[Vertical AI Cases: Restaurants] AI Cuts Reconciliation From 3 Hours to 10 Minutes
Today's Highlights: In Restaurant AI, the Back Office Comes First
In the past two weeks, English-language media published two very concrete restaurant AI case studies: one is a franchisee story compiled by OpenAI, and the other is usage data for the AI assistant from restaurant POS vendor Toast. Both come from companies that sell AI, so take the numbers with a grain of salt. But they happen to answer the same question from two angles, the owner's and the manager's:
Where are restaurants actually using AI?
Case 1: A three-store franchisee hands back-office chores to AI
Ted Taylor spent 25 years in the semiconductor industry before switching careers to run restaurants. He is now an independent franchisee of three combined Schlotzsky’s and Cinnabon locations in New Mexico. According to a case study OpenAI published on October 20, he uses ChatGPT Work for several of his most time-consuming back-office tasks:
Menu pricing: He gives AI his ingredient costs, item sales, nearby competitors' prices, and the different costs of dine-in versus delivery all at once. His annual pricing analysis, which used to take 5 days, now takes 1 to 2 hours and produces suggested price lists for dine-in and delivery.
Monthly reconciliation: He uploads the cash transaction records and bank deposit records for all three stores and has AI match them line by line. A 3-hour job now takes 5 to 10 minutes.
Payroll: AI compiles the payroll input data, saving about 4 hours of manual work each pay period.
Scheduling trial: He gives AI each employee's position, training, availability, and hourly wage, plus the staffing needed at different sales levels, and it produces a full weekly schedule, which managers review and adjust before entering it into the scheduling system. He estimates this could eventually cut hourly labor costs by 1 to 2 percentage points and add $50,000 to $100,000 a year in cash flow, but that is still an estimate and hasn't happened yet.
More worth learning than the numbers is how he went about it: the first time he used AI for pricing, he also did the math by hand as a check, and only after confirming that AI's comparison was more detailed did he let AI produce the first draft. When analyzing labor costs, he gave AI read-only access.
Case 2: What managers ask AI during the rush
Toast analyzed how U.S. restaurants used its AI assistant from April to June 2026 and published the results via Stacker at the end of October. A few numbers stand out:
The most common topic is sales and revenue, at 30% of conversations; menu management and labor hours each account for 8%.
The dinner rush accounts for 27% and the lunch rush for 26%, so more than half of all questions come during the busiest hours, and more questions come from phones (54%) than from computers.
Of the requests asking AI to actually make a change, more than 70% involve the menu, most often taking sold-out items off it.
On loss prevention, checking discounts accounts for 62% and checking voided orders for 25%.
In other words, the moment AI gets called on most isn't when someone is sitting in an office, but when they're standing at the pass with a phone in hand. Toast is also candid that these numbers only show what managers asked and did, not how much money was saved, and that topics were classified by keyword matching, so the figures are approximations.
Separately, Forbes columnist Gene Marks, in his October 25 weekly tech roundup, discussed both this franchisee and a survey from headset brand Jabra: of more than 6,000 knowledge workers, 85% said AI had increased at least one type of work, while only 35% said AI had reduced some of their workload. His conclusion: what matters isn't whether you use AI, but how you use it.
Jesse's Take: Let AI Do the Work You'll Actually Check
After reading these two cases, my biggest takeaway is this: the real time savings don't come from writing menu copy or making Instagram graphics, but from reconciliation, pricing, and scheduling, the monthly jobs owners dread most and can't afford to get wrong. That said, an honest caveat: both are success stories hand-picked by the vendors, so there may be selection bias. You don't get to see the restaurants where it didn't work, so treat the numbers as a direction, not a guarantee.
1. Start with work that has a right answer and that you'll check
Reconciliation is the best example: either the POS cash records match the bank deposits or they don't, and a quick comparison tells you which. Mistakes in this kind of work are easy to spot, which makes it ideal for letting AI practice. By contrast, open-ended asks like "help me come up with new dishes for next quarter" have no right or wrong answer, save little time, and make it hard to judge whether AI did a good job.
2. Get your data out of LINE and out of the owner's head
A common scenario: at many small restaurants in Taiwan, supply orders get called out in LINE groups (LINE is Taiwan's go-to messaging app), the manager writes the schedule by hand and sends a photo of it, and only the owner has a feel for menu costs. Ted could have AI work out his pricing because he had spreadsheets of item sales and supplier prices. Without spreadsheets, even the smartest AI can only guess. Toast's data is also a reminder that managers mostly ask questions on their phones during the rush, so your data needs to be organized well enough that a single question on a phone gets an answer.
3. Run old and new side by side once, then gradually let go
Follow Ted's approach to his first pricing run: do the math yourself, have AI do it too, and compare where they differ. For scheduling, have the manager check AI's version first and add any missed rules to the instructions. Start with read-only access, and a person always makes the final call.
A first step you can take today: Export last month's POS cash report and bank deposit statement for one of your stores to Excel, hand them to Claude or ChatGPT, and ask it to match them line by line and list anything that doesn't reconcile. Then check its work against your own reconciliation. In one evening, you'll know whether AI is worth using in your restaurant.
Further Reading
Related posts on JCXO
Original sources
Comments