Personal agents like Meta Muse and Instinct are teaching people to let AI shop for them. That is good for everyone. But the checkout they run is not the checkout most surfaces will need.
September was the month consumer agents got distribution
In the span of three weeks, the personal-agent category went from a handful of early adopters to the front door of Meta and Shopify.
Meta launched Muse on September 8, free to U.S. adults, with years of social activity behind its recommendations. On September 21, Shopify turned on Shop Pay checkout for Muse across every Shopify store. Three days later PayPal did the same across its merchant network. Amazon blocked it, which tells you the same thing from the other side.
Instinct, the text-first personal agent, added Stripe Link one-time cards in August and a Shopify Shop Pay partnership on September 28. Its founder says 35% of users already shop through it. The same day, Shopify opened WebMCP checkout to any browser-based agent across its eligible merchants: three tools, get_checkout, update_checkout, complete_checkout, and the order goes in once the buyer authorizes it.
And on September 29, OpenAI introduced dots: always-on personal agents with their own cloud computer and browser, connected to over 4,000 apps, that sign into websites with saved passwords and run an auto-review layer to decide which actions proceed on their own and which wait for your approval. Included in Pro and Business Premium plans. That is a third personal agent with a browser and a payment path, shipped by the company with the largest consumer AI audience on earth, in the same month as the other two.
That last clause is the whole story. Once the buyer authorizes it.
Watch enough of these transactions and a pattern shows up. There are two flows of agentic checkout being built right now, and they are not the same product. One keeps a human in the loop from first prompt to final tap. The other has no human in it at all after the order is placed. Muse, Instinct and dots are the first kind, and they are going to be very good at it. CartAI was built for the second kind from day one.
Flow #1: the human stays in the loop
This is what Muse and Instinct do, and it is worth being precise about how much of it is still the human.
You tell the agent what you want. It goes and finds options. It shows you a shortlist and you pick one, or ask for more, or change your mind. Then the checkout starts, and you are still there for it. Confirm the size. Confirm the address. Pick the shipping tier. Approve the amount. Tap to authorize the payment, usually through a wallet you linked ahead of time: Shop Pay, a Stripe Link one-time card, PayPal. Meta says Muse asks for approval before any purchase. Instinct's terms have you explicitly review and approve what it can spend. Shopify's new checkout tools place the order only after the buyer authorizes it.
The agent is doing real work in that flow. It is not doing the transaction. It is running a conversation that ends in a transaction, and the human is signing off at every step that matters.
That is not a criticism. It is the right design for a consumer who is handing their card to a piece of software for the first time. 72% of consumers say they will not share card details with AI. Approval gates are how you earn that trust back one tap at a time. And the infrastructure for this kind of checkout is arriving fast: WebMCP on the merchant side, wallet-linked one-time cards on the payment side, Visa and Mastercard agent credentials at the network level.
It is also a market with Meta, OpenAI and a $10 billion startup in it as of this month, plus Google, Amazon and every wallet provider that wants to be the tap at the end of the conversation. Human-in-the-loop checkout is going to be solved at scale, by companies with billions of users.
Flow #2: programmatic non-human-in-the-loop checkout
Now think about every surface that is not a chat assistant.
A gifting platform sends 4,000 holiday gifts for a corporate client. A bank's rewards app lets a cardmember redeem points for a product. A loyalty program fulfills a redemption. A procurement tool places 1,000 orders a day against a corporate card. A recipe app has a "buy the ingredients" button. A creator storefront, a shopping search result, a social post with a product tag.
None of these surfaces has a conversation in it. None of them will have a human sitting on the other side of the checkout, picking a shipping tier and tapping approve. The user expects what they have always expected from e-commerce: show me the choices, I pick one, I hit buy, it arrives. Everything after "buy" is the platform's problem.
That changes the shape of the transaction completely.
In Flow #1, the agent can be wrong and recover, because the human is right there to say "no, the blue one" or "use my other address." In Flow #2, there is no recovery. Every decision the human would have made in the loop has to be collected up front, before the order is placed: the exact variant, the delivery address, the shipping method, the payment credential, the final price including tax and shipping. Then it has to execute, in the background, on a merchant site that was built for a human, and come back with a confirmation. Not a cart. Not a "looks like it went through." An order number.
And it has to do that when the merchant's fraud engine sees a headless session with no human behavior signals. When the site throws a CAPTCHA. When the coupon field changes the total. When the merchant has no protocol, no UCP, no WebMCP, just a checkout page. When the same corporate card is used for the thousandth time today and the issuer wants to know why.
Human-in-the-loop checkout is a UX problem. Programmatic non-human-in-the-loop checkout is an infrastructure problem. The first one gets easier as merchants adopt protocols and wallets. The second one only gets easier if someone builds the execution, payment, identity and fraud-cooperation layers together and makes them deterministic.
In Flow #1 the human acts at four of six steps. In Flow #2 the human acts once, and everything after that tap has to run without them.
Even this week's big announcement is Flow #1
Shopify's move to open checkout to browser agents is the most agent-friendly thing a major platform has done. We have spent the past month running every route it offers against live stores, and the result is consistent: an agent can build the checkout, set the address, choose shipping, select payment and get the order ready to submit. The buyer still has to authorize the purchase. With Checkout WebMCP, the buyer remains in the checkout session, handles interactions such as Shop Pay login or payment challenges, sees the final order and total, and gives permission before the agent can place the order.
Shopify also has a server-side Checkout MCP path where eligible agents can call complete_checkout directly with a payment credential instead of handing the buyer back to the merchant checkout. Access to that capability is gated, rather than generally available. Shopify has announced direct agentic-commerce relationships with platforms including Muse and Instinct.
That is not a gap Shopify forgot to close. It is the design. The buyer is in the tab, the buyer's wallet pays, the buyer consents. It is the right shape for personal agents, and it leaves gifting, rewards redemption, procurement and every embedded buy button exactly where they were: one step short of an order, with nobody there to take it.
The left column is a handful of very large companies. The right column is everyone else who wants a buy button, and none of them has a conversation to put an approval gate in.
Where CartAI plays
We built CartAI for Flow #2 from the first commit. Everything about the product follows from the assumption that once the order is submitted, no human is coming to help.
Everything is resolved before the order. The catalog API returns variant-level pricing and availability with live shipping and tax already computed, so the surface can show a final buyable price and collect the choice up front. What the user approves is what gets charged.
Execution is a single API call with an order number at the end. POST /checkout with the product, the address and the payment credential. CartAI runs the checkout on the merchant's own site, in the background, and fires a webhook when the merchant confirms. If we do not get a definitive confirmation, we treat it as failed rather than guessing. Retries are idempotent, so a flaky merchant does not turn into a duplicate order.
It works where there is no protocol. Direct integrations through UCP, ACP and merchant APIs where they exist. Verified browser execution on the merchant's checkout page where they do not, or where the protocol stops short of completion, which as of this week includes Shopify. The surface does not have to care which path was taken.
Payments are built for no-human-present. Rail-agnostic by design: Visa and Mastercard agent credentials, virtual cards, Stripe wallets, or PCI-compliant direct card entry for corporate and buy-on-behalf flows where a business is paying with its own card at volume. Card intake and vaulting sit inside CartAI, so the surface never touches a PAN.
It cooperates with fraud and bot mitigation instead of evading it. Web Bot Auth, Skyfire agent identity, and integrations with Cloudflare, Akamai, HUMAN, Fingerprint and Forter. A programmatic transaction with no human behavior signals gets flagged by default. The only durable answer is to show up identified, with a payment reputation, on the merchant's terms.
The merchant stays merchant of record. Pricing, fulfillment, returns and the customer relationship never move. The bank, the gifting platform or the app becomes the place the purchase starts, with no merchant-side liability.
Can CartAI support the Flow #1 experience too? Yes. The same API works under an assistant that wants a human to confirm each step. But that is not what we optimized for.
A tide that lifts all boats
Here is the part that should be said plainly, because it is easy to read the above as "the personal agents are not our market." They are not our market. They are the best thing that could happen to it.
A year ago the hard part of agentic commerce was not the technology. It was that almost nobody had let software spend their money. Muse changes that at Meta's scale, for free, with a Shopify and PayPal checkout behind it. Instinct is doing it with a text thread and a reported $1,300 a month in spend per buyer. Every time one of those transactions completes, a person gets a little more comfortable with the idea that an agent can buy things for them and the right thing shows up.
That comfort does not stay inside the chat window. Once people trust AI to shop for them in a conversation, they will expect it everywhere: the bank app, the rewards portal, the gifting site, the search result, the recipe. Every one of those surfaces will get asked "why can't I just buy it here?" The answer will not be to bolt on a chat assistant. It will be to embed a checkout that works the way the surface already works, with the decision made up front and the order executed in the background.
The personal agents are training the demand. The surfaces that do not have a human in the loop will need the supply. That is the checkout CartAI was built for, and the tide coming in behind Muse, Instinct and dots is going to lift it.
If you run a platform that needs to place orders without a human babysitting each one, get in touch with us here.
Manil Uppal
Founder
