The $87 decision
I can already see the situation.
It’s 10:47 p.m. You need a new pair of running shoes. Your current ones are worn through, tomorrow’s a workday, and you absolutely do not want to spend another hour opening twelve browser tabs to compare cushioning, price, reviews, return policies, and whether the shoes actually come in your size.
So you tell an AI assistant:
“I need comfortable running shoes under $100. I have wide feet, I run three times a week, and I don't care about the brand. Find me the best option and buy them if they're under $90.”
That last part changes everything.
For years, AI shopping meant help me decide.
Now we're moving toward help me decide, then actually do it.
That's the interesting part of AI shopping agents. They aren't just glorified search boxes. The goal is for an AI agent to understand what you want, compare products, check prices and availability, put something in a cart, and eventually complete the transaction with much less human involvement.
Some of this is already happening.
OpenAI introduced Instant Checkout in ChatGPT in September 2025, allowing eligible U.S. users to purchase certain products directly inside ChatGPT rather than being sent to a retailer's website. Google has been building agentic shopping around Gemini, its Universal Cart and Universal Commerce Protocol. Amazon has been developing its own AI shopping experience, now called Alexa for Shopping, with features that can automate parts of product discovery, deal finding, cart building and recurring purchases.
But we're not quite at the point where I would hand an AI unrestricted access to my credit card and say, “Buy whatever you think I need.”
Not even close.
And that's actually what makes this technology interesting.
Shopping is changing
The biggest mistake people make with AI shopping is thinking the technology is mainly about finding cheaper products.
That's only one piece of it.
The bigger change is that the interface for shopping could move from websites to conversations with agents.
Think about how shopping normally works.
You search for a product.
You open a retailer.
You apply filters.
You read reviews.
You compare five nearly identical products.
You check shipping.
You look for a coupon.
You wonder whether the cheaper seller is legitimate.
You abandon the purchase.
Then three days later you remember you still need the thing.
An AI agent can potentially handle much of that annoying middle section.
OpenAI's shopping research experience, for example, can take a natural-language request, ask follow-up questions about budget and preferences, research products across the web, compare specifications and tradeoffs, and create a personalized buying guide. OpenAI explicitly says the system can still make mistakes with details such as price and availability, which is a pretty important warning when you're talking about purchases.
Google is taking the concept further. Its Universal Cart is designed to work across retailers and services, while Google's Universal Commerce Protocol gives agents a standardized way to interact with retailer catalogs, including product information, pricing and inventory.
That's the infrastructure underneath the experience.
The consumer-facing version is much simpler:
“Find it.”
“Compare it.”
“Buy it.”
Eventually:
“Keep buying it.”
That last one is where things get really interesting.
What an agent actually does
Let's say you tell an AI:
“I'm almost out of coffee. Find my usual kind, make sure it's cheaper than $20, and order two bags.”
A basic chatbot might tell you where to buy it.
A shopping agent is supposed to do more.
It might identify the product from your previous purchases or your description. It could check several retailers. It could compare the current price against previous prices. It could look at shipping costs. It might notice that one retailer has a two-pack for less than buying two individual bags elsewhere.
Then it could ask:
“Do you want me to place the order?”
That's still human oversight.
A more advanced agent might eventually be allowed to complete the transaction automatically if the purchase falls inside rules you've already established.
For example:
* Household supplies under $50: automatic.
* Groceries under $100: ask before checkout.
* Electronics: never buy without approval.
* Recurring purchases: automatic unless the price rises more than 10%.
* Anything over $200: always ask me.
That's much closer to having a digital shopping assistant than having a chatbot.
Amazon has already been experimenting with this kind of automation. Its shopping AI can help find products, add items to carts, identify deals and automate certain routine purchases. Amazon says its current shopping experience combines these capabilities under Alexa for Shopping.
And that distinction matters.
AI recommendation says: “Here are some options.”
AI agent says: “I've handled it.”
The second one is dramatically more useful.
It's also dramatically more dangerous when it gets something wrong.
The part I wouldn't automate
Here's where I disagree with some of the breathless predictions about AI shopping.
I don't think most people want an AI making every purchasing decision.
I certainly wouldn't want an agent deciding that a $700 laptop is “close enough” to the $500 limit I gave it.
There are purchases where specifications matter in ways that aren't obvious from a product page.
Take a mattress.
An AI can compare firmness ratings, dimensions, materials, reviews and prices. Great.
But if I sleep differently from the average customer, none of those numbers completely answer the question.
Same thing with clothing.
An AI might know my size, preferred colors and budget. It can narrow down the choices. But it doesn't know whether a particular shirt feels cheap when you actually touch it.
And then there are the products where a tiny difference matters.
A phone case that doesn't fit the exact model.
A replacement filter with the wrong dimensions.
A charger with incompatible specifications.
A pair of shoes that technically matches your size but doesn't fit your foot properly.
An agent can be incredibly good at reducing the number of decisions you have to make.
That doesn't mean it understands every decision as well as you do.
There's a difference.
Where AI is already useful
The sweet spot right now is research-heavy shopping.
This is the part I think AI handles surprisingly well.
Imagine you're buying a $1,200 laptop.
You could spend Saturday reading reviews from ten websites, watching comparison videos, checking processor specifications, looking at battery tests and trying to understand why two models that look almost identical have a $300 price difference.
Or you can give an AI your actual requirements.
“I need a Windows laptop for photo editing and everyday work. Minimum 16GB RAM, good screen, at least eight hours of realistic battery life, under $1,200. I don't care about gaming.”
That's a much better shopping query than:
“Best laptop 2026.”
The AI has constraints to work with.
OpenAI's shopping research is specifically designed around this type of task. It can ask clarifying questions and research product information based on constraints such as price, features, category and personal preferences.
That's where I would use AI first.
Not to blindly press Buy.
To eliminate the junk.
If an AI can turn 300 possibilities into four genuinely relevant choices, that's already a huge improvement.
My first mistake
The easiest way to get burned by AI shopping is to give it a vague instruction and assume it understands what you meant.
“Find me a good one under $100.”
That's not enough.
Good according to whom?
Good for what?
New or refurbished?
Lowest price or best value?
Do I care about shipping?
What about warranty?
Does the $100 include tax?
Is $101 acceptable if the product is substantially better?
These aren't tiny details. They're the actual shopping decision.
I've seen enough AI product recommendations to know that an apparently sensible recommendation can fall apart once you inspect the details.
That's why I'd change the way I give shopping instructions.
Instead of:
“Find me a good coffee maker.”
I'd say:
“I want a coffee maker under $150. I make two to four cups each morning. I don't want a pod machine. It needs to be easy to clean, available from a reputable U.S. retailer, and I care more about reliability than having lots of smart features. Show me three options and explain the biggest downside of each before recommending one.”
That's a completely different task.
The AI has boundaries.
It knows what “good” means in this particular situation.
And most importantly, I haven't told it to purchase anything yet.
That's the part I'd keep.
The checkout problem
Getting AI to recommend a product is relatively easy.
Getting an AI to safely purchase that product is much harder.
A purchase involves identity, payment information, shipping addresses, refunds, returns, taxes, subscriptions and fraud protection.
There are also platform rules.
And those rules are already becoming a battleground.
Just look at what's happening right now.
In September 2026, Amazon blocked Meta's Muse AI shopping agent from making purchases on Amazon, citing its conditions of use. Reporting around the dispute has focused on issues including unauthorized automated access, privacy and how third-party agents interact with Amazon's platform.
At the same time, Shopify has moved in the opposite direction by working with Meta on agentic checkout through Shop Pay.
That's a fascinating split.
One company says, essentially, “Don't let outside agents control this shopping experience.”
Another says, “Let's build infrastructure that lets agents participate.”
So the future of AI shopping isn't just a technology problem.
It's a platform problem.
The agent has to be allowed to act.
The retailer has to trust it.
The payment provider has to authenticate it.
And the customer has to understand what happened.
Who gets paid?
Here's another question people don't talk about enough.
If an AI chooses the product for you, who influences that choice?
Imagine you say:
“Find me the best air purifier under $250.”
An AI gives you three choices.
You assume those choices were selected because they're objectively the best match for your requirements.
But what happens if retailers start paying for placement?
What happens if brands pay for preferential treatment?
What happens if the shopping agent has commercial relationships you don't fully understand?
This isn't hypothetical as a general concern. The financial sector is already raising questions around AI shopping agents, including transparency, privacy, payment security and whether agents will reliably act in consumers' interests.
That's why transparency matters so much.
I'd want to know whether the AI's recommendations are based on my requirements, advertising, commercial partnerships, or some combination of those things.
And I'd want that information before spending money.
OpenAI says its product results are designed to be organic and unsponsored and that merchants cannot pay to influence those results, while Instant Checkout merchants pay a fee on completed purchases.
Those policies are useful.
But consumers still need to pay attention as this market evolves.
The incentives around shopping are enormous.
Billions of dollars are spent trying to influence what people buy.
If AI becomes the person—or rather, the agent—standing between shoppers and retailers, every company selling something will want access to that agent.
The privacy tradeoff
There's another uncomfortable tradeoff.
The better your shopping agent knows you, the more useful it can become.
But that also means it needs information about you.
Your clothing sizes.
Your preferred brands.
Your household.
Your budget.
Your purchase history.
Your allergies or dietary preferences, if you're shopping for food.
Your address.
Potentially even information about other people in your household.
That creates a weird situation.
The most helpful shopping assistant would know an enormous amount about what you buy.
The safest shopping assistant might know much less.
Those goals don't naturally point in the same direction.
That's one reason I wouldn't give a new AI shopping service unrestricted access to everything on day one.
I'd start with low-risk purchases.
Toilet paper.
Cleaning supplies.
Office supplies.
A replacement cable.
Something boring.
Let the system prove it can follow instructions before giving it permission to spend hundreds of dollars.
And even then, I'd use purchase limits.
The $500 experiment
Here's how I'd actually test an AI shopping agent.
Start with something you already understand.
Say you're looking for a $500 monitor.
Don't tell the agent:
“Buy me a monitor.”
Tell it:
“Find three monitors under $500. I want a 27-inch or larger display, USB-C connectivity, good text clarity and a stand with height adjustment. Check current prices and tell me which requirements each model satisfies. Don't purchase anything.”
Then inspect the results yourself.
Look at the retailer.
Check the exact model number.
Check shipping.
Check the return policy.
Check whether the price is actually current.
Then ask the AI:
“Now compare these two specifically. What am I giving up by choosing the cheaper one?”
That's where the agent earns its keep.
Only after you've verified its behavior would I consider giving it purchasing authority.
And even then:
“Purchase only if the final total is below $500 including shipping and taxes. Ask me before checkout.”
That's a much more sensible first step than handing an AI unrestricted buying authority.
What happens to Amazon?
Amazon is probably one of the clearest examples of why this transition won't be simple.
Amazon already has enormous amounts of shopping data, a huge product catalog and an established payment and fulfillment system.
It doesn't need another company to tell it how to build a shopping assistant.
That's why Amazon has been investing in its own AI shopping experience. Its Alexa for Shopping combines conversational product discovery with features such as shopping guides, price history, deal finding, cart building and automated routine purchases.
But third-party agents create a different problem.
If you shop through an independent AI, the retailer potentially loses part of the customer relationship.
The shopper isn't necessarily thinking:
“I need to go to Amazon.”
They're thinking:
“I need dishwasher detergent.”
The AI decides where to buy it.
That could fundamentally change how online retail works.
Today, retailers fight to become the place where you search.
Tomorrow, they may be fighting to become the place your AI chooses.
That's a very different business model.
Google wants the middle
Google has an obvious interest in this transition too.
For decades, Google has been the place people go before they shop.
Search something.
Click a result.
Visit a store.
Buy something.
Agentic shopping could compress that entire process.
Google's Universal Commerce Protocol is designed to let AI agents communicate with retailers and handle commerce interactions more consistently. Google has also introduced Universal Cart, intended to bring products from different retailers and services into a more unified shopping experience.
That's potentially a major change.
Instead of me visiting five websites, I could have an agent assemble the options in one place.
But again, the important question isn't simply whether it works.
It's who controls the experience.
If Google, Amazon, OpenAI, Meta and retailers all build competing shopping agents, consumers could end up with several different versions of the same idea.
And retailers will have to decide which agents they're willing to work with.
The current Amazon-Meta dispute shows that compatibility isn't guaranteed.
What about scams?
This is probably the question I'd ask before giving any AI permission to buy things.
What happens if the agent picks the wrong seller?
Or encounters a fake website?
Or misreads a product listing?
Or gets tricked by a malicious instruction embedded somewhere on a webpage?
Or buys a counterfeit product because it looked like the legitimate listing?
The more autonomy an AI gets, the more opportunities there are for something to go wrong.
Banks and financial institutions have already raised concerns about AI agents handling shopping transactions, including fraud, scams, sensitive financial information and unclear consumer protections.
So I wouldn't think of “AI can buy things for me” as a single feature.
It's really a chain of permissions.
Can it search?
Can it compare?
Can it add to cart?
Can it enter my shipping information?
Can it use my payment method?
Can it actually place the order?
Can it reorder automatically?
Each step deserves its own level of trust.
Where this is heading
The most interesting future isn't an AI that randomly buys everything for you.
It's an AI that understands your shopping rules.
Imagine telling it:
“I need groceries for five dinners this week. Keep the total under $130. Prioritize ingredients I already have. Don't buy anything with peanuts. If an item is unavailable, suggest a substitute but don't purchase the substitute without asking.”
That's genuinely useful.
Or:
“Watch the price of this laptop. If it drops below $900 from a reputable retailer with free returns, tell me.”
That's useful too.
Then maybe:
“If it drops below $850, buy it.”
Now the AI is acting more like a personal purchasing system.
It doesn't need to make every decision.
It needs to understand which decisions you've already made.
That's the real promise of agentic commerce.
The human stays involved
I'm not convinced the winning shopping experience will be completely automated.
I think it'll be selectively automated.
You let AI handle repetitive decisions because you don't care enough to spend time on them.
You keep control over decisions that matter.
I don't need to personally compare twelve brands of paper towels.
I do want to approve a $1,400 television.
That's not a contradiction.
It's exactly how a good assistant should work.
The goal isn't to remove humans from shopping.
It's to remove the pointless parts.
And there's a huge difference.
What I'd do today
If you're curious about AI shopping agents, don't start by giving one permission to spend $1,000.
Start with research.
Give it a real shopping problem you've been putting off. Tell it your budget, requirements, deal-breakers and priorities. Let it compare products. Then check the actual retailer pages yourself.
ChatGPT's shopping research is already designed for this type of product comparison, while Google's shopping tools and Amazon's Alexa for Shopping are pushing toward increasingly agentic experiences.
Once you're comfortable with the recommendations, try a low-cost purchase where the consequences of a mistake are minor.
Set a hard spending limit.
Require approval before checkout.
And don't assume the AI knows what you mean just because the sentence sounded clear in your head.
Give it rules.
That's the biggest lesson I'd take from this whole shift.
AI shopping agents aren't really about making shopping disappear. They're about turning shopping from a series of clicks into a conversation with something that can eventually take action.
That could save an absurd amount of time.
It could also create a whole new category of mistakes if we give these systems too much authority too quickly.
So yes, I think AI will buy things for people.
Some of it is already happening.
But the version I'd actually want isn't an AI that says, “Don't worry, I've got it.”
It's an AI that says, “I found exactly what you asked for. Here's what it'll cost. Here's what could go wrong. Want me to place the order?”
That's the point where I'd start trusting it.
And honestly, that's probably the version most of us will end up using anyway.