AI shopping agents: 61 real purchases, from research to refunds
We read 61 real purchases (138 counting looser matches), and the pattern is a funnel with the risk in the middle. Agents are strongest before the purchase and after it, and weakest at the cart. All four clear failure stories happen at or around the point of buying: a wrong item, a misread list, a payment check the agent couldn’t pass, and a seller it couldn’t contact. None happen during research or refunds. So let the agent research, watch prices and chase refunds alone. At the cart, specify the item tightly, approve the total and pay with something limited.
Muse appears in 33 of the 61 core rows and Instinct in 21. Buying, checkout and orders account for 49 of them, so most people are already letting agents complete purchases, not just suggest them.
Research and compare: the spec does the work
Research is the safest place to start, because the agent only reads and recommends. The good results come from a checkable spec. One user asked Muse for the cheapest noise-cancelling headphones under $150 with good mic reviews and got a ranked list with links. They bought the winner in about five minutes. That works because a budget and one hard requirement make the ranking easy to verify. An Instinct user set a standing rule to penalize the top five sponsored items on every purchase. That works because a persistent preference beats ad-ranked results every time.
Research also covers jobs people would never do by hand. Muse called five Tesla locations at once to find a Model YL ready for a test drive. Instinct compared three mice and found 10-minute delivery on Blinkit when Amazon couldn’t deliver the same day.
A clear spec also makes price-watching possible, which is the next step.
Deals and price tracking: patience as a feature
Deal-hunting is still low-risk, since nothing is bought until you say so. Muse appears in most of the deal stories we read. One user ran Grok Bot and Muse on the same iPhone 14 trade-in search, and Muse found one for $230 against Grok Bot’s $245. That works because the same task on two tools gives a real comparison. Another had Muse browse iPhone 17e carrier promos while they worked, returning a chat widget of every promo and a recommendation.
Secondhand markets follow the same rule. A parent had Muse hunt Facebook Marketplace for winter baby gear, and “all I did was approve the messages to sellers.” That works because the one step that speaks for the user is the one that’s gated. Another user’s agent couldn’t message Marketplace sellers yet, which is the funnel’s middle showing up early. Restock alerts get no section, because only one row in the whole dataset mentions “back in stock”.
A found deal still has to get through the cart.
Grocery and reorder runs: ambiguity is the risk
Repeat buying moves the agent into the risky middle on a schedule. The failures here come from ambiguity, not effort. One user photographed a messy whiteboard shopping list and had Muse order it from Sam’s Club. The first picks weren’t what they meant. It failed at first because the list was written for one person, and one round of feedback fixed it. Others had Muse order Instacart groceries without opening the app and keep a master file of skincare and household items stocked.
Planning the week’s meals belongs in the home management guide. Every run here ends at payment.
Checkout and payment: limit the blast radius
Checkout is where the thesis is clearest. Eight stories describe how payment was handled, and 7 of them succeeded. Each of those seven used a limited instrument or a confirmed total. The eighth failed on a payment check.
Grok Bot ordered moving boxes from Home Depot, caught a wrong address and paid with a one-time virtual card the user approved, so it never saw the real card. That works because the card caps what can go wrong. An Instinct user bought a three-machine haircut set in three separate orders for $385 instead of $654. The agent sent the checkout total and the user confirmed it. That works because the user approves one number, not every step. Muse users paid through Link, ordering an F-150 part with a virtual card and setting up Link to buy RAM. An Instinct user chose cash on delivery for a Blinkit order so they wouldn’t have to share card details.
The failures show what’s missing. An Instinct agent looped on a Shop Pay verification check while buying a bike helmet. It failed because it didn’t know how to ask the user for the code, and the user notes it also stored a CVV. Another Instinct user reports simply that it bought the wrong product from the drug store. Explicit approval appears in only a handful of the 52 shopping stories we read, so the limit on payment, not a step-by-step sign-off, is doing most of the protective work.
After the order, the agent’s patience becomes useful again.
Returns, refunds and order tracking: where agents win
Refunds are the strongest part of the funnel. Every one of the 8 return or refund stories we read reports success. Muse opened an Amazon support request and won the refund, getting through to a human. That works because support queues reward persistence. Instinct chatted with IKEA about broken bowls and nudged the user when the human agent joined, saving about 30 minutes. Another Instinct user got a Blinkit refund without lifting a finger. Muse found a book subscription the user’s wife forgot to cancel and recovered a year of payments, and other Muse users had it fix double-charged concert tickets in about 30 minutes and a double-charged bag fee.
Order tracking is thin but real. Muse bought a garage freezer, tracked the delivery and submitted the registration.
How to set it up
Move through the funnel one stage at a time.
- Research only: ask for a ranked shortlist with links and buy it yourself.
- Let the agent buy, but have it send the final total first.
- Pay only with a one-time or virtual card, or cash on delivery.
- Send any verification code request back to you.
- Let it handle refunds and return deadlines on its own.
Find me a cordless stick vacuum under $250 that weighs under 3 kg, has a removable battery, and has at least 4.3 stars with 500+ reviews. Check Amazon, Best Buy, Target, and Walmart and skip sponsored listings. Give me a ranked top 3 with price, shipping date, return window, and a link. Wait for me to pick one. Once I do, add it to the cart, apply any coupon you can find, and send me the final total with tax and shipping before you pay. Pay only with the one-time virtual card I approve, and ship it to my saved home address. If anything at checkout asks for a code or verification, stop and ask me. Don't retry. Save the order number and return deadline, and remind me 3 days before the return window closes.Where to start
The funnel tells you where to trust the agent. If you’re cautious, start at either end: research, deal-watching and refunds are where the stories succeed and where a mistake costs only time. When you let it buy, protect the middle with two controls: confirm one final total and pay with something limited, such as a one-time card, a payment link or cash on delivery. Make sure verification codes come back to you rather than stalling the agent. Write the spec like a return reason. If you couldn’t return the item for being “not what I asked for,” the agent can’t reliably buy it either. Refunds that start in your inbox are covered in the email guide.