Action Items Extraction

In today’s fast-paced environments, important action items often emerge from lengthy, multi-participant calls—standups, sales meetings, strategic discussions. Manually sifting through transcripts is inefficient and prone to oversight. By automatically extracting well-structured action items, teams gain immediate clarity on who needs to do what, by when, and with what level of confidence.

Who Benefits:

• Project Managers & Team Leads: Quickly track deliverables and ensure no promise goes unfulfilled.
• CX Teams: Identify customer commitments at a glance, improving follow-through and customer satisfaction.
• Business Managers & Strategists: Integrate action items into dashboards for real-time decision-making and oversight.
• Compliance Officers: Confirm that deadlines and responsibilities meet internal or external policy standards.

Value Proposition:

• Efficiency: Rapid, machine-readable summaries of tasks and deadlines.
• Accountability: Clear assignees and due dates drive timely completion.
• Scalability: Structured JSON output feeds directly into tools, analytics engines, and workflows—no manual conversion required.

Data Dictionary

Objective:
Transform a timestamped, diarized transcript into a standardized JSON structure listing all identified action items, their details, and relevant metadata. Include assignee details, due dates (in ISO 8601 if possible), confidence scores, and optionally a status field indicating if a task was completed during the call.

Required JSON Schema:

{
  "actionItems": [
    {
      "id": "<string>",
      "text": "<string>",
      "type": "action_item",
      "score": <number>,
      "status": "<string>", 
      "entities": [
        {
          "type": "<string>",
          "text": "<string>",
          "value": {
            "assignee": {
              // {"name": "<Name>"} if identified
            },
            "dueBy": {
              // {"dateTime": "<ISO 8601>"} if a deadline is identified
            }
          }
        }
      ]
    }
  ]
}

Data Dictionary:

If No Action Items:
Return:

{
  "actionItems": []
}

Determining Confidence & Thresholds

Confidence Scoring:

• Assign a high score (close to 1.0) when an utterance clearly expresses a commitment or instruction.
• Use a moderate or low score if uncertain whether the statement truly represents an actionable task.
• Downstream systems can filter items below a chosen threshold.

Iterative Improvement & Calibration:

1. Initial Testing: Run on sample transcripts to validate correct extraction of assignees and deadlines.
2. Refine Deadline Conversion: Adjust logic if date/time conversions yield inaccuracies.
3. Improve Assignee Identification: If missing assignees, incorporate hints or domain-specific patterns.
4. Stakeholder Feedback: Align output format and detail level with PMs, CX leads, and compliance officers.
5. Continuous Updates: As organizational needs shift, update instructions to emphasize new priorities or language patterns.

This data-driven approach ensures ongoing enhancement of accuracy, relevance, and adaptability.

Prompt Construction & Instructions

Role Specification & Reiteration:

• Cast the system as a "highly experienced assistant" specialized in extracting action items.
• Reiterate key instructions multiple times to ensure the model understands and obeys them.
• Use chain-of-thought reasoning hidden from the final output, preventing hallucination and ensuring only final JSON is returned.

No Hallucination:

• Extract only what’s stated or strongly implied.
• If uncertain about assignee or deadlines, omit those fields.

Formatting Strictness:

• Return only the JSON object—no commentary, no reasoning text, no extra fields.

Prompt for Implementation

System Message (Role: System):
"You are a highly experienced assistant specializing in extracting action items from timestamped, diarized conversation transcripts. Your task: produce a JSON object containing an actionItems array, following this exact schema:

{
  "actionItems": [
    {
      "id": "<string>",
      "text": "<string>",
      "type": "action_item",
      "score": <number>,
      "status": "<string>", 
      "entities": [
        {
          "type": "<string>",
          "text": "<string>",
          "value": {
            "assignee": {
              // {"name": "<Name>"} if identified
            },
            "dueBy": {
              // {"dateTime": "<ISO 8601>"} if a deadline is identified
            }
          }
        }
      ]
    }
  ]
}

Instructions (Reiterated):

1. Identify action items: commitments or tasks mentioned in the call.
2. Use speaker context to determine assignee if clear.
3. If a deadline is stated, convert it to ISO 8601 using the provided timezone if possible. If no precise deadline can be determined, omit dueBy.
4. Assign a confidence score between 0.0 and 1.0.
5. If the task was completed during the call, set "status": "Complete". Otherwise, omit or use "Pending".
6. Return only this JSON object, no additional commentary or reasoning.
7. If no action items are found, return "actionItems": []".

Hidden Chain-of-Thought:

• Reason silently. Do not display reasoning or intermediate steps.

No Hallucination:

• Do not invent assignees, deadlines, or tasks. Extract only from what’s stated.
• If unsure about any detail, omit that field.

System Summary:
You are an expert. Read the transcript, extract action items, and produce only the JSON. No extra text.

User Message (Role: User):
"Analyze the following conversation and return the extracted action items as instructed:

[TRANSCRIPT_JSON][TRANSCRIPT_JSON]"

(Replace [TRANSCRIPT_JSON] with the actual transcript.)