The task sounded simple: “Find me a replacement filter for my kitchen water jug, the cheapest reliable option, and put it in my basket.” I typed it into an AI assistant that could browse websites on my behalf and watched it work. It opened a shopping site, searched, compared a few listings, read some reviews, and added an item to my basket. The whole thing took about two minutes. Then I noticed it had chosen a pack of six filters instead of the single one I needed.
That moment summed up my week-long experiment with letting an AI agent handle my everyday online shopping. These agents, which can click, scroll, compare, and fill in forms on websites, are being promoted as the next big step for AI assistants. I wanted to know what it is actually like to hand over a boring chore to one, and whether I would trust it with more. The answer turned out to be more nuanced than either the hype or the skepticism suggests.
The rules I set for myself
Before starting, I set a few ground rules. The agent could search, compare, and add things to my basket, but it could never complete a purchase. I would always review the basket and pay myself. I also kept it away from any saved payment details and did not let it log in to accounts with stored cards. This was partly about safety and partly about curiosity: I wanted to see how good its choices were before trusting it with my money.
For the week, I gave it every small online shopping task that came up: household supplies, a birthday gift, a replacement phone cable, groceries for a dinner, and a few comparison questions like “which of these two vacuum cleaners is better for a small flat?” I kept notes on how long each task took, how good the result was, and how much I needed to correct.
Where it was genuinely helpful
The agent shone at tedious comparison work. When I asked it to find a USB-C cable of a specific length, with a braided cover and fast-charging support, under a certain price, it checked several listings and summarized the differences clearly. That is exactly the kind of fiddly task I usually avoid because it involves opening a dozen tabs and squinting at specifications.
It was also good at reading reviews at scale. For the vacuum cleaner question, it summarized common complaints from dozens of reviews: one model was praised for suction but criticized for a short battery life, the other was quieter but had a small dust container. That summary saved me at least half an hour of reading, and it matched what I found when I spot-checked the reviews myself.
For repeat purchases, like the groceries for a dinner I make regularly, it was efficient. I gave it a list, and it found everything and built a basket. The only thing I had to adjust was a brand preference for olive oil, which it had no way of knowing.
Where it made strange choices
The six-pack of filters was the first of several odd decisions. When I asked for the “cheapest” option, it sometimes interpreted that as the lowest price per unit rather than the lowest total price. Technically defensible, practically annoying. I did not want six filters.
For the birthday gift, it struggled more. I described my friend’s interests and a budget, and the agent suggested a reasonable but generic set of options: a candle, a book about gardening, a fancy mug. None were wrong, but none felt personal. Gift-giving relies on knowing someone, and an agent working from a two-sentence description simply does not have that knowledge. In the end, I chose something myself.
It also occasionally got distracted by sponsored listings. Shopping sites often place paid products at the top of search results, and the agent sometimes treated them as the most relevant options. When I asked it to ignore sponsored results, it did, but I had to know to ask.
Twice, it misread product details. Once it added a phone case for the wrong model, because the listing title was confusing. Another time, it picked a lamp that turned out to need a bulb not included in the box. These are mistakes a careful human might also make, but they show why reviewing the basket was essential.
Writing good instructions turned out to be the real skill
By the middle of the week, I realized that most of the agent’s mistakes came from my instructions, not its abilities. “Cheapest filter” was ambiguous. “A gift for a friend who likes gardening” was vague. When I started writing more precise requests, the results improved dramatically.
A good request, I learned, includes the exact quantity, the total budget rather than a vague idea of cheapness, any brands or features to avoid, and what to do when there is no perfect match. For example: “Find one replacement filter compatible with this jug model, total under a set price, from a seller with good reviews, no multipacks, and no sponsored listings. If nothing matches, tell me instead of choosing something else.” That last part was especially important. Without it, the agent tended to pick the closest option rather than admit failure.
Writing instructions like this took a little effort, but it is effort I only had to make once for recurring tasks. I saved my best instructions as templates and reused them. In a way, the experiment taught me less about AI and more about how unclear my own everyday requests usually are.
It was slower than I expected
One surprising thing was speed. I had imagined an AI agent would be instant. In reality, watching it work was a bit like watching a careful but slightly hesitant person use a website for the first time. It scrolled, paused, clicked, went back, and tried again. Simple tasks took a couple of minutes; more complex comparisons took ten or more.
For tasks I could do myself in thirty seconds, like reordering something I buy regularly, the agent was actually slower than doing it by hand. Its real value was in tasks that would take me a long time or that I found boring. Once I understood that, I stopped giving it trivial jobs and focused on the tedious ones.
The trust question
Throughout the week, I kept thinking about trust. Allowing software to act on websites on my behalf is very different from asking it a question. An agent that can click buttons could, in principle, click the wrong ones. It could be confused by a misleading page, or even manipulated by hidden instructions planted on a website, which is a real concern security researchers have raised.
Keeping purchases in my own hands made me comfortable. I treated the agent like a helpful assistant who prepares things for me to approve, not like someone with access to my wallet. I think that is the right model for now. The convenience of fully automatic purchasing would be nice, but the cost of a mistake, or of a compromised agent, is too high to ignore.
I also noticed that I had to be careful about what personal information I shared. When filling in a delivery address for a gift, I did it myself rather than letting the agent type it. It is a small thing, but it felt better to keep personal details out of the loop whenever possible.
How it changed the way I shop
An unexpected effect of the experiment was that I bought less. Because the agent prepared baskets but did not complete them, I had a natural pause before every purchase. Looking at a basket someone else had built, I often asked myself whether I really needed the item at all. A couple of times, the answer was no, and I simply emptied the basket.
The agent also made me more aware of how shopping websites are designed. Watching it navigate pages full of banners, pop-ups, suggested items, and sponsored listings made me realize how much of that clutter I normally ignore, and how much of it is designed to steer my choices. Seeing the site through the eyes of a slightly confused robot was oddly illuminating.
The bigger picture for shops and shoppers
Thinking beyond my own experiment, it is interesting to imagine what happens if many people start shopping through AI agents. Online shops are designed for human eyes: attractive images, persuasive descriptions, prominent deals. An agent does not care about a pretty photo. It reads specifications, prices, and reviews. If agents become common, shops may need to compete more on clear information and genuine quality, which could be good for shoppers.
There are also risks. Sellers might try to optimize listings specifically to attract agents, the way websites learned to optimize for search engines, and not always honestly. Fake reviews could become even more damaging if agents summarize them without noticing they are fake. As with search engines before them, a new kind of gatekeeper creates new incentives to game the system.
What I will keep using it for
After a week, I have a clear sense of where an AI shopping agent fits in my life. I will keep using it for comparing technical products, summarizing reviews, finding specific items that match detailed requirements, and building baskets for regular grocery orders. These are tasks where it saves real time and where its mistakes are easy to catch.
I will not use it for gifts, for anything where personal taste matters a lot, or for completing purchases on my behalf. Those decisions still benefit from a human who knows the context, and the risks of automating payment are not worth the small convenience.
A useful assistant, not a replacement
The best way I can describe the experience is that the AI agent behaved like an eager new assistant on their first week. Capable, fast at reading, good at following clear instructions, but lacking context, occasionally literal-minded, and in need of supervision. With clear instructions and a final check, it was genuinely helpful. Without them, it made the kind of small mistakes that could quietly cost money.
The filters, by the way, were fine. I returned the six-pack and ordered a single one myself. It took me forty seconds. Sometimes the fastest assistant is still the one who already knows exactly what you need.
