Incremental AI Adoption for E-commerce
Source: arcturus-labs.com
Four Levels of AI Integration
Level 0: Traditional Search
Keyword box + sidebar filters + results. Burden entirely on the user to know terminology and filter structure. High bounce rates for non-expert users.
Level 1: Beginner AI
Keep traditional search unchanged. Add an async “Did you mean this?” suggestion bar after results load.
- AI interprets natural language → suggests better filter/sort combo
- Zero latency risk (async, appears after page loads)
- Takes ~50px vertical real estate
- Measures: impression rate, click-through, conversion
Implementation: define a search tool with well-specified arguments (keyword text + facet selections + sort), call it asynchronously on every search.
Level 2: Intermediate AI
AI automatically executes the rewritten query. “Interpreted as: Property Type: Condo, Terms: ‘Downtown’”
- Latency increases 2-3s (acceptable because AI saves user 5+ seconds of manual reformulation)
- Adds result summary + recommended next queries above results
- A/B test: does auto-rewriting improve or hurt conversion?
Level 3: Advanced AI (Conversational)
Replace search box with chat interface. Stateful conversation — agent remembers prior turns, user can course-correct.
- Users express intent naturally (“Show me some big ass houses”)
- Agent clarifies, asks follow-ups, discusses results
- Can perform aggregate analysis (“what’s the typical price distribution?“)
- Traditional metrics + LLM conversation analysis to understand user journeys
Architecture Philosophy
Nothing magical: an HTML request (text in, text out) + a couple of for-loops = an agent.
- Inner loop: LLM + tool calls until sufficient
- Outer loop: multi-turn user conversation
Your existing search infra (Elasticsearch, Algolia) stays in place. The AI is a thin, decoupled layer that translates intent into search calls.
Key Principle
Start with Level 1 (zero risk), gather data, proceed only if metrics improve. “AI is saving [users] a lot of effort” is the bar for accepting increased latency.