The figure comes from early tests in English and Spanish. As of June 2026, Archy was reported in five US restaurants. For comparison, IBM's system ran about 85% order accuracy at 10 Chicago restaurants in 2021, with roughly one order in five needing a crew member to step in. A 2022 BTIG analyst survey of franchisees put a 24 store Illinois pilot in the low 80s, against the 95% or better that operators said a wide rollout would need.

Above 90% beats IBM. It is still short of that bar. Neither the article nor any coverage I found defines accuracy. Orders finished with no human help, orders with zero item errors and items correct per order are three different measurements, and they can sit ten points apart. Five restaurants is a pilot.

That is per restaurant, per Restaurant Technology News's account of the same presentation. Fifty hours is a little more than one full time position. On stage the claim carried a qualifier, "as Archy scales," and Restaurant Dive quotes it that way.

It is a forecast, sitting next to test results. Taco Bell, which now runs voice AI in more than 890 US restaurants, describes its payoff differently: better staff retention, because crews spend less time on the headset and more on the customer. Whether McDonald's restaurants bank the hours or move them to the counter is a franchisee decision, and franchisees operate about 95% of the 46,000 restaurants.

Intelligent scales weigh the bag against what the order should weigh. More than 10,000 restaurants already have them. McDonald's says crews catch and correct 10% of verified orders this way. Unlike claims 1 and 2, this is a deployed system reporting a measurement.

If a scale flags one bag in ten, roughly one order in ten was leaving the window wrong before anyone weighed it. That is the closest thing in the announcement to a human baseline, and it is the reference point Archy's 90% is missing. A person on a headset appears to run near 90% as well. The accuracy a customer experiences belongs to the whole chain: an order taken, checked by weight, then handed over. Weight has limits. It catches a missing burger. It does not catch a Coke that should have been a Diet Coke.

Bluetooth Low Energy tags on stock let a restaurant count inventory automatically, which McDonald's says saves about five labor hours a week. Connected fryers and grills, plus computer vision, feed predictive maintenance.

Neither figure comes with a baseline, a test size or a time frame. A 50% improvement in availability could mean half as many outages or availability rising by half, and those describe very different restaurants. The best known public measure of McDonald's equipment uptime is McBroken, a crowdsourced map of broken ice cream machines that has run since 2020. The presentation did not say whether soft serve counts as key equipment.