Incremental AI Adoption for E-commerce

Source: arcturus-labs.com

Four Levels of AI Integration

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.