AI agents barely place orders yet. They already shape what gets ordered.

This week, on DoorDash's third-quarter earnings call, its chief executive was asked about AI agents ordering food. His answer was the most honest thing said about agentic commerce in weeks.

If you look at where AI agents are today, we're seeing maybe several hundred orders in total from some of these third-party agents.

TechCrunch put that next to a number from DoorDash's 2025 report: an average of 3 million deliveries a day. Several hundred orders in total, against millions a day. Xu named Muse, Instinct and Fo among the agents and summed up his view: "you never want to fight gravity. But at the same time, you don't want to be delusional."

The same week, Amazon's Andy Jassy was similarly cool. Fortune reported him saying third-party shopping agents are "still not great, nor more convenient than coming to a retailer directly."

If you only read those two stories, you might conclude AI shopping is mostly hype. Then you read the other data published the same week.

The same week, the other numbers

Adobe reported on October 29 that AI traffic to US retail sites grew 71% year over year in September, and that those visitors converted 90% better than visitors from other sources. Adobe expects AI traffic to grow 159% on Thanksgiving.

The National Retail Federation found that 47% of consumers are using AI platforms and tools to support their Halloween shopping, mainly for brainstorming and for recommendations during product research.

Meta said Muse, its assistant, has more than 10 million users, with shopping a top activity, as TechCrunch reported on October 21.

Table: Two kinds of AI shopping data from the last two weeks

Signal	Number	Source	Stage it measures
Orders placed by third-party agents on DoorDash	Several hundred, in total	DoorDash via TechCrunch	Checkout by an agent
DoorDash deliveries, for scale	About 3 million a day (2025 report)	DoorDash via TechCrunch	All orders
Consumers using AI tools for Halloween shopping	47%	NRF	Research and planning
AI-referred retail visits vs. others, conversion	90% higher	Adobe	Research, then human checkout
Year-over-year growth in AI traffic to retail sites	71% (September)	Adobe	Research, then human checkout
Muse users	More than 10 million	Meta via TechCrunch	Mixed

The table sorts itself. Where an AI does the buying, the numbers are tiny. Where an AI does the advising and a person does the buying, the numbers are large and growing.

Advisers, not buyers

We think the clearest way to read this week is that AI has become a shopping adviser well before it has become a shopping agent.

That matches what McKinsey found in its June survey, which we covered in (post:mckinsey-agentic-commerce-survey-cpg-2026): 48% of US adults had used AI for at least one shopping task in the previous three months, while the share of US AI users who completed a purchase with AI was 17%. It also matches Forter's data in (post:forter-agent-orders-screen-scraping-data), where agent orders grew fast from a small base.

Show the data

Use	Value
Using AI tools for Halloween shopping	47%
Brainstorming ideas while planning	34%
Recommendations during product research	32%

Source: NRF and Prosper Insights & Analytics, September 22, 2026. Survey of 7,889 consumers, September 1-9, 2026. NRF does not report how many let an AI complete a purchase.

Notice what the top uses have in common. Brainstorming and recommendations are both about deciding what to buy, not about the mechanics of buying it. A shopper who asks an assistant which allergy-friendly candy to hand out still walks into a store or opens a retailer app to buy it. The assistant's influence is real. It just does not show up as an agent order.

Why the checkout numbers are small

There are good reasons agent checkout lags, and most of them have little to do with whether the technology works.

Retailers are choosing sides. TechCrunch noted that Amazon has been blocking personal agents while Walmart has opened up to them. We followed that split in (post:retailers-pick-sides-on-ai-agents). An agent that cannot buy at the store a household prefers is not much use.

People like their apps. TechCrunch reported Xu saying users still predominantly order through DoorDash's app. Grocery and takeout are repeat purchases where the app already remembers what you like. An agent has to be clearly better than two taps to win.

Trust takes time. Handing a payment method to software is a bigger step than asking it a question. Walmart's new agent spending limits, which we covered in (post:walmart-agent-spending-limits-cpg-baskets), exist because shoppers want control before they let an agent spend.

There is also a simpler point about the data itself. An agent order is easy to count: a platform can see which software placed it. Influence is hard to count. When a shopper reads an assistant's comparison of three granola brands, closes the app and buys one at Target that evening, no system records that the assistant made the choice. Adobe can see the click when a shopper follows a link, and surveys like NRF's can ask people what they did. Everything else is invisible. So the checkout numbers are not only small. They are also the easiest numbers to find, which is why they get quoted.

That creates a measurement trap for brands. A team that asks "how many orders did AI agents place for us?" will get an answer close to zero and may decide to wait. A team that asks "what do assistants tell shoppers about our category, and how often do they name us?" will learn something it can act on this month.

None of this means agent checkout will stay small. DoorDash itself has launched its own agents and says it wants a healthy relationship with third-party agent makers, possibly with advertising inside them. It means the timing is uncertain.

What this means for packaged goods brands

For CPG brands, the useful conclusion is about priorities.

The adviser stage is where the money is now. If 47% of Halloween shoppers used AI tools and AI-referred visitors convert 90% better on retail sites, then what an assistant says about your product already affects sales, regardless of who presses the buy button. Getting that answer right is this quarter's work. Tip (tip:12) explains how to check what ChatGPT says, and (tip:25) covers what to do when it is wrong.

Agent checkout is a project to prepare, not to bet on. The decisions about letting agents buy from your own store, covered in (tip:33), are worth making deliberately. But a brand that spends the holiday season on agent checkout plumbing while its products are misdescribed in assistant answers has the order backwards.

Measure influence, not just orders. If you only count agent-placed orders, you will see almost nothing and conclude AI does not matter. If you track AI referral traffic, branded search after AI exposure and what assistants say about you, you will see where it does. Some teams track assistant answers with tools such as SearchPanel from XENA Intelligence; a weekly manual check works for a small range.

The honest middle

The loudest predictions about AI agents doing all our shopping were premature, and DoorDash's several hundred orders show it. The loudest dismissals are wrong too, and Adobe's conversion data shows that.

We would also be careful about reading too much into a single company's numbers in either direction. DoorDash's figure covers third-party agents on one platform, in a category, restaurant delivery, where the app is already fast. Adobe's figure covers retail sites across 18 categories. Neither tells us how many grocery or personal care purchases an assistant shaped this month. What they show together is the shape of the market, not its size.

The accurate picture is less dramatic and more useful. AI is already the place many shoppers go to decide. It is not yet the thing that buys. For brands, that means the most valuable work this season is the least futuristic: making sure that when someone asks an assistant what to buy, the answer about your product is accurate, current and easy to act on.