Introduction
This guide will help you think strategically about which LLM is best for each use case. While it's tempting to use the same model for everything, a multi-model approach can be:
More cost-effective
Faster and more accurate
A better overall experience
Looking for technical details? See the companion document:
Before even considering LLMs, think about the workflow you are doing.
As an example, if you are just curious whether a specific phrase or word is popping up in a set of DSAT tickets, using a LLM to find this would be an inefficient and costly approach to using LLMs.
What LLM Model to Try for Different Workflows
1. Monthly Trend Analysis & VOC Reports
What you're doing:
Analyzing 1000+ tickets per month
Identifying top customer pain points
Creating executive summaries
Finding patterns across conversations
Recommended Model: Claude Sonnet 4
Why this works best:
✅ Superior at identifying patterns across large datasets
✅ Better at explaining why trends exist, not just what they are
✅ Can process hundreds of tickets together to find connections
✅ Generates comprehensive, well-structured reports
✅ Nuanced understanding of customer frustration vs. minor complaints
Example prompt:
"I'm providing 500 support tickets from March. Identify the top 5 customer pain points, explain what's causing each issue, provide 3 example tickets for each pain point, and recommend actions we should take."
2. Automated Ticket Categorization
What you're doing:
Tagging tickets by issue type (billing, technical, returns, etc.)
Routing tickets to appropriate teams
Organizing tickets for analysis
Processing thousands of tickets daily
Recommended Model: GPT-4o Mini OR Gemini 2.5 Flash-Lite
Why this works best:
✅ Very cost-effective for high-volume processing
✅ Accurate for straightforward classification tasks
✅ Dramatically lower cost than premium models
Example prompt:
"Categorize this ticket into one of these categories: billing_issue, technical_support, product_question, shipping_inquiry, return_request, general_feedback. Return only the category name."
3. Root Cause Analysis
What you're doing:
Understanding why certain issues keep happening
Analyzing customer journeys that led to escalations
Investigating spikes in specific complaint types
Creating detailed analysis for stakeholders
Recommended Model: Claude Sonnet 4.5
Why this works best:
✅ Best reasoning capabilities for complex analysis
✅ Excellent at connecting patterns across multiple data points
✅ Can generate very detailed reports (up to 48 pages)
✅ Strong at explaining causal relationships
✅ Methodical in breaking down complex problems
Example prompt:
"We saw a 40% increase in cancellation requests from customers who had billing issues in the past 60 days. Analyze these 200 tickets to understand: what's the connection between billing issues and cancellations? What happened in the customer journey? What could we have done differently?"
4. Customer Journey Mapping
What you're doing:
Tracking individual customers across multiple interactions
Understanding touchpoints before churn or escalation
Analyzing long-term customer relationships
Creating cohort analysis
Recommended Model: Claude Sonnet 4 OR Gemini 2.5 Flash
Why this works best:
✅ Large context window to see entire customer history
✅ Good at temporal understanding (what happened when)
✅ Can track sentiment changes over time
✅ Identifies patterns in multi-touch journeys
Example prompt:
"This customer has had 8 interactions with support over 6 months before canceling. Analyze the complete journey: what was the sentiment trend? When did things go wrong? Were there warning signs we missed?"
5. High-Volume Text Summarization
What you're doing:
Creating executive summaries of hundreds of tickets
Condensing long tickets into brief overviews
Generating daily/weekly digest reports
Processing large volumes efficiently
Recommended Model: Gemini 2.5 Flash
Why this works best:
✅ Can process massive amounts of text in one pass (1M tokens)
✅ Fast processing for time-sensitive reports
✅ Cost-effective for volume work
✅ Good quality summaries
Example prompt:
"Summarize these 800 support tickets from this week into a 1-page executive summary covering: volume by category, top 3 issues, sentiment distribution, and notable customer feedback."
Quick Decision Framework
Start with these questions:
1. Do you need this done in real-time?
Yes → GPT-4o or Gemini Flash
No → Consider Claude Sonnet for better depth
2. How many tickets/items are you processing?
1-50 at a time → Any model works
100-500 → Use models with 128K+ context
1000+ → Use Gemini Flash or Claude extended context
3. How complex is the analysis?
Simple (categorization, basic sentiment) → GPT-4o Mini or Flash-Lite
Moderate (quality scoring, trend identification) → GPT-4o
Complex (root cause, strategic insights) → Claude Sonnet 4.5
4. What's your budget?
Very cost-conscious → Mini/Lite models
Balanced → GPT-4o, Gemini Flash
Quality is priority → Claude Sonnet models
5. How long of a response do you need?
Short (few paragraphs) → Any model
Medium (2-10 pages) → Any standard model
Long (10+ pages) → Claude Sonnet 4.5 or GPT-5
Testing Recommendations
Before committing to a model:
Test with your actual tickets - model performance varies by content type
Compare 3-5 models on the same task and prompt
Measure:
Accuracy/quality of results
Processing time
Cost per ticket
Output completeness
Start small - test with 20-50 tickets before scaling
Use filters on your Worksheet to narrow down results to a set of tickets you have full understanding of
Example Multi-Model Approach:
GPT-4o Mini → Daily ticket categorization (high volume, simple)
GPT-4o → Daily quality scoring (moderate complexity)
Claude Sonnet 4 → Weekly trend reports (deep analysis)
Claude Sonnet 4.5 → Monthly executive VOC reports (strategic insights)
Cost optimization: This approach uses expensive models only where they add real value
FAQs
Can I use different models for different tasks?
Yes! This is actually recommended. Use budget models for simple tasks, premium models for complex analysis.
Should I always use the largest context window?
No - larger contexts can be slower and more expensive.
Can these models handle industry-specific terminology?
Yes, but don't leave room for LLMs to interpret your organization. Control details to avoid LLMs from making assumptions that can overtly derail your analysis.