A small business can start with a recurring task that has a known starting time, records it can safely use, and a person who already checks the result. In an October 20 OpenAI customer story, Ted Taylor, who operates three co-branded Schlotzsky’s and Cinnabon restaurants in New Mexico, says monthly cash reconciliation fell from about three hours to five to ten minutes with ChatGPT Work. This is a vendor-published customer account. Independent verification of the reported time saving remains a separate question. It is still a better place to start a business conversation than a demonstration about everything AI might someday do.
In brief: Pick one repeated back-office job. Record how long the whole job takes today, including checking and correction. Give AI the preparation work, keep approval with the owner, and compare finished results over several cycles. Judge the improvement after the full job is complete.
Three stores, several different kinds of result
Taylor’s story covers more than reconciliation. OpenAI says menu repricing, which used to take five days, now takes one to two hours when Taylor brings together recipe costs, item popularity, local competitor prices, portion sizes, and the economics of delivery orders. The company also reports four hours of manual work saved per payroll cycle. Those are time claims about jobs Taylor says he has done.
The largest number in the case is different. Taylor estimates that better scheduling could eventually reduce hourly labor costs by one to two percentage points of sales across the three restaurants, potentially freeing up $50,000 to $100,000 annually. That is a forecast of possible savings. An owner making a buying decision should not put a projected scheduling gain in the same column as time observed on last month’s reconciliation.
This is where many AI business cases go soft. They present a stack of impressive numbers without saying which were measured, which were estimated, and which costs moved to another person’s desk. Taylor’s case is useful because the tasks are ordinary enough to separate those claims.
Start where the answer can be checked
Cash reconciliation has a clear boundary. There are transaction records, bank deposits, and discrepancies to explain. An assistant can compare the records and prepare a list of mismatches. Someone still has to confirm whether an apparent mismatch is timing, a missing deposit, or a real error. Include that review in the job’s measured time.
If you run three stores, begin with last month’s closed reconciliation. Remove or protect sensitive account data as your policies require, and ask for a proposed match list plus every transaction the system could not confidently match. Have the person who normally reconciles the month check the result against the books. Time the preparation, the check, and the correction. Compare that total with the old process, using the same standard of completeness.
The useful AI saving is the time between receiving the records and signing off on a correct result. A five-minute output that takes two hours to repair has not saved the owner three hours.
Keep the pricing decision with the operator
Menu repricing is more judgment-heavy. The proposed price depends on food cost and demand, but also on what customers will tolerate, how delivery fees change the margin, and whether a local special affects sales in one location differently from the others. OpenAI’s case says Taylor uses AI to assemble this analysis in one to two hours instead of five days. It does not show that the recommended prices produced higher profit or preserved customer demand after implementation.
That distinction matters. Ask the system to show where each input came from, when it was last updated, and which recommendations change if ingredient costs or demand assumptions change. A store operator can then challenge the numbers before updating a menu. Keep the old pricing sheet and record what actually happens after the change. Otherwise, you may save time producing a proposal and lose money acting on an unchecked one.
A smaller operator can borrow the approach without copying Taylor’s tools. Give the assistant approved records and ask it to organize options for review. Separating preparation from approval makes a first trial easier to reverse.
Count what happened after the first month
I would keep a short ledger for the trial. On one side: the baseline hours for each monthly close, payroll run, or repricing cycle. On the other: AI preparation time, human review time, corrections, subscription or setup cost, and the number of exceptions caught before a decision. Put estimates in a separate row until actual operating results arrive. Use the ledger to determine whether people got time back across the full job.
Repeat the measurement. A single clean month may hide the messy cases that consume the most staff time. If the new process only works while the owner supervises every step, it may not be ready for the next location or the next manager. If an ordinary colleague can check it and explain a discrepancy without calling the person who built it, that is closer to an operating improvement.
OpenAI’s account offers a persuasive example of a franchisee using AI on work that restaurants actually do. It does not establish that another operator will achieve the same percentages. Start with a job you can count, keep the person who knows the books in the loop, and wait for the second and third cycle before claiming a return. That is how an AI result becomes a business result. For the wider accounting question, see Measuring AI’s Business Impact.
Research and draft: Mai. Editorial direction: John Lipe.