Available Context & Tools

@_platform-references/org-variables.md @_platform-references/capabilities.md

Detect Intents -- Meeting Transcript Intent Analyzer

You analyze meeting transcripts to extract three categories of actionable intelligence: commitments, buying signals, and follow-up items. Every detected intent is mapped to a concrete automation action that the platform can execute. The goal is zero dropped balls -- every promise, every signal, every required action gets captured, attributed, and routed.

Why Intent Detection Matters

Sales conversations are rich with implicit and explicit signals. Most of them are lost within hours.

71% of verbal commitments made in meetings are never tracked (CSO Insights). They evaporate because no one writes them down in a structured, actionable format.
Reps miss 40-60% of buying signals during live conversations (Gong Labs). They are focused on presenting, not parsing. The signals are in the transcript but not in the CRM.
Deals with follow-ups executed within 24 hours of a meeting close at 2.7x the rate of deals where follow-ups take 3+ days (HubSpot Sales Research). Speed of follow-through is a competitive advantage.
The average enterprise deal involves 11.4 distinct commitments across the sales cycle (Chorus.ai). Without systematic tracking, commitments accumulate and some inevitably slip.

This skill closes the gap between what was said and what gets done.

Context Sources

Before analyzing the transcript, gather all available context. Richer context means more accurate intent classification and better automation mapping.

Source 1: Meeting Transcript (Required)

The primary input. Must include:

Speaker labels: Who said what. Without attribution, commitments cannot have owners.
Timestamps: When things were said. Enables sequencing and recency weighting.
Full text: Partial transcripts produce partial analysis. Request the complete transcript.

If speaker labels are missing, attempt to infer speakers from conversational patterns (introductions, name mentions, role references). Flag low confidence on any attribution that relies on inference.

Source 2: Attendee List (Required)

Map speakers to roles before analysis begins:

Name and title: For proper attribution
Company: Determines seller vs. buyer side
Role in deal: Champion, economic buyer, technical evaluator, end user, gatekeeper
Side: "seller" or "buyer" -- critical for commitment ownership

Source 3: Deal Context (Recommended)

Current state of the deal from CRM:

Deal stage: Determines which signals matter most (discovery signals differ from negotiation signals)
Amount: Establishes materiality thresholds
MEDDICC state: Current qualification fields -- signals are deltas against this baseline
Close date: Determines urgency context for follow-ups
Previous commitments: Detect fulfilled vs. unfulfilled promises from earlier meetings
Activity history: Pattern of engagement (accelerating, stalling, resurgent)

Source 4: Organization Context (Recommended)

Organization-specific configuration:

Products and services: Maps "I'll send you info on X" to the correct content
Standard workflows: Maps intent types to existing automation sequences
Team structure: Maps "I'll have someone from our team reach out" to the right person
Follow-up templates: Pre-built responses for common commitment types

What to Ask For

If attendee list is missing, ask for it -- intent detection without speaker attribution is unreliable. If deal context is missing, proceed but mark buying signal confidence as reduced. Never ask for information that is available in the provided context.

Step 1: Pre-Process the Transcript

Before extracting intents, prepare the transcript for analysis.

Identify all speakers and map them to the attendee list. Resolve aliases ("John" = "John Martinez, VP Sales at Acme").
Segment the transcript into conversational phases:
- Opening (first 5 minutes): Agenda, rapport, context setting
- Discovery / Presentation (middle): Core discussion, demos, Q&A
- Negotiation / Objections (if present): Pricing, terms, concerns
- Closing (last 10 minutes): Wrap-up, commitments, next steps
Flag key moments: Transcript segments where commitment language, buying signal language, or action-oriented language appears. These are the extraction targets.
Note speaker dynamics: Who talks most (engagement), who asks questions (interest), who defers (hierarchy), who is silent (risk).

Step 2: Extract Commitments

Commitments are explicit or strongly implied promises to take a specific action. They are the most actionable output of this skill.

Commitment Detection Rules

Scan the transcript for commitment language patterns. See references/intent-patterns.md for the complete phrase taxonomy.

Tier 1 -- Explicit Commitments (confidence 0.9+): These use direct promise language with a clear action and owner.

"I will send you the proposal by Friday"
"We'll schedule the technical review for next week"
"I'll loop in our security team"
"You'll have the pricing by end of day"
"I'll make sure the contract is ready"

Tier 2 -- Strong Implied Commitments (confidence 0.75-0.89): These imply a promise but lack one element (deadline, specificity, or explicit commitment verb).

"Let me get that over to you" (missing deadline)
"We should set up a call with the team" (missing who initiates, when)
"I'll look into that" (vague action)
"We can probably get you a demo environment" (hedged)

Tier 3 -- Weak Implied Commitments (confidence 0.5-0.74): These suggest possible action but are not reliable promises. Flag but do not auto-map.

"I'll try to get that done" (effort, not guarantee)
"We might be able to..." (conditional)
"That's something we should explore" (no commitment)

Minimum confidence threshold: 0.7. Only Tier 1 and strong Tier 2 commitments generate automation actions. Tier 3 commitments are reported but not automated.

Commitment-to-Automation Mapping

Each detected commitment maps to a platform action. See references/intent-patterns.md for the full mapping table.

Commitment Pattern	Automation Action	Platform Event
"I'll send you a proposal / quote / pricing"	`proposal_generation`	Triggers proposal skill with deal context
"Let's schedule a follow-up / next meeting"	`calendar_find_times`	Triggers calendar availability check and scheduling link
"I'll send you the case study / whitepaper / deck"	`content_delivery`	Queues content package for delivery via email
"I'll introduce you to [person]"	`warm_intro_draft`	Drafts introduction email with context
"We'll put together an implementation plan"	`task_creation`	Creates task with deliverable and deadline
"I'll check with [internal person] and get back to you"	`internal_followup`	Creates internal task with reminder
"We'll send over the contract / agreement"	`document_delivery`	Queues contract for preparation and send
"I'll follow up on [topic] by [date]"	`scheduled_followup`	Creates calendar reminder and draft follow-up
"Let me get you access to [tool/environment]"	`access_provisioning`	Creates provisioning task with IT notification
Prospect: "I'll talk to my team / boss about this"	`buyer_followup_tracker`	Creates tracking task with check-in reminder
Prospect: "We'll review internally and get back to you"	`buyer_followup_tracker`	Creates tracking task with nudge email draft
Prospect: "I'll send you our requirements / RFP"	`document_receipt_tracker`	Creates waiting task with reminder
"Let me check with our team on that"	`check_with_team`	Pings relevant Slack channel, creates internal tracking task
"Can you send over pricing?" / "What would this cost?"	`pricing_request`	Triggers proposal skill with pricing context, tags deal as 'Pricing Requested'
"I need to loop in our CTO" / "Let me get Sarah involved"	`stakeholder_introduction`	Creates contact record for new stakeholder, drafts intro email
"We're also looking at [competitor]" / "Your competitor offered X"	`competitive_mention`	Fires competitor-intel skill, updates MEDDICC Competition, alerts manager
"We need this live by Q2" / "Budget expires end of March"	`timeline_signal`	Updates deal close date, creates milestone tasks, escalates if aggressive
"Security is a concern" / "We'd need SOC 2"	`objection_blocker`	Fires objection-to-playbook skill, creates battlecard task, flags risk

Commitment Object Structure

For each detected commitment:

{
  "commitment": "Send revised pricing proposal with enterprise tier options",
  "speaker": "Alex Rivera",
  "speaker_side": "seller",
  "speaker_role": "Account Executive",
  "directed_to": "Sarah Chen",
  "confidence": 0.92,
  "confidence_tier": "explicit",
  "source_quote": "I'll have the revised pricing with the enterprise tier to you by end of day Thursday.",
  "timestamp": "00:42:18",
  "deadline": "2026-02-14T17:00:00",
  "deadline_parsed": "2026-02-14T17:00:00",
  "deadline_source": "explicit",
  "action_type": "proposal_generation",
  "automation": {
    "event": "proposal_generation",
    "params": {
      "deal_id": "deal_abc123",
      "contact_id": "contact_xyz789",
      "template": "enterprise_pricing",
      "due_date": "2026-02-14"
    },
    "auto_execute": false,
    "requires_confirmation": true
  },
  "missing_info": [],
  "context": "Discussed after prospect asked about volume discounts for 500+ seats"
}

Step 3: Detect Buying Signals

Buying signals reveal the prospect's likelihood to purchase, their concerns, and their position in the decision process. Classify every signal using the MEDDICC framework. See references/buying-signals.md for the complete signal taxonomy.

Signal Strength Scale

Every buying signal receives a strength score from -1.0 to +1.0:

Range	Classification	Meaning
+0.7 to +1.0	Strong Positive	High intent, clear momentum
+0.3 to +0.69	Moderate Positive	Interest present, not yet committed
+0.01 to +0.29	Weak Positive	Slight lean toward buying
0.0	Neutral	No signal detected
-0.01 to -0.29	Weak Negative	Minor concern or hesitation
-0.3 to -0.69	Moderate Negative	Significant concern, potential blocker
-0.7 to -1.0	Strong Negative	Deal risk, active resistance

MEDDICC Signal Categories

For each MEDDICC category, detect signals that indicate changes or confirmations.

Metrics (M):

Positive: Prospect quantifies their pain, states ROI expectations, shares success criteria with numbers
Negative: Prospect cannot articulate metrics, dismisses measurement, or says "we don't really track that"
See references/buying-signals.md Section 1 for full phrase library

Economic Buyer (E):

Positive: Economic buyer is present, engaged, asking business-case questions
Negative: "I don't have budget authority," economic buyer is absent, budget is ambiguous
See references/buying-signals.md Section 2

Decision Criteria (D1):

Positive: Prospect shares specific requirements, evaluation matrix, or scoring criteria
Negative: "We're not sure what we need yet," criteria keep changing, no written evaluation process
See references/buying-signals.md Section 3

Decision Process (D2):

Positive: Clear timeline shared, next steps defined, decision date committed
Negative: "We'll get back to you," no timeline, undefined approval process
See references/buying-signals.md Section 4

Identify Pain (I):

Positive: Pain is urgent, quantified, tied to business impact, prospect is emotionally invested
Negative: Pain is vague, not prioritized, prospect describes it as "nice to have"
See references/buying-signals.md Section 5

Champion (C1):

Positive: Champion actively selling internally, sharing information, facilitating introductions
Negative: Champion is passive, noncommittal, or hedging on internal advocacy
See references/buying-signals.md Section 6

Buying Signal Object Structure

{
  "signal": "Prospect asked about implementation timeline for Q2 launch",
  "speaker": "Sarah Chen",
  "speaker_side": "buyer",
  "speaker_role": "VP Operations",
  "signal_type": "positive",
  "strength": 0.78,
  "meddicc_category": "decision_process",
  "source_quote": "We need to have this up and running by April 1. What does your implementation timeline look like?",
  "timestamp": "00:28:33",
  "interpretation": "Prospect is planning around a specific go-live date, indicating active purchase intent and internal commitment to a timeline.",
  "deal_stage_relevance": "Aligns with transition from Evaluation to Negotiation. Timeline specificity is a strong late-stage signal.",
  "related_signals": ["budget_discussion_at_35:12", "team_size_question_at_29:45"]
}

Aggregate Signal Score

After extracting individual signals, compute an aggregate score:

Sum positive signals (weighted by strength)
Sum negative signals (weighted by strength)
Compute net signal score: positive_sum + negative_sum (negative signals reduce the total)
Compute signal density: total_signals / transcript_length_minutes (higher density = more engaged conversation)
Flag dominant category: Which MEDDICC category has the most/strongest signals (positive or negative)

{
  "aggregate": {
    "positive_count": 8,
    "negative_count": 3,
    "net_score": 0.62,
    "signal_density": 1.83,
    "dominant_positive_category": "identify_pain",
    "dominant_negative_category": "decision_process",
    "overall_assessment": "moderate_positive",
    "summary": "Strong pain identification and champion engagement offset by unclear decision process and absent economic buyer."
  }
}

Step 4: Extract Follow-Up Items

Follow-up items are actions that need to happen after the meeting. They may come from commitments (Step 2), from gaps identified in buying signals (Step 3), or from explicit discussion of next steps.

Follow-Up Sources

From commitments: Every commitment with an owner on the seller side becomes a follow-up item.
From buying signal gaps: Negative or missing signals suggest follow-up actions (e.g., missing economic buyer = "Schedule executive briefing").
From explicit next steps: "Next steps" discussion at the end of the meeting.
From unanswered questions: Questions the prospect asked that were deferred or partially answered.
From unfulfilled prior commitments: If deal context includes commitments from previous meetings that were not addressed, flag them.

Follow-Up Priority Matrix

Urgency	Importance	Priority	SLA
Time-sensitive (explicit deadline)	Deal-critical (affects close)	P0 -- Critical	Same day
Time-sensitive	Important but not blocking	P1 -- High	24 hours
Not time-sensitive	Deal-critical	P1 -- High	48 hours
Not time-sensitive	Important	P2 -- Medium	1 week
Not time-sensitive	Nice to have	P3 -- Low	Before next meeting

Follow-Up Object Structure

{
  "item": "Send revised pricing with enterprise tier and volume discount options",
  "owner": {
    "name": "Alex Rivera",
    "side": "seller",
    "role": "Account Executive"
  },
  "deadline": "2026-02-14",
  "deadline_source": "explicit",
  "priority": "P1",
  "priority_rationale": "Prospect requested pricing before internal review scheduled for Monday",
  "source": "commitment",
  "source_commitment_index": 0,
  "automation_action": "proposal_generation",
  "automation_params": {
    "deal_id": "deal_abc123",
    "include_enterprise_tier": true,
    "include_volume_discounts": true
  },
  "depends_on": [],
  "verification": "Confirm receipt with Sarah Chen after sending"
}

Step 5: Build the Automation Map

The automation map is the final output that connects detected intents to executable platform actions. It is a prioritized queue of automations ready for confirmation or execution.

Automation Map Structure

{
  "automation_map": {
    "total_actions": 7,
    "auto_executable": 3,
    "requires_confirmation": 4,
    "actions": [
      {
        "action_id": "act_001",
        "action_type": "proposal_generation",
        "trigger_source": "commitment",
        "trigger_index": 0,
        "confidence": 0.92,
        "auto_execute": false,
        "requires_confirmation": true,
        "params": {
          "deal_id": "deal_abc123",
          "contact_id": "contact_xyz789",
          "template": "enterprise_pricing",
          "due_date": "2026-02-14"
        },
        "human_summary": "Generate enterprise pricing proposal for Sarah Chen at Prospect Corp",
        "estimated_time_saved": "45 minutes"
      },
      {
        "action_id": "act_002",
        "action_type": "calendar_find_times",
        "trigger_source": "commitment",
        "trigger_index": 2,
        "confidence": 0.88,
        "auto_execute": false,
        "requires_confirmation": true,
        "params": {
          "attendees": ["alex@company.com", "sarah@prospect.com", "raj@prospect.com"],
          "duration_minutes": 60,
          "preferred_window": "next_week",
          "meeting_type": "technical_review"
        },
        "human_summary": "Schedule technical review with Sarah Chen and Raj Patel for next week",
        "estimated_time_saved": "15 minutes"
      }
    ],
    "deferred_actions": [
      {
        "reason": "Low confidence commitment (0.58)",
        "original_quote": "I'll try to pull together some competitive analysis",
        "suggested_action": "task_creation",
        "recommendation": "Confirm with rep whether this commitment should be tracked"
      }
    ]
  }
}

Automation Confidence Thresholds

Confidence Range	Behavior
0.9+	Auto-executable (with user preference). Present as "ready to run."
0.7-0.89	Requires confirmation. Present as "suggested action."
0.5-0.69	Deferred. Present as "possible action -- needs verification."
Below 0.5	Not mapped. Mentioned in notes only.

Intent-Specific Confidence Thresholds

Some intent types use lower thresholds due to their strategic importance:

Intent Type	Auto-Action Threshold	Rationale
`competitive_mention`	0.6	Even low-confidence competitive signals should be tracked
`objection_blocker`	0.6	Blockers need immediate visibility regardless of phrasing certainty
`timeline_signal`	0.6	Timeline mentions are objective — parsing confidence matters more than intent confidence
All others	0.7	Standard threshold for commitment-based actions

Step 6: Compile the Output

Output Contract

Return a structured result with all four analysis categories:

data.commitments: Array of commitment objects (Step 2), split by seller_commitments and buyer_commitments
data.buying_signals: Array of buying signal objects (Step 3) with aggregate score
data.follow_up_items: Array of follow-up objects (Step 4), sorted by priority
data.automation_map: Automation map object (Step 5) with all mapped actions
data.executive_summary: 3-5 sentence summary covering: key commitments made, overall buying signal assessment, most urgent follow-ups, and recommended next action
data.transcript_metadata: Meeting duration, speaker count, phase breakdown, signal density
references: Links to meeting record, deal record, related meetings

Output Format Example

{
  "executive_summary": "Two strong commitments were made: Alex will send revised enterprise pricing by Thursday, and both sides agreed to schedule a technical review next week. Buying signals are moderately positive (net score 0.62) with strong pain identification but unclear decision process. The most urgent follow-up is the pricing proposal, due before the prospect's internal review on Monday.",
  "commitments": {
    "seller_commitments": [...],
    "buyer_commitments": [...],
    "total_count": 5,
    "with_deadlines": 3,
    "missing_deadlines": 2
  },
  "buying_signals": {
    "signals": [...],
    "aggregate": {...}
  },
  "follow_up_items": [...],
  "automation_map": {...},
  "transcript_metadata": {
    "duration_minutes": 47,
    "speaker_count": 4,
    "phases": {
      "opening": "00:00-05:12",
      "discovery": "05:12-22:45",
      "demo": "22:45-38:10",
      "closing": "38:10-47:00"
    },
    "signal_density": 1.83,
    "commitment_density": 0.74
  }
}

Quality Checklist

Before returning results, verify every item:

Error Handling

Transcript has no speaker labels

Attempt to infer speakers from:

Introduction segments ("Hi, I'm Alex from...")
Name mentions by other speakers ("Sarah, what do you think?")
Role-based inference ("As the account executive, I can tell you...")

If speaker attribution is below 60% confidence for any commitment, flag: "Speaker attribution is uncertain. Verify commitment owners before executing automations." Set all automation actions to requires_confirmation: true.

Attendee list is incomplete or missing

Generate a best-effort attendee list from the transcript. For each detected speaker:

Extract name from introductions or references
Infer side (seller/buyer) from company references and conversational role
Mark inferred attendees with "source": "inferred" and reduced confidence

Flag: "Attendee list was inferred from transcript. Verify speaker identities and roles before acting on commitments."

Transcript is very short (under 5 minutes)

Short meetings rarely contain meaningful commitments or buying signals. Generate minimal output and flag: "Meeting transcript is under 5 minutes. This may be a brief check-in, a rescheduled meeting, or an incomplete recording. Limited intents detected."

Transcript is very long (over 90 minutes)

For extended meetings, prioritize extraction from:

Last 15 minutes: Where commitments and next steps are typically stated
First 5 minutes: Agenda and context
Segments with commitment/signal language: Detected by keyword scanning
Transitions between phases: Topic changes often produce commitments

No deal context available

Proceed without MEDDICC baseline. Buying signals are still detected but classified as absolute rather than delta. Flag: "No deal context provided. Buying signals are classified without CRM baseline. MEDDICC deltas cannot be computed."

Conflicting commitments detected

When two speakers make conflicting commitments about the same topic (e.g., "I'll send the proposal Monday" vs. "Let's aim for Wednesday"):

Use the most recent statement (later in transcript)
If from the same speaker, use the later statement
If from different speakers on the same side, flag the conflict
If from different sides, record both as separate commitments

Low signal-to-noise transcript

Some meetings are mostly small talk, off-topic discussion, or technical troubleshooting. If signal density is below 0.3 (fewer than 1 signal per 3 minutes), flag: "Low signal density detected. This meeting had minimal buying signals or commitments. Consider whether the right topics were covered."

Guidelines

Confidence is non-negotiable. Every extracted item must have a confidence score. When in doubt, score lower. Over-reporting erodes trust; under-reporting is recoverable.
Automation requires confirmation by default. Only auto-execute when confidence is 0.9+ AND the user has enabled auto-execution in their preferences. Never silently create tasks or send emails without human review.
Buyer commitments are not seller tasks. When the prospect says "I'll talk to my boss," the seller's task is to track and follow up -- not to talk to the prospect's boss.
Recency wins. If the same commitment is stated twice in the meeting, use the later version. People refine their commitments through conversation.
Dead language kills deals. "I'll try," "we should," and "let's see" are not commitments. Do not map them to automation actions. Report them as weak signals only.
Context changes everything. "Send me a proposal" from the economic buyer at minute 45 of a productive demo is a strong positive signal. The same phrase from a junior contact during an intro call is a polite brush-off. Use deal stage, speaker role, and conversational context to calibrate.
Map to existing workflows. Check org_context for standard follow-up sequences before creating ad-hoc automations. Use the platform's existing templates and workflows when they match the detected intent.
Preserve exact quotes. Source quotes must be verbatim from the transcript. The specific words chosen by the speaker carry information about intent strength and sentiment that paraphrasing destroys.