conversational-intelligence

Conversational Intelligence Extractor

Purpose

Turn call transcripts into structured, actionable intelligence. Works from any call transcript source: Fireflies, Gong, Chorus, or plain transcript files exported locally. Outputs JSON-structured insights that drive handoff automations, competitive displacement campaigns, expansion plays, and product feedback loops.

Required Tools

• A call transcript source: any of the following works
  • Fireflies (API or MCP, if connected)
  • Gong or Chorus exports
  • Plain transcript files on disk (.txt, .vtt, .srt, JSON, markdown)
• Optional: an enrichment tool of your choice (Clay, Apollo, or any data provider) to enrich mentioned companies and contacts. The skill works fully without it.

Inputs

Required

• extraction_type: One of: handoff | competitive | expansion | closed-lost | full (runs all)
• target: One of:
  • A transcript ID or file path (specific call)
  • A company name/domain (pulls all recent calls mentioning that company)
  • A date range + keyword (searches across all calls)

Optional

• output_label: A label (e.g. the account name) used in output filenames
• enrich: Boolean, default false. Enrich mentioned companies/contacts via your enrichment tool of choice
• output_format: json | markdown | both (default: both)

Process

Step 1: Retrieve Transcripts

Based on the target input, pull the right transcripts from your transcript source.

If transcript ID or file path provided:

→ Fetch that transcript (and its AI summary, if your source provides one)

If company name/domain provided:

→ Search your transcript source for the company name (limit ~10 results)
→ For each result: fetch the full transcript

If date range + keyword:

→ Search your transcript source by keyword within the date range (limit ~20 results)
→ For each result: fetch the full transcript

With local files, "search" means grepping the transcript directory for the company name or keyword and filtering by file date.

IMPORTANT: When your source provides both a summary and a full transcript, always pull both. The summary gives quick orientation; the transcript gives verbatim quotes and speaker attribution.

Step 2: Extract Structured Intelligence

Run extraction based on extraction_type. For full, run all four in sequence.

2A: Handoff Extraction (handoff)

Extract from ALL pre-sale conversations to build complete customer context.

Schema:

{
  "handoff_intelligence": {
    "account_name": "string",
    "extraction_date": "YYYY-MM-DD",
    "calls_analyzed": "number",
    "icp_alignment": "string: fit assessment with specific evidence",
    "org_structure": "string: reporting lines, team sizes, relevant departments",
    "primary_challenges": [
      {
        "challenge": "string: specific problem",
        "severity": "critical | high | medium",
        "quote": "string: verbatim from transcript",
        "speaker": "string: name and role"
      }
    ],
    "red_flags": [
      {
        "flag": "string: risk description",
        "source": "string: which call, who mentioned it",
        "mitigation": "string: suggested approach"
      }
    ],
    "expansion_potential": "string: adjacent use cases or teams mentioned",
    "buying_motivations": [
      "string: specific reasons they chose us (verbatim language)"
    ],
    "key_stakeholders": [
      {
        "name": "string",
        "role": "string",
        "influence": "decision_maker | influencer | end_user | champion | blocker",
        "concerns": ["string: specific concerns raised"],
        "communication_style": "string: brief note on how they communicate"
      }
    ],
    "implementation_context": {
      "timeline_expectations": "string",
      "technical_requirements": "string",
      "success_criteria": "string: what does 'working' look like to them"
    }
  }
}

Edge cases:

• If org structure is unclear, note "Org structure not discussed: confirm in first CX call"
• If no red flags found, set to empty array (don't invent risks)
• For influence level, default to "end_user" if role is ambiguous

2B: Competitive Intelligence Extraction (competitive)

Scan all transcripts for competitor mentions, budget data, renewal timelines.

Schema:

{
  "competitive_intelligence": {
    "account_name": "string",
    "extraction_date": "YYYY-MM-DD",
    "calls_analyzed": "number",
    "competitors_detected": [
      {
        "competitor_name": "string",
        "competitor_domain": "string: if identifiable",
        "current_spend": "string or null: '$X annually/monthly'",
        "contract_end_date": "string or null: 'YYYY-MM-DD' or 'QX YYYY'",
        "renewal_timeline": "imminent | within_3_months | within_6_months | beyond_6_months | unknown",
        "primary_frustrations": ["string: use exact customer language"],
        "satisfaction_level": "satisfied | neutral | frustrated | actively_looking_to_switch",
        "mentioned_by": {
          "name": "string",
          "role": "string",
          "influence_level": "decision_maker | influencer | end_user",
          "specific_quote": "string: verbatim"
        },
        "displacement_opportunity": {
          "probability": "high | medium | low",
          "rationale": "string: why this is/isn't a good displacement target",
          "recommended_approach": "string: specific campaign angle"
        }
      }
    ],
    "tool_stack_mentioned": ["string: all tools/vendors mentioned"],
    "budget_signals": [
      {
        "amount": "string",
        "context": "string: what it's for",
        "speaker": "string"
      }
    ]
  }
}

Edge cases:

• If no budget mentioned, set current_spend to null: NEVER estimate
• Vague timelines ("sometime next year") → extract as full year range "Q1-Q4 YYYY"
• Only mark "actively_looking_to_switch" if they explicitly mention evaluating alternatives
• If multiple people mention the same competitor, create ONE entry with multiple quotes

2C: Expansion Signal Extraction (expansion)

Detect upsell/cross-sell signals from customer conversations.

Schema:

{
  "expansion_signals": {
    "account_name": "string",
    "extraction_date": "YYYY-MM-DD",
    "calls_analyzed": "number",
    "signals": [
      {
        "signal_type": "new_use_case | new_team | new_geography | increased_usage | feature_request | integration_need",
        "description": "string: what they want",
        "current_state": "string: what they do today",
        "desired_state": "string: what they're looking for",
        "urgency": "immediate | near_term | exploratory",
        "mentioned_by": {
          "name": "string",
          "role": "string",
          "quote": "string: verbatim"
        },
        "recommended_action": "string: specific next step"
      }
    ],
    "product_feedback": [
      {
        "type": "feature_request | improvement | bug_report | praise",
        "description": "string",
        "speaker": "string",
        "business_justification": "string: why they need it"
      }
    ]
  }
}

2D: Closed-Lost Analysis (closed-lost)

Process transcripts from lost deals to find systematic failure patterns.

Schema:

{
  "closed_lost_analysis": {
    "account_name": "string",
    "extraction_date": "YYYY-MM-DD",
    "calls_analyzed": "number",
    "primary_loss_category": "missing_capability | pricing | competitor | timing | internal_politics | no_budget | champion_left | bad_fit | other",
    "detailed_reason": "string: specific explanation with evidence",
    "stakeholder_objections": [
      {
        "stakeholder_role": "string",
        "objection": "string: specific concern",
        "severity": "deal_killer | major_concern | minor_concern",
        "quote": "string: verbatim",
        "was_addressed": "yes | partially | no",
        "how_addressed": "string or null"
      }
    ],
    "competitive_factor": {
      "competitor_chosen": "string or null",
      "why_they_won": "string: specific reasons",
      "could_we_have_won": "string: honest assessment"
    },
    "deal_stage_lost": "string: where in pipeline",
    "recovery_possibility": "high | medium | low",
    "recovery_rationale": "string: what would need to change",
    "lessons": [
      {
        "category": "sales_process | product_gap | positioning | pricing | timing",
        "insight": "string: what we should learn from this",
        "actionable": "boolean: can we act on this now"
      }
    ]
  }
}

Edge cases:

• If rep marked "pricing" but transcripts reveal a different issue, note the discrepancy
• Don't simplify: capture the full chain of objections that led to the loss
• For recovery_possibility, only mark "high" if there's a clear trigger event (e.g., contract renewal, champion returning)

Step 3: Enrich (optional, if enrich = true)

For every company and key stakeholder mentioned in the extracted intelligence, enrich mentioned companies via your enrichment tool of choice:

Company enrichment:

→ Enrich each company domain for data points like: Tech Stack, Open Jobs, Recent News, Annual Revenue, Competitors

Cross-reference enrichment data with transcript mentions:

• Do their open jobs validate the challenges mentioned? (e.g., hiring data engineers → confirms data quality pain)
• Does their tech stack confirm or contradict tool mentions from calls?
• Does revenue/headcount context help size the opportunity?

Contact enrichment (for key stakeholders):

→ Find and enrich contacts at the company matching the stakeholder roles mentioned

Append enrichment results to the extracted intelligence under an enrichment key.

Step 4: Generate Output

Based on output_format:

JSON output: Clean, validated JSON matching the schemas above. Save to conversational-intelligence-[output_label]-[date].json in your working directory.

Markdown output: Human-readable report with:

• Executive summary (3-5 bullet points of highest-value findings)
• Detailed findings organized by extraction type
• Verbatim quotes with speaker attribution
• Recommended next actions (prioritized)
• Enrichment insights (if applicable)

Save markdown to conversational-intelligence-[output_label]-[date].md in your working directory.

Step 5: Identify Actionable Workflows

Based on extracted intelligence, recommend specific workflows to trigger:

Signal Found	Recommended Action
Competitor with renewal < 3 months + frustration	Displacement campaign targeting the frustrated stakeholder
New use case mentioned by decision maker	Expansion play with specific use case positioning
Red flag on implementation concerns	Proactive mitigation in first CX call
Champion identified across multiple calls	Champion enablement sequence with supporting materials
Budget amount + timeline confirmed	Fast-track opportunity with specific pricing
Product gap caused deal loss	Product feedback ticket with verbatim customer language

Quality Checks

Before outputting, validate:

• Every quote is verbatim from the transcript (not paraphrased)
• Speaker attribution matches the transcript (correct name + role)
• No fields are estimated: null if not mentioned
• Predefined options use ONLY allowed values
• Logical consistency (e.g., if renewal_timeline is set, competitor_name must exist)
• Enrichment data cross-referenced with transcript mentions (flag contradictions)

Example Usage

Mine a specific call for competitive intelligence

extraction_type: competitive
target: [transcript ID or file path]
enrich: true

Build handoff package for closing deal

extraction_type: handoff
target: "acme.com"  (pulls all calls mentioning Acme)
output_label: acme
enrich: true
output_format: both

Analyze lost deals this quarter

extraction_type: closed-lost
target: keyword:"closed lost" from:2026-01-01 to:2026-03-31
enrich: false
output_format: markdown

Full intelligence sweep for account planning

extraction_type: full
target: "example.com"
output_label: example
enrich: true
output_format: both

Common Workflows

Workflow: Deal Handoff Automation

1. conversational-intelligence (handoff) → Extract all pre-sale context
2. Optional enrichment → Add firmographic + technographic depth
3. Output → Structured handoff doc for CX team

Workflow: Competitive Displacement Campaign

1. conversational-intelligence (competitive) → Find competitor mentions + frustrations
2. Optional enrichment → Validate competitor data, find decision makers
3. Write a displacement email targeting the frustrated stakeholder
4. Build a follow-up sequence timed to the renewal date

Workflow: Quarterly Account Intelligence

1. conversational-intelligence (full) → Sweep all calls for account
2. Optional enrichment → Current company data
3. Output → Account intelligence brief for CS/sales planning