Hear from CIOs, CTOs, and other C-level and senior execs on data and AI strategies at the Future of Work Summit this January 12, 2022. Learn more
DoorDash has been using forms of AI since its founding four years ago, namely advanced machine learning. And now that AI is contributing significantly to its business model, says Tony Xu, CEO of the food delivery service.
Xu — who is appearing at MobileBeat July 11-12 — cites results like the reduction of food delivery times, and personalized offers that are more 25 percent or more likely to get customers to buy than other offers.
Moreover, other restaurant chains that offer DoorDash to their customers are starting to cite DoorDash as becoming a contributing factor in their own earnings. Cheesecake Factory and Buffalo Wild Wings both mentioned DoorDash in their recent earnings calls.
The reason, Xu says, is because the shift toward convenience is happening so rapidly. Delivery is a central part of the reorganization of local and e-commerce.
“There’s a confluence of factors causing consumers to want more and more convenience,” Xu said. “But first and foremost, it’s driven by a change in consumer behavior, where last-minute purchases are the norm.”
Over the last decade, and especially in the last five years, people are more and more comfortable — to the point of habitual — with buying things on their phones or any internet-connected devices.
And delivery has become a company priority for both restaurants and retailer.
“AI and machine learning become very useful in this convenience economy, because there’s so much information that we have to process to bridge the online to the offline world,” he explains.
The company receives orders through the app or website — but in in order to fulfill it, lots of things have to happen in the physical world: food prep, food pickup, food delivery. For all of this, they’re able to collect information that helps deliver predictions that mean happily fulfilled deliveries and full bellies.
So DoorDash needs to perform millions of computations per second, Xu says, because of the scale and complexity of their operations. They’re factoring in variables including how long it takes a restaurant to whip up a meal, what kind of vehicle is best to complete the delivery, traffic patterns, and how easy it is to snag a parking spot.
Multiplied by the volume of deliveries happening per second gives DoorDash the kind of granular data that continually improves predictions around delivery times, punctuality, customer service, and consumer satisfaction.
“One of the things that we have to always predict and make better estimations on with every delivery we perform is, what types of deliveries are best suited for the types of vehicles on the platform?” Xu says.
Their algorithms are specifically able to predict which dashers – their delivery persons – will be successful or not. From cars to bicycles to scooters to motorcycles and vans, each is more suited to particular types and sizes of orders, and each for types of distances. For example, a bicycle traveling in San Francisco may have some trouble getting up the hills, but in New York that biker will be flying.
With a just-in-time system, the platform also always needs to balance the timing so that when a driver shows up at a restaurant, there’s no wait time for the food. Too soon, efficiency is lost. Too late, cold food and probably no tip (or repeat business).
“Knowing when to send the driver is a very computationally intense problem, where you’re constantly estimating, at all hours of the day, across different-sized orders,” Xu says. “We can use human heuristics to make those judgments, but to know the precise distances and the specific capacities that these vehicles can carry, we have to run these simulations in order to make those decisions.”
The challenge is that AI and machine learning predicate on the ability to collect high volumes of information with high fidelity. Because they’re collecting a lot of information that’s not directly in their control, or necessarily even visible to them, there can be a tremendous amount of noise in the information.
A restaurant gets really busy on a Friday night, but when does it get busy? When does that spike actually occur? Is it exactly the same hour every single week at the same restaurant, or does that change? How does that vary based on events happening inside the city? And how does that impact traffic as well?
They need to use a number of different signals, plus a variety of techniques, to sanitize the data before it can actually be used — otherwise, no matter how sexy your AI and machine learning strategies are, you’ve got the classic garbage-in, garbage-out scenario and delivery people are driving off the end of Pier 39 and everyone’s dim sum is cold.
However, he says, “There are instances where human heuristics dominates even machine learning or AI.”
Meaning no amount of AI is going to solve a situation where restaurant staff didn’t show up for work, accidents that cause traffic closures, or powerlines going down. Life, in general.
“There has to be a sense of pragmatism where you’re also building products to solve the issue even if there are going to be ‘surprises’ inside of your system,” Xu says. “Yes, a lot of it uses prediction to try to make sure that we can estimate the way the day is going to play out. But we’re absolutely also building a lot of products to support the case where there are surprises, both positive and especially negative surprises, so we can solve those problems in real time as well. ”
He does admit that AI and machine learning, or frankly any technological innovation, is not a silver bullet.
“It doesn’t all of a sudden magically solve our problems of fulfilling deliveries on time and fulfilling as many of them on time as possible, or as accurately as possible,” Xu says. “But a company that does not believe in the power of using information regularly and with extreme rigor in making every decision, I think, is going to struggle regardless of which machine learning or AI expert they’re able to attract to the business.”
Companies require three specific skills working together: the ability to use, collect and interpret data; the ability to take that data and build it into software products; and then strong operations, with local teams on the ground helping to integrate the online with the offline, helping troubleshoot when predictions and reality don’t match.
“While machine learning and AI is the topic of today, it’s been a part of our DNA since day one, and it will always be that way,” Xu says.