closely: switching on five modules in one month, which guarantees nobody adopts any; leaving the supplier catalog unnormalized, the same ingredient spelled four different ways, which wrecks any costing; and measuring savings in hours without converting them to money, the reason so many initiatives die the moment the manager changes.

The mistakes that sink implementation, and how to dodge them

Personalized email opens 26% more than generic email (Stripo 2025), and the lesson carries over whole: automation that respects context pays off, automation that flattens everything into one mold does not. Fix the process first, connect it afterward. You are done when you can answer five things without opening a file: how many weekly hours you recovered against the step-one baseline, what share of invoices reaches costing without anyone typing them, what last month's prime cost was and on which day you knew it, who holds the credentials and the CSV backup, and which task is next on the short list. Put a number on every answer and a date on the last one. If you recovered 6 of those original 11 hours and those 6 now live at the pass and at table 14, the project paid for itself; if you recovered them and spent them babysitting the system, you did not automate, you swapped chores.

How to know it all landed: the closing checklist?

Time another week ninety days from now and compare both sheets. The first difference is ORDER.

Automating marketing before costing is the most expensive decision I see owners make, because marketing amplifies a business whose unit economics are still unmeasured, and amplifying a dish that loses 1.40 USD per unit simply loses money faster. Cost data first, then the dashboard, then the agent that acts on the data, and only at the end AI content generation that brings more people through the door. The second is data OWNERSHIP. A stack where the POS will not export and the inventory system keeps everything in its closed cloud turns every future improvement into a commercial negotiation. Before signing any licence, ask three concrete questions: is there a read API, does the full history download as CSV, does the contract allow connecting a third party. A no on any of the three pushes the real cost of that tool far above its monthly fee.

The four differences that decide whether this works

The third is THRESHOLD. Automation pays when the task repeats more than 20 times a month and eats more than 4 minutes per repetition; below that line the maintenance of the automation costs more than doing it by hand. It is a boring calculation and that is precisely why almost nobody runs it, yet it separates the project that survives the second quarter from the one abandoned in March with a licence paid for a year. The fourth difference is cultural, and here I was wrong for years: I believed installing the tool and showing the report would be enough. It is not. If the head chef does not understand why the shrink alert reaches him and not the owner, he mutes it within three weeks. Algorithmic hospitality works when every alert has a NAMED owner, a threshold agreed with that person and a clear consequence; without that it is noise with a pretty chart.