Meeting transcription

A meeting transcript earns its keep when you need a written record you can search, quote, and attach to a decision, and when everyone who will be captured has agreed to that. It is not the default for every huddle. You pay for it in consent, a recording that actually captured the room, cleanup of names and speakers, and a human who still owns what goes out as the notes.

Use transcription when the conversation itself is the artifact: a design review you will cite later, a client call with commitments, a hiring debrief, a policy discussion where the exact wording matters. Skip it when the meeting is short, informal, or already producing a live document the group is editing together. A standup that exists only to unblock people rarely needs a machine-written record.

Transcription is not the same job as dictation AIs. Dictation captures one person's speech into a draft they already intend to write. Meeting transcription captures several speakers, overlapping talk, and decisions nobody typed. Mixing the two leads to the wrong tool and the wrong consent story. A personal dictation workflow does not authorize recording a room.

If you cannot get consent, or the recording failed, the output stays empty. Do not backfill from memory and label it a transcript. Memory can become a note you wrote. It cannot become a transcript you did not capture.

Get consent before anyone hits record. That includes people on the invite, people who join late, and guests added in the moment. A calendar footer or a bot that announces that the meeting is being recorded is a start, not a complete practice. Someone still has to pause when a new person joins, say that a recording is on, and wait for a clear yes, or stop the recording.

Recording without consent is the failure no cleanup step can repair. You cannot transcribe your way out of a capture you were not allowed to make. If a participant refuses, or you never asked, you do not produce a transcript, a summary, or an action list from that session. Empty stays empty. If the group still needs notes, a designated human writes them from what they heard, without pretending those notes came from a recording.

Policy sits above the tool. Some rooms ban recording outright. Some allow it only on company-managed systems. Some require storage in a specific region or deletion after a set period. Tools in this class, including Otter, Fireflies, Granola, Microsoft Copilot, and Whisper-based pipelines, differ in where audio lives and who can see it. Treat those differences as a procurement and legal question, not as a ranking on this page. If policy says the audio cannot leave the tenant, a consumer notetaker is the wrong class of tool.

When one person opts out, the honest options are: do not record, record only after they leave, or take human notes that do not claim to be a transcript of the full room. Do not record anyway and leave their name out of the export. The audio still captured them.

Record only after consent is in place. Then transcribe from that recording. Do not skip the source file and ask an assistant to write up what you discussed as if that write-up were a transcript. That path invents phrasing and invents certainty.

Recording quality is the practical bottleneck. A laptop microphone in a lively room will drop speakers, flatten names, and turn a hard date into a vague one. Prefer the meeting platform's own recording when everyone is remote. In a hybrid room, put a dedicated microphone on the table and repeat decisions out loud so they exist as complete sentences. If the recording is missing, corrupted, or inaudible, you do not have a transcript. Empty stays empty. You can still send a short human note that says the recording failed.

Otter, Fireflies, Granola, and Microsoft Copilot sit in the notetaker class: audio in, timestamped text out, usually with a generated summary. Whisper is the speech-to-text engine many local or custom pipelines wrap. In every case, use the transcript as the source of truth. Use any summary as a map of the file, not as the minutes.

After the transcript exists, spend a few minutes on names, speakers, and the handful of lines you will cite. Fix a wrong name before you extract actions.

For a recurring series, keep a glossary and a roster of who is usually in the room inside AI Projects so product terms and speaker names do not reset every week. That is setup. It does not replace reading the new transcript.

Extract actions only from lines you can point to. Each action should cite a transcript span: a timestamp, a speaker, and the words that created the commitment. If nobody said they would do it, there is no action item. Do not invent one to make the list look complete. Do not invent an owner because the topic sounds like someone's job.

The quality bar is notes that cite the transcript span. A line such as Alex at 14:02 saying "I'll send the revised timeline by Thursday" is an action. A line such as "Follow up on timeline" with a guessed owner and no quote is a wish. Leave the owner blank rather than guessing. Leave the action out rather than paraphrasing a maybe into a must.

A design review includes two teammates and a contractor who joins for the second half. The group agrees the navigation should drop an extra tab. Priya says, at 22:18, "I'll open the ticket and put the screenshots in it today." No one else volunteers. The contractor asks whether legal has signed off on the empty-state copy. Someone answers "we should check," and the talk moves on. A faithful extract has one action: Priya, ticket plus screenshots, citing 22:18. It does not create "Legal sign-off, owner: contractor," because that sentence was never a commitment. It does not create "Jordan owns empty-state copy," because Jordan was not in the meeting. The navigation decision can appear as a decision with a cite, not as an action with a fake owner.

AI assistants can draft that extract if you paste the transcript and require citations. They will still invent owners unless you forbid it. The instruction that matters is quote or omit. If the model cannot find a span, the field stays empty.

Do not treat the model summary as the minutes. Summaries compress, drop dissent, and turn "we might" into "we will." The notes you send are a human-selected set of decisions and cited actions. If a claim in the summary cannot be found in the transcript, it does not ship.

Transcription does not send the notes. The meeting owner does. That is the last step. Skip it and a private transcript sits in a tool inbox while the team operates on folklore.

The owner reviews the cited extract, deletes anything that lacks a span, restores anything the model missed that they can cite, and sends a short note to the people who need it. The transcript can be linked or attached for anyone who must verify a quote. The message or document the owner sends is the record the team will use. The auto-summary is not.

If consent was missing or the recording was missing, the owner may still send notes, but those notes are labeled as human notes, not as a transcript-backed record. They should not include invented action items to fill the silence.

If you need duration for billing or capacity, pull it from the calendar or from time tracking, not from how long a topic felt. Transcript timestamps help only when the recording exists.

The usable loop is consent, record, transcribe, extract actions with cites, then the owner sends. Break any step and the quality outcome fails. Recording without consent is not a transcript workflow. Inventing an owner is not extraction. Treating the summary as the minutes is not sending notes. The written record is worth the overhead only when those steps stay intact.