Anthropic Claude AI

Claude is used for AI-powered call summarization in Client Portal. When a sales call produces a transcript, Claude analyzes it to extract disposition, sentiment, key points, buying signals, and objections -- replacing manual note-taking and providing consistent, structured intelligence.

Configuration

Setting	Value
API URL	https://api.anthropic.com/v1/messages
Auth Header	x-api-key: {ANTHROPIC_API_KEY}
API Version	anthropic-version: 2023-06-01
Model	claude-sonnet-4-5-20250929 (configurable via ANTHROPIC_MODEL env)
Max Tokens	1024
Env Variable	ANTHROPIC_API_KEY (Supabase secret)

Edge Function: summarize-sales-call

Trigger

Called automatically by ghl-call-webhook (or ghl-sales-webhook) when a call transcript meets these conditions:

1. Transcript text is longer than 50 characters (after flattening)
2. Transcript does not match the voicemail greeting pattern

Can also be triggered manually via the "Summarize" button in the Sales Call Detail drawer.

Request

{
  "call_id": "uuid",
  "transcript": "Full call transcript text..."
}

Process Flow

AI Output Schema

Claude returns a JSON object with the following structure:

{
  "disposition": "Future Potential",
  "key_points": [
    "Prospect runs a 12-person legal search firm in NYC",
    "Currently uses in-house researchers for candidate sourcing",
    "Expressed interest in patent litigation market coverage",
    "Asked about pricing and timeline for validation sprint"
  ],
  "sentiment": "interested",
  "action_items": [
    "Send pricing deck for Market Validation Sprint",
    "Schedule follow-up call for next Tuesday"
  ],
  "engagement_assessment": "Strong prospect -- actively looking for market intelligence solution. Decision timeline is 2-4 weeks.",
  "conversation_quality": "warm",
  "interest_level": 7,
  "objections": [
    "Concerned about overlap with existing research team"
  ],
  "buying_signals": [
    "Asked about pricing unprompted",
    "Mentioned specific market they want covered",
    "Asked about onboarding timeline"
  ]
}

Field Definitions

Field	Type	Values	Description
disposition	string	One of 9 canonical	The AI's best-fit disposition for the call
key_points	string[]	3--5 items	Most important facts from the conversation
sentiment	string	interested, skeptical, not_interested, excited, neutral	Prospect's emotional tone
action_items	string[]	Variable	Specific next steps needed
engagement_assessment	string	Free text	1--2 sentence commercial assessment
conversation_quality	string	cold, neutral, warm, hot	Temperature of the conversation
interest_level	number	1--10	Numeric interest rating
objections	string[]	Variable	Concerns or pushback raised
buying_signals	string[]	Variable	Positive indicators of purchase intent

Disposition Rules

The system prompt instructs Claude to choose from the 9 canonical dispositions with specific guidance:

Disposition	When to Use
Voicemail	Transcript is a voicemail greeting or automated message
Gatekeeper	Receptionist/assistant answered, decision-maker not reached
Incorrect Number	Wrong person, number not in service
Not Interested	Prospect explicitly declined or shut the conversation down
Callback Requested	Prospect literally said "call me back", "try me later" -- do NOT use as default
Future Potential	Any engaged conversation without a meeting booked. Correct choice for neutral-to-positive calls. When in doubt between Callback Requested and Future Potential, choose Future Potential.
Meeting Booked	A meeting or demo was scheduled during the call
Do Not Contact	Prospect explicitly asked not to be contacted again
No Answer	Only if transcript shows no real interaction

Disposition Priority

AI disposition overwrites the regex-inferred disposition from the webhook handler. The ghl-call-webhook handler awaits the AI response and uses the AI disposition as the final value. BDR-confirmed dispositions from GHL's disposition workflow take highest priority and are never overwritten by AI.

Disposition Validation

The AI response is validated against the canonical list before being applied:

const VALID_DISPOSITIONS = [
  "No Answer", "Voicemail", "Gatekeeper", "Incorrect Number",
  "Not Interested", "Callback Requested", "Future Potential",
  "Meeting Booked", "Do Not Contact",
];

const aiDisposition = VALID_DISPOSITIONS.includes(summary.disposition)
  ? summary.disposition
  : null;

If the AI returns an invalid disposition, it is ignored and the existing disposition is preserved.

Database Updates

sales_calls Table Updates

Column	Source	Description
ai_summary	Full JSON	Complete AI analysis object (JSONB)
ai_summary_text	Constructed	Plain-text summary for display
disposition	summary.disposition	Canonical disposition (if valid)
disposition_reason	summary.engagement_assessment	Prefixed with "AI: "
conversation_quality	summary.conversation_quality	cold/neutral/warm/hot
interest_level	summary.interest_level	1--10 rating
objections	summary.objections	Array of objection strings
buying_signals	summary.buying_signals	Array of buying signal strings
notes	Plain-text summary	Key points + sentiment + interest
ai_processed_at	Timestamp	When AI processing completed

sales_prospects Table Updates (Auto-Disposition)

After updating the call, the function automatically adjusts the prospect's engagement state:

AI Disposition	Engagement Level	Follow-Up
Meeting Booked	meeting_set	Cleared (meeting is the follow-up)
Future Potential	engaged	7 days from now
Callback Requested (warm/hot)	engaged	2 days from now
Callback Requested (cold/neutral)	contacted	3 days from now
Not Interested	nurturing	14 days from now
Do Not Contact	closed_lost	Cleared, do_not_contact = true
Gatekeeper	contacted	2 days from now
Voicemail	contacted	1 day (first 3 VMs), 7 days (after)
Incorrect Number	closed_lost	Cleared
No Answer	contacted	None

The follow-up includes context in next_follow_up_notes:

"Future potential. Interest: 7/10. Follow up in 1 week."
"Voicemail #3. Retry tomorrow."
"warm conversation. Interest: 8/10. Follow up in 2 days."

System Prompt

The full system prompt provides Claude with business context about 1 Hour Recruitment's service model:

You are an AI assistant for 1 Hour Recruitment, a company that sells
market intelligence services to boutique legal search firms.

You are analyzing a sales call transcript between a 1HR sales rep and
a prospective client (a legal search firm owner/partner).

Extract the following in JSON format: { ... }

Disposition rules: ...
- When in doubt between Callback Requested and Future Potential,
  choose Future Potential.

Be concise and specific. Focus on commercial signals relevant to
closing the deal. Return ONLY valid JSON, no other text.

Error Handling

Error	Behavior
ANTHROPIC_API_KEY not set	Returns 500 with descriptive error
Claude API error (4xx/5xx)	Returns 500 with Claude's error message
JSON parse failure	Tries regex extraction of JSON from response
Invalid disposition	Preserves existing disposition, still saves summary
Auto-disposition failure	Logged, does not fail the summary response

Manual Trigger

Users can trigger summarization from the UI via two paths:

1. SummarizeCallButton (src/components/sales/SummarizeCallButton.tsx) -- one-click on a call that already has a transcript.
2. PasteTranscriptDialog (src/components/sales/PasteTranscriptDialog.tsx) -- paste a transcript manually for calls without one.

Both call the edge function with the same { call_id, transcript } payload.

Related Documentation

• GHL Webhooks -- Auto-triggers AI summarization when transcripts arrive
• GHL Overview -- Architecture of the GHL integration that feeds transcripts