QSR drive-thru AI: what a franchisee should actually buy in 2026.

Drive-thru AI is three procurements, not one. Buy the wrong workload and you are recreating the McDonald's IBM story on your own dime.

Three drive-thru AI workloads, and only two are procurement-ready in 2026

"Drive-thru AI" as a buying category collapses three very different systems into one line item. The first is order-taking voice AI — the automated voice that greets a car at the speaker, parses the order, and pushes it to the POS. This is the McDonald's IBM use case. The second is upsell prompting — an AI layer that whispers suggested add-ons to the human employee taking the order, based on cart composition, day-part, and weather. The third is back-of-house workflow AI — computer vision on the make line for order accuracy, throughput prediction on the ticket queue, and camera-based drive-thru timing. In our last several QSR engagements the operator who conflated these three walked in wanting to "buy drive-thru AI" and walked out having priced three separate contracts against three different vendor pools. Order-taking is the risky one. Upsell prompting and back-of-house have measurable ROI right now. A franchisee who scopes them together is scoping them wrong.

Order-taking voice AI is where the industry got burned in 2024, and 2026 has not cleanly resolved it

McDonald's ended its Automated Order Taker test with IBM in June 2024 after two and a half years and roughly 100 locations. Restaurant Business broke it; CNBC confirmed the shutoff by July 26, 2024. The proximate cause was accuracy in the low 80% range and viral videos of 260-nugget orders and bacon on the McFlurry. The structural cause was that a fast-food menu, a bad speaker, and a stressed customer are still, in 2026, an edge-case generator that beats general-purpose speech models more often than the pitch decks admit. Two years on, the field has consolidated but not cracked it. Wendy's FreshAI on Google Cloud is running in 500-plus locations with pilot accuracy reported at 86% without human intervention. Yum Brands and Nvidia rolled a Riva-based stack across roughly 500 Taco Bell, Pizza Hut, KFC, and Habit Burger locations in Q2 2025; Taco Bell had expanded to nearly 900 locations by July 2026. Presto Voice reports around 95% on its deployments. Chick-fil-A is running trials but its Elevated Drive-Thru concept is a human-in-the-loop bet, not an AI-in-the-loop bet. The independent vendor pool — SoundHound, Kea, Hi Auto, ConverseNow, ClearFlame, Vistry, OpenCity — is real, but a 14% failure rate at the speaker translates into a real human on headset for every seventh car. If your labor model already assumed the AI held, that gap is where the ROI evaporates.

Upsell prompting to a human employee is quietly delivering the measurable lift

The workload almost nobody talks about is the one paying its way in 2026. An upsell-prompt AI does not touch the guest. It listens to the order the human employee is entering, checks it against cart composition, day-part, weather, and prior visit patterns, and prompts the employee to offer a specific add-on — a shake at 2 p.m. on a hot day, a specific side that pairs with a combo. Nvidia's Yum stack has this baked in. Deep North and RestaurantSpaces sell it as a discrete camera-and-analytics layer. The lift is small per transaction and huge per store per year, and because the AI is not on the customer-facing path there is no failure mode that ends up on TikTok. In our engagements this is the workload franchisees keep skipping because it does not sound like the future. It is the one with the cleanest 12-month payback in the current vendor pool.

Buyer-side. Supplier-paid. Buyers pay zero. Compensation has zero weight in the Cardinal Index scoring. We scope the three workloads separately under the Cardinal Method, then match each to vendors that actually own the IP for it — not to a bundled pitch that promises all three.

Back-of-house AI is the least glamorous spend and the most reliable one

The third workload is cameras and analytics inside the four walls. Agot AI — acquired by HM Electronics in April 2025 and now bundled with the drive-thru headset that most operators already own — watches the make line and catches missing or wrong items before the bag closes. Kitchenary, BrightLoom, Deliverect, and PAR Technology sit around this same box: throughput prediction on the ticket queue, delivery aggregation into a single kitchen view, drive-thru timing and queue vision from an overhead camera. None of it is voice. None of it talks to the customer. All of it moves the two numbers a franchisee actually reports to the brand: order accuracy and service time. This is where the reliable spend is in 2026, and it is the workload the industry press covers least.

Five questions to ask any drive-thru AI vendor before signing

Paste these straight into the RFP:

1. Which workload do you actually do? Order-taking voice, upsell prompting to the employee, or back-of-house workflow. If the pitch spans all three, ask which one is your own IP and which is a reseller relationship.
2. What is your accuracy rate on our brand's menu, not on your reference brand's menu — measured as orders completed without human intervention, over at least 30 days of live traffic in our region.
3. What is the human-agent fallback path when the AI cannot resolve? Where does the escalation route to, how long does the handoff take, and what does the customer hear during it?
4. How does the system integrate with our incumbent POS — Toast, Square for Restaurants, NCR Aloha, Oracle MICROS, or Revel? Is that a certified integration or a middleware bridge, and who owns the ticket if the POS is down?
5. How are you priced? Per drive-thru unit, per completed order, per human-labor hour saved, or fixed monthly. And what is the contractual exit if the accuracy SLA misses for two consecutive quarters.

If a vendor cannot answer question 2 with a number on your menu, they are selling you a reference implementation and hoping your menu behaves.

What breaks

Buying order-taking AI to fix a throughput problem. Throughput is a back-of-house problem — kitchen capacity, order sequencing, expo staging — not a speaker problem. Order-taking voice AI removes one labor cost at the speaker and adds a new cost when the bag is wrong. Franchisees who bought voice AI to move their drive-thru time by 15 seconds usually found the 15 seconds in the kitchen instead.

POS integration as an afterthought. Most of the order-taking failures we have unpicked in the field were POS integration failures dressed up as accuracy problems. The AI heard the order fine. The POS took the wrong modifier because the middleware could not represent it. Scope the POS integration before the voice model.

What to do this week

Sort your drive-thru AI question by workload before you take a single vendor call. If the pain is speed of service and labor at the speaker, you are shopping order-taking, and the 2024 verdict still applies — pilot small, hold a hard accuracy SLA, keep the human path live. If the pain is average check or add-on attach, you are shopping upsell prompting, and the ROI is available now. If the pain is order accuracy or throughput inside the four walls, you are shopping back-of-house vision, and that is the least risky spend in the category. This Field Note nests under Hospitality, where the same integration-depth logic governs every guest-facing system a multi-unit operator sources. Sibling read: Hotel WiFi: the guest-experience math no one runs.

Sources

• Restaurant Business, “McDonald's is ending its drive-thru AI test with IBM,” June 2024. restaurantbusinessonline.com
• CNBC, “McDonald's to end IBM AI drive-thru test by July 26,” June 17, 2024. cnbc.com
• Wendy's, “Drive-thru innovation: Wendy's FreshAI.” wendys.com
• Restaurant Dive, “Wendy's to deploy FreshAI and digital menu boards at 500+ restaurants in 2025.” restaurantdive.com
• Nvidia, “Yum Brands and Nvidia partnership announcement.” blogs.nvidia.com

All linked sources were live at time of publish (July 2026). Verify before quoting in a procurement document.