Handing a stuck drive-thru order to a human lifted accuracy from 81% to 95%

Orders taken by the machine alone were wrong 19 times in 100 ^[18]. Orders that moved to an employee after the system got stuck were wrong 5 times in 100 ^[19]. The handoff removes about three quarters of the errors ^[20].

Both figures come from the same set of visits, to three chains already running automated voice ordering ^[13], and they are not two arms of an experiment. The orders that got handed off are the orders the system could not finish, so they were harder on average. An observational split like that cannot show that escalation caused the improvement. It can show that the recovery path performed better on the cases the model could not handle. CIO's account does not include a cost per order for the McDonald's system, nor the number of visits behind the 81% and 95% figures ^[21].

The economics are the part CIO states flatly: the voice system had to be cost-effective compared with having a person at the counter, and it failed to offset its operating costs ^[6]. That claim is single-sourced and unquantified, and it sits awkwardly beside the remedy. An escalation path costs staff attention at the window, the same expense the automation was bought to remove. For three years employees stood within reach of those speakers and did not step in, CIO reports, because nobody and nothing called them ^[16].

Franchisee frustration was on the record before the parody videos were. Analysts following the project noted discontent at the lack of progress, and updates were infrequent ^[5]. In August, after lower than expected US results, chief executive Chris Kempczinski asked analysts to sympathize with restaurant managers handling several product launches in a few weeks, each with its own training ^[10]. Service times lengthened and customer satisfaction fell ^[11]. Drive-thru is around 70% of the company's US business ^[12].

The accent problem is real and, by CIO's account, unfinished: dialects and ambient noise at the order window remain difficult ^[8]. A model that mishears is still workable if something notices. On this record, nothing did, and the counter labour the system was meant to displace stayed on the payroll ^[6]. Archy is now in limited testing at a handful of drive-thrus ^[1]. The figures I would want out of that test are the rate at which it asks for help and the wait the customer absorbs while it does, not another word-accuracy score ^[17].