McDonald's Archy AI Drive-Thru Returns With a 90% Accuracy Claim

McDonald's Archy AI drive-thru now takes orders in English and Spanish, with the company claiming accuracy above 90%. That figure puts voice automation back inside McDonald's growth plan, two years after its troubled IBM pilot ended.

The return matters because McDonald's is no longer presenting automated ordering as an isolated technology experiment. Archy sits inside ArchIQ, a broader restaurant operating system developed with Google. McDonald's wants that platform to connect ordering, inventory, scheduling, equipment, and restaurant data.

The timing also raises a difficult question. Can McDonald's turn a promising test metric into consistent performance across more than 46,000 restaurants? Its earlier system struggled with accents, background conversations, and unusual order changes. Those problems become harder, not easier, when software expands across regions and franchise operators.

McDonald's says Archy can release at least 50 labor hours per restaurant each week as it scales. CFO Ian Borden says the company does not intend to use those savings simply to remove workers. Instead, employees would move toward food preparation, order fulfillment, and customer service.

That promise places McDonald's between two competing versions of restaurant automation. One treats voice AI as a controlled assistant with humans ready to intervene. The other treats automation rates as evidence that human order-taking can largely disappear.

Archy must prove the first model can produce the economic gains expected from the second.

McDonald's Archy AI Drive-Thru Becomes Part of a Bigger Operating System

Archy is no longer just a voice at the speaker box. It is becoming the customer-facing layer of McDonald's broader restaurant technology strategy.

The assistant can accept drive-thru orders in English and Spanish. It can also process changes during an order before displaying the final selection and total. McDonald's says its accuracy rate exceeds 90%, although the public descriptions do not fully define that calculation.

Earlier reports offered a related but different metric. A McDonald's franchisee said roughly 90% of orders in five test restaurants were completed without human escalation. The franchisee also said the system had processed more than 1 million transactions.

An order completed without escalation is not necessarily an entirely correct order. A customer might accept a small transcription error, while an employee might correct another mistake without taking over the conversation. McDonald's has not publicly released a detailed test methodology that connects automation rates with final fulfillment accuracy.

That distinction matters because Archy belongs to a larger platform called ArchIQ. The Google-backed system is designed to support inventory management, employee scheduling, equipment monitoring, and order accuracy. Voice ordering is its most visible feature, but the platform's economic case depends on many restaurant workflows.

McDonald's placed that case inside a 10-year modernization program announced on September 23, 2026. The company plans to provide $8.5 billion in support for restaurant upgrades through 2036. Other changes include revised kitchen layouts, delivery lockers, larger play areas, and more visible coffee preparation.

The company says large-scale ArchIQ deployment can contribute to roughly $100,000 in annual cash-flow benefits for a typical franchisee. That is a company projection rather than an independently measured result. It combines several operational improvements, not only automated ordering.

McDonald's also plans to expand order-verification scales from 10,000 restaurants to 20,000 by 2028. Those scales help confirm that the packaged food matches the order before handoff. Their inclusion shows why accurate speech recognition alone cannot guarantee an accurate customer experience.

Archy creates the digital order. Kitchen systems must route it correctly, employees must prepare it, and the handoff process must deliver every item. An error at any stage can erase the benefit of a flawless conversation.

The new platform therefore represents a change in scope. McDonald's previous voice trial tested whether software could understand a customer. ArchIQ tests whether AI can coordinate more of the restaurant around that customer.

That expansion creates greater potential value. It also gives McDonald's more ways to fail.

Why McDonald's Is Returning After Ending Its IBM Trial

The central story is not that McDonald's discovered voice AI. It is that the company returned after publicly withdrawing an earlier version.

McDonald's began testing automated drive-thru ordering with IBM in 2021. The trial eventually reached more than 100 restaurants in the United States. In June 2024, McDonald's confirmed that it would end the test and remove the system from participating locations.

The company did not abandon voice ordering altogether. It said the IBM work had increased its confidence that an automated solution would eventually become part of its restaurants. McDonald's planned to evaluate other options after the test ended.

The failed pilot still left a visible credibility problem. Customers posted videos showing the system adding unwanted items or misunderstanding modifications. Some clips involved excessive quantities of chicken nuggets or unusual additions such as butter and ketchup.

Accents and dialects also caused difficulty, according to reporting cited by the Associated Press. Drive-thru audio adds engines, passengers, rain, poor microphones, and conversations from nearby vehicles. Each factor can distort speech before the software interprets an order.

Restaurant ordering is also structurally difficult for language systems. Customers interrupt themselves, use local menu nicknames, combine promotions, and change their minds. A person might say "make that two" without repeating the item that "that" refers to.

An automated system must connect those fragments while tracking prices, availability, and menu rules. It must also know when uncertainty is high enough to call a worker. Waiting too long creates errors, while escalating too quickly removes the labor benefit.

McDonald's began a multiyear Google Cloud partnership before the IBM pilot ended. The companies said the arrangement would bring distributed cloud technology into restaurants and support generative AI use cases. The Google Cloud agreement provided a new infrastructure path for later systems.

ArchIQ builds on that relationship. Edge computing, which processes data near the restaurant instead of relying entirely on a distant data center, can reduce response delays. Centralized data can also give software a more consistent view of menus and operating conditions.

Those improvements address important technical weaknesses, but they do not settle the customer-experience question. A quick wrong answer remains wrong. A highly accurate system can also feel frustrating if it cannot recover naturally from an ambiguous request.

The bilingual capability is particularly important. Spanish is not a single uniform speech pattern, and customers often switch between English and Spanish during one order. Menu terms, regional accents, and borrowed words can complicate recognition further.

McDonald's says Archy can handle both languages. The company has not yet published separate performance results for English orders, Spanish orders, and conversations that mix them.

That missing breakdown is central to evaluating the comeback. The IBM trial showed that average performance can hide the situations most likely to frustrate customers. Archy needs more than a stronger average if its weakest interactions remain highly visible.

The 90% Claim Does Not Answer the Hardest Accuracy Questions

A headline accuracy figure is useful only when McDonald's explains what was measured, which orders were included, and what counted as success.

McDonald's describes Archy's accuracy as exceeding 90%. Earlier demonstrations described about 90% of transactions as completed without human escalation. Those statements sound similar, but they measure different outcomes.

Accuracy can refer to recognized words, interpreted items, completed orders, or correctly fulfilled meals. Each level introduces new failure points. McDonald's has not publicly provided a confusion matrix, error distribution, or independent validation for Archy.

The denominator also matters. Simple orders can inflate an overall success rate if they dominate the test set. A customer requesting one standard meal presents less difficulty than a family changing sauces, sizes, substitutions, and promotional combinations.

Testing across one million transactions provides volume, but transaction count alone does not establish representativeness. The five reported test restaurants might not capture regional speech, weather, traffic, menu differences, or microphone quality across the wider system.

The phrase "without human escalation" also leaves room for several interpretations. It might mean a worker never joined the conversation. It does not necessarily mean every requested item reached the customer correctly.

Independent drive-thru measurements provide a useful baseline. The 2025 drive-thru study from Intouch Insight reported an industry average of 87% for fulfilled-order accuracy. McDonald's scored 86% in that study.

Those figures cover the final order, including the main item, side, drink, and special requests. They should not be directly compared with Archy's voice metric. However, they show why McDonald's ultimately needs to connect conversational performance to the meal delivered.

A drive-thru assistant can transcribe every word correctly while still selecting the wrong menu configuration. The restaurant can then prepare a correct digital ticket but omit an item during packaging. Customers experience the complete chain, not the software layer separately.

McDonald's is adding physical scales to more restaurants partly because it recognizes this issue. The scales provide another verification point before employees hand over an order. Archy and the scales could reinforce each other if the order data remains consistent throughout production.

McDonald's must also show performance during stressful conditions. Lunch and dinner peaks bring longer lines, more simultaneous tasks, and greater pressure to avoid pauses. A system that works during quiet testing might escalate more frequently when background noise and order complexity increase.

False confidence presents another risk. Employees can compensate for a system that openly requests help. They have less opportunity to catch an order that the software confidently misunderstood.

The most useful future disclosure would separate several metrics. McDonald's could report full-order recognition, human intervention, customer correction, kitchen-ticket accuracy, and final fulfillment accuracy. It could then divide those results by language and order complexity.

Without those details, the 90% claim remains a promising company-reported benchmark. It does not yet establish that Archy can outperform employees or maintain its result across McDonald's global network.

Wendy's and Taco Bell Show Why Scale Changes the Test

McDonald's is entering a live competition in which rivals have already learned that deployment scale exposes weaknesses hidden by demonstrations.

Wendy's developed FreshAI with Google Cloud and moved from a limited pilot toward broader deployment. The company said in 2025 that the system was processing tens of thousands of orders daily. It presented the assistant as a way to let employees focus on food and customer interaction.

FreshAI uses large language models to interpret customized orders and respond to menu questions. Wendy's says the system must handle regional dialects, background noise, and millions of possible menu combinations. Those are the same conditions Archy will face.

Wendy's planned to deploy voice AI at hundreds of restaurants. Its FreshAI update framed nationwide expansion as the next stage after testing in two states.

Taco Bell also expanded voice ordering across more than 100 locations in 13 states. Its parent company, Yum Brands, later partnered with Nvidia on restaurant AI systems. Taco Bell's experience showed both the potential reach and public-relations risk of automated ordering.

Customers posted videos of unusual requests designed to confuse the system. Other users described trouble with substitutions and mobile-order pickup. Taco Bell subsequently emphasized a more selective approach to when and where voice AI should operate.

These cases demonstrate why a restaurant chain cannot evaluate AI only through average accuracy. Public failures attract attention because the interaction is easy to record. A rare error can reach millions of viewers and shape perceptions of the entire deployment.

McDonald's carries an additional disadvantage because customers remember its earlier mistakes. Archy will not receive the benefit of being treated as a first experiment. Every error will be compared with the IBM system that McDonald's removed.

Its advantage is operational scale and data. McDonald's serves a wide range of customers, markets, and ordering patterns. A common platform can feed those interactions into product improvements, provided the company manages privacy and data quality carefully.

Scale also strengthens the financial argument. Saving 50 labor hours each week at one restaurant is meaningful. Applying that target across thousands of restaurants would reshape staffing and operating budgets.

However, labor hours "freed" by software do not automatically become better hospitality. Managers must decide how to schedule those hours and which tasks should receive them. Franchise economics can create pressure to reduce payroll even when corporate leaders emphasize redeployment.

Borden addressed that concern during McDonald's investor presentation. He said the objective was not to eliminate staffing but to redirect work toward hospitality and food preparation. The full investor-day transcript includes examples of employees preparing drinks while Archy handles an order.

That operating model treats AI as a task allocator. The software manages a repetitive conversation, while employees work on physical tasks requiring judgment and movement. It resembles the strategy described by Wendy's.

The model succeeds only if employees can trust the orders entering the kitchen. Frequent corrections would pull them back to the speaker while disrupting their other tasks. The labor savings would then exist in a presentation but not during a busy shift.

McDonald's therefore faces pressure from two directions. Rivals are building their own voice systems, so waiting carries a competitive cost. Its previous failure means moving too quickly carries an unusually high reputational cost.

Fifty Labor Hours Can Mean Better Service or Leaner Staffing

The real test for Archy is not whether it removes a headset. It is whether restaurants convert the recovered time into measurable operating improvements.

McDonald's says Archy will eventually release at least 50 labor hours per restaurant each week. Spread across seven days, that equals slightly more than seven hours daily. The figure is large enough to affect scheduling, kitchen throughput, and franchise cash flow.

McDonald's has paired that estimate with a clear public message. The company says it does not intend to reduce restaurant staffing because of the system. Workers can instead focus on hospitality and hands-on preparation, including new chicken products.

That claim deserves careful treatment. McDonald's corporate organization does not make every staffing decision inside franchised restaurants. Individual operators respond to sales, local wages, hiring conditions, and service targets.

A franchisee might use the released hours to add capacity during peak periods. Another might reduce scheduled order-taking time to offset rising operating costs. Both outcomes can emerge from the same technology.

The $8.5 billion modernization plan increases the economic pressure surrounding those choices. McDonald's says it will support franchisees, but operators will also make substantial investments. The company expects technology and restaurant changes to produce financial returns.

McDonald's says the broader program can add approximately $100,000 in annual cash flow for a restaurant. That estimate covers ArchIQ and other operational improvements. It should not be interpreted as a measured return from Archy alone.

The most persuasive result would connect labor reallocation with customer outcomes. McDonald's could show shorter service times, fewer missing items, fresher food, or better satisfaction at comparable sales volumes. Those measures would support its hospitality argument.

The 2025 Intouch Insight study found a strong relationship between friendliness and overall satisfaction. Visits rated friendly produced 97% satisfaction and 89% order accuracy. Visits rated unfriendly produced 22% satisfaction and 70% accuracy.

That finding does not prove causation, but it highlights the role of human interaction. Removing an employee from the ordering conversation can create value if that employee becomes more available elsewhere. It can weaken the experience if customers simply encounter fewer people.

Automation can also change the emotional texture of a mistake. Customers often forgive a person who quickly recognizes confusion. They may react differently when software repeatedly rejects a modification or fails to understand an accent.

A human handoff must therefore remain easy. Customers should not need a hidden phrase or repeated failed attempts before reaching an employee. The escalation process is part of Archy's product design, not evidence that the product has failed.

McDonald's also needs to protect workers from alert overload. If Archy escalates every unusual order without providing structured context, employees must reconstruct the conversation. That can take longer than answering the speaker from the beginning.

A useful handoff should show what the customer requested, what the system understood, and where uncertainty arose. The employee can then resolve one detail instead of restarting the order.

Archy's labor promise rests on hundreds of small design choices like this. Speech recognition attracts attention, but workflow integration determines whether those 50 hours exist inside a real restaurant.

What McDonald's Must Prove Before Archy Scales

Three signals will show whether Archy is becoming dependable infrastructure or another limited pilot with an impressive headline number.

The first signal is a clear deployment schedule. McDonald's has described Archy as part of its modernization strategy, but the known test began with five restaurants. Investors and franchisees need to see how quickly the company adds locations and how it chooses them.

A controlled expansion across different regions would provide stronger evidence than a rapid rollout within similar markets. Tests should include varied accents, bilingual communities, weather conditions, and restaurant layouts.

If McDonald's expands while maintaining performance, the case for a scalable McDonald's Archy AI drive-thru becomes stronger. If the rollout remains limited, the 90% claim will look more like pilot data than operational proof.

The second signal is metric transparency. McDonald's should separate order completion from order accuracy and final fulfillment. It should also report intervention rates for English, Spanish, and mixed-language conversations.

The distinction matters because Archy's bilingual capability is one of its most important improvements. A single combined percentage cannot show whether both language groups receive comparable service.

Independent testing would further strengthen the result. Mystery-shopping studies can examine complicated substitutions, background conversations, and deliberate corrections. They can also compare AI lanes with human-operated lanes at similar restaurants.

The third signal is what happens to labor and service after installation. McDonald's says Archy can free at least 50 hours each week without reducing staff. Restaurant-level evidence can test that claim.

Useful measures include scheduled labor, service time, kitchen throughput, employee turnover, and customer satisfaction. McDonald's should also track how often workers interrupt another task to resolve an Archy order.

If service improves while staffing remains stable, McDonald's can support its argument that AI redirects labor toward higher-value work. If payroll falls without visible customer benefits, Archy will look more like a cost-cutting system.

The company's previous IBM trial gives these signals unusual importance. McDonald's already knows that voice ordering can work in selected situations. Its unresolved challenge is building a system that fails safely across thousands of unpredictable restaurants.

That is why the Google partnership matters, but infrastructure cannot substitute for evidence. Faster processing, centralized data, and shared restaurant software create a stronger foundation. Customers still judge the final interaction one order at a time.

Archy is also part of a wider contest over restaurant automation. Wendy's, Taco Bell, White Castle, and other chains are testing different combinations of language models, edge computing, and human oversight. No company has established a universally reliable formula.

McDonald's has the scale to set an industry standard if its system works. It also has enough locations to amplify every unresolved weakness.

The company has made its initial case with a bilingual assistant, a reported accuracy rate above 90%, and a labor-saving target. Now it must publish results that connect those claims to completed meals and better service.

Watch the next expansion closely. Does McDonald's disclose performance beyond pilot averages, and do Spanish-language orders hold up under wider use? Most importantly, do restaurant workers gain time for customers, or do they spend that time correcting the machine?

Those answers will determine whether the McDonald's Archy AI drive-thru represents a durable operating change or simply a better-managed second attempt.