People have been asking AI chatbots for shopping advice for a while now.
Comparing two laptops, summarising reviews, or finding a cheaper alternative has become a normal way to use ChatGPT, Gemini, or Amazon's own AI tools.
Recently, the pitch has changed.
Amazon, Google, and OpenAI have all launched tools that promise to go further: not just help someone decide what to buy, but complete the purchase on their behalf.
The headlines call it "agentic commerce," and some of them make it sound like AI shopping agents are already doing the buying.
The reality is messier than that, and it splits into three questions that tend to get blurred together.
The first is what the technology is capable of, and agents that buy something on their own now exist where a year ago they did not.
The second is how much of it a normal shopper can actually use, and much of that is still US-only or halfway through a rollout.
The third is how many people are actually buying anything off the back of it, and that number is the smallest of the three.
Most coverage answers the first question and leaves the other two to be assumed.
The rest of this guide works through them in turn.
It covers what Amazon, Google, and OpenAI have actually built for shoppers, how many people are actually using them, and why OpenAI's version was pulled back within six months of launching.
It ends with the practical part: what is worth checking before letting an agent spend someone's money.
1. What an AI shopping agent actually is
An AI shopping agent is different from AI shopping help.
Instead of helping someone decide what to buy, it is built to complete the purchase itself.
That can mean adding the item to a cart and entering payment details.
Checkout might get confirmed with a person watching, or without anyone checking at all.
The difference between AI shopping help and an AI agent buying something
Most people are already using AI shopping help, even if they would not call it that.
Someone asks whether one air fryer is better than another, or whether a £70 jacket is worth it, and the AI talks it through with them.
It might summarise a few reviews, spot a cheaper listing somewhere else, or explain what the spec differences actually mean day to day.
Then the conversation with the chatbot ends.
The person still needs to open the retailer's website and click the buy button themselves. It is the same split that shows up when using AI to plan a holiday, where a chatbot can suggest the trip but cannot actually book it.
An AI agent removes that final click, or at least offers to make it for them.
The tool holds saved payment details and keeps watching in the background, waiting for a price to drop or a condition to be met.
Nobody has to remember to check back, because something else is doing the checking and is allowed to act on what it finds.
At that point the AI is not comparing products anymore. It is doing the shopping.
That is a genuine shift in what buying something online involves.
The person is trusting the agent with money, not just an opinion.
Most of what gets called "AI shopping" in 2026 is still the first kind, the research assistant that stops short of the checkout.
The second kind, an AI that actually finishes the purchase, is newer, more limited, and still being figured out by the companies building it.
Why "agentic commerce" is suddenly everywhere in 2026
The phrase "agentic commerce" describes AI systems that can take real actions on a person's behalf, shopping being one of the clearest examples.
It has become a buzzword because three of the biggest names in tech have all shipped their own version of it within roughly the same twelve month period. Amazon, Google, and OpenAI have each built something that fits the description.
That timing is not a coincidence.
Each company has its own reason to want people buying through its AI instead of a rival's.
Amazon wants to keep shoppers inside its ecosystem even when the product is not sold on Amazon.
Google wants its Search and Gemini tools to be where a purchase actually finishes, not just where it starts.
OpenAI wanted ChatGPT to become a place people buy from directly, not just research from.
The result is a wave of new features that all promise some version of the same thing: let AI handle more of the buying, not just the deciding.
Whether that promise holds up in practice is the honest question this guide tries to answer.
2. What AI can actually buy right now
The clearest way to understand where things stand is to look at what the three biggest players in the field have actually produced for users, rather than what they have talked about creating.
Amazon: Buy for Me and Alexa for Shopping
Amazon's Buy for Me feature began as a pilot with a subset of US customers from around April 2025.
It lets a shopper tap "Buy for Me" on products Amazon itself does not sell, shown inside the Amazon Shopping app from the brand's own website.
Amazon's agent completes the checkout on another company's website using the customer's saved payment and shipping details, so fulfilment, delivery, returns, and customer service happen with the other company and not Amazon.
On 13 May 2026, Amazon said on its own news site that it was folding that feature into something bigger.
It retired its standalone Rufus chatbot and replaced it with Alexa for Shopping, a single agentic assistant built into the main search bar.
Alexa for Shopping can auto-buy an item once it hits a target price, run scheduled repeat purchases such as a regular grocery order, and use Buy for Me to shop other retailers on the customer's behalf.
It is free to any signed-in US Amazon customer, with no Echo device or Prime membership required.
That US scoping matters, because Amazon's announcement covers American customers only and says nothing about when, or whether, the merged assistant reaches other marketplaces.
That is a genuine step up in shopping autonomy.
Auto-buying at a target price and running scheduled repeat purchases both mean the agent can complete a transaction without someone actively watching in the moment.
Google: Universal Cart and the Agent Payments Protocol
Google announced Universal Cart at its I/O developer conference on 19 May 2026, in a post on its own Google Shopping blog.
The simplest way to describe it is one shopping basket that works everywhere Google does.
Normally, buying three things from three different shops means three separate baskets, three checkouts, and three lots of card details and delivery addresses to fill in.
Universal Cart is Google's attempt to collapse all of that into one.
Someone searching Google for a desk finds one they like on a furniture retailer's site, and adds it to the cart.
A few days later, a YouTube review of a monitor stand catches their eye, and that goes into the same cart, even though it is sold by a completely different shop.
That evening, they ask Gemini to reorder the usual coffee and washing-up tablets, and those land in the cart too, this time from a supermarket.
Three shops, three completely different products, one basket.
At the end there is a single checkout through Google Pay, instead of three.
The cart is not just a list sitting there either. It keeps an eye on things while it waits:
- it watches prices and says when something in the basket gets cheaper
- it checks stock, so an item does not quietly sell out before checkout
- it flags when two products would not actually work together, like a stand that will not fit the monitor already in the basket
- it points out loyalty points, discounts, or card perks that could apply to the order
Underneath all of this sit two pieces of Google plumbing: its Universal Commerce Protocol, and the one that matters more here, its Agent Payments Protocol, usually shortened to AP2.
The name sounds technical, but the idea behind AP2 is simple. It is the rulebook that decides when an AI is allowed to spend someone's money.
The person sets those rules in advance.
They might say the agent can buy one specific model of headphones, from two named retailers only, and only if the price drops below £80.
If every one of those conditions is met, the agent can go ahead and buy. If any single one is not met, it cannot.
This is the part that usually gets lost in the coverage.
As things stood at the I/O 2026 announcement, Gemini does not go off and buy things by itself.
Checkout is quicker and more joined up through Google Pay, but there is no general setting that turns Gemini loose with a saved card.
What actually exists today is the rules-based version above, where a person decides the conditions first and the agent can only act inside them.
Google said in that May post that Universal Cart would roll out across Search and the Gemini app in the US over the summer, with YouTube and Gmail to follow.
That window has now passed without a wider availability update from Google, so the honest position outside the US is that Universal Cart has been announced rather than handed to everyone.
The same post puts the checkout technology behind it in Canada and Australia in the coming months, and the UK later than that.
OpenAI: what happened to Instant Checkout in ChatGPT
OpenAI launched Instant Checkout in ChatGPT on 29 September 2025, built on its own Agentic Commerce Protocol developed with Stripe.
It let US ChatGPT users buy directly from Etsy sellers without leaving the chat.
Shopify's roughly one million merchants, including brands like Glossier, Skims, Spanx, and Vuori, were added shortly after.
Around 200,000 Walmart products followed that.
This was the boldest version of agentic commerce that any of the three companies shipped: a full, in-chat checkout, not just a cart or a guardrailed spending cap.
It did not last in that form.
In March 2026, OpenAI said it would stop handling payment inside ChatGPT, a retreat CNBC reported on 24 March 2026.
Shoppers would instead be passed out to the retailer's own checkout page, through apps built with individual retailers that sit inside the chat.
The stated reason was weak performance.
Walmart's own EVP of AI Acceleration said items sold through in-chat Instant Checkout converted three times worse than the same items did when a shopper clicked through to Walmart's own website.
Etsy said it saw no large volume of sales from Instant Checkout, though it still valued ChatGPT as a channel for people discovering products, even if they did not buy there and then.
Three things were reported as the cause.
The first was choice. There simply was not that much available to buy.
The second was accuracy. OpenAI was copying product details from retailer websites instead of getting a live feed from them, so a price or a stock level could already be wrong by the time someone tried to buy.
The third was the checkout itself. Paying for items from several different shops in one go was unusual, and those items then turned up in separate deliveries.
That difference is worth sitting with.
OpenAI is the company that created the most autonomous version of AI shopping at real scale, and then scaled it back within six months, not because it was hacked or unsafe, but because people did not use it the way OpenAI hoped.
3. How many people actually let AI buy for them?
There are now real numbers behind this, and they are more lopsided than the headlines suggest.
Riskified, a fraud-prevention firm, surveyed 5,400 shoppers across nine markets in September and October 2025 and published the results on 20 October 2025.
It found that 73% of shoppers already use AI somewhere while shopping, whether that is asking for product ideas, reading a summary of reviews, or checking prices.
Only 13% have actually completed a purchase after being referred by an AI agent or assistant.
That is the real gap.
People are using AI heavily at the start of the shopping journey and rarely letting it finish the job.
The same study found that 70% of shoppers say they are at least somewhat comfortable with the idea of an AI agent making a purchase on their behalf.
That is a striking number next to the 13% who have actually bought something an AI pointed them to.
People are broadly open to the idea in theory.
Far fewer have taken the step to complete a purchase with one.
Why most people stop researching before they let AI pay
This gap makes sense once the two halves of shopping are separated out.
Researching a purchase is low-stakes: comparing prices, reading about a product, deciding what to buy.
Getting an answer wrong costs a few minutes.
Handing over a payment method and letting an AI complete a transaction is a different category of decision entirely, because a mistake there costs actual money.
It also lines up with what happened to OpenAI's Instant Checkout.
Even where the technology existed and was genuinely easy to use, most people either did not try it or tried it once and went back to buying from the retailer's own site.
Comfort with an idea and trust to actually hand over money are not the same thing, and the numbers above show exactly how far apart they still sit.
4. Why AI shopping agents keep stumbling in practice
The pattern is not that these tools do not work at all.
It is that they keep running into the same handful of problems once real shoppers use them.
Walmart's conversion problem
The clearest evidence comes from Walmart's own data on Instant Checkout.
Items sold in-chat converted three times worse than the same items did when a shopper clicked through to Walmart's website directly.
That is not a small gap.
It suggests something about buying inside a chat window itself made people less likely to follow through, even when the product and the price were the same.
Stale product data and fragmented checkout
Part of the problem traced back to how the product information reached ChatGPT in the first place.
OpenAI was scraping retailer websites rather than pulling live product feeds, which meant prices, stock levels, or product details could be out of date by the time someone tried to buy. It is the same underlying issue covered in where AI tools actually get their information from.
Fragmented checkout added to the friction.
Buying several items from different retailers in one chat conversation is more complicated than a single retailer's own basket and delivery options, and shoppers noticed the complication more than the convenience.
5. How to try an AI shopping agent safely, if at all
None of this makes AI shopping agents a bad idea.
They are just early, and early means the mistakes are still being found by the people using them.
The practical version of that caution is a spending limit and a short list of things the agent is allowed to buy.
Setting spending limits and guardrails
Guardrails are the bit of setup most people click straight past.
They are also the bit that decides how much a mistake can cost.
Google's AP2 model, for example, is built around letting someone name specific brands or products and set a spending cap in advance, so an agent can only act inside those limits.
Amazon's Alexa for Shopping works a similar way when auto-buying at a target price: the person sets the condition, and the agent only acts once it is met.
That is a long way from handing an agent a card and letting it decide.
What to check before letting AI complete a purchase
A quick check before the agent buys anything catches most of what has gone wrong with these tools so far.
Four things are worth confirming:
- the product is the right one, in the right size and colour
- the price shown is current, not a figure scraped days ago
- the delivery address and payment card are the right ones
- the retailer is one worth buying from on its own website
None of that takes more than a few seconds.
Limitations
AI agents cannot judge style, quality, or whether something is actually a good deal
An AI shopping agent can compare listed prices and match a description to a request.
It can still get details wrong, and it cannot try on a pair of shoes, judge whether a fabric feels cheap, or know that a "deal" is only a deal because the retailer inflated the price beforehand.
Those judgments still need a person, and they are exactly the kind of thing that gets skipped when a purchase happens automatically in the background.
Handing over a payment method is a trust decision, not a convenience toggle
Letting an AI complete a purchase means giving it access to saved payment and shipping details. That sits apart from the general rule about what not to share with AI, because a retailer using a card already on file inside a preset cap is a different arrangement from typing card details into a chat box.
That should be treated as a real decision, not a setting to switch on because it sounds convenient.
The gap between the 70% who say they are comfortable with the idea and the 13% who have actually bought something an AI pointed them to suggests most people already sense this, even if they have not put it into words.
Returns, mistakes, and who is responsible when an agent buys the wrong thing
If an AI agent buys the wrong size, the wrong colour, or the wrong item entirely, someone still has to sort out the return.
With Amazon's Buy for Me, for example, that responsibility sits with the brand being bought from rather than Amazon itself.
It is worth knowing whose returns process applies, and how straightforward it actually is, before the wrong parcel turns up.
Final thoughts
The honest answer to "can AI buy things for me right now" is yes, but only just, and only within limits somebody has to set first.
Amazon, Google, and OpenAI have each shipped a version of this in the twelve months to mid-2026.
Not one of them is a hands-off system that people are genuinely using to do their shopping.
The clearest evidence for that caution is OpenAI's own experience.
It built the most ambitious version, a full in-chat checkout, and pulled it back within six months once the data showed people converting three times worse than on a retailer's own site.
That is not a company being cautious out of principle. That is a company responding to shoppers voting with their behaviour.
The comfort is real.
70% of people are at least open to the idea of an AI agent buying on their behalf.
But comfort with an idea and acting on it are different things, and only 13% have so far bought anything an AI pointed them to, let alone let one complete the purchase itself.
For anyone curious to try one of these tools, the guardrail features, spending caps, named products, target prices, are the sensible starting point, not full autonomy.
Set the limits, check what the agent picked before it becomes a habit, and treat the whole category as still early rather than settled.
AI shopping agents are worth watching. They are not yet worth handing the credit card to and walking away.
Frequently asked questions
Is it safe to let AI buy things for me?
It comes down to the limits set beforehand more than the tool itself. Google's Agent Payments Protocol and Amazon's target-price auto-buy both let someone name specific products and a spending cap before anything gets bought, so the agent can only act inside those limits. Giving an agent open access to a saved card is a different proposition entirely, and even with limits in place the orders are still worth checking afterwards.
Which AI shopping agents can be used today?
Amazon's Alexa for Shopping, launched 13 May 2026, is available to any signed-in US Amazon customer and can auto-buy at a target price, run scheduled repeat purchases, and use Buy for Me to shop other retailers. Google's Universal Cart, unveiled 19 May 2026, was announced for Search and the Gemini app in the US over the summer, with no wider availability confirmed since. ChatGPT no longer hosts full in-chat checkout the way it briefly did, and now points shoppers toward dedicated retailer apps that hand off to the retailer's own checkout.
Does ChatGPT let me buy things directly?
Not in the way it did when Instant Checkout first launched. OpenAI scaled that feature back in March 2026 after retailers reported weak results, Walmart saw conversion three times worse than its own site, and Etsy saw no large volume of sales. ChatGPT now works with retailers on in-chat apps that hand shoppers off to the retailer's own checkout rather than completing the purchase itself.
Can any of this be used outside the US?
Mostly not yet. Amazon's Alexa for Shopping launch covers US customers, and OpenAI's Instant Checkout was US-only before it was scaled back in March 2026. Google's own announcement is the only one with a stated international path, putting the checkout technology behind Universal Cart in Canada and Australia in the coming months and the UK later than that, without naming a date. Shoppers elsewhere can still use AI to research and compare, which is the part that works in any country.
What happens if an AI agent buys the wrong item?
Whoever sold the item, not Amazon, in the case of Buy for Me. Fulfilment, delivery, and customer service responsibility sits with the brand selling the product rather than the platform that arranged the purchase. Before relying on an agent to buy something, it is worth knowing which company's returns process would apply if the order turns out to be wrong.
Why are so few people actually completing purchases through AI agents?
The Riskified study found that while 73% of shoppers use AI somewhere in their shopping journey, only 13% have completed a purchase after being referred by an AI agent, even though 70% say they are at least somewhat comfortable with the idea. Researching with AI is low-risk. Letting it complete a payment is a bigger step, and most people are still choosing to finish that part themselves.
Should spending limits be set before trying an AI shopping agent?
Yes, and setup is the easiest time to do it. A first cap does not need to be clever, a low ceiling and one or two named products will do, just enough to see how the agent behaves before anything expensive goes through it. Without one, there is nothing between a wrong product match and a completed order.
