The most common failure in AI meeting summary deployments: teams confuse transcripts for summaries. The two are different products that solve different problems. Get the distinction wrong and you end up with output nobody reads. Get it right and you save your team hours per week.
A transcript is what was said (8,000+ words, useful for search and reference, nobody reads it). A summary is what mattered (200 words of decisions, actions, and unresolved questions, people read and forward it). Most 'AI meeting summary' tools default to compressed transcripts. Configure them for actual summaries — or pipe transcripts into Claude with a structured prompt.
A transcript is the verbatim record of dialogue: every 'um,' every tangent, every time someone spent two minutes looking for a file. They're typically 5,000-15,000 words for a 30-60 minute meeting.
Transcripts are useful when:
• You need to search across hundreds of meetings later
• You need to verify what someone specifically said (legal, HR, contract negotiation)
• You're analyzing patterns across many calls (sales call coaching, customer research)
• You'll feed them into other tools (an AI for further analysis, a CRM for record)
Transcripts are NOT useful for: anyone catching up on what happened. Nobody reads transcripts.
A summary is the structured distillation of what mattered: decisions, action items, unresolved questions, key context. 150-400 words for the same 30-60 minute meeting.
Summaries are useful when:
• Someone who wasn't there needs to know what happened
• Decisions or commitments need to be archived as the record
• Action items need clear ownership and deadlines
• You're keeping cross-functional teams aligned without forcing them into more meetings
Summaries are NOT a search index. They're not a verbatim record. They're a distillation.
Look closely at most AI meeting summary output and you'll find it's structured but redundant. 'Topics discussed: A, B, C' followed by paragraphs on each topic that re-narrate the dialogue. This is a compressed transcript, not a summary.
The reason: it's easier to programmatically produce compressed transcripts than to identify what mattered. 'Identify the decisions' requires semantic judgment. 'List what was talked about' requires only sequence detection.
The fix: configure the tool (or write the prompt) to ask for specific structured output, not 'a summary.'
Three questions to ask before you set up any AI meeting summary tool:
1. Who reads this? Attendees? People who weren't there? Future-you in three months? The audience changes the output.
2. What action does it enable? Decision logged in a CRM? Task added to a project tracker? Discussion topic for next meeting? Different actions need different formats.
3. What can be omitted? If a section will never be read, don't generate it. 'Topics discussed' is the most-skipped section in meeting summaries — usually safe to omit.
Most teams skip these questions, get default tool output, and wonder why nobody uses it.
Transcripts aren't bad — they're just for a different job. Times you should keep transcripts:
• Sales calls (CRM record, coaching review, deal intelligence)
• Customer interviews (quote extraction, pattern analysis)
• Legal/HR meetings (verbatim record matters)
• Board meetings (formal record alongside summary)
• Long strategy sessions where you'll want to revisit specific exchanges
Most dedicated AI meeting tools (Fathom, Otter, Gong) keep both the transcript and the summary. Use both — they're for different jobs.
Times you should focus on summary, not transcript:
• Internal team syncs where nobody re-reads after
• Cross-functional updates posted to Slack
• Personal 1:1s where the structured takeaways are what matter
• Decision meetings archived for institutional memory
• Any meeting whose value is in the structured outputs, not the dialogue itself
For these, save the transcript only as a fallback; lead with the summary.