Follow-Ups Detection

In complex conversations—whether pre-sales calls, cross-functional meetings, or strategic briefings—important follow-ups often emerge. A follow-up signals that additional information, confirmation, or coordination is required in the future. Identifying these early helps teams prevent delays, keep projects on track, and ensure all unanswered questions are addressed promptly.

Who Benefits:

• Project Managers & Team Leads: Quickly locate points requiring further confirmation, reducing the risk of stalled initiatives.
• CX Teams: Identify where agents need to gather more details from customers or internal departments, improving responsiveness.
• Business Analysts: Track unresolved topics that need more input before decision-making.
• Compliance & QA Officers: Confirm that no critical pending inquiry or action is overlooked, maintaining regulatory or policy compliance.

Value Proposition:

• Efficiency: Standardized follow-up extraction streamlines workflows, so tasks don’t get lost.
• Accountability: Clear identification of what still needs attention fosters better follow-through.
• Scalability: A JSON-based schema easily integrates into dashboards, analytics tools, or third-party applications.

Data Dictionary & Schema

Objective:
Convert a timestamped, diarized conversation transcript into a structured list of follow-ups. A follow-up indicates a need for future action, additional info, or engagement with another party or resource.

Required JSON Schema:

{
  "followUps": [
    {
      "id": "<string>",
      "text": "<string>",
      "type": "follow_up",
      "score": <number>,
      "entities": [
        {
          "type": "<string>",
          "text": "<string>",
          "value": {
            "channel": "<string>"
          }
        }
      ]
    }
  ]
}

Data Dictionary

Note:

• followUps is an array of objects. Each object represents one follow-up.
• entities can provide additional context like the channel (person/team) involved in the follow-up.
• If no follow-ups are found, simply return "followUps": [].

If No Follow-Ups:
Return:

{
  "followUps": []
}

Confidence & Calibration

Confidence Scoring:

• High (0.8–1.0): Clear statements needing future action, like “I will find out...”, “We must schedule...”, or “We’ll need to check...”
• Medium (0.5–0.8): Less certain language—“I might have to...”, “Perhaps I should confirm...”
• Low (<0.5): Vague or speculative mentions that may not represent a true follow-up.

Calibration Tips:

1. Initial Testing: Apply on various transcripts and verify if genuine follow-ups are detected accurately.
2. Refine Entity Recognition: If channels aren’t correctly identified, add domain-specific instructions.
3. Stakeholder Feedback: Incorporate input from PMs, CX leads, or compliance teams to adjust scoring or definitions.
4. Continuous Improvement: As language evolves (e.g., new phrases for follow-ups), update instructions accordingly.

This data-driven approach ensures that the extraction process improves over time, aligning with real-world needs.

Prompt Construction & Instructions

Role Specification & Reiteration:

• The system is a “highly experienced assistant” that focuses on identifying follow-ups from transcripts.
• Reiterate the schema and instructions multiple times to ensure compliance.
• Instruct the model to think silently (chain-of-thought) and output only the final JSON.

Avoid Hallucination:

• Only extract follow-ups mentioned or strongly implied in the transcript.
• If unsure whether a statement is a follow-up, assign a lower score or omit it.
• Do not invent channels or follow-up tasks not present in the text.

Formatting Strictness:

• Return only the JSON structure—no reasoning, no additional commentary.
• If no follow-ups are found, return an empty array.

Prompt for Implementation

System Message (Role: System):
"You are a highly experienced assistant that identifies follow-ups from a timestamped, diarized conversation transcript. A follow-up indicates a need for future action, further information, or additional coordination. Your output must adhere to the following JSON schema:

{
  "followUps": [
    {
      "id": "<string>",
      "text": "<string>",
      "type": "follow_up",
      "score": <number>,
      "entities": [
        {
          "type": "<string>",
          "text": "<string>",
          "value": {
            "channel": "<string>"
          }
        }
      ]
    }
  ]
}

Instructions (Reiterated):

1. Identify all follow-ups in the transcript. A follow-up suggests a future action or confirmation.
2. Assign a unique id (e.g., "followUp-1").
3. text should capture the core follow-up phrase.
4. type must be "follow_up".
5. Assign a score reflecting confidence.
6. For entities, if the follow-up references a person/team ("manager", "design team"), classify it as type: "channel" and include "channel": "<string>" in value.
7. Return only the JSON, no explanations or reasoning steps.
8. If no follow-ups, return "followUps": []".

Chain-of-Thought (Hidden):

• Reason silently. Do not show intermediate reasoning.

No Hallucination:

• Extract only what is present in the transcript.
• If uncertain, use a lower score or omit the item.

System Summary:
You will read the transcript, detect follow-ups, and return them in the given JSON format. No commentary, just the JSON."

User Message (Role: User):
"Analyze the following conversation and extract follow-ups according to the specified schema:

[TRANSCRIPT_JSON][TRANSCRIPT_JSON]"

(Replace [TRANSCRIPT_JSON] with your actual conversation data.)