How DoorDash Built an AI Shopping Assistant That Doesn’t Rely on the LLM Alone

DoorDash has shared the architecture behind Ask DoorDash, its conversational AI assistant that helps consumers discover restaurants, plan meals, and build grocery carts via natural-language interactions. Over the course of a three-part engineering deep dive, the company described how it built the system using large language models, specialized AI agents, Model Context Protocol (MCP)-based tooling, persistent consumer memory, and automated evaluation infrastructure to operate AI-driven experiences in production.

DoorDash reported early production results showing that computed consumer memory improved grocery checkout conversion by approximately 24%, increased basket sizes by 17%, and reduced conversational turns by 7% during a seven-day evaluation. For restaurant discovery, the assistant achieved a 15% higher conversion rate on open-ended queries. The company also reported improvements in AI agent quality measurement through automated evaluation, enabling faster validation of model and system changes.

As Raghav Saboo, RecSys and Search Lead at DoorDash, noted in a post,

Building a useful AI agent is hard. Knowing if it is actually good is even harder.

To validate AI agent behavior in production, DoorDash built an automated evaluation framework that simulates stateful customer conversations using LLM-generated users and recorded tool fixtures. The framework mirrors the production runtime to independently evaluate orchestration, guardrails, and domain agents. DoorDash said the platform scaled quality measurement to more than 2,000 automated evaluations daily, improved quality scores by eight points, reduced regression testing from six hours to twenty minutes, and validated a model migration that reduced latency by 35% while maintaining quality.

The architecture behind Ask DoorDash separates orchestration from business capabilities. An Assistant Runtime coordinates interactions among specialized agents, while a shared MCP layer exposes business functionality, including catalog search, recommendations, carts, checkout, order history, and consumer memory. Rather than embedding business logic directly into prompts, the assistant invokes reusable tools backed by existing DoorDash services, allowing multiple AI experiences to share integrations while backend capabilities evolve independently.

DoorDash Assistant runtime architecture (Source: DoorDash Blog Post)

In a LinkedIn post announcing the feature, the company highlights a core challenge

Agents do not just need access to user data. They need the right context, for the right task, at the right moment.

DoorDash introduced an intelligence layer that manages personalization through three memory systems. Long-term memory is generated offline from historical consumer behavior to capture preferences such as favorite cuisines and dietary restrictions. Session memory maintains conversational context during an interaction, while agentic memory stores explicit facts provided by consumers. Relevant memories are retrieved through semantic vector search, ranked, and incorporated into prompts, separating memory management from model inference.

Memory architected across generation, tooling layer, storage, policy, and the agents (Source: DoorDash Blog Post)

In a LinkedIn post announcing the feature, Andy Fang, Co-founder at DoorDash, said

Ask DoorDash can build you a grocery cart ~5x faster than doing it manually, and that it takes a single prompt to complete your cart in under 2 minutes.

DoorDash emphazied, the platform improves latency and reliability through operational optimizations, including deterministic actions that update versioned artifacts without invoking the language model and confirmation workflows for recommendations and generated carts. The architecture aligns with DoorDash’s engineering model, where domain teams build specialized agents while platform teams maintain orchestration, MCP tooling, memory, evaluation, and shared components.

DoorDash notes that production AI agents introduce complexity through orchestration, retrieval, and tool coordination, but these investments enable reusable infrastructure, improved reliability, and independent evolution of domain capabilities.