DoorDash Reduces Grocery Reorder Processing Time by 8-Fold

Food delivery major DoorDash has significantly reduced the processing time for grocery reorders from 16 seconds to 2 seconds. The approach adopted by the company utilizes a technique called AI-powered batch processing, which pre-calculates purchase patterns before users interact with the app. As a result, the number of shopping list views has doubled, and the order conversion rate from that flow has improved by approximately 34%, according to the company's report.

This initiative was born from a structural challenge inherent in online grocery ordering. Since groceries are everyday items, many users repeatedly purchase the same products. Meanwhile, product lineups at supermarkets and similar stores constantly fluctuate based on inventory levels and seasonality. This makes fulfilling the simple need of "I want to order the same product as last time" surprisingly difficult from a technical standpoint.

The system DoorDash has built consists of three major components. First, it uses batch inference with large language models (LLMs) to pre-calculate which products each user is likely to purchase regularly. Second, it goes through a deterministic verification process to ensure the reliability of those calculations. Third, when users open the app, it conducts real-time verification against the product catalog and updates to display only products that are currently available for purchase. Additionally, by providing an interface where users can edit their orders directly, the system reduces the need to invoke the AI model repeatedly, thereby improving overall system efficiency.

The technical key lies in combining "pre-computation" and "real-time updates." By completing heavy processing (AI-driven purchase prediction) in the background beforehand and performing only lightweight processing (inventory verification) at the moment of user interaction, the perceived speed was dramatically improved. This represents a practical architectural design choice that incorporates "when to run" AI processing into the overall design.

What this case demonstrates is that beyond simply implementing AI, "designing when to run AI" directly impacts actual service quality. While LLMs have high processing capacity, running them in real-time for each request inevitably takes time. DoorDash's approach attempts to resolve this trade-off through batch processing and pre-computation, and this design philosophy is applicable as a practical solution for e-commerce and reservation services facing similar challenges.

In the grocery delivery market, convenience and speed of ordering are increasingly becoming major factors in service selection. The results of doubled view counts and improved order conversion rates demonstrate that improvements in processing speed directly influence user behavior. Going forward, whether similar pre-processing designs utilizing AI agents will expand beyond the delivery sector will be a key point to watch.