There is one invoice Xavier Mariezcurrena Vega still talks about.
At first glance, it is exactly what every accounting automation platform hopes to receive: clean, typed, organized, easy for software to process. A $10,000 wine delivery. Every SKU matches. Every line item is legible.
Then, almost as an afterthought, there is a handwritten note across the bottom.
Returned. Case broken.
A bookkeeper notices it immediately. The invoice gets adjusted. The business pays only for what it actually received. An AI model may never see it. The software approves the full amount, the payment goes through, and thousands of dollars quietly disappear, not because the technology failed, but because the document told only part of the story.
That small handwritten note has become his favorite example because it captures a much larger truth. The hardest part of restaurant accounting is not processing information. It is recognizing the difference between what happened on paper and what happened in reality.
As co-founder of Over Easy Office and ChouxBox, Mariezcurrena Vega has spent years helping restaurants make sense of operational chaos. What began as two founders manually scanning invoices in Philadelphia has grown into a back-office operation with roughly 400 employees across Philadelphia, Bogotá, and Manila. Along the way, he has watched wave after wave of automation promise to eliminate accounting work altogether. On the meez Podcast episode with Josh Sharkey and Michael Jacober, he explained why he is not buying it, and the reason is math.
"AI is great at getting you 80 to 85 percent of the way there... but sometimes 100 is not valuable in accounting when two to five percent is what you're scratching the surface on in terms of your margins."
Most conversations ask whether AI will replace restaurant accountants. Mariezcurrena Vega asks something more useful: what happens when the remaining 5 percent is the only part that decides whether a restaurant makes money?
The Last 5 Percent Is the Whole Game
In many industries, 85 percent accuracy is good enough. Restaurants do not have that luxury, because they do not have 15 points of margin to give away.
The numbers make that literal. Full-service restaurant profit margins sat at just 2.8 percent last year, with limited-service at 4 percent, according to the National Restaurant Association's 2026 State of the Restaurant Industry report. Total operating expenses have climbed 36 percent in six years, and 42 percent of operators reported their restaurant was not profitable in 2025.
Those numbers change the whole conversation about automation. If your business runs on a 3 percent margin, a 5 percent error is not insignificant. It can erase the entire profit.
That is what Mariezcurrena Vega means by "the last 5 percent." Outside hospitality, those final points of accuracy are diminishing returns. Inside restaurants, they are where the business survives or does not. The difference between an invoice processed correctly and one processed almost correctly is not administrative. It is financial.
Restaurants do not lose money only because food costs rise or labor climbs. Sometimes they lose it because small inaccuracies accumulate quietly until the month-end financials expose them. By then, there is nothing left to fix.
"When your margin is 3 percent, a 5 percent error in your numbers is not a rounding error. It is the difference between profit and loss."
Restaurants Generate Messy Data Because Restaurants Are Messy
So why doesn't automation simply close that last gap? Because restaurant data was never built to be clean in the first place.
Mariezcurrena Vega describes restaurant operations as systems built around survival, not structure. Anyone who has spent time in a kitchen understands immediately.
- A dishwasher cuts a hand during service, so someone from another station jumps over to cover, and invoices that were supposed to be scanned stay on the counter.
- A manager creates a temporary "open fish" button on the POS because service is moving too fast to build a proper menu item, and later nobody knows whether that sale was salmon, halibut, or swordfish, or whether it was taxed correctly.
- Third-party drivers leave with the wrong order. Inventory gets sampled during prep. An ingredient arrives in an unexpected pack size because the distributor substituted products.
None of these situations is unusual. They are what restaurants look like on an ordinary Tuesday.
That is exactly why automation struggles. Technology expects consistency. Restaurants specialize in adaptation.
He learned how wide that gap runs while building Over Easy Office's teams overseas. Teaching invoice processing was not enough; his teams first had to understand how restaurants actually functioned. He spent time in Manila sitting beside new employees, explaining that mustard might arrive in gallon containers, glass jars, squeeze bottles, or individual packets, and that every one of those could be the same ingredient in an entirely different form.
The invoice could not explain that. Only operational context could. His goal was not to teach accounting. It was to teach people how chefs think.
That challenge extends well beyond one department. As MarginEdge's leadership told Modern Restaurant Management, operators pull information from POS systems, invoices, payroll, and reservations, but most of those systems speak completely different languages, and for AI to be useful the data has to be standardized, connected, and consistently labeled first.
That is the catch. Restaurant data reflects real life, real life is messy, and AI inherits every inconsistency it is given.
AI Isn't Replacing People. It's Changing Where People Add Value.
None of this makes Mariezcurrena Vega an AI skeptic. He sees enormous potential for automation, especially in the repetitive administrative work that consumes back-office teams. What he rejects is the idea that automation eliminates human judgment. He sees the two as complementary: AI excels at volume, people excel at nuance, and restaurants happen to make or lose money in the nuance.
He is not alone in drawing that line. At the 2026 National Restaurant Association Show, when Restaurant Dive asked what people get wrong about the sector, tech leaders near-universally pointed to overhyped expectations for artificial intelligence, even as they stayed enthusiastic about the technology itself. Handwritten notes, delivery variances, substitutions, missing documentation, and two-factor logins are exactly the edge cases where an unsupervised model quietly gets it wrong.
This is the same logic that governs recipe data and food costing. Automation should carry the repetitive load so a person can spend their attention on the 5 percent that actually moves the margin. The technology is the floor, not the ceiling.
Instead of asking whether AI can replace the accounting team, operators can ask sharper questions:
- Is the data feeding our automation accurate enough to trust the output?
- Where does a human still need to review before money moves?
- Which mistakes would not surface until the P&L arrives weeks later, too late to fix?
Those questions shift the conversation from technology to decision-making. That is where AI becomes genuinely useful.
Bad Data Doesn't Begin in the Back Office. It Begins in the Kitchen.
Mariezcurrena Vega's examples all start with invoices, but they do not end there. Every accounting error eventually finds its way into an operational decision. A missed credit, an incorrect pack size, an outdated ingredient cost: none of them stays an accounting problem for long. They become food cost problems, pricing problems, margin problems.
The back office can only work with the information it receives. If the data entering the system is incomplete or inaccurate, every report built on top of it inherits the same flaws. The kitchen is where that process begins.
Every invoice introduces new information into the business. Prices fluctuate, suppliers substitute, pack sizes change without warning. If those changes are not reflected in the recipes that drive food costing, operators start making decisions on numbers that no longer describe reality.
As meez lays out in its guide to restaurant invoice management, a restaurant may believe it is running a 29 percent food cost when the true number is closer to 33 percent, not because anyone made a dramatic mistake, but because ingredient costs drifted while recipe data stayed still. By the time the month-end P&L reveals the variance, thousands of covers have already gone out at the wrong margin.
The invoice itself is another trap, because it does not represent what reaches the plate. As meez explains in its work on ingredient yield, the purchase price printed on an invoice is only the starting point. Trim loss, prep loss, and usable yield determine what an ingredient truly costs once it is ready to cook. A whole chicken bought at invoice cost ignores 20 to 25 percent butchery loss. Fresh herbs lose weight as they are cleaned. Feed that into any system, automated or not, and the answer is wrong before service begins.
"Garbage in, garbage out is not a slogan in restaurant accounting. It is the mechanism by which margins disappear."
The Last Mile Isn't About Better AI. It's About Better Data.
For all the attention AI receives, Mariezcurrena Vega's argument is surprisingly modest. He is not asking restaurants to adopt more technology. He is asking them to improve the quality of the information flowing into the technology they already have.
That is where his thinking aligns with meez. The problem is not that accounting platforms cannot calculate food costs. It is that they depend on accurate recipe data to calculate them correctly. Recipes are where ingredients connect to invoices, where unit conversions happen, where yields are accounted for, and where menu pricing begins. When those building blocks are structured correctly, downstream automation becomes dramatically more reliable.
A supplier price changes, and every recipe that uses that ingredient updates automatically instead of waiting for someone to re-cost by hand. A pack size changes, and the built-in conversions stay accurate. Ingredient costs rise, and menu profitability reflects reality instead of last month's assumptions. The result is a theoretical food cost you can actually trust, which is the only kind worth measuring against actuals.
That is the role meez plays. Not replacing a back-office platform like Restaurant365, but strengthening the data those systems depend on so the numbers finance relies on rest on an accurate foundation instead of a guess. Recipe data becomes the connective tissue between purchasing, culinary, and finance, giving the back office the operational context it has always needed. Mariezcurrena Vega spent years teaching accounting teams how restaurants work because spreadsheets alone could not explain restaurant operations. Structured recipe data solves much of that translation automatically.
The Better Question Behind Every AI Conversation
Strip away the headlines predicting AI will transform restaurant operations, and Mariezcurrena Vega's view is grounded. The real question is not whether AI is powerful enough. It is whether restaurants are giving it information accurate enough to act on.
That distinction changes everything. Instead of asking what AI can automate, operators might ask whether they can trust the information they are feeding it, where human judgment still matters most, and which decisions depend on absolute accuracy rather than reasonable estimates. Those questions produce far better technology decisions than chasing the newest feature. Because restaurants are not running on software. They are running on margins, and margins do not reward approximation.
AI will almost certainly become a permanent part of restaurant finance. It will process invoices faster, reduce repetitive work, and flag anomalies that once took hours of manual review.
But it will not eliminate the need for judgment. If anything, it raises the value of judgment by removing everything else, finally letting people spend their attention on the few decisions that actually determine profitability. In an industry where margins live between two and five percent, those decisions are still made by people, and they will only ever be as good as the data they are given.
"You're dealing with nature, and you're dealing with people, and it's what makes it beautiful."
See how clean recipe data keeps your food costs accurate before they ever reach your P&L, with meez and Restaurant365.
Listen to the full episode: Xavier from OEO on restaurant bookkeepers vs AI and how to charge corkage fees
Will AI replace restaurant bookkeepers?
Not in the near term, according to Over Easy Office co-founder Xavier Mariezcurrena Vega. AI reliably handles about 80 to 85 percent of an accounting process, but restaurant margins run just 2 to 5 percent, so the remaining nuance, like a handwritten "returned, case broken" note on an invoice, still requires a human to verify before money moves. AI reduces the manual load; it does not remove the need for review.
Why is AI accounting harder for restaurants than other industries?
Because restaurant data is built around survival, not structure. Invoices arrive with handwritten notes and inconsistent pack sizes, POS buttons get miscategorized, deliveries come up short, and staff consume inventory. As MarginEdge's leadership told Modern Restaurant Management, most restaurant systems "speak completely different languages," and AI needs clean, consistent data to be useful, which restaurants rarely have out of the box.
How does a small accounting error affect restaurant profit?
Enormously, because there is so little margin to absorb it. Full-service restaurant profit margins are around 2.8 percent and limited-service around 4 percent. When your entire margin is a few points, a 5 percent error in your cost data can erase the profit entirely, which is why accuracy matters more in restaurant accounting than in almost any other industry.
What is the most common source of inaccurate food cost data?
Latency and bad source data. Ingredient prices change on invoices but recipe costs do not update, so the numbers drift. meez's invoice management guide explains how you can believe you are running a 29 percent food cost while actually running 33 percent, undetected until the month-end P&L. Connecting invoice data directly to recipe costs keeps the number current.
Does the invoice price equal my true ingredient cost?
No. The invoice price is almost never the real cost of what reaches the plate, because trim loss, prep loss, and yield reduce usable product. A brunoise onion yields about 68 percent, and picked mint can lose more than half its weight. Costing to the raw invoice price understates your true cost from day one, as meez details in its breakdown of ingredient yield.
How do I make my restaurant's financial data accurate enough to automate?
Start at the source, in the kitchen. Clean recipe data with built-in yields and live invoice-linked pricing produces a theoretical food cost you can trust, which is the foundation everything downstream depends on. meez's guide to actual vs. theoretical food cost puts it plainly: if the source recipe data is wrong, your variance is not measuring bad operations, it is measuring bad math.
Does meez replace my accounting software?
No. meez is the culinary layer that feeds clean, recipe-level cost data into back-office systems like Restaurant365 rather than replacing them. It makes the systems you already use more accurate by ensuring the food cost data flowing into your ERP reflects real ingredient prices, real yields, and real recipes. meez frames this for culinary leaders as making your existing systems more valuable, not adding another one to maintain.
