DoorDash's Agentic Commerce Reality: A Few Hundred Orders, Not a Few Million
On DoorDash's third-quarter 2026 earnings call, an analyst asked CEO Tony Xu how much business the company is actually seeing from third-party AI shopping agents, the kind of autonomous software that OpenAI, Perplexity and a wave of startups have promised will browse, decide and buy on a consumer's behalf. Xu's answer was blunt: orders placed through agents like Muse, Instinct and Fo currently total "maybe several hundred... in total." DoorDash's own 2025 disclosures put its average delivery volume at roughly 3 million orders a day. Measured against that baseline, agent-driven commerce isn't a rounding error so much as a statistical null result.
Infrastructure built ahead of demand
What makes the admission notable isn't that the number is small — agentic commerce is genuinely new — but how much DoorDash has built in anticipation of it being big. In the space of about four months, the company shipped Ask DoorDash, a chatbot that takes orders from text prompts and photos; a command-line interface that lets developers script orders directly; a text-to-order agent that works inside Apple Messages; and a Model Context Protocol (MCP) connector that lets outside AI tools place orders on a company's behalf. That is a rapid, well-funded buildout of rails for a form of commerce that, by the company's own account, almost nobody is using yet.
Xu framed the gap carefully rather than dismissively, telling investors he wants to avoid two opposite mistakes: fighting a real technological shift, or being "delusional" about where agents actually stand today. That framing matters for how the rest of the industry should read the number. DoorDash isn't walking away from agentic ordering — it's telling investors not to mistake infrastructure readiness for market existence.
The corporate niche that's actually working
If there's a place where agentic ordering is producing real usage rather than a few hundred novelty transactions, it's not the consumer "AI will shop for you" use case getting most of the press. At its Dash Forward event on October 1, DoorDash launched its MCP connector specifically for corporate ordering — letting companies route lunch catering and office-supply restocking through AI agents and workplace tools like Slack. The early partners it named are telling: SpaceX, Cognition, Vercel, Mercor and Tempo — AI-native and engineering-heavy companies integrating agentic ordering into their own internal tooling, not mainstream retail customers. Access is still gated by waitlist, a signal that even this narrower B2B use case remains pre-scale.
Co-founder Andy Fang described the goal as letting companies "put their office orders and office supply restocks on autopilot" and bring ordering "into tools their teams are already working in." That's a modest, operational framing — closer to internal workflow automation than to the consumer-facing agentic shopping revolution that OpenAI's Instant Checkout and Google's Gemini task-automation tests have been marketed around. It suggests the first durable production use of agentic commerce may be enterprise procurement automation among AI companies themselves, rather than mass-market autonomous shopping.
A pattern of downgraded expectations
This isn't a one-quarter data point; it's the fourth consecutive quarter in which Xu has quietly recalibrated the agentic narrative downward. In the first quarter of 2026 he said DoorDash would "absolutely" build agentic ordering experiences and that outside platforms would drive "a small percentage" of traffic. By the second quarter, he was describing agentic order volume as simply "low," framing the near future as a contest between the "battle for attention" (chat assistants) and the "battle for atoms" (physical fulfillment) — and betting DoorDash's value sits in the atoms. By the third quarter, "low" had become a literal headcount: a few hundred orders, full stop. Each quarter, the public-facing product roadmap accelerated while the usage figures DoorDash was willing to disclose shrank in scope and grew more specific.
That divergence is the real story for anyone tracking enterprise AI deployment rather than AI marketing. Retailers are currently split on how to respond to agentic traffic — Walmart is opening its catalog to outside agents to win share from rivals, while Amazon is blocking personal shopping agents outright — but neither posture is yet being tested by meaningful transaction volume. DoorDash's numbers suggest the industry is building checkout APIs, MCP connectors and agent partnerships for a market that, measured in actual completed orders rather than pilot announcements or referral-traffic growth, has not yet shown up. The lesson echoes a pattern recurring across enterprise AI more broadly: the infrastructure is often production-ready well before the demand is, and the companies most willing to say so out loud are the ones worth watching most closely.