Why I’m Not Sold on Agentic Commerce
I’m going to say something that may sound strange coming from payments, but I’m not sold on agentic commerce yet, for four major reasons.
1. Someone has to pay for it. No one will want to.
This is probably the biggest barrier to widespread adoption, and it’s bigger than any one company wants you to think. I asked ChatGPT, Gemini, and Claude to estimate the compute cost of an agentic purchase. Yes, this is like asking a kid how fast they can ride their bike. Still, it gives us a starting point. Depending on the model and assumptions, estimates for a relatively straightforward purchase landed between roughly $0.50 and $1.35 per attempt. More complicated browser-based transactions could cost considerably more. Agents don’t always succeed, so the cost per completed purchase can easily exceed $1. That will get cheaper over time. But it’s not nothing, and someone has to pay it. Look at the players:
Consumers: A June 2026 study of UK banking transactions found that only 5.1% of adults were paying for an AI assistant. All of humanity’s knowledge in your pocket for $20 (ethical issues notwithstanding), and most people want the free version. People do pay for convenience when it saves them a trip, like delivery fees. An agent mostly saves a few clicks over one-click checkout.
Merchants: They’re already paying for websites, advertising, payment processing, and bot protection. Now we’re asking them to fund another intermediary that may be working against their interests.
Card networks: My read, not inside knowledge: card networks care about one thing above all, which is volume on the network. The more money moves across it, the more they make. But why subsidize a purchase that would happen anyway?
AI companies: Like Uber and Airbnb in the boom years, they’re subsidizing adoption. Uber didn’t die when the subsidy ended. They just…raised prices. And also decided to keep somewhere between 5% and 95% of the ride fee and not tell drivers until after the ride.
That said, the three players I’d watch closely are Meta, Google, and Amazon. Meta is pushing Muse HARD and offering it for free. But Meta isn’t exactly known for behaving ethically with user data, so it’s going to have to monetize it somehow. Google has a different incentive. Until generative AI, it was the gatekeeper of the internet, and companies paid handsomely for access. Those companies have no loyalty to Google; they’re loyal to eyeballs (remember newspapers?). Google has every incentive to subsidize AI shopping if that’s what it takes to keep consumers from defecting to OpenAI, Anthropic, or Meta.
Amazon has its own moat to protect. In September of 2026, Amazon blocked Muse. The stated reasons are around unauthorized access and security. Sure. But Amazon makes money by charging to recommend products. Muse is a potential existential threat to Amazon’s ad business.
The companies with the most reason to pay for this are the ones whose business is ads. Which brings us to trust.
2. It’s built on trust
Depending on how you use it, your chatbot knows a LOT about you, which can make it excellent at finding what you want. But riddle me this, Batman: I play Schecter and Fender basses. I’m not a big Gibson fan (beautiful instruments, but too heavy for my taste). What happens if I ask for a recommendation and it convinces me that Gibson isn’t as heavy as I think? I buy one. Then I find out Gibson paid for it (note: this didn’t actually happen). Why would I ever trust it again?
The ad model isn’t even a hypothetical. ChatGPT started showing ads in February, based on what you’re talking about. This is for free or low-tier accounts, but even my Plus account tries to upsell me on higher tiers whenever I talk about anything sales team related (so… daily).
Now, OpenAI claims the ads don’t influence its answers. I’m willing to skeptically trust that today, but I don’t see it lasting. Anthropic is having an IPO shortly. OpenAI will probably follow next year. Once shareholders enter the picture, company leadership no longer gets the only vote. We’ve seen this movie: Google search is public and ad-funded, and over twenty years, more and more of the results page turned into paid placement. There was an incentive, and shareholders got rewarded.
And marketers aren’t waiting. There’s already a whole industry popping up around generative engine optimization, dedicated to getting brands into chatbot answers. It’s not just content writers trying to game Google anymore. It’s AI gaming AI, with infinite capacity to do it. Think the “what time is the Super Bowl” articles were bad? Just you wait.
3. What are we actually solving?
Behavioral economics has spent decades showing we don’t optimize the way traditional economics expects. So while agents may save time, the truth is we LIKE buying stuff.
Be honest: how many products do you have on a subscription? I should put my dog’s food on one. She eats the same amount and goes through the bag in the same amount of time. But I still buy it when she runs out. It takes me three minutes, and it’s not a big deal.
Remember Amazon Dash? It was a little button you could push when you were next to the laundry detergent and realized you were out, and it would buy more. Amazon killed the physical buttons. My read: it solved a problem that barely mattered with a product hardly anyone wanted.
That said, there is a use case for B2B procurement. But even then, it means threatening people’s jobs and authority. I once worked with a company trying to sell an ERP implementation to a business where one person ran the whole thing in BASIC. Yes, the programming language that powered the random RadioShack TRS-80 computer that was in my basement as a kid (I still don’t know why). In the company’s case, that woman was the only one who could run it, and that was her job security.
Everyone in B2B sales knows change scares people and threatens jobs. And people really don’t like training their replacement, especially when that replacement is an AI.
4. Who pays when it goes wrong?
Let’s say my agent buys something I don’t want, or the wrong thing. Sure, I could authorize the last step, but now we’ve moved from agentic into AI-assisted purchasing, which isn’t the same thing.
Depending on the model, AI hallucinations can range somewhere from “student daydreaming in class” to “Dr. Timothy Leary is intrigued.” If it makes up a random fact during a conversation? Not ideal. If it imagines you suddenly want the suitcase in Meet the Parents? That’s going to be an awkward dinner conversation.
So, my agent buys the wrong thing, and I want to return it. Merchants already hate returns. There’s a fee to process the refund, they rarely get all their original fees back, and they often eat the shipping. Then the replacement is another round of fees and shipping.
Merchants can put terms and conditions around agentic purchases and refuse refunds, but that creates friction. Consumers hate friction. So consumers will start disputing with their card companies. Disputes are worse. The merchant pays the chargeback fee and spends time fighting it, and the issuers and networks pay people to process it. And too many disputes change how a card network looks at a merchant. Short version: everyone is pissed.
We already have a version of this problem. Someone’s kid takes a credit card and spends $1,000 on Minecraft. The transaction might be properly authenticated, but that doesn’t mean the cardholder intended to authorize it. Now imagine an agent acting within a set of permissions nobody fully understood. More importantly, credit card companies don’t have “AI hallucinated” as a dispute code. Let me know if you figure that one out.
Think this is messy now? We’re just getting started.
What happens next
Cheaper compute may just create more ways to influence what we buy. As compute gets cheaper, marketers can subsidize shopping agents. Or they can build their own and try to keep consumers inside their ecosystems. And if you think companies won’t try to influence an agent’s purchasing decisions, look at a B2B RFP process. Back in my MarTech days, I could read an RFP and tell you which vendor wrote it. The requirements just happened to describe that vendor’s product perfectly. Companies have been trying to influence purchasing decisions forever. Why would they stop because the buyer is an AI?
Merchants may end up teaching their next competitor how to beat them. Agentic commerce will require a lot of data from merchants: SKUs, inventory, pricing, shipping, policies, availability. Every piece of data they give an AI platform is another piece of information they have less control over. And the API that enables it is just another API someone has to build, maintain, and fix when it breaks.
And what happens when companies get that kind of merchant data? Two words: Amazon Basics. Amazon is a direct seller and a marketplace in the same environment. And they know exactly what you search for, buy, and don’t buy. Suddenly, Amazon Basics starts competing in categories from batteries to foam rollers, and they have a hell of a competitive advantage (there’s been regulatory scrutiny over this).
Why wouldn’t an AI company do something similar?
There will be rules to govern this sort of behavior, but rules only work when they have teeth. Meta has paid billions in privacy fines and keeps growing anyway. When you make tons of money breaking the rules, and the fines cost less than the money you’re making, it starts to look a whole lot like Edward Norton’s recall formula from Fight Club: if the expected out-of-court settlement cost is less than the cost of a recall, they don’t do one.
Merchants will have to decide if they want to get in the ecosystem, or maintain control at the risk of reduced sales. Rock, meet hard place.
Consumers will pay for this one way or another. Companies are already passing on as many costs as they can through surcharges, service fees, and whatever else they can come up with. Tipping has reached the point where e-commerce companies are asking for a tip like Tim Curry in Home Alone 2. And if consumers don’t pay with money, they’ll pay with attention or data. That’s Meta’s entire business model. Which brings us straight back to the problem of trust.
So what’s the future of agentic commerce?
I genuinely don’t know. I could be wrong about all of this, but I don’t think I am. Here’s what would change my mind: consumers paying for agents directly, an ad-free agent that works at scale, or dispute rules written for agents instead of around them.
More importantly, this is how I look at problems. They’re rarely “here’s the problem, here’s the solution.” Every fix changes someone’s incentives, which changes their behavior, which changes the problem. And now you have to figure out what happens next. That’s what Follow the Incentives means, and it’s how I diagnose teams, not just markets.