Do human shopping tactics work the same way on AI agents?

In a Substack essay on whether behavioral pricing tactics still work now that AI agents shop on consumers' behalf, Utpal Dholakia reports a study by Sabbah and Acar, published in Harvard Business Review, that ran "thousands of simulated shopping rounds with four AI models covering four different product categories." The study found that marketing tactics designed for human buyers — "scarcity messages, countdown timers, strikethrough prices, vouchers, and bundles" — "had inconsistent effects on machine purchases, varying by model." By contrast, "only star ratings based on reviews consistently increased choice by AI agents," while "higher prices consistently reduced it."

Sabbah and Acar's own takeaway, as Dholakia quotes it, is to "treat AI models as distinct segments": pricing and marketing teams can no longer assume one tactic generalizes across agents the way it once generalized, imperfectly, across human shoppers. A tactic that moves one model's choices may do nothing, or something different, in another. Dholakia sets this against the premise behind classic behavioral pricing: charm prices (ending in 9 or 99) and partitioned pricing (splitting a total into a base price plus smaller add-ons) work on humans because of how people read and weight digits and anchor on a base figure — mechanisms that presume a human reading a price, not a model parsing structured product data.

Dholakia's own reading cautions against inferring from this inconsistency that machine buyers are therefore rational: "the fact that marketing tactics designed for humans fail with machine buyers does not mean that the machine buyers are rational." That caution sits well against When do pricing cues actually mislead AI shopping agents?, which likewise locates agent susceptibility in prompt and task conditions rather than fixed irrationality, and against Do shopping agents respond more strongly to price and rating cues? — the "one comprehensive study tested 17 AI models" that Dholakia cites later in the same essay, reporting the same roughly 89%-vs-4% position effect and 30%-to-80%-vs-5% rating effect. Sabbah and Acar's finding adds a complication to that picture: not every manipulation that works on humans, or that moves one model, moves agents consistently as a class — ratings and price read as robust signals, but scarcity and bundling tactics do not.

The excerpt does not report which four models were tested, the product categories, effect sizes, or whether the inconsistency reflects differences in training data, system prompts, or retrieval tooling across models — Dholakia's summary of the HBR piece gives only the qualitative pattern. It also leaves his own follow-on questions — what agents respond to instead, and who decides an agent's decision calculus — unanswered within this excerpt (see Does AI agent resistance to marketing prove buyer rationality?). The implication the evidence supports is narrower but still useful: a pricing team optimizing for the AI shopper as a single persona is working from a false premise, since tactic effects need to be tested per model rather than assumed to transfer from human research, or from one AI agent to the next.