Agentic shopping assistants increasingly move beyond recommending products to executing actions in users’ workflows (e.g., adding items to cart, applying coupons, selecting shipping). This shift from advice to delegation raises questions about appropriate reliance, perceived control, and how interface cues support oversight when systems can act. We report a laboratory eye-tracking experiment using a chat-only e-commerce prototype in a mixed 2 × 2 design: action autonomy varied within participants (recommend-only vs. act-on-behalf, with undo/edit), and transparency cues varied between participants (minimal statements vs. preview + rationale describing what will happen and why). Three standardized shopping tasks were completed by 72 participants. Results included behavioral logs (task time, overrides), areas-of-interest (AOI)-based eye-tracking (chat attention and verification indicators), and post-task self-reports (trust, control, uneasiness, perceived transparency). Act-on-behalf autonomy reduced completion time, but it also increased unease, decreased trust and perceived control, and increased the likelihood of an override, suggesting a trade-off between efficiency and oversight. The autonomy-related penalties for trust and perceived control under act-on-behalf execution were lessened by preview + rationale transparency, which additionally enhanced perceived transparency, trust, and unease. This mechanism coincided with eye-tracking: transparency decreased verification latency during agent actions and redirected attention toward information supplied by assistants. Transparency did not reliably reduce overrides, suggesting that minimal effective transparency can streamline supervision and improve evaluations without eliminating corrective behavior.
Balaskas et al. (Sun,) studied this question.