Hello,
Two decades ago, I used to accompany my mom to the vegetable market. I’d try to help by picking a few tomatoes and some onions. They never made it to the final basket. She would quietly set aside my picks and choose her own. Each one that made the cut was pressed, turned over, held up against some standard only she could see.
For most people, grocery shopping is a chore. For the chef in my mom, it was a high-precision job. The right tomatoes meant the curry would taste the way she intended. Choosing wasn’t merely a step before cooking. It was where the cooking began.
That’s the thing about shopping, it’s not one activity. Some of it we’d delegate to anyone who’d take it. Some of it we won’t hand over even to our own kids. And any attempt to delegate how we buy things has to start by understanding which is which, not by assuming that automated robots are the default solution to everything.
In today’s guest op-ed, Anderl takes apart the “give the agent a wallet” thesis and shows why paying was never the hard part of shopping.
On to Anderl’s story,
Prathik
There is a line going around the AI and crypto world right now: give an agent a wallet, let it shop for you. The ultimate use case. It sounds clean, it sounds like the future, and it still doesn’t feel quite right. The mismatch is worth unpacking because it shows that the line confuses the hard part of shopping with the easy part and then puts the easy part at the centre.
Let’s start with shopping itself, not with payment.
Shopping Splits into Two Things
Shopping is really two activities that today only look fused. One is search, the other is judgment. Search (gathering, filtering, comparing, pre-sorting) is mechanical and delegable, almost entirely. Judgment (is it any good, does it suit me, do I trust the seller) is the part that a human is attached to.
It is measurable that search is already migrating. Adobe Analytics reported a roughly 4,700% rise in traffic from generative AI referrals to US retail sites between July 2024 and July 2025. The wallet thesis now quietly assumes the agent takes on both search and judgment. This is where the sleight of hand sits. One can be handed off, the other only under conditions, and sometimes not at all.
Judgment Splits Again
The interesting move is that judgment isn’t monolithic either. It breaks into two parts. One is evaluation, which involves checking options against a utility function. The other is authorship, where you set that utility function in the first place: which dimensions count, how they are weighted, which values bind, and what “good” ends up meaning.
Evaluation automates. Authorship doesn’t. And it doesn’t sit as a single gate up front that you pass through once. It runs as a thin thread through every layer. Conformance is about which specifications matter to you. Quality is about whether a broken zipper is a deal-breaker or beside the point. And the choice of merchant turns on which values bind you. Every layer does (human criterion × agent evaluation). In every case, what automation eats is the evaluation half. What’s left is the authorial half.
This has a consequence that everyone underestimates who thinks the human writes one spec sheet and walks away. You can’t. It’s one of the most robust findings in decision research. Since Paul Slovic’s work on decision-making under risk, we speak of constructed preferences. The American professor and psychologist found that we don’t carry a finished liking inside us that we simply retrieve; we form it in the act of choosing (Slovic 1995; Lichtenstein & Slovic 2006). The evidence is preference reversals. Normatively equivalent elicitation methods like choosing versus pricing produce systematically different rankings, violating the basic axiom of rational choice. Ariely, Loewenstein and Prelec (2003), in their work on coherent arbitrariness, showed that even an arbitrary number (the last digits of a Social Security Number) anchors willingness to pay for familiar products, and that the effect disappears neither with experience nor through market forces. The “stable preferences” picture is largely an illusion of order.
The realistic interface is therefore not the spec sheet but an iterative one. The agent surfaces, at each layer, the decision a criterion needs (”you said durable, at what premium does durability stop being worth it?”), and the human writes exactly that.
The Real Axis: Chore or Pleasure
Now, the distinction that orders everything. We like to talk about “standardised” versus “individual” products, as if it were about specifiability. It isn’t. The real axis is different. On one side, choosing is a pure chore. On the other hand, choosing is part of the point.
For a commodity like printer paper, batteries or the recurring reorder, the act of choosing has zero consumption value. Nobody enjoys deciding between two identical toner cartridges. That’s why commodities are the natural home of full delegation: nothing is lost when the agent silently reorders the right thing.
With pleasure goods, it’s the reverse. The wine, the piece of furniture, the coat, and the book. Here, choosing is part of the good itself. Whoever hands off the judgment not only saves effort but also destroys part of what they wanted to consume. And that holds even if the follow-up questions were free. You wouldn’t want to delegate it even at zero friction.
With pleasure goods, the agent therefore doesn’t withdraw but should switch roles. From buyer to scout. It still does the search, the shortlisting, the spec matching, the merchant verification, and the pulling of recurring weaknesses out of a thousand reviews. It narrows 200 to 5. And then it stops. The human savours the final choice.
The Friction Trilemma
“But the agent could just ask for my criteria.” It could, and that’s exactly the annoying variant that people resort to doing themselves. Worse, interrogating degrades the choice. Wilson and Schooler (1991) had participants rate jams. Those asked to analyse the reasons for their feeling beforehand ended up with rankings that matched expert judgment less well than the control group’s. In a follow-up study, people who spelt out their reasons chose different posters and were less happy with them weeks later. Articulating directs attention to what is easy to put into words, not to what actually carries the preference. Taste is recognised, not verbalised.
So there is a trilemma:
Interrogate. Faithful to your current self, but high friction, and it can even worsen the choice.
Infer from the past. Low friction, but it locks you into your past self and shuts down discovery. You only get more of what you already want.
Judge yourself. Full autonomy, full labour.
There is a fourth mode that avoids the bad corners: recognition instead of interrogation. “Show me three, I’ll point to one.” Low friction and accurate at once, because it doesn’t require naming criteria in the abstract that you often can’t name. Here, the most famous study on choice overload helps too. Iyengar and Lepper (2000) sold jam at a tasting stand. With six varieties, far more people bought than with 24. Still, the effect is disputed, and honesty requires saying so. A meta-analysis by Scheibehenne and colleagues (2010) found, averaged over many studies, almost no general choice-overload effect. A second one across 99 studies (Chernev, Böckenholt & Goodman, 2015) showed that it appears only under certain conditions, such as high complexity, a hard decision, and an unclear preference. Telling is the original authors’ self-criticism. Perhaps, faced with 24 varieties, customers simply didn’t have enough time to “determine their preferences”. The line where autonomy actually lives is the one between “the agent applies my criteria” and “the agent invents them from my history”.
Back to the Wallet
With all that behind us, we can finally name why the wallet line sounds off. It throws together three things that are separable: who decides, who executes and who holds the funds. “Give the agent a wallet” answers only the third, and that only matters if the first falls to the agent, too.
Three cases. Human decides and pays themselves. Here, the agent doesn’t pay but remains a scout. In the second, the human decides and delegates execution (”yes, buy that one”). Now, the agent runs the checkout, but it needs no custody of money, only a narrow, revocable authorisation for exactly this approved purchase. Only in the third case, when the agent decides and pays autonomously, with no human at checkout, does the wallet itself become load-bearing.
And here it gets interesting because the payments industry built exactly this separation in 2025. Not custody, but authorisation. The Agentic Commerce Protocol, developed by OpenAI and Stripe, issues a “Shared Payment Token” bound to a single merchant, with a fixed amount, that is time-limited and single-use. Mastercard’s Agent Pay (April 2025) issues “Agentic Tokens” scoped to a specific agent, a merchant scope and a consent policy. The agent never sees the raw card number. Google’s Agent Payments Protocol (AP2, September 2025) even cleanly separates an Intent Mandate (what the human wants) from a Cart Mandate (what the agent proposes to buy), both as signed verifiable credentials, which is exactly the authorship-versus-evaluation axis cast in cryptography. Visa’s Trusted Agent Protocol (October 2025) takes the same route. None of these approaches hands the agent funds. All of them hand agents a narrowly bounded authority. The industry independently confirms the article’s thesis.
So where does the agent-owned wallet become native? Precisely in the commodity tier and in the machine room. The genuinely strong crypto cases don’t lie in consumer shopping. The x402 protocol, backed by Coinbase and Cloudflare, fills exactly the gap that card rails don’t close: fully autonomous agent-to-agent payments with no human in the loop, billed per API call, around the clock. Over 100 million transactions in the first months. Mastercard, in turn, announced “Agent Pay for Machines” for high-frequency, low-latency micro-payments between machines. That is infrastructure for a machine economy, not shopping. So the thesis isn’t wrong. But its importance is inverted. The wallet of its own becomes load-bearing where the goods are most interchangeable and the amount per transaction is often the smallest.
Where the Wallet Really Belongs
This explains, in passing, why crypto has shifted its centre of gravity over the past two years, away from the consumer-driven culture of liberation, toward rails for companies: stablecoin settlement, tokenisation, and institutional infrastructure. That isn’t a turn away from agent shopping but is a homecoming to the one tier where the funds-holding wallet actually fits. Corporations are its anchor tenants for structural reasons. A procurement department is the institutionalisation of “choosing is a pure chore”. Its purpose is to replace taste and gut feeling with specification, price, reliability and contract terms. Procurement is agentic shopping, carried out by rule-following humans. So the agent wallet teaches corporations no new behaviour, and instead automates one they perfected long ago.
You can see it in an entire industry layer built for exactly this. Many companies have outsourced the purchasing of C-items, such as office supplies, to specialised platforms, including Mercateo and Amazon Business. Their promise is not the lowest price. It’s the lowest effort: one catalogue instead of twenty suppliers, one invoice instead of many. The company accepts a slightly higher unit price and saves the process costs, which for C-items often exceed the value of the goods themselves. The agent attacks exactly this logic. It pushes the ordering effort toward zero, while simultaneously searching the fragmented market nobody had time for before. What remains for the platforms is only conformance: approved suppliers, a clean invoice, protection against counterfeits. This verification, further down, turns out to be the real bottleneck.
The recommendation is therefore more precise than “wallets for firms instead of consumers”. It reads: ‘match the instrument to the tier’. The autonomous, custodial wallet (the agent decides, holds, pays) has its market in the procurement of standard goods and in machine-to-machine payments. The buying there is high-volume and repetitive, driven by specifications, so largely B2B and M2M. The consumer side, by contrast, wants authorisation instead of custody, meaning a narrowly bounded, single-use token on existing rails, only when the human approves a purchase. “Give the agent a wallet”, as a consumer headline, therefore, aims at the weakest slice. The product’s actual bridgehead is the machine room of companies.
A caveat that stays honest. Corporations don’t buy only commodities either. They have their own non-delegable tier, just one driven by consequence rather than pleasure: the choice of a law firm, an acquisition target, or a strategic supplier. There, the human stays, and the autonomous wallet has as little business as it does with the consumer’s wine.
The rule is domain-independent. The wallet rides the commodity layer wherever it sits, and that layer is simply the largest and densest inside companies.
What the Bottleneck Really Is
Paying was never the bottleneck. Moving money is solved, several times over. The two real bottlenecks are different.
One is trust in the inputs. Automating judgment is only a gain if the data ground holds. If it doesn’t, a judging agent is worse than a sceptical human. It launders manipulated signals at machine speed. That this ground is shaky today is official. The US trade regulator, the FTC, issued a rule in 2024 (effective 21 October 2024) banning fake reviews, noted that such fakes are widespread, and explicitly named generative AI as a tool that eases their mass production. By late 2025, the first warning letters followed, with penalties of around $51,744 per violation. Add the physical counterpart. Per OECD/EUIPO (2025), global trade in fakes amounted to roughly $467 billion in 2021, or 2.3% of world trade, and in the EU to 4.7% of imports, with clothing, shoes and handbags at the top. Pleasure goods, of all things. Anchored product identity, verifiable reviews, a verifier independent of the seller, and provenance records at the unit level (as the EU anti-counterfeiting directive and the US DSCSA have long required per pack in pharma) are the precondition that tips judgment automation from dangerous to useful.
The other bottleneck is authorship. Everything downstream of a defined wanting becomes increasingly automatable. Defining the wanting stays human, not out of nostalgia but structurally. To automate it, the agent would have to want for you, and a wanting fabricated for you is not yours.
Give the agent a wallet, and you’ve solved the easiest part. The interesting part is making judgment partly and safely automatable, and leaving the human where they belong: as the author of the criteria and as the one who savours the final choice.
For Anyone Building Marketplaces
One consequence to close, for pleasure goods. Once agents commoditise sourcing, the goods are findable everywhere. Then your product is no longer the goods. Choosing is. Make yourself maximally legible to the scout agent (verifiable provenance, clean data) so it routes the human to you. And on top of that, occupy the one thing automation can’t take: the discovery that widens taste rather than confirming it, and the pleasure of the final decision. Hand sourcing to the agents. Build your business around the better choices.
That’s all for today.
Until next time,
Anderl
We will be back with another guest essay soon.
P.S. This piece was originally published here.
We will be featuring good writing and writers we love from time to time. If you have recommendations, send them our way.
Token Dispatch is a daily crypto newsletter handpicked and crafted with love by human bots. If you want to reach out to 170,000 subscriber community of the Token Dispatch, you can explore the partnership opportunities with us 🙌
📩 Fill out this form to submit your details and book a meeting with us directly.
Disclaimer: This newsletter contains analysis and opinions of the author. Content is for informational purposes only, not financial advice. Trading crypto involves substantial risk - your capital is at risk. Do your own research.
 | A guest post by
|