When AI Personalization Helps—and When It Starts to Feel Creepy

This is an executive digest of the scientific paper “When AI Advice Backfires: Field Experimental Evidence on Personalization, Intrusiveness, and Purchase Behavior,” published in the Journal of Retailing and Consumer Services in 2026. We did not conduct this experiment. Our purpose is to translate the researchers’ findings into a concise, practical format for marketing leaders.

Personalization has become one of marketing’s default promises. Give an AI system enough customer data, the logic goes, and it will deliver more relevant recommendations, smoother experiences, and higher conversion rates.

But there is a point where relevance begins to feel like surveillance.

What the researchers tested

The researchers wanted to understand how the communication style of an AI shopping assistant affects actual purchases—not merely whether consumers say they might buy something.

The project combined two methods: 46 semi-structured interviews and a randomized field experiment involving 409 customers of a US retailer. Participants interacted with an AI shopping assistant and received one of two versions of a product recommendation: personalized or humorous.

The recommended product and core rationale remained the same. Only the message framing changed. This helped isolate the effect of communication style rather than the quality of the recommendation itself.

Importantly, the researchers connected survey responses to transaction records. The primary outcome was therefore a completed purchase, not self-reported purchase intention—a common limitation in marketing studies.

Personalization increased purchasing—but not without a cost

Personalized responses produced a statistically significant increase in purchase behavior compared with humorous ones. The logistic regression coefficient was 0.42, with a p-value of 0.005. The corresponding odds ratio was 1.52, meaning that personalized messages increased the odds of purchase by approximately 52%.

The finding was consistent across different conversion windows:

• same-session purchase: p = 0.018;
• purchase within three days: p = 0.009;
• within seven days: p = 0.006;
• within fourteen days: p = 0.011.

The effect was not limited to conversion. Among the additional behavioral outcomes, personalized communication was associated with an approximately 6.8% higher order value (p = 0.024) and a 7% increase in the expected number of items purchased, expressed as an incidence rate ratio of 1.07 (p = 0.041).

But the most important result is not simply that personalization “works.” It works through two competing psychological mechanisms.

When customers interpreted the AI’s message as genuinely helpful, their odds of purchase increased. The odds ratio for perceived helpfulness was 1.78 (p < 0.001). In contrast, perceived intrusiveness reduced the odds of purchase: the corresponding odds ratio was 0.63 (p < 0.001).

In plain language, personalization can create two simultaneous reactions:

“This system understands what I need.”

and:

“Why does this system know so much about me?”

The final commercial outcome depends on which reaction becomes stronger.

The same message does not work equally well for everyone

Context played an important role. Shopping urgency amplified both sides of personalization. When customers needed to make a decision quickly, tailored advice felt more useful—but the implied use of personal data could also feel more intrusive.

The customer’s existing relationship with the retailer changed the response as well. Customers with a stronger relationship were more likely to interpret personalization as legitimate assistance. Relationship strength also significantly reduced intrusiveness concerns (p = 0.008).

This has an important implication for segmentation. A message that feels convenient to a loyal customer may feel unsettling to a first-time visitor. Personalization is evaluated relative to the level of trust the brand has already earned.

What CMOs should take from the study

The findings challenge the assumption that more customer knowledge should automatically produce more visible personalization. A company may have access to dozens of behavioral signals, but displaying them can make the technology feel less helpful, not more intelligent.

The objective should therefore be appropriate personalization, not maximum personalization.

That means using information customers can reasonably expect the brand to possess, making the reason for a recommendation understandable, and matching the depth of personalization to the maturity of the customer relationship. Current-session context may feel natural; an unexpected reference to historical behavior may require transparency, permission, or an explanation.

There is one important limitation. The experiment compared personalized messages with humorous ones, not with a neutral control condition. It therefore does not prove that personalization will outperform every communication style. It shows that, in this retail setting, personalized framing performed better than humorous framing—and that its effectiveness depended on the balance between helpfulness and intrusiveness.

The strategic lesson is straightforward: AI personalization should not demonstrate how much a company knows about a person. It should demonstrate how well the company understands what kind of help is appropriate.

Original paper: When AI Advice Backfires: Field Experimental Evidence on Personalization, Intrusiveness, and Purchase Behavior.