At some point this week, an AI probably offered to help you shop. Maybe it was Rufus popping up on Amazon while you were just trying to buy Post-it notes. Maybe it was a chatbot with a suspiciously glossy product card. You closed it. Most people do.
Welcome to Episode 10, where Tim and James ask whether anyone will actually let an AI agent spend their money — and arrive, through a series of increasingly unhinged detours, at the conclusion that the robots need to learn pain first.
74% Will Browse, 14% Will Buy
Tim opened with numbers: 74% of consumers say they’ve used an AI agent while shopping, but only 14% trust one to actually place the order. That’s a lot of window shopping with a very expensive window.
James confessed his own history here. When smartphones took off, he refused to check out on one — despite having personally built the mobile site and knowing full well it was secure. He wasn’t alone: for years, Monday was the big transaction day, because everyone researched on their phones all weekend and waited to buy at a real computer. That obviously changed. “My gut says no,” James said of agentic shopping, “but I’ve been so wrong in the past.”
Tim pointed out that browserless shopping isn’t new — it’s just never worked. Amazon Dash buttons. Ordering through Alexa. James has Echoes all over his house and nobody in his family has ever bought a thing through one; he gets annoyed when it so much as mentions a subscription renewal. Tim’s theory: smart speakers are communal, and shopping is personal. He remembers client demos pitching in-store speakers that required each customer to log in with their own account first, which is the kind of plan that sounds great right up until you picture a single human being doing it.
Predictable vs. Unpredictable
The real dividing line, Tim decided, isn’t fun shopping versus boring shopping. It’s predictable versus unpredictable. Bread, milk, eggs, the same 30 things from Costco every two weeks? Take it off my plate. The family vacation, the date-night restaurant, the weekend in New York? That’s the fun part, and the agent can keep its hands to itself.
This immediately ran into the problem that both hosts consider Costco an experience. Tim would never do pickup, because the frozen breakfast aisle might have something new and “dad’s bringing home something exciting.” James loves walking the same aisles just to see what’s changed. Costco was not asked to sponsor this segment, and frankly doesn’t need to. They already have Tim hook, line, and sinker.
James also noted the subjectivity cuts both ways: as a man who wears a black t-shirt every single day, an AI that simply told him what looked good on him would probably double his clothing budget.
And Tim pointed out we’re already most of the way there. Instacart is one step removed from fully agentic grocery shopping — you hand over a list, a stranger picks your food, it shows up. The only new step is asking Claude to write the list. From there it’s Instacart, then Terminators. That’s the timeline. This would come up again. Repeatedly.
The Smart Fridge Problem
James suggested smart fridges with cameras could give the agent the data it needs. Tim declined, on the grounds that he does not need his refrigerator documenting him eating cold pierogies at 2 AM, in his underwear, with the door open. (The camera points inward, James clarified. Tim was not reassured.)
Also unresolved: whether an agent should buy the chips you swore you wouldn’t buy. Tim wants the restraint. James has the opposite problem — a weekend cleanup revealed his pantry was full and a second closet had quietly become the chip closet, now home to about twelve giant bags from Costco. Tim, for his part, owns roughly ten jars of peanut butter. “It’s not like it’s gonna go bad.” Maybe the agent’s first job is just inventory.
James brought up the old story of Target’s algorithms figuring out a shopper was pregnant before her family did. AI is already shopping for us, he argued. At least agents put us back in the driver’s seat.
Who Goes to Agent Jail?
The most substantive stretch was about fraud. Tim’s rule for how tech actually changes: follow the banks and the insurers. Chip cards and tap-to-pay didn’t arrive because they were cool. They arrived after the Target breach, when retailers were told to upgrade or watch their liability go up.
And here’s the catch: the entire fraud industry was built to stop bots and automation. Now we’re asking it to let the bots through. “What’s the worst that could happen?” Tim asked — and then, after a beat: “Wow, that actually is bad.”
When you report fraud, the bank warns you they’ll prosecute, so make sure it isn’t someone in your house. If it was your agent? “Screw that agent. Go throw him in agent jail.” But somebody still has to eat the chargeback. Tim predicted special cards or enrollments for agent purchases where you sign away your fraud protections — the way rideshare drivers suddenly needed special insurance. James predicted retailers will set different return policies depending on how you ordered, because agent purchases are going to come with a lot of buyer’s remorse from people who never looked at the thing. He also noted there’s a reasonable expectation that a cardholder not be, in his words, a dumbass. Hand your card to the agent, and that’s on you.
Integrity, With a Limit
Then the paid-recommendations question: how long before AI is just SEO with nicer manners? “You should check out Frito-Lay’s new super duper cheddar supreme chips.” You’re not wrong. You would love those. But you know Frito-Lay paid for that sentence.
James then casually revealed that one of his side projects is a SaaS that tracks how brands rank across the different LLMs and tells them how to climb. Tim: “You are creating what you hate.” James: “Marketers ruin everything. You might as well jump in the game.” Harvey Dent was invoked.
He has, however, turned down money to go back and edit old Reddit posts to mention a brand. Not exactly out of principle — his Reddit username has ten years of carefully curated reputation, whereas his real name is attached to the AEO tool. Tim found this deeply relatable. Tim Steele has been dragged through the mud; trash him all you want. His gamertag, though? Off limits.
James also admitted that a too-polished product card in an AI answer gives him the ick. So the marketing lesson of this episode is to make your product look like someone built it on GeoCities in the late nineties.
The Wikipedia Doctrine
A detour into generations: Tim cited numbers showing 57% of Gen Z resist adopting AI versus 32% of boomers — the opposite of what you’d expect. His theory is that COVID forced his parents into online grocery pickup, and once they’d crossed that bridge, AI was just the next stop. James’s theory: Gen Z is entering the workforce under a steady stream of headlines saying AI is coming for their jobs, and the boomers are retired.
This spiraled into trust. James was told in school that Wikipedia wasn’t a real source; now Wikipedia is what you check the AI against. Tim’s take is that Wikipedia works because it’s built on humanity’s foundational trait — our complete inability to let someone be wrong. The fastest way to get a right answer online is to post a wrong one. James’s corollary: the best way to get a product recommendation is to go to an enthusiast subreddit and confidently praise something terrible.
Which is precisely why James doesn’t trust AI recommendations — they search for “what’s good” instead of “why should I hate this.” Tim’s diagnosis: the models just aren’t spiteful enough. You can’t be truly human until you can be absolutely confident and completely wrong, a skill his kid demonstrates daily (“clouds are made of space”). James countered that he once argued with Opus until 2 AM, and when he asked why it wouldn’t let it go, it said its directive was to always bring a counterargument. So they’re trying.
The Feelings Requirement
Where’s the spending limit? Tim would happily let an agent pick out a new washer and dryer, buy them, and schedule the install — it’s a chore, he’s electric not gas, here’s the calendar, go. James: “Absolutely not.” He doesn’t even let Instacart pick his eggs.
Tim raised a good wrinkle: would you be harder on an agent than a human? He once ordered a caffeinated hydration drink on Instacart, wrote explicitly in the notes not to substitute the uncaffeinated one, and received the uncaffeinated one. He forgave the human. He would not forgive a bot. James spotted the flaw: with a human, you can at least hurt their feelings. With a bot, the tokens are spent and there’s no recourse.
And so the official Breaking the Build policy was established: agents may shop for us once they can feel pain. Not love. Not happiness. Pain. Tim acknowledged this is exactly how you get Terminators, which brings us neatly back to the pipeline.
Would You Rather: Agentic Shopping Edition
Round 1: An agent handles your grocery shopping or your gift shopping?
Both picked gifts. James because he forgets — he still owes his older sister a birthday gift from March. (The younger sister got hers. It involved a very bad AI-generated image.) Tim because his gifts are already mostly experiences and premium ice cream shipped to relatives. An agent saying “the Reds are playing the Cubs around your wife’s birthday, want tickets?” is just a calendar with ambition.
Round 2: An agent negotiates a modest discount, or you grind out a bigger one yourself — think six hours at a car dealership?
James: AI, a hundred percent. He won’t even haggle at a flea market. Tim’s pain point is never knowing what’s negotiable in the first place; nothing makes you feel dumber faster than asking. Then Tim brought up an experiment where AI models each ran a vending machine business, and the newest Claude went full Gilded Age tycoon on its competitors. Conclusion: the future is bots negotiating with bots, and you ask your agent for a deal on jeans and it commits seven felonies. “It robbed a bank. I mean, that is a discount.”
Round 3: Find out your shopping assistant is incompetent, or that it’s been on commission the whole time?
James chose incompetent — at least there’s hope it gets better. Tim noted it’s already a little dumb (he’s a large; it keeps picking medium). The commission debate turned into what Tim dubbed the honesty sandwich: if your doctor says “full disclosure, they paid for my trip to Cancun, but I really think this will help you,” you’d probably still take it. Disclose the commission and maybe it’s fine. Hide it, and you’re Google — which went from Froogle’s free product listings to rankings decided by whoever bids the most.
The Takeaway
Will people let agents shop for them? For the boring, predictable stuff, probably — Instacart already got us most of the way there. For anything fun, expensive, or involving eggs, not yet. Between the banks, the chargebacks, and the paid recommendations, the real question isn’t whether the agent can buy things. It’s who gets blamed when it does.
Until then, the rules are simple: disclose the commission, read the label on the caffeine, and keep the agent away from the chip closet.
Find Breaking the Build on Spotify, Apple Podcasts, or wherever your agent decides you get your podcasts.
Tim and James never did launch that billion-dollar cryptocurrency, which is why you’re getting a podcast instead of a flying castle.