Sales Call CRM Field Extraction Prompt Template

This prompt is purpose-built for revenue operations engineers who need to convert unstructured sales call transcripts into structured CRM update payloads. The core job-to-be-done is reducing manual data entry by automatically extracting opportunity stage, key contacts, discussed products, objections, and next steps from recorded conversations. The ideal user has access to completed transcripts, a defined CRM schema, and a clear understanding of which fields must be explicitly stated versus inferred. Use this when transcripts are already available and the goal is to accelerate CRM hygiene—not to guide live calls or analyze conversation quality.

The prompt distinguishes between explicitly stated facts and inferred information by requiring confidence flags on every extracted field. For example, if a prospect says 'we're evaluating your enterprise tier,' the product field is explicit; if the transcript only mentions 'scaling concerns,' inferring a product interest requires a confidence score and a note explaining the inference chain. This design is critical for CRM integrity because downstream workflows—forecasting, territory assignment, deal scoring—break when inferred data is treated as ground truth. The prompt also handles speaker attribution, so extracted contacts are tied to named individuals rather than generic roles, and next steps are assigned to specific owners when the transcript supports it.

Do not use this prompt for real-time call guidance, legal deposition analysis, or scenarios where transcript quality is too poor to identify speakers or topics reliably. If the transcript contains garbled speech-to-text artifacts, overlapping speakers, or missing segments, extraction accuracy degrades sharply and the confidence flags will reflect that uncertainty. In those cases, invest in transcript cleanup or human review before running extraction. For high-stakes deals or regulated industries, always route low-confidence extractions to a human review queue rather than writing directly to the CRM. The next section provides the copy-ready prompt template you can adapt to your specific CRM field schema and confidence thresholds.

This prompt template is designed to be pasted directly into your AI harness. It instructs the model to process a raw sales call transcript and produce a strict JSON payload suitable for ingestion into a CRM system. The prompt enforces a specific output schema, distinguishes between explicitly stated and inferred information, and requires the model to flag low-confidence extractions for human review. Before using this template, ensure you have a clean transcript and have defined the target CRM fields your downstream system expects.

textCopy
You are an AI assistant for a Revenue Operations team. Your task is to analyze the provided sales call transcript and extract a structured CRM update payload. You must follow the output schema exactly. Do not add fields that are not in the schema.

[INPUT]
The full transcript of the sales call, with speaker labels.

[OUTPUT_SCHEMA]
{
  "opportunity_stage": {
    "value": "string | null", // e.g., 'Discovery', 'Demo Scheduled', 'Negotiation'. Use null if not discernible.
    "confidence": "high | medium | low",
    "evidence": "string" // Quote from the transcript supporting this stage.
  },
  "key_contacts": [
    {
      "name": "string",
      "title": "string | null",
      "role_in_deal": "string | null", // e.g., 'Champion', 'Decision Maker', 'Influencer'
      "confidence": "high | medium | low"
    }
  ],
  "discussed_products": [
    {
      "product_name": "string",
      "interest_level": "high | medium | low",
      "confidence": "high | medium | low"
    }
  ],
  "objections": [
    {
      "objection": "string",
      "status": "open | resolved",
      "confidence": "high | medium | low"
    }
  ],
  "next_steps": [
    {
      "action": "string",
      "owner": "string | null",
      "deadline": "string | null", // ISO 8601 date format if possible
      "confidence": "high | medium | low"
    }
  ],
  "overall_summary": "string" // A 2-3 sentence summary of the call's outcome.
}

[CONSTRAINTS]
1. **Explicit vs. Inferred:** For every field, the 'confidence' must be 'high' only if the information is explicitly stated by a speaker. If you are inferring the value from context or tone, set confidence to 'medium' or 'low'.
2. **Evidence:** For the 'opportunity_stage', you must provide a direct quote from the transcript as 'evidence'. For other fields, evidence is optional but recommended for 'medium' or 'low' confidence values.
3. **Null Handling:** If a top-level field like 'opportunity_stage' cannot be determined at all, set its 'value' to null and 'confidence' to 'low'. Do not omit the field.
4. **Schema Adherence:** The final output must be a single, valid JSON object conforming to the [OUTPUT_SCHEMA]. Do not include any text outside the JSON object.
5. **Data Minimization:** Only extract information relevant to the schema. Ignore small talk, scheduling logistics, and other non-commercial conversation.

[RISK_LEVEL]
HIGH. Incorrect CRM data can lead to missed forecasts and poor customer experience. All 'low' confidence fields and any 'objections' with a 'status' of 'open' must be routed for human review before CRM insertion.

To adapt this template, replace the [INPUT] placeholder with the actual transcript text in your application code. You can also extend the [OUTPUT_SCHEMA] to include custom fields specific to your CRM, such as lead_source or budget_range. The critical part to preserve is the [CONSTRAINTS] section, which teaches the model the difference between explicit and inferred data—a common failure point in production extraction. After receiving the model's output, always validate the JSON structure against your schema and route low-confidence records to a human review queue as specified in the [RISK_LEVEL] directive.

This prompt is designed to sit inside a revenue operations pipeline that ingests call transcripts and produces structured CRM update payloads. The typical integration pattern is: transcript ingestion → chunking (if needed) → prompt execution → JSON validation → confidence threshold check → CRM write or review queue. Because the output includes confidence flags for inferred versus explicitly stated information, the harness must treat low-confidence fields differently from high-confidence fields. Do not blindly upsert every extracted field into your CRM—use the confidence signals to decide which fields update the system of record automatically and which require a sales rep to confirm.

Start with a strict output schema validator. The prompt template includes an [OUTPUT_SCHEMA] placeholder that should be populated with a JSON Schema definition matching your CRM's field contract. After the model responds, validate the JSON structure before any downstream processing. If validation fails, retry once with the error message appended to the prompt context—but stop after two attempts and route the transcript to a human review queue. For model choice, use a model with strong structured output support (GPT-4o with response_format, Claude 3.5 Sonnet with tool use, or equivalent). Avoid models that struggle with long-context adherence, since sales call transcripts can exceed 10k tokens. If transcripts are longer than the model's effective context window, chunk by speaker turn boundaries and run extraction per chunk, then merge results with a deduplication pass.

The confidence flags in the output (explicit vs inferred vs not_discussed) should drive your write logic. Fields marked explicit can be written to the CRM directly if the overall extraction confidence score exceeds your threshold (start with 0.85 and tune based on audit results). Fields marked inferred should be written as draft values with a review flag, or held in a staging table for rep confirmation. Fields marked not_discussed should never overwrite existing CRM values. Log every extraction with the original transcript ID, model version, prompt version, confidence scores, and final disposition (auto-written, staged, or escalated). This audit trail is essential when sales reps challenge the extracted data. For high-stakes fields like opportunity stage or close date, consider requiring human approval regardless of confidence score—the cost of a wrong stage update outweighs the automation benefit. Wire the review queue into your existing CRM workflow (Slack notification, Salesforce task, or internal tool) so rejected extractions create a feedback loop for prompt improvement.