Meeting Minutes AI: From Recording to Action Items

Transcription, transcript cleanup, LLM summaries and action item extraction

Meeting minutes are a high-frequency but low-efficiency administrative task in the enterprise. Manually writing up a 60-minute meeting often takes more than 30 minutes of effective time, and quality varies significantly with how well the note-taker understood the discussion. An AI-assisted automated workflow can compress that time to a few minutes, but workflow design determines output quality, and installing a tool does not deliver automation by itself.

The Four Stages of the Full Workflow

Speech-to-Text (Transcription)

Transcription is the first stage and the foundation. The main options include Zoom’s built-in AI transcript, Otter.ai, Fireflies.ai, and the general-purpose Whisper API. The main factors affecting accuracy are audio quality (remote video calls capture sound far worse than in-person recording), accent diversity (non-native English pronunciation in multi-person meetings raises the error rate), and recognition of proper nouns and industry terminology.

For Hong Kong businesses, meetings that mix Chinese and English are a common challenge. Mainstream transcription tools differ significantly in how they handle mixed Chinese-English speech, so before purchase it is advisable to test with actual meeting recordings rather than relying solely on the accuracy figures supplied by vendors.

Mixed-language content is also a common source of failure in document processing; the related system design approach is covered in AI processing of mixed Chinese-English documents.

Transcript Cleanup

Raw transcripts contain many spoken filler words (“that one”, “I mean”, “um”), repeated phrases and interrupted sentences. Basic cleanup before the transcript goes to the LLM for summarisation significantly improves summary quality. This step can be handled by rules (regular expressions filtering common fillers) and needs no additional AI call.

LLM Summary Generation

The cleaned transcript goes to the LLM, which generates a summary from a structured prompt. An effective prompt states: the required output format (list of discussion topics, decisions, open items), the handling rules (keep specific figures and commitments, and generate no inferences absent from the source), and speaker identification (if the transcript carries speaker labels, keep the attribution in the summary).

A summary without a declared format tends to produce a running-account paragraph that needs a second round of editing, which cancels the efficiency advantage of automation. For more on prompt design, see Prompt Engineering in Practice: Five Fixes for AI Output Quality.

Action Item Extraction

Action items must be identified from the summary or the raw transcript: who committed to do what, and by when. LLM accuracy on this task depends on how clearly the commitments were stated in the transcript. If a commitment in the meeting was vague (“sort it out before next time”), the extracted action item is equally vague. When designing a production system, require human confirmation after action item extraction rather than syncing directly to a task management tool.

For each stage of a production workflow, keep a traceable record of inputs and outputs; the design approach is covered in AI Application Log Design: What to Record, What Must Go to Database.

Tool Cost Structure

A self-built workflow offers the most flexibility, with full control over summary format and action item extraction logic, but carries the integration engineering cost.

Summary

The most common failure in meeting minutes automation is assuming that installing a tool completes the automation. In practice, an effective workflow requires separate decisions at transcription tool selection, transcript cleanup, summary prompt design and action item confirmation. The design quality of each stage determines whether the final output is genuinely usable, or only a quickly generated draft.