Abstract
How does the timing of an AI assistant's approach affect consumer behavior in the Metaverse? This study investigates the impact of Immediate vs. Delayed interference on user experience, decision confidence, and purchase intention.
The Experimental Scenario
Participants were placed in a Virtual Reality showroom and tasked with choosing a 3D printer. The complexity of the product data (specs, price, kit vs. assembled) created a realistic cognitive load, making the role of the AI assistant crucial.
Product Options
1. Explorer
- Price: €229.99
- Type: Kit (Bausatz)
- Build Size: 350x350x250mm
- Heated Bed: No
- Materials: PLA, PETG
- Entry level kit. Requires assembly. Lacks heated bed, limiting material options.
2. Solid
- Price: €499.99
- Type: Kit (Bausatz)
- Build Size: 400x400x405mm
- Heated Bed: Yes
- Materials: PLA, PETG, ABS
- Mid-range kit. Large build volume and heated bed allow for more durable materials.
3. Plus
- Price: €659.99
- Type: Ready to Use
- Build Size: 430x400x435mm
- Heated Bed: Yes
- Materials: PLA, PETG, ABS
- Premium experience. Comes assembled. Cited as 'Good value for money' in the study.
Task: Evaluate these options. In the study, the AI agent would approach either immediately upon entry or delayed (after 5+ seconds of browsing).
The GenAI Pipeline
The sales agent wasn't a static script. It utilized a real-time Generative AI pipeline to converse naturally with participants.
Pipeline Steps
-
User Speech — Voice Input
- The participant speaks naturally to the avatar in the VR environment, asking questions like "Which printer is best for beginners?"
-
Speech-to-Text — Azure STT
- The audio input is captured and converted into text string using Azure Cognitive Services, handling ambient noise and voice clarity.
-
LLM Processing — GPT-4 / OpenAI
- The text query is sent to a Large Language Model (GPT-4). The model has system prompts defining it as a 'helpful sales assistant' aware of the product catalog.
-
Text-to-Speech — Azure TTS
- The LLM's text response is converted back into synthesized speech (Azure Neural TTS) and lip-synced by the 3D avatar.
Quantitative Results
Comparing the Immediate (Agent interrupts instantly) vs. Delayed (Agent waits for browsing) conditions across 100 participants.
Note: Interactive charts are not available in this version. View the original Hugo site for visualizations.
Key Finding: Timing Matters
The Delayed condition significantly outperformed Immediate interference across all positive metrics. Users preferred autonomy before assistance.
Hypothesis Validation
- ✓ H1: Delayed ↑ Usefulness
- ✓ H2: Delayed ↓ Intrusiveness
- ✓ H3: Delayed ↑ Confidence
- ✓ H4: Delayed ↑ Purchase Intent
Research Data Summary
User Perception (Likert Scale 1-5):
- Perceived Usefulness: Delayed (3.86) vs Immediate (2.93)
- Perceived Intrusiveness: Delayed (1.92) vs Immediate (3.87)
Decision Outcomes (Likert Scale 1-5):
- Decision Confidence: Delayed (4.12) vs Immediate (3.45)
- Purchase Intention: Delayed (3.78) vs Immediate (3.10)
Implications for V-Commerce
Respect Autonomy
Users in VR environments value the ability to explore independently. Premature intervention by AI agents breaks immersion and is perceived as annoying rather than helpful.
Context Awareness
Sales agents should be programmed to detect "browsing behavior" (e.g., viewing multiple products) before offering assistance, mirroring effective human sales strategies.
Conclusion
This study demonstrates that the timing of GenAI sales agent intervention is critical in VR commerce environments. A delayed approach that respects user autonomy leads to:
- Higher perceived usefulness
- Lower perceived intrusiveness
- Greater decision confidence
- Increased purchase intention
For VR commerce practitioners, the recommendation is clear: implement browsing-aware AI agents that detect user readiness before offering assistance, rather than interrupting the exploration experience.