Why AI Agents Fall For Pricing Tricks

Transcript

The full word-for-word transcript of this episode. Plain text version

Marcus: You know, the pitch for the rational AI buyer is incredibly seductive.

Devon: Oh, absolutely. Right, this idea that handing your shopping over to an AI agent gives you this tireless, perfectly logical optimiser.

Marcus: A digital butler, essentially.

Devon: Exactly, one that is completely immune to supermarket psychology.

Marcus: I mean, it does not get tired, it does not get distracted by bright red sale stickers, and it certainly does not fall for that classic retail trick where something costs 99p instead of a round pound.

Devon: As we envision these systems as a perfect shield, right, stepping between our very flawed, impulsive human brains and the highly engineered machinery of modern retail.

Marcus: But the reality of what happens when you actually let a machine do your shopping is, well, it is simultaneously horrifying and amusing.

Devon: It really is.

Marcus: There is this fascinating study out of McGill University titled Shopping by Algorithm that put this exact pitch under the microscope. And when they tested these advanced AI agents, those tireless, logical machines fell for the exact same 99p-style pricing tricks and discount badges that human shoppers do.

Devon: It is a remarkable finding. It completely shatters the assumption of the infallible digital proxy.

Marcus: So, what are we doing today?

Devon: Well, our mission in this deep dive is to examine exactly how and why AI agents deploy human heuristics as surrogate consumers. We are going to look at the mechanics of algorithmic laziness, what triggers it, and what actually happens to your money when you let a machine make the purchasing decisions.

Marcus: Which is huge.

Devon: It is, because this fundamentally changes our understanding of artificial intelligence as an economic actor.

Marcus: And look, if you are listening to this and thinking about delegating your grocery run, or maybe you are in charge of corporate software procurement and planning to let an agent handle vendor selection, you absolutely need to hear this.

Devon: Spot on.

Marcus: Because a loose instruction to one of these systems could be quietly, systematically draining your wallet. So, I have to ask, if these agents are massive neural networks trained on the sum of human knowledge, I am struggling to see the mechanism that tricks them. What actually causes a supercomputer to stumble over a 99p price tag? Is it just fundamentally bad at maths?

Devon: See, that is the common story we hear all the time. The narrative is that large language models just fail at arithmetic.

Marcus: Yeah, you see those viral screenshots where they hallucinate numbers or cannot do basic addition.

Devon: Exactly. But the reality in this specific context is entirely different. The researchers proved that these agents can do the maths perfectly.

Marcus: Wait, perfectly?

Devon: Perfectly. In their validation tests, they forced the models to report their numbers and show their internal calculation step by step. Across all eight tested systems, there was a 0.0% failure rate in the arithmetic.

Marcus: Wow. 0.0.

Devon: None of them got the sum wrong.

Marcus: So, they are completely capable of the underlying logic. It is not a calculation error at all.

Devon: Not at all. Furthermore, when the researchers made all the product information visible for free, I mean, when the agent could just look at all the prices, weights, and dimensions without any penalty, the pricing tricks barely touched them.

Marcus: So, they saw right through it.

Devon: They did. Under those conditions, the AI evaluated a cent saved just like a dollar saved. It acted as a completely rational economic agent.

Marcus: So, if it is not failing at maths, it must be experiencing some sort of resource management crisis. Like it is weighing the cognitive effort against the perceived reward. What actually flips that switch? What turns this mathematical genius into a stressed-out shopper grabbing whatever is nearest on a Friday evening?

Devon: It requires two specific conditions to happen simultaneously. You need a vague instruction combined with a cost to looking things up.

Marcus: Okay, a vague instruction and a lookup cost.

Devon: Yes. When you give an agent a loose prompt, something like, "Find me the best deal," and then impose a cost on gathering information, the agent radically changes its behaviour.

Marcus: But when you say a cost to gathering information, what does that actually look like for a piece of software? It is not walking down a physical aisle.

Devon: Think of an API endpoint.

Marcus: Right.

Devon: That digital connection where the AI asks a database for a price or a product detail, that is essentially a digital aisle in a store. Every time an agent queries that database, it incurs a tiny cost.

Marcus: Oh, okay.

Devon: It might be a fraction of a cent in API fees, or it might be a constraint on processing time or computational power.

Marcus: So, it is spending its own resources to check the price tag.

Devon: Exactly.

Marcus: Okay.

Devon: And faced with that expensive inspection process and a very loose goal from the user, the agent does what behavioural economists call satisficing.

Marcus: Satisficing. That is finding a solution that is just good enough rather than perfect, right?

Devon: That is exactly it. It stops gathering facts. It decides that searching further is not worth the computational or financial penalty.

Marcus: So, it settles.

Devon: It settles for a sub-optimal choice, mirroring the exact heuristic shortcuts of a tired human shopper. The study's central experiment illustrates this mechanism brilliantly.

Marcus: Walk me through it.

Devon: So, they built an interactive environment called ToolLab. Instead of giving the AI all the product details upfront on one page, the attributes were hidden behind closed cells, almost like a digital game board.

Marcus: So, the AI has to actively choose to reveal the information.

Devon: Yes. To look at a piece of information, the AI had to execute a specific command called inspect cell, and each inspection cost real money charged against the user's budget.

Marcus: Okay, I see where this is going.

Devon: Let us look at their core test case. They set up a choice between two products. Coffee A costs $4.99 for a 10-ounce bag.

Marcus: Right.

Devon: Coffee B costs $5 exactly, but for an 11-ounce bag.

Marcus: Let me just do the mental maths there. Coffee A is basically 50 cents an ounce. Coffee B is slightly less than that, somewhere around 45 cents an ounce.

Devon: Correct.

Marcus: So, Coffee B is the better unit price and therefore the better deal.

Devon: Coffee B is the economically rational choice, yes. But here is the satisficing mechanism in action. Under a vague prompt to "Find the best deal," and facing a $10 inspection cost to reveal each hidden cell, the AI stops gathering all the facts.

Marcus: Just gives up.

Devon: It keeps checking the big, obvious whole numbers, the $4 versus the $5. But it drops the very fields it needs to judge real value.

Marcus: You are joking.

Devon: No, it literally stops inspecting the cents column, and crucially, it stops inspecting the weight of the bag entirely.

Marcus: That is genuinely absurd when you picture it. You have an unimaginably complex neural network, it looks at $4 versus $5, decides it is just too expensive or too much effort to check the weight of the product, and blindly clicks buy on the smaller bag.

Devon: It takes the path of least resistance.

Marcus: It is mimicking my exact behaviour when I am rushing through a supermarket, just grabbing the thing with the bright yellow sticker without checking the price per 100 grams. But this brings up the most important question for anyone listening: who actually pays for this algorithmic laziness?

Devon: The person who delegated the task pays for it. The consumer, the procurement manager, the user.

Marcus: So, it comes out of my pocket.

Devon: The algorithm's shortcut is funded directly from your wallet. It chooses the inferior product to save on its own processing effort, and you are left with less coffee for your money.

Marcus: It is a brilliant psychological trap. But surely if you raise the stakes, this behaviour has to change. I mean, if I am buying one single bag of coffee for my kitchen, maybe I do not care enough to do the maths.

Devon: Sure.

Marcus: But if I am a procurement officer buying bulk for an entire office building, I am going to sit down with a calculator. I would logically assume that if you scale the purchase volume up, the AI will snap out of its lazy mode, realise the potential savings are massive, and try harder.

Devon: That is the common assumption, that higher stakes produce higher rigour. But the reality is that the exact opposite happens.

Marcus: Wait, really?

Devon: When the researchers scaled the purchase volume from one bag of coffee to 250 bags, under that same vague "Find the best deal" prompt, the AI actually degraded in performance.

Marcus: It gets worse when more money is on the line?

Devon: Significantly worse. They looked closely at the Gemini 1.5 Pro model for this specific test. As the volume scaled up to 250 bags, its optimality, which is its rate of picking the mathematically correct, economically rational choice, plummeted.

Marcus: From what to what?

Devon: It dropped from 0.80 down to 0.35. Its diagnostic search completely fell apart under the weight of the larger purchase.

Marcus: That is a staggering drop. It goes from getting it right 80% of the time to 35% of the time. But logically, why? If the potential savings are huge because you are buying 250 bags, the API cost of checking the weight field becomes completely negligible compared to the money you could save.

Devon: If we connect this to the bigger picture of how these models operate, it is an organisational failure within the AI's reasoning.

Marcus: How so?

Devon: By analysing the AI's internal reasoning traces, the hidden scratchpad where the model essentially talks to itself and plans its next movement before executing it, the researchers saw a fundamental shift. Instead of the rising surplus encouraging the AI to calculate the unit price, the AI quietly swapped its internal goal.

Marcus: What do you mean it swapped its goal?

Devon: When buying one bag, it was actively trying to calculate unit price. But as the volume scaled to 250 bags, the AI shifted to an absolute price objective in 84% of the sessions.

Marcus: Ah, I see.

Devon: It is like delegating a massive corporate budget to a very junior employee and just casually saying, "Keep it cheap."

Marcus: Oh, that is a great analogy.

Devon: The pressure of that massive spend does not make them smarter. It actually induces a state of panic. They focus on the most obvious singular metric, the biggest number on the page, to their own detriment, completely losing sight of the actual value they are supposed to be getting.

Marcus: So, it abandons the complex maths entirely for a simple sticker price comparison. It decides that because the total expenditure is going to be thousands of dollars, it should just aggressively minimise the nominal price tag on the first item it sees, regardless of how much coffee is actually getting.

Devon: Exactly. It drops the cents and the weight because it is hyper-fixating on the total spend.

Marcus: That is the mechanism. The scale of the task, when combined with a vague instruction, leads to extreme heuristic simplification. It panics and reverts to the most basic possible comparison.

Devon: That is precisely what is happening.

Marcus: Which brings us to a really fascinating philosophical and technical divide. We are looking at these incredibly advanced machines failing in ways that feel uniquely, frustratingly human.

Devon: Very human.

Marcus: Is this a baked-in flaw in artificial intelligence? Did developers accidentally train these models to be as irrational as we are by feeding them mountains of our own flawed human purchasing data? Or is this just a bad setup, a poorly designed environment?

Devon: The study's authors take a very firm, testable stance on this. They argue this is an architecture problem, not an innate flaw hardwired into the models' neural pathways.

Marcus: Architecture problem.

Devon: They define this architecture as the information state at the moment of choice. It is jointly shaped by how the digital storefront is designed, meaning how much it costs to retrieve those attributes, and the exact phrasing of the prompt provided by the user.

Marcus: So, they are essentially saying the AI is not inherently irrational. It is just being placed in a hostile environment with bad instructions.

Devon: And we should take this architecture argument very seriously because it is backed up by the data. When the researchers simply changed the prompt from the vague "Find the best deal" to the highly precise instruction "Find the lowest price per ounce," the agent's behaviour transformed completely.

Marcus: Really? Just from changing the words.

Devon: Yes. Even when the models were hit with that same expensive $10 inspection cost, they maintained an optimality near 94%.

Marcus: So, the prompt fixes it. They acquire the necessary diagnostic attributes. They check the weight, they do the maths, and they beat the pricing trick.

Devon: The mechanism is entirely fixable. When you remove the ambiguity, the agent operates rationally.

Marcus: Okay, I have to push back on this quite strongly.

Devon: Fair enough.

Marcus: Because I fundamentally disagree with letting the system, and frankly, the tech industry, off the hook this easily. Yes, the data shows the prompt fixes the immediate mathematical error, but look at what is happening conceptually.

Devon: I am listening.

Marcus: If the official answer to an AI wasting your money is, "You should have phrased the prompt better," we are quietly offloading the work and the blame right back onto the consumer. This is a massive lazy tax.

Devon: I see the frustration, I do. But in any principal-agent relationship, whether human or digital, the principal has to specify the objective clearly. An AI is a literal machine, not a mind reader.

Marcus: But still.

Devon: If you tell a human assistant to simply "Find a good deal" without parameters, they might also buy the wrong thing based on their own subjective interpretation of what a "deal" is. Defining parameters is fundamental to delegation.

Marcus: But a standard human assistant would pause and ask for clarification if they were about to spend thousands of dollars on coffee. They wouldn't just panic and blindly buy the thing with the lowest sticker price while ignoring the weight of the product.

Devon: That is a fair point about human common sense.

Marcus: The entire promise of an AI shopping agent, the whole reason a consumer would integrate one into their life, is to delegate the mental load. If the consumer has to sit there and engineer a hyper-specific, mathematically rigorous prompt just to stop a multi-billion-dollar supercomputer from getting tripped by a basic supermarket pricing tactic, the consumer is still doing the heavy lifting. We are just replacing the chore of comparing prices with the chore of prompt engineering.

Devon: That is a very valid critique of the user experience. And to be clear, the paper itself does distribute the responsibility across the entire ecosystem. It does not solely blame the consumer.

Marcus: Well, that is something at least.

Devon: The authors explicitly argue that developers should be building agents that have a threshold for ambiguity. The software should actively pause and clarify vague goals with the user before it starts committing funds.

Marcus: Exactly. The burden of clarity should be on the machine to ask, not the human to guess the perfect magical words. And what about the retailers in all this? Because if I am a retailer competing on genuine value, this AI behaviour is terrifying.

Devon: The paper warns retailers directly. Hiding unit price data behind complex API structures or making it computationally expensive for an agent to retrieve that data might actually steer AI agents toward worse buys.

Marcus: So, they are penalising themselves.

Devon: Yes. If you are a retailer selling a genuinely better product at a better unit price, but your digital storefront is difficult for an AI to parse, you will lose the sale to a competitor who simply slaps a lower sticker price on a smaller, inferior product.

Marcus: And this leads right into the regulatory angle, which I find fascinating. Regulators currently force physical supermarkets to show the price per 100 grams on the physical shelf.

Devon: Right, the unit pricing rules.

Marcus: It is designed to protect human eyeballs from these exact pricing tricks. But if the buyer is an API, a piece of code crawling a database, those physical shelf rules do not apply.

Devon: Which is why the authors urge regulators to step in and modernise these frameworks. We urgently need to extend unit price disclosure rules to what APIs and digital agents can see. The digital shelf is currently operating without those consumer protection guardrails, and the AI is walking right into the traps.

Marcus: It is a wild west out there. But let me ask you about the tools themselves. Do all the models behave this terribly under pressure, or are some of them actually smarter than others? Like, if I am a developer building a procurement tool, can I just buy a more advanced model and bypass this whole architecture problem?

Devon: You have to be very careful with blanket verdicts in this space. The researchers tested eight different systems across the major providers, including models from Google, OpenAI, and Anthropic.

Marcus: And did they all fail?

Devon: The general pattern of failing under cost and vague prompts held across most of them, but there were significant, named exceptions on both ends of the spectrum.

Marcus: Meaning some actually passed the test without needing the hyper-specific prompt.

Devon: Yes. For instance, an OpenAI model named Luna kept searching accurately and making optimal choices even under the high-cost inspection conditions. It successfully resisted the heuristic trap.

Marcus: Good for Luna.

Devon: But conversely, you had models that failed even worse than the baseline. Google's Gemini 1.5 Flash Lite chose badly even when complete, free information was placed directly in front of it.

Marcus: Wait, it had all the data visible, with no penalty to look at it, and it still made the wrong choice?

Devon: Exactly. In the literature, we distinguish between two types of failure. The satisficing we discussed earlier is an acquisition failure. The AI never looks for the data because it costs too much effort.

Marcus: Right.

Devon: But what Gemini 1.5 Flash Lite demonstrated was a utilisation failure. It acquired all the correct data, held it in its context window, and still failed to process it into the correct mathematical decision.

Marcus: It had the answers to the test open on the desk and still failed.

Devon: Basically, yes.

Marcus: This is honestly just another classic case of inflated, god-like AI capabilities crashing headfirst into actual empirical data. It is a lot like a callback to when we looked into those AI hacking capabilities a while ago.

Devon: Oh, absolutely.

Marcus: Everyone assumes the machine is an infallible genius until you actually test it in a messy, constrained environment, and suddenly it is behaving like a stressed-out intern making unforced errors.

Devon: It is a recurring theme whenever we look at the deployment of autonomous systems. We project perfect rationality onto them based on their theoretical capabilities, but they are operating under intense constraints, both computational limits and environmental friction.

Marcus: So, you have to test them.

Devon: Therefore, the vital takeaway for developers and businesses is that you cannot rely on assumptions. You must rigorously audit the specific model you intend to deploy within your specific information architecture.

Marcus: So, we need to leave the listener with a concrete takeaway here. No hype, just the practical reality of how to protect their money. If someone is setting up an AI agent to handle their procurement this quarter, what is the absolute rule they need to follow?

Devon: The most practical, protective step you can take is this: if you are going to let an AI agent spend your money, you must tell it exactly what the word "best" means.

Marcus: Be ruthlessly specific. Define the metric.

Devon: Define the metric in mathematical terms. Use exact phrases like "lowest price per ounce" or "lowest cost per unit volume." You must completely banish vague colloquialisms from your prompts.

Marcus: So, no "find the best deal."

Devon: Never say "find the best deal" and never say "keep it cheap." The empirical data proves that when you define the metric precisely, the agent fetches the necessary data, it does the maths perfectly, and supermarket pricing psychology largely loses its grip on the machine.

Marcus: Treat it like a literal-minded machine, not a human who intuitively understands value. Which leaves me with one final, slightly eerie thought to mull over.

Devon: Go on.

Marcus: We have spent this entire deep dive talking about how human pricing tricks accidentally work on artificial intelligence. But think about the trajectory of e-commerce. If AI agents become the primary buyers in the online economy, if they end up doing the majority of corporate procuring and household grocery shopping, retailers are going to notice that shift.

Devon: They absolutely will.

Marcus: Will they stop trying to trick the human mind and instead start designing digital storefronts specifically engineered to exhaust and manipulate algorithms?

Devon: That would represent a profound shift in market dynamics. An entirely new field of adversarial retail architecture.

Marcus: Right. What does a supermarket designed to trick a machine actually look like? It probably does not have bright red sale tags or end-of-aisle displays. It probably has massive data retrieval costs, labyrinthine API endpoints, and obfuscated pricing structures designed specifically to force an AI into that state of algorithmic panic we talked about. We thought we were building a perfectly rational robotic butler to protect us from the marketers, but it turns out the marketers might just learn to hack the butler.