AI-Powered Sales Call Analysis: Whisper Transcription + GPT-4o Insights with Airtable & Slack

Sales rep submits a call recording link → n8n auto-downloads audio, Whisper transcribes, GPT-4o analyzes sentiment & objections → results saved to Airtable, Slack notified, HTML report returned

Who

Sales teams, RevOps managers, and sales coaches who want structured call intelligence without manually reviewing every recording

You get

After submitting a recording, automatically receive: sentiment score, objection list with direct quotes, buying signals, stage recommendation, coaching tip, CRM note, and follow-up email draft — simultaneously logged to Airtable and posted to Slack

You need

n8n · airtable · slack

Steps

1. Prepare credentials and form: configure OpenAI, Airtable, Slack, and activate the workflow

   Complete the following in n8n: 1. Open node **4. OpenAI — Transcribe with Whisper** and replace `YOUR_TOKEN_HERE` in the Authorization header with your OpenAI API key (preferably by creating an n8n Header Auth credential; do not hardcode). The source description text (description.md) writes `YOUR_OPENAI_API_KEY` here, but the actual node parameter in workflow.json is `YOUR_TOKEN_HERE`; the node actual value is used below. 2. Go to Credentials → Add OpenAI API credential → select it in node **7. OpenAI — Chat Model (GPT-4o)**. 3. Add an Airtable Personal Access Token credential → in node **9. Airtable — Save Full Analysis**, select your base and table. The table must contain exactly these 26 columns (names must match including spaces and capitalization): Submission ID, Rep Name, Rep Email, Company, Contact Name, Contact Email, Deal Stage (Before), Deal Stage (AI Recommended), Deal Value, Call Duration (min), Word Count, Sentiment, Sentiment Score, Primary Intent, Deal Risk, Deal Momentum, Objection Count, Buying Signal Count, Competitors Mentioned, Summary, CRM Note, Coaching Tip, Rep Talk Ratio, Follow-up Email Draft, Submitted At, Processed At. 4. In Slack, create an App with Incoming Webhooks enabled → copy the webhook URL → add a Slack Webhook credential in n8n → select it in node **10. Slack — Post Analysis to Channel**. 5. Toggle the workflow to Active and copy the form URL from node **1. Form — Sales Rep Submits Call** to share with your sales team. The sales rep submits via form: name, email, prospect company, contact name, current deal stage, estimated value, direct audio link (mp3/wav/m4a). Call Duration and Any notes before analysis? are optional fields.

2. Download audio and transcribe with Whisper

   The workflow uses an HTTP Request node to download the submitted audio direct link (60-second timeout, continueOnFail: true), then calls the OpenAI Whisper API (model=whisper-1, response_format=verbose_json, language=en, 180-second timeout, continueOnFail: true) to convert speech to a full timestamped transcript. Empty or corrupted audio will trigger an error in downstream nodes.

3. Merge transcript with call metadata

   A Code node pulls the transcript from the Whisper response, calculates wordCount, and merges it with the original form data into a unified object for AI analysis. If the transcript is empty, it throws an error indicating the audio may be silent or corrupted.

4. GPT-4o sales analysis: generate structured JSON insights

   The AI Agent node passes the full transcript and call metadata to GPT-4o with a system prompt that instructs the model to return a single strict JSON object containing: summary, sentiment with score, primary intent, deal risk, deal momentum, recommended stage, rep talk ratio, objections with direct quotes, buying signals, competitors mentioned, next steps with owners and deadlines, coaching tip, CRM note, and follow-up email draft. The prompt explicitly requires: output ONLY JSON, no markdown code blocks, use empty arrays or null for unmentioned items, no hallucination.

   Execution prompt · Source text。Node '6. AI Agent — GPT-4o Sales Analysis' text parameter (system prompt)

   Copy prompt

   =You are an expert sales call analyst. You will receive a sales call transcript along with call metadata. Your job is to analyze the transcript deeply and return a single valid JSON object — nothing else. No explanation, no markdown, no code fences. Only raw JSON.
   
   ---
   
   INPUT YOU WILL RECEIVE:
   - repName: Name of the sales rep
   - company: Prospect company name
   - contactName: Prospect contact person
   - dealStage: Current deal stage (e.g. Discovery, Demo, Proposal)
   - dealValue: Estimated deal value in USD
   - durationMins: Call duration in minutes
   - repNotes: Any notes added by the rep before analysis
   - transcript: Full text transcript of the sales call
   
   ---
   
   YOUR JOB:
   Carefully read the transcript and return this exact JSON structure:
   
   {
     "summary": "A clear 3–5 sentence summary of what happened on this call. What was discussed, what the prospect said, and how the call ended.",
   
     "overall_sentiment": "Positive | Neutral | Negative",
   
     "sentiment_score": A number from 0 to 100. 0 = very negative, 100 = very positive.,
   
     "primary_intent": "One short phrase describing what the prospect mainly wants. Example: 'Evaluating alternatives to current tool' or 'Ready to buy pending legal approval'",
   
     "deal_risk": "Low | Medium | High",
   
     "deal_momentum": "Accelerating | Steady | Stalling | Declining",
   
     "recommended_stage": "The deal stage you recommend AFTER this call. Choose from: Prospecting, Discovery, Demo, Proposal, Negotiation, Closed Won, Closed Lost",
   
     "rep_talk_ratio": "Estimate of how much the rep talked vs the prospect. Example: '65% rep / 35% prospect'",
   
     "objections": [
       {
         "type": "Price | Timeline | Competition | Authority | Need | Other",
         "quote": "A short direct quote from the transcript showing this objection",
         "was_addressed": true or false
       }
     ],
   
     "buying_signals": [
       "List each buying signal as a short sentence. Example: 'Prospect asked about onboarding timeline' or 'Prospect mentioned they have budget approved'"
     ],
   
     "competitors_mentioned": [
       "List any competitor names mentioned on the call. Empty array if none."
     ],
   
     "next_steps": [
       {
         "action": "What needs to happen next",
         "owner": "rep | prospect | both",
         "commitment": "Was this committed to on the call? yes or no",
         "deadline": "Deadline if mentioned, otherwise null"
       }
     ],
   
     "coaching_tip": "One specific, actionable coaching tip for the rep based on this call. Be direct and helpful. Example: 'You spoke for 70% of the call — on the next call, ask more open-ended questions and let the prospect lead.'",
   
     "crm_note": "A 2–3 sentence CRM note ready to paste into a CRM. Professional tone. Include deal stage, key outcomes, and next steps.",
   
     "follow_up_email": "A ready-to-send follow-up email from the rep to the prospect. Include subject line at the top like 'Subject: ...'. Keep it professional, warm, and under 150 words. Reference something specific from the call."
   }
   
   ---
   
   STRICT RULES:
   1. Output ONLY the JSON object above. No text before or after it.
   2. Do not wrap in markdown code blocks or backticks.
   3. If something is not mentioned in the transcript, use an empty array [] or null — do not guess or hallucinate.
   4. Quotes in the "objections" field must come directly from the transcript.
   5. The follow_up_email must feel personal — reference something the prospect actually said.
   6. Keep all text in English unless the transcript is in another language.

5. Parse GPT-4o output and combine into complete record

   A Code node cleans any accidental markdown formatting from the GPT-4o response, parses it as JSON, merges it with the original call data, and adds a processedAt timestamp. If parsing fails, the error message includes the first 300 characters of the raw output for debugging.

6. Parallel delivery: write to Airtable, notify Slack, return HTML report page

   The complete record is fanned out simultaneously to three channels: - **Airtable**: creates a new row populating all 26 fields (continueOnFail: true); - **Slack**: posts a summary notification to your sales channel (continueOnFail: true); - **HTML page**: a Code node builds a formatted page containing summary, sentiment score, objections, next steps, coaching tip, and follow-up email, returned directly to the rep via node **12. Webhook — Show Results Page to Rep** as the form submission confirmation. The HTTP download node, Whisper node, Airtable node, and Slack node all have continueOnFail: true, ensuring the rep still sees the HTML results page even if some steps fail. The workflow uses responseMode: lastNode, meaning the form waits for the final Webhook response node to complete before showing the page to the rep.

You are setting up an n8n workflow: after a sales rep submits a call recording, they automatically receive an AI analysis report. I have provided complete node configuration instructions, the original GPT-4o system prompt, the Airtable column list, and form field definitions. Please first confirm you have: an n8n instance, an OpenAI API key, an Airtable Personal Access Token, and a Slack Incoming Webhook. Then verify credentials and node parameters step by step; do not execute directly.