An AI meeting summarizer takes a transcript and produces a structured overview: key topics, decisions made, action items and sometimes a verbatim quote or two. Most tools use a large language model (LLM) like GPT-4 or Claude for the summarisation step, after the audio has been transcribed with a speech-to-text model like Whisper.
The two-step pipeline
Meeting AI involves two separate AI systems: (1) a speech-to-text model converts audio to a word-for-word transcript, and (2) an LLM processes that transcript and extracts structure. Whisper handles the first step for most tools; the second step is where most products differentiate themselves, with custom prompts trained on meeting content.
What AI summaries get right
Modern AI meeting summaries are quite good at: extracting a bulleted list of topics discussed, identifying explicit decisions (when speakers say 'we decided', 'we will', 'agreed'), capturing action items when phrased directly ('John will send the report by Friday') and producing a readable paragraph summary of each section.
Where AI summaries still miss
- Implicit decisions: things decided by consensus without an explicit statement.
- Tone and subtext: the AI cannot capture that a topic was contentious or that a stakeholder seemed unenthusiastic.
- Proper nouns and names: still the most common error; review these manually.
- Long meetings: summaries of 2+ hour meetings can lose fidelity on content from the middle section.
- Non-English speakers: accented English reduces transcription accuracy, which cascades to summary quality.
Local vs cloud summarization
Cloud tools send your transcript to an LLM API for summarisation. Ledgeur' summary step also uses an LLM, but can be configured to run locally using an open-source model on your device. This keeps both the audio and the summary generation off any external server — important for privileged or confidential meetings.