What to do with a transcript

Getting a transcript feels like finishing, and it is closer to starting. A transcript is raw material: complete, searchable, and almost never the thing you actually needed. The value comes from what you turn it into, and from handling it sensibly once it exists.

This is the step most guides skip. Here is a practical sequence, from the minute the transcript appears to the point where you can safely forget about it.

First, correct it while the audio is fresh

Do this before anything else, because it gets harder every hour. Speech models are reliably wrong about the same categories of thing, and those are exactly the categories that matter:

• Proper nouns.People's names, company names, product names, and place names. A model has no way to know how your colleague spells theirs.
• Technical and domain terms. Anything specific to your field, especially acronyms.
• Numbers. Dates, figures, and quantities. These are easy to misread and consequential when wrong.
• Negations.A dropped "not" inverts a sentence completely, and it reads perfectly naturally, so nothing flags it. This is the single most dangerous error class.

Fix these in the editor before you download anything, so the corrected version is the only one that ever exists as a file. If you are also renaming speakers, rename first and then correct the wording, because applying names rebuilds the transcript and would discard edits made beforehand.

Decide what you actually need

This is where people waste the most time, by producing a summary when they needed a record, or a full transcript when they needed three action items. Pick one:

• Notes, if colleagues need to know what happened. Short, decision-focused, and written for whoever was not there. See our guide to meeting notes.
• Minutes, if this was a formal meeting with a governance requirement. A different document with different rules, covered in meeting minutes.
• Action items, if the point was to leave with commitments. Often the only thing anyone will read. See turning a transcript into action items.
• Quotes, if you are writing something and need exact words. Verify each one against the audio before it goes anywhere public.
• The transcript itself, if the point is a searchable record you will come back to. This is legitimate, and it is rarer than people assume.

Being honest about which of these you need usually reveals that the transcript was a means, not the deliverable.

Using an AI assistant, without it inventing things

Turning a long transcript into a summary is exactly the kind of task an AI chat is good at, with one serious caveat: models will happily produce a confident summary containing things nobody said. The defenses are straightforward:

• Instruct it to use only the transcript and to omit sections where information is absent, rather than filling gaps.
• Ask for attribution. Requesting who said what makes fabrication easier to spot, because invented claims tend to have no clear owner.
• Spot check the specifics. Every name, number, date, and commitment in the output should be findable in the transcript. If it is not, it was invented.

After transcribing, the Copy notes prompt button gives you a prompt built along these lines, which you paste into whichever assistant you use. Nothing is sent by the copy itself. What you paste, and where, is your decision, and it is a real one: pasting a confidential transcript into a consumer AI service is a disclosure to a third party regardless of how the interface feels. For sensitive material, writing the summary yourself is the conservative choice.

The transcript is as sensitive as the recording

This is the part that gets handled badly most often. If the recording was confidential, the transcript is confidential. It is arguably worse, because text is searchable, copyable, and trivially forwarded in a way audio is not.

Practical handling:

• Save it where the sensitive material already lives, the matter file, the client record, the secure project folder. Not the downloads folder, which is where files go to be forgotten.
• Name it usefully. Date first in year-month-day order, then subject, so files sort chronologically and are findable in two years.
• Redact before sharing. If a version is going to a wider audience, make a separate redacted copy rather than editing the original.
• Check whether your working folder syncs to a cloud drive. Transcribing locally to protect a recording, then saving the transcript into an automatically synced folder, undoes most of the benefit.

Then delete the audio

Once the transcript is corrected and the artifact you needed is written, the recording is usually pure liability. It is larger, more sensitive, more identifying, and less useful than the text.

Many organizations have a retention policy that requires destroying meeting recordings after minutes are approved, precisely for this reason. If yours does, follow it. If it does not, a sensible default is to delete the audio once you have verified the transcript, keeping it only where there is a specific reason: an unresolved dispute about what was said, an ongoing matter, or a documented retention requirement.

Remember the copies. The original on the recording device, the copy on your laptop, and anything that synced in between are all separate deletions.

A short sequence

1. Rename speakers, then correct names, terms, numbers, and negations.
2. Decide which single artifact you need.
3. Produce it, verifying any specifics that came from an AI assistant.
4. Save the transcript with the sensitivity it inherits, named to be findable.
5. Delete the audio, everywhere it exists, unless you have a reason not to.

The whole sequence takes a few minutes and turns a transcript from a file you have into work you have finished. If you have not made one yet, the transcriber runs in your browser and your audio never leaves your device.