My AI Wants to Go Shopping. Retailers Would Like to Know Whose Side It’s On.

I used to think the most complicated part of shopping was figuring out whether I actually needed something. Apparently, that was an embarrassingly primitive concern. Now I am supposed to consider whether an artificial intelligence agent should evaluate my needs, compare the options, negotiate the digital obstacle course, and spend my money while retailers argue over whether to let it through the door. We have finally reached the stage of civilization where buying socks requires a debate about sovereignty.

The headline that set me off was The Wall Street Journal’s “Retailers From QVC to Kohl’s Are Split on Agentic Shopping.” According to its October 2 reporting, QVC and HSN welcome purchasing agents, Kohl’s prevents agentic purchases, and Tapestry allows browsing but not buying. Retailers’ concerns include customer experience and control over data. Those are the reported positions; what interests me is the argument underneath them. Everyone wants to help me shop, but I suspect we have different definitions of “help.” [1]

Before anyone sends me a presentation featuring a glowing shopping cart, let me translate the terminology. Agentic shopping means software can act on a shopper’s instructions, potentially moving beyond recommendations to taking purchasing steps. The exact permissions and capabilities vary. Asking for three decent blenders is one thing. Authorizing a system to choose one and place the order is another. Somewhere between those two requests, a helpful assistant acquires access to consequences.

I understand the appeal immediately. Give me something that can compare unit prices, notice shipping charges, reject suspicious sellers, and remember which replacement filter fits the appliance I already own. That sounds useful. I do not need a digital visionary to reinvent consumption. I need an assistant that understands a bargain stops being a bargain when delivery costs more than lunch and the return policy reads like a hostage negotiation.

But the sales pitch tends to rush past those small, practical improvements toward a much grander promise: soon I will barely need to participate. My preferences will become instructions, my instructions will become transactions, and my doorstep will become the receiving department for decisions made elsewhere. This is described as liberation. I can also imagine it becoming a very efficient way to discover that my software and I have irreconcilable differences about throw pillows.

What fascinates me about the QVC side of this debate is the apparent contradiction. A business associated with demonstrations, personalities, enthusiasm, and the emotional momentum of watching something being sold is open to a buyer that might experience none of those things. I find that commercially sensible, at least in principle. If a customer wants to arrive through a different entrance, there is a reasonable argument for making sure the entrance does not lead into a brick wall.

Still, I enjoy imagining a shopping agent encountering a beautifully staged product demonstration. The host is conveying possibility, comfort, perhaps a slightly better version of domestic life. The agent is looking for dimensions, delivery costs, and whether the lid goes in the dishwasher. One side is selling the feeling of finally getting organized. The other wants to know whether the storage container actually fits the shelf. I would pay a modest admission fee to watch that conversation.

Kohl’s taking a more restrictive position does not automatically strike me as foolish. I can want better shopping tools without assuming every company should grant every outside system purchasing access immediately. A retailer has to think about mistakes, identity, service, and what happens after a transaction. Letting software shop sounds wonderfully straightforward until the software orders the wrong variation and three different companies start explaining why the misunderstanding belongs to somebody else.

The distinction matters because technology debates love assigning costumes. Early adopters become brave pioneers. Cautious companies become confused villagers threatening the automobile with a rake. I have no interest in either costume. A fast decision can be intelligent or reckless. A slow decision can be disciplined or defensive. The important question is what the decision protects, what it costs, and whether anybody revisits it after the conditions change.

My suspicion is that part of the tension comes from what a capable shopping agent could remove. Imagine a system that goes directly from my actual requirement to a suitable product at an acceptable total price. No wandering. No dramatic countdown. No unnecessary upgrade because the premium version has a photograph suggesting its owner drinks better coffee and maintains healthier boundaries. A machine that reliably resists persuasion would be convenient for me and potentially inconvenient for whoever budgeted for the persuasion.

That is my interpretation of the commercial stakes, not a claim about any particular retailer’s private motives. Shopping interfaces can serve several purposes simultaneously. They help me find things, but they also give a business opportunities to introduce more things. An agent instructed to buy one replacement charger might decline every invitation to turn the errand into a personal electronics renaissance. I can see why the person buying the charger and the person selling accessories might evaluate that efficiency differently.

Yet I refuse to assume the agent automatically belongs on my side. A friendly conversational interface does not establish loyalty. If I ask for the best option, I want to understand how “best” was determined. Did it prioritize price, durability, delivery, an existing commercial relationship, or the ease of completing checkout? Those factors can point in different directions. A recommendation delivered in reassuring prose is still a recommendation that deserves an explanation.

Imagine hiring a personal shopper who insists they work entirely for you but becomes evasive whenever you ask how they get paid. I would not be soothed by their excellent vocabulary. I would become more interested in the compensation arrangement. My standard for software should be at least as demanding. If there is a paid placement, a referral relationship, or a restricted set of participating sellers, I want that information where the recommendation happens.

Otherwise, we risk inventing a remarkably intimate advertising channel and calling it independence. A conventional advertisement at least gives me some chance to recognize the performance. An assistant can wrap a commercial suggestion in the language of knowing me. It can sound patient, thoughtful, even protective. I am not alleging that every shopping assistant does this. I am saying that the possibility makes transparency a basic requirement, not a decorative feature reserved for the settings menu.

Then there is the information I would need to share for the system to become genuinely useful. My shoe size is simple enough. My budget is more revealing. My household routines, gift recipients, delivery address, brand preferences, and explanations for avoiding certain products can create a much richer picture. The better the assistant understands my circumstances, the more useful it might become. That same understanding is also something I would prefer not to scatter around casually.

I would like to know what gets passed to the retailer, what stays with the assistant, and what is kept after the purchase. I do not need an uplifting paragraph about commitment to privacy. I need a comprehensible answer. If buying a kettle requires me to become a part-time investigator of data flows, we have misplaced the convenience somewhere between the demonstration and the paperwork. Saving six clicks should not create an evening of homework.

There is another problem that feels less technical and more recognizably human: I do not always know what I want. Sometimes I know the category and the budget but cannot explain why one option feels right. Sometimes I change my mind when I see something. Sometimes the sensible purchase loses to the object that makes me smile. An assistant can help me explore those reactions, but I am skeptical of any promise that my preferences can be neatly completed in advance.

Consider buying a gift. I could provide the recipient’s age, interests, and my spending limit. That might produce reasonable suggestions. But a good gift can depend on a private joke, a difficult year, an old conversation, or the delicate understanding that someone enjoys a hobby without wanting every birthday to become a shipment of equipment for it. I would welcome help finding possibilities. I would hesitate before outsourcing the gesture itself.

The same distinction applies to shopping for myself. I might happily delegate ordering the identical household item I buy repeatedly. Choosing clothes for an important occasion feels different. I want to see them, consider them, and occasionally make an aesthetically questionable decision that belongs entirely to me. Convenience is valuable, but so is participation. I do not want the future designed around the assumption that every moment spent choosing is a malfunction.

Of course, the opposite assumption can be smug, too. There is nothing spiritually enriching about spending forty minutes locating a compatible cable. People have jobs, children, disabilities, exhaustion, and a finite amount of patience. For someone who finds conventional shopping difficult, a capable assistant could make ordinary tasks substantially easier. I take that possibility seriously. My impatience is with inflated promises, not with tools that could give someone useful time or independence.

The standard I keep returning to is whether the system removes work or merely changes its costume. If I have to write an elaborate brief, check every specification, inspect the seller, confirm the price, and correct the order, I may have become an unpaid supervisor. The process can still be worthwhile, but let us count honestly. Watching software attempt my errand is not automatically better than doing the errand myself.

I would start delegation with narrow instructions and small consequences. Reorder this exact item, below this total price, from these acceptable sellers, without substitutions. Ask me before anything changes. That is less glamorous than a system autonomously curating my lifestyle, but I find glamour overrated in transactions. I want a predictable result. The toaster does not need to understand my journey; it needs to arrive in the right color without a surprise subscription attached.

Spending limits should also mean what ordinary people think they mean. If I authorize fifty dollars, that should include the costs I will actually pay. I should not discover that the system interpreted my budget as an inspirational starting point before shipping, taxes, and other charges joined the celebration. An assistant that cannot complete the task within the limit should stop and ask. Disappointment is cheaper than creative interpretation.

Substitutions deserve equally plain treatment. “Similar” can conceal a small inconvenience or a complete failure. A different package size may be fine. A different ingredient, material, connector, or fit may defeat the entire purpose. I want the system to recognize when an uncertainty matters rather than smoothing it over with confident language. There are circumstances in which the smartest purchasing decision is to leave the cart alone and admit that clarification is needed.

Returns are where I would look hardest for evidence of competence. Buying is the attractive demonstration. Unwinding a bad purchase is the actual relationship test. Can the assistant find the applicable policy, preserve the receipt, track the deadline, and help arrange the return? Can I reach a person when the process breaks? I have little interest in a futuristic purchase followed by a customer-service experience that feels like being lost inside a telephone tree.

I would also want a readable record of what happened. Show me the instructions I gave, the product selected, the final cost, and the approval used. If something goes wrong, I want to reconstruct the decision without arguing with a cheerful paragraph that insists it was trying to help. Good intentions are pleasant. A clear transaction history is more useful when I am staring at six of something I requested once.

From a retailer’s perspective, I can imagine the operational challenge being substantial. Product information would need to be clear enough for the systems doing the shopping. Stock information, variants, delivery estimates, and restrictions would need careful handling. These are design demands I would expect, not evidence that a particular company has solved them. A beautiful storefront cannot compensate for an agent misunderstanding which size is available or what is included in the box.

There is a potential benefit here that I actually find encouraging. If businesses have stronger reasons to publish accurate specifications and straightforward terms, human shoppers could benefit as well. I would love a retail internet where dimensions are easy to locate and product photographs do not require archaeological interpretation. We should not need an artificial intelligence revolution to learn whether a table is solid wood, but I am willing to accept progress through an embarrassing route.

I also wonder what this means for smaller sellers. An assistant could introduce me to an excellent business I would never otherwise encounter. That would be a meaningful improvement over repeatedly seeing familiar names. But it could also overlook sellers whose information is harder to access or whose systems are not compatible. I would want to know how broad the search really was before treating a tidy shortlist as a comprehensive view of the market.

The danger I see is another layer of dependency. If customers begin asking a handful of assistants what to buy, appearing in those assistants’ answers could become extremely important. Sellers might find themselves adapting to rules they do not control, while shoppers mistake a limited selection for the whole landscape. This is a possible outcome, not an established destiny. It is also precisely the kind of possibility worth considering before convenience becomes difficult to leave.

My preferred assistant would therefore be comfortable revealing its limits. Tell me which stores were checked and which were unavailable. Tell me when information might be outdated. Tell me when two products cannot be compared confidently. I do not need my shopping software to project executive certainty. I need it to distinguish between finding the best option it could evaluate and finding the best option that exists. Those are very different accomplishments.

I keep thinking about the word “friction,” which tends to get treated as an enemy whenever someone is designing a checkout. Some friction is pointless: repeated forms, confusing menus, missing information. But some friction is the moment I look at the total and reconsider. I want the pointless obstacles removed without losing that useful pause. There is a difference between making it easier to carry out my decision and making it easier to avoid noticing I am making one.

The interface should make that distinction obvious. I would want to choose between research, preparing a cart, and actually purchasing, with permissions that do not quietly bleed into one another. Asking for advice should not feel like starting a conveyor belt. And revoking access should be as straightforward as granting it. If turning the assistant on takes one cheerful button and turning it off requires a scavenger hunt, I have learned something important about the relationship.

Even the language matters to me. “I found this” should not imply that every plausible alternative was examined. “You might like this” should not become “You need this” simply because the system has accumulated enough details to sound familiar. I can appreciate a helpful tone while still wanting modest claims. The assistant can be personable without pretending that a record of my purchases amounts to a complete understanding of my life.

For retailers, I think the sensible response is experimentation with measurable boundaries. Test particular categories. Watch errors, returns, service demands, and whether customers actually come back satisfied. Make decisions based on what happens after the demonstration. Announcing that a company has embraced the future sounds impressive, but the future does not answer support tickets. Someone with a schedule and an increasingly cold cup of coffee eventually inherits the practical details.

I would apply the same discipline to the technology companies. Show me completed tasks that customers wanted completed, at costs they understood, with mistakes they could resolve. Show me why the process was easier. I do not need another prediction that shopping will never be the same. Shopping has been declared permanently transformed often enough that I am impressed it still occasionally involves discovering an item is unavailable after entering my address.

The more I think about it, the more I believe the most useful shopping agent might sometimes recommend buying nothing. It could notice that the replacement is unnecessary, that I already own something suitable, or that the supposedly urgent purchase can wait. I would trust an assistant more after it saved me from a transaction than after it delivered a polished explanation for spending extra. Restraint would be unusually persuasive evidence of whose interests mattered.

That creates an uncomfortable test for the business model. If the assistant becomes more valuable to me when I purchase less, will its incentives support that outcome? I do not assume there is only one possible answer. Subscriptions, transaction income, and other arrangements can produce different pressures. But I want the question asked openly. “We help you make better decisions” should be compatible with decisions that do not produce a sale.

I can imagine myself becoming a regular user of this technology under those conditions. I would delegate the boring, repeatable errands and keep control over the purchases where taste, uncertainty, or meaning mattered. I would expand permissions only after reliable results. That feels less like resistance to progress than ordinary adult judgment. I do not hand a new employee unrestricted authority on the first morning because they delivered an excellent introduction.

What I will not do is confuse speed with trust. A system can make a poor decision faster than I can, and the shipping notification may arrive with breathtaking efficiency. I want technology that respects hesitation, asks useful questions, and accepts a firm no. If the assistant treats every pause as an obstacle to conversion, it has misunderstood the assignment. My uncertainty is part of the decision, not a defect in the customer.

So I understand why retailers are divided, and I expect my own enthusiasm to remain conditional. There is real value in getting help with tedious shopping. There is also real value in knowing who selected the product, why it was selected, and who remains accountable when the box arrives. I am perfectly willing to let an agent carry the basket. Before I let it carry my wallet, I would like a straight answer about who signs its paycheck.

[1] Source: Nat Ives, The Wall Street Journal, October 2, 2026, “Retailers From QVC to Kohl’s Are Split on Agentic Shopping.” https://www.wsj.com/cmo-today/retailers-from-qvc-to-kohls-are-split-on-agentic-shopping-b75c3419

This is original opinion commentary. Hypothetical shopping scenarios illustrate the author's argument; they are not reported incidents involving the named retailers.