Skip to content

Comparing the Three Use Cases

Overview Matrix

Aspect Product Search Support Tickets Recommendations
Frequency Per user query (1-5/user/month) Per issue (5/user/year) Often (multiple/session)
Latency requirement <100ms p95 <2s <500ms
Cost per call $0.001 $0.001 $0.001
Effective cost $0.0006 (40% cache) $0.0001 (90% simple) $0.0004 (60% cache)
Complexity Ranking Text generation Ranking + explanation
Quality impact Revenue +15% Cost -90% Revenue +15%
Risk if AI fails Mediocre results Escalated to human Falls back to ML
Best for E-commerce Support Any marketplace

Cost-Benefit Comparison

Annual Impact (1M users, 10M queries served)

Product Search

Text Only
Effective cost: 10M * $0.0006 = $6K
Incremental revenue (15% * $50 AOV): +$7.5M
ROI: 1,250,000x

Support Tickets

Text Only
Small query volume but huge per-ticket savings
Effective cost: 100K tickets * $0.0001 = $10
Cost savings (92% reduction * $6.67/ticket): +$617K
ROI: 61,700x

Recommendations

Text Only
Effective cost: 10M * $0.0004 = $4K
Incremental revenue (15% * $50 AOV): +$7.5M
ROI: 1,875,000x

Implementation Complexity

Text Only
Simplicity Ranking (simplest to hardest):

1. Support Tickets (★★)
   - Input: User's issue (text)
   - Output: Helpful response (text)
   - Validation: Easy (just check it's relevant)
   - Error handling: Simple (escalate if bad)

2. Product Search (★★★)
   - Input: Query + product catalog
   - Output: Ranked list
   - Validation: Must check accuracy
   - Error handling: Fallback ranking

3. Recommendations (★★★★)
   - Input: User profile + huge catalog
   - Output: Personalized ranked list
   - Validation: Must align with business rules
   - Error handling: Fallback to ML model

When to Use Each

Good for: - E-commerce sites - Dense product catalogs (>10K items) - Users searching with natural language - Sites with high AOV

Avoid: - Small catalogs (<1K items) - Users click through lists anyway - Keyword search is good enough

Support Tickets

Good for: - High support volume - Many FAQ-style questions - Need immediate customer response - Want to reduce support costs

Avoid: - Complex support (needs specialized agent) - Legal/compliance heavy - Complaints that need empathy (LLM can't always deliver)

Recommendations

Good for: - Marketplaces (Amazon-like) - Users with diverse preferences - Want to increase AOV - Have user history/profile

Avoid: - Cold-start users (no profile) - Simple stores (<100 products) - Algorithmic ranking works well

Decision Tree

Text Only
Should I use AI for this feature?

    START
       ↓
   Does it involve
   ranking/ordering?
       │
       ├──Yes→ Use AI? (Search or Recommendations)
       └──No
           ↓
       Is it generating
       customer-facing text?
           │
           ├──Yes→ Use AI (Support Tickets)
           └──No
               ↓
           Probably don't use AI
           (simpler solution exists)

Learning Path

Start with Support Tickets: - Simplest to implement - Easy to demonstrate value - Quick wins build confidence

Then Product Search: - More complex ranking - Better ROI on e-commerce - Good for interview discussion

Then Recommendations: - Most complex - Highest ROI - Demonstrates full system design

Interview Prompt

"Design an AI system for a new use case of your choice."

Suggested: Product Search (middle complexity)

Structure: 1. Explain the problem (traditional search is generic) 2. Show AI solution (LLM-based ranking) 3. Discuss architecture (service layer) 4. Analyze cost-benefit 5. Handle edge cases (cache, fallback, errors) 6. Monitor in production