Extract Win/Loss Insights from Transcripts

Use the Claude API to systematically analyze win/loss interview transcripts and survey responses, extracting structured insights categorized by theme.

Prerequisites

• Anthropic API key from console.anthropic.com
• Interview transcripts (from Fireflies, Grain, or manual notes)
• n8n instance for automation (recommended)

Steps

1. Prepare the transcript for analysis. Retrieve the transcript from Fireflies API or your storage. Clean it: remove filler words sections, trim speaker labels to first name only, and split into chunks of 50,000 tokens max if the transcript exceeds Claude's efficient processing range.

2. Call the Claude API with a structured extraction prompt. Use claude-sonnet-4-20250514 for cost-effective analysis:

   POST https://api.anthropic.com/v1/messages
   Headers: x-api-key: {ANTHROPIC_API_KEY}, anthropic-version: 2023-06-01
   Body: {
     "model": "claude-sonnet-4-20250514",
     "max_tokens": 4096,
     "temperature": 0.2,
     "messages": [{"role": "user", "content": "TRANSCRIPT:\n{transcript}\n\nAnalyze this win/loss interview transcript. Extract:\n1. OUTCOME: Won or Lost\n2. PRIMARY_REASON: The single biggest factor in the decision (1 sentence)\n3. COMPETITORS_MENTIONED: List of competitor names mentioned\n4. PRODUCT_FEEDBACK: Specific product strengths or gaps mentioned (list each with a direct quote)\n5. SALES_PROCESS_FEEDBACK: What the buyer said about the sales experience (list each with a direct quote)\n6. PRICING_FEEDBACK: Any comments on pricing, packaging, or value perception\n7. DECISION_CRITERIA: What mattered most to this buyer (ranked list)\n8. ACTIONABLE_INSIGHTS: 2-3 specific, concrete things we should change based on this interview\n9. SENTIMENT_SCORE: 1-10 rating of overall buyer sentiment toward us\n10. KEY_QUOTES: 3-5 verbatim quotes that are most revealing\n\nReturn as JSON."}]
   }
   

3. Parse the JSON response. Extract the structured fields from Claude's response. Validate that all 10 fields are present. If any field is null or missing, re-prompt with a follow-up asking specifically for that field.

4. Categorize insights by theme. Group the ACTIONABLE_INSIGHTS into standard categories:

   • product-gap — Missing feature or capability
   • product-strength — Feature that drove the win
   • pricing — Price, packaging, or discount-related
   • sales-process — How the sales experience affected the outcome
   • competitive — Competitor-specific intelligence
   • messaging — How our positioning/messaging landed
   • timing — Budget cycle, urgency, or timeline factors

5. Store insights in your CRM. Push the structured insights to Attio as a note on the deal record via the Attio MCP. Include: outcome, primary reason, competitors mentioned, sentiment score, and each actionable insight as a separate line. Tag the note as "win-loss-insight" for later querying.

6. Batch processing for multiple transcripts. When processing a backlog, loop through transcripts sequentially with a 1-second delay between API calls to respect rate limits. Track processing status: queue each transcript, mark as "processing", then "complete" or "failed". Log any failures for manual review.

7. Quality check the extraction. For the first 5 transcripts, manually compare Claude's extraction against the raw transcript. Check: Are the key quotes actually verbatim? Are the insights genuinely actionable (not vague)? Is the sentiment score reasonable? Adjust the prompt if extraction quality is below 90% accuracy.