Reply Analyzer

Reply Analyzer

Classifies the sentiment and intent of replies to outreach messages. Determines the recommended next action, generates draft responses for positive and neutral cases, and outputs CRM stage updates. Used by the evening-scan playbook after inbox-scanner returns responses.

Prerequisites

• agency.config.json at repo root with outreach, services, case_studies
• Reply data from inbox-scanner or manually provided

Phase 0: Intake

1. Read agency.config.json from the project root.
2. Extract:
   • services -- list of services offered (for matching interest areas)
   • case_studies -- available case studies (for draft responses)
   • outreach.tone -- voice guidelines for draft responses
   • outreach.banned_phrases -- phrases to avoid in draft responses
   • outreach.sign_off -- signature line for draft responses
3. Accept parameters:
   • reply_text -- the full text of the reply (required)
   • from_name -- name of the person who replied
   • from_email -- their email address (if available)
   • company -- their company name (if known)
   • channel -- where the reply came from (email, linkedin, instagram)
   • lead_context -- object with cadence day, previous touches, lead score, tier

Phase 1: Sentiment Classification

Analyze the reply text and classify it into exactly one of these categories:

Classification rules:

• If multiple signals are present, choose the strongest positive category. A reply that asks a question AND requests a call is POSITIVE_DEMO, not NEUTRAL_QUESTION.
• Short replies ("Thanks", "Got it", "OK") without additional context: classify as NEUTRAL_QUESTION with LOW confidence.
• Replies that are clearly automated (vacation responders, ticket confirmations): classify as NEUTRAL_OOO.
• Any mention of "unsubscribe", "remove", "stop", or "spam" overrides all other signals and forces NEGATIVE_UNSUBSCRIBE.

Phase 2: Intent Extraction

Beyond the top-level sentiment, extract these structured details from the reply:

Questions asked

List every question the person asked, verbatim or paraphrased:

• "What are your rates?"
• "How long does a CRO audit take?"
• "Do you work with brands outside India?"

Objections raised

Identify any objections or hesitations:

• Price: "Too expensive", "What's the budget?", "We don't have budget right now"
• Timing: "Not the right time", "Maybe next quarter", "After our funding round"
• Existing vendor: "We already work with an agency", "We just signed with someone"
• Platform: "We're not on Shopify", "We use WooCommerce", "Thinking of migrating to Shopify"
• Trust: "Never heard of you", "Can you share references?", "How many stores have you done?"
• Scope: "We only need X, not Y", "This is too comprehensive"

People mentioned

Extract any person references that could be referrals:

• Name (if mentioned)
• Title/role (if mentioned)
• Company (if different from the replier's company)
• Contact info (if shared)

Timeline indicators

Extract any timing references:

• Specific dates: "after March", "in Q2", "next month"
• Relative: "in 2 weeks", "next quarter", "after funding"
• Vague: "sometime later", "eventually"
• Urgent: "ASAP", "this week", "immediately"

Interest areas

Map the reply to specific services from agency.config.json:

• Which service(s) they seem interested in
• Any specific problems they mentioned (e.g., "our conversion rate is low" maps to CRO)

Phase 3: Recommended Action

Based on the classification, recommend a specific next action and CRM stage update:

Draft response generation

Generate a draft response for POSITIVE_DEMO, POSITIVE_INTERESTED, NEUTRAL_QUESTION, NEGATIVE_NOT_NOW, NEGATIVE_NOT_INTERESTED, and NEGATIVE_UNSUBSCRIBE categories.

Draft response rules:

• Follow the outreach.tone from config: direct, helpful, zero fluff
• Never use any phrase from outreach.banned_phrases
• Keep it under 100 words for email, under 50 words for LinkedIn/Instagram
• Use the replier's first name
• For positive replies: reference the most relevant case_studies entry based on their industry/interest
• For demo requests: include a calendar link placeholder [CAL_LINK]
• Sign off with outreach.sign_off
• Match the channel tone: email can be slightly longer, LinkedIn/Instagram should be conversational and brief

Phase 4: Output

Return structured JSON:

{
  "classification": "POSITIVE_INTERESTED",
  "sentiment": "positive",
  "confidence": "HIGH",
  "summary": "Interested in CRO audit services, asked about pricing and timeline",
  "questions_asked": [
    "What are your rates?",
    "How long does a CRO audit take?"
  ],
  "objections": [],
  "referrals": [],
  "timeline_mentioned": null,
  "interest_area": "Conversion Rate Optimization",
  "recommended_action": "Reply with CRO audit pricing and Kibi Sports case study. Suggest 15-min call.",
  "crm_stage_update": "REPLIED",
  "pause_cadence": true,
  "reactivation_date": null,
  "draft_response": "Hey Sarah, glad this caught your eye. For CRO, we typically start with a full audit (takes about a week) -- for reference, our Kibi Sports audit uncovered issues causing 15-25% revenue loss, and the redesign drove measurable uplift in conversions.\n\nHappy to walk through what a quick audit of your store would look like. Got 15 min this week?\n\nEkata | Plasho (plasho.com)",
  "new_leads_to_create": []
}

For REFERRAL classification, the new_leads_to_create array is populated:

{
  "classification": "REFERRAL",
  "sentiment": "neutral",
  "confidence": "MEDIUM",
  "summary": "Redirected to their Head of Ecommerce, Priya Sharma",
  "questions_asked": [],
  "objections": [],
  "referrals": [
    {
      "name": "Priya Sharma",
      "title": "Head of Ecommerce",
      "company": "Brand X",
      "email": null,
      "source": "Referral from Sarah Chen"
    }
  ],
  "timeline_mentioned": null,
  "interest_area": null,
  "recommended_action": "Thank Sarah for the referral. Research Priya Sharma at Brand X. Add as new lead with referral source.",
  "crm_stage_update": "REPLIED",
  "pause_cadence": true,
  "reactivation_date": null,
  "draft_response": "Thanks Sarah, really appreciate the pointer. Will reach out to Priya.\n\nEkata | Plasho (plasho.com)",
  "new_leads_to_create": [
    {
      "name": "Priya Sharma",
      "title": "Head of Ecommerce",
      "company": "Brand X",
      "email": null,
      "source": "Referral from Sarah Chen",
      "priority": "HIGH"
    }
  ]
}

Confidence levels:

• HIGH: Reply is unambiguous, classification is clear from the text
• MEDIUM: Reply has mixed signals or is short/ambiguous, classification is best guess
• LOW: Reply is very short (1-3 words), automated, or context-dependent

Example Usage

Trigger phrases:

• "Analyze this reply"
• "Classify this response"
• "What should I do with this reply?"
• "Run reply analyzer on these responses"

User: Analyze this reply: "Hey, interesting work on Kibi Sports. We're actually looking at overhauling our Shopify store. What would something like that cost for a mid-size D2C brand? Could do a call next week."
Assistant: [classifies as POSITIVE_DEMO, extracts pricing question + timeline "next week", recommends sending calendar link with 2-3 slots, generates draft response referencing Kibi Sports case study]

User: What do I do with this: "Thanks but we just hired an agency last month."
Assistant: [classifies as NEGATIVE_NOT_NOW, extracts objection "existing vendor", recommends polite acknowledgment with 90-day reminder, generates short gracious reply]