Instacart's agentic commerce strategy: you do not have to own the agent to win

TL;DR

Instacart's agentic commerce strategy makes three things plain. You can win without owning the agent, and you can get paid without owning it: Instacart's shelf sells through ChatGPT, Claude, Gemini and AI Mode, none of which it owns, and every cart built there becomes an Instacart order. Nobody knows what the shopper's agent will look like, so Instacart also built one of its own, a hedge that 1.6 billion orders of data can fund and a shop cannot. Your move is the first half of that playbook: make your shelf readable by any agent and let the order complete with you.

Where Instacart stands

Two things happened at one company, in this order. Between December 2025 and July 2026, Instacart put its shelf inside four assistants it does not own: ChatGPT, Claude, Gemini and AI Mode in Google Search. Then, on 9 September 2026, it introduced Clementine, an assistant of its own, and named seven grocers taking Cart Assistant, the same assistant under the grocer's name on the grocer's own site.

Those are two different plays. One is to plug into every agent there is. The other is to own the agent. Instacart is doing both, and the order it did them in says which one it trusts.

Four outside hosts before its own. A company that thought the assistant was the prize would have built it first.

You can win without owning the agent

Instacart does not own ChatGPT, Claude, Gemini or Google Search, and it does not need to. In each of them a shopper says what they want to cook, the assistant builds a cart from live Instacart inventory at a store near them, and the order completes on Instacart, per Progressive Grocer on the Claude connector and TechCrunch on AI Mode. Anirban Kundu, Instacart's chief technology officer, said in April what that makes the company:

What Instacart brings to each of those doors is the same thing, on its own figures from the 9 September release:

That is what an assistant cannot do without: what is on the shelf, in which store, at what price, right now. The host owns the conversation. Instacart owns the shelf, and the shelf is the part every host needs.

You can get paid without owning the agent

Follow the order. In all four hosts the conversation ends on Instacart, where users "can complete their order for delivery or pickup from their preferred retailer", in the release's words. In ChatGPT the shopper pays inside the chat through Instant Checkout, built on the Agentic Commerce Protocol with Stripe, and the order is still "shopped and delivered through Instacart's trusted platform", per Instacart's December release. The host gets the conversation. Instacart gets the order, the delivery and the account. Both of its own assistants are to carry advertising by the end of the year, per Modern Retail.

None of that money depends on the shopper opening the Instacart app first. The money is in the order, not in the chat window.

Nobody knows what the shopper's agent will look like

This explains the second play. If Instacart knew ChatGPT would win, it would put everything there. If it knew shoppers would keep a grocery assistant the way they kept a grocery app, it would build only Clementine. It does not know, and its own line says so: present wherever the journey starts. So it hedges. It plugs into every large assistant, builds its own, and sells a rebranded copy to grocers who want their own name on the door. Ryan Hamburger, its chief commercial officer, told Modern Retail why grocers take that copy: "retailers want to build on brands they've developed over 100 years."

Then the tell, from the same interview: "We're not sharing the same level of data with OpenAI as we would with the shopping assistant embedded on our platform." The assistant goes to every host. The best data stays at home. And the assistant is the part everyone has. Shipt launched Ask Shipt the same day, per The Shelby Report. Uber Eats and DoorDash already had theirs, and TechCrunch reads the wave as delivery apps keeping users in their own apps "rather than having them turn to popular AI chatbots like ChatGPT." A week earlier Anthropic gave away a free blueprint for a retailer-embedded shopping agent.

So here is where Instacart stands. Bet one: whichever agent the shopper uses, the shelf under it is Instacart's. Bet two: the shopper will also keep a grocery assistant, and it had better be Instacart's, or a grocer's running on Instacart's data. Bet two is a bet that AI goes the way apps went: one per vertical, ten on a phone, a grocery one among them, and no single personal agent doing all of it. Instacart can afford both. It placed the first one first, and it feeds the second one better.

The approach that holds whichever agent wins

Nobody knows what the shopper's agent will look like in six months, in a year, in two years. Instacart's own timeline says so. The list of places a shopper can send a grocery order changed four times in nine months, and the fifth, Instacart's own, arrived last week. The next change is as likely to come from a host nobody has named yet as from any of those five.

That is the demand side, and it moves first. People are handing their daily tasks to a personal agent, and the agent that shops for them will be whichever one they already have open. Nobody gets to pick it for them. Not Instacart, not a grocer running a rebranded copy, not a shop with a chat window on its home page.

The supply side moves second, and it has started to move. A rebranded assistant on a grocer's site is a store-side agent, and so is the retailer-embedded one Anthropic gave away a week earlier. Stores will run more of them: an agent that answers for the catalog, for stock, for price, for the order. At some point the two sides meet in the middle, the shopper's agent talks to the store's agents, and the order settles between them.

Until then, a shop does not get to bet on which agent wins. It works with whatever the demand side adopts. Whichever agent the shopper shows up with, the shelf has to be readable by it, and the order has to complete with the shop. That is the part of Instacart's playbook it built first, and the one part that does not need Instacart's money.

That is the approach we take as well. We do not own the agent. SIL is a shopping capability the person's own agent loads, on a host they run, and it stays theirs when they change hosts. We are building toward the day the two sides meet, and we are not guessing which agent gets the shopper there.

What this means if you run a shop

You are who this is for: you have a shop, and you want agentic commerce to work for you. Instacart's playbook has two halves. You can afford one.

The owned half is a bet that a shopper opens an AI for your shop on purpose, the way Instacart bets they will open one for groceries. Instacart funds that bet with 1.6 billion orders, 2 billion catalog items and a brand people already open. Three rivals placed the same bet in the same week. Even Instacart placed it second.

The other half is the one Instacart built first, and it works whichever agent wins. Anthropic said it from the other side, in the closing line of its blueprint: "some of the traffic to your storefront will come from agents that shop on behalf of users." A store today publishes a feed and hopes. What Instacart sells to every host is the opposite of hope: stock, store, price, right now, in a form an agent can act on. Machine legibility already has a price tag, and the bill now arrives through every door.

What to do on Monday

1. Count how often your shelf leaves the building. Instacart takes in more than 10 million inventory signals a day. Ask how many times a day your stock and prices reach anything outside your own site, and in what form. A feed refreshed weekly is a number too.
2. Make the shelf readable before you build the window. Structured product data, stock and price per location, and a checkout that completes when the cart was built somewhere else. Every host on the timeline above reads that layer. None of them needs your assistant.
3. If you still build an assistant, build it on the shelf. Instacart's is grounded in real-time inventory because the inventory came first. Bolt the shelf on later and the assistant gives confident answers about things you do not have.

If you run your own agent, the same rule holds from the other side: send it to shelves it can read, and hold what it finds against the spec you gave it.

FAQ