Ai / Business | Arc has raised $10.76 million in seed funding led by Andreessen Horowitz as AI drive-thru ordering gets another chance. The question now is whether better models and real-time testing can clear the reliability bar that earlier fast-food pilots missed.
AI drive-thrus are getting a second hearing, this time with better models, sharper data, and a startup backed by Andreessen Horowitz.
The drive-thru has become one of the hardest places to prove whether voice AI can actually work. It is noisy, rushed, full of accents, interrupted orders, engine sounds, and customers who change their minds halfway through a combo meal. That is exactly why Arc thinks the market is ready for another try.
Fortune reported on May 26 that Arc, a voice AI startup founded by Square and Cash App veterans Mike MacLennan and Ali Hussain, has raised $10.76 million in seed funding led by Andreessen Horowitz. The company says it is already working with two major fast-food chains, each with hundreds of locations, and claims its system can complete orders on its own more than 95% of the time while lifting average order value by 4% to 5% through smarter upselling.
That matters because the last wave of AI drive-thru experiments left plenty of doubts behind. McDonald’s ended its IBM automated ordering test in 2024 after running it in more than 100 restaurants, even though the company said voice ordering still had a future in its restaurants. Customers had seen enough wrong orders and viral mistakes to make the technology look unfinished. For restaurant operators, unfinished is not good enough. A bad order slows the line, frustrates the customer, and often pulls a human worker back into the process anyway.
Arc is not selling the drive-thru as a simple labor replacement story. That is the more interesting part. The company is pitching it as a data layer for a channel that still behaves, in many ways, like an analog business. A drive-thru lane can generate most of a quick-service restaurant’s revenue, yet operators often have far less visibility into why some orders move quickly, why some upsells work, or where accuracy breaks down.
The company says its models are trained on real drive-thru conversations and tuned for brand-specific menus, scripts, accents, and order patterns. It also gives operators the ability to test different model variants, measuring speed, accuracy, and ticket size in real time. That turns the lane into something closer to an e-commerce funnel, where a restaurant can see whether a milkshake upsell works at the end of an order or whether it only adds friction.
This is where the economics become practical. Restaurants do not need AI to sound futuristic. They need it to keep cars moving, reduce errors, and increase the average check without irritating customers. In a business where margins are often thin, a few percentage points of order lift can matter, especially if the system also frees workers to focus on food preparation, payment, and customer problems that still need human judgment.
Incumbents now have to prove their own staying power
Arc is not entering an empty market. Presto, SoundHound AI, IBM, Google Cloud partners, and restaurant chains themselves have all been circling automated ordering. SoundHound has pushed voice AI across drive-thru, phone, text, and in-car ordering. Presto has marketed itself around restaurant automation and voice ordering. Yum Brands has expanded voice AI across Taco Bell drive-thrus and has also explored broader restaurant automation through its Nvidia partnership.
The difference now is that newer systems are arriving after the market has already seen what does not work. Early deployments showed that speech recognition alone is not enough. The system has to understand menu logic, handle interruptions, recover from confusion, know when to hand off to staff, and avoid forcing customers into awkward conversations. A person ordering food does not care whether the model is technically impressive. They care whether the order is right.
That creates pressure on incumbents. Better-funded startups can come in with newer large language models, stronger testing infrastructure, and a cleaner story around measurable outcomes. Public companies and older vendors have customer relationships and deployment experience, but they also carry the burden of earlier expectations. If Arc can show consistent results across different brands and geographies, the competitive argument changes from who has the most pilots to who can run the lane with the fewest mistakes.
There is still a hard question sitting underneath all of this. Restaurants may talk about labor augmentation, but workers will hear automation. If AI can take enough orders accurately, some chains will use it to reduce staffing pressure, especially at peak times or in locations where hiring is difficult. Others may use it to move workers away from headsets and toward fulfillment. The outcome will vary by operator, but the direction is clear: the drive-thru job is becoming more software-mediated.
For startups, the larger lesson is that physical-world AI is becoming investable again when it has a narrow job, a measurable return, and a painful customer problem. The drive-thru is a useful test case because failure is visible immediately. There is no long enterprise pilot to hide behind. The customer either gets the right order or they do not.
Arc now has capital, a clear wedge, and a market that wants the technology to work but remembers why it failed before. The next phase will be less about demos and more about operating proof. If the company can hold accuracy above 95% in real restaurants, across rush hours and regional accents, AI drive-thrus may finally move from experiment to infrastructure.
Also read: Uber says AI spending needs a clearer link to useful features • ByteDance raises the price of keeping its AI talent • China's AI travel curbs show the real doomsday risk is fragmentation