Plenty of teams run meetings the same way now. Somebody props a phone against a coffee mug and hits record, an app like the one at www.getvoicerecorder.com spits out a transcript with timestamps plus an AI summary with action items, and by lunch everyone has read the twelve-line version instead of sitting through the hour again. Handy. Mostly. The trouble is the stuff that gets squeezed out on the way from sixty minutes to twelve lines.
Here’s a small one. Summary: “Maria will send the revised budget by Friday.” What Maria said, forty-ish minutes in, was more like “I’ll try to get it out Friday, if finance signs off on the headcount thing.” Three people read the summary. Nobody opened the recording. Come Friday, two of them were refreshing their inbox, waiting for a budget she never actually promised.
Nobody lied there, and the model didn’t hallucinate anything wild. It just chopped off an “if”. That’s the usual pattern, in my experience. Summaries don’t make up meetings. They sand off doubt and conditions and who-said-what, and you’re left with something that sounds a lot surer than the room was.
Two Machines, Two Ways to Get It Wrong
Under the hood, there are usually two separate steps. First, speech-to-text: audio becomes words. Then a language model reads those words and writes the summary. Step one makes its own kind of errors, step two inherits them, and then (this is the annoying part) polishes them into tidy confident sentences you’d never think to question.
Speech-to-text is way better than it used to be. Just not equally better for everybody. Back in 2020, a PNAS paper ran the big five systems (Amazon, Apple, Google, IBM, Microsoft) and found word error rates of 0.35 for Black speakers vs 0.19 for white speakers. Nearly double. Heavy accents, people who mumble, people who speed up when they’re nervous, all of that drags accuracy down too.
And then there’s the weird stuff. A 2024 study on OpenAI’s Whisper found around 1% of transcripts had entire phrases nobody said, and 38% of those made-up bits were harmful in some way, like false associations or invented authority. One percent sounds tiny. If your team records twenty meetings a week, though, do the math.
Where It Usually Goes Sideways
On top of transcription errors, the summary step adds its own. Five spots cover most of what actually bites:
- The “if” and the “maybe”. Hedges are the first thing to go. Maybe becomes will.
- Not vs. now. One misheard word, and “we’re not moving the launch” turns into “we’re now moving the launch.” On a bad Zoom line those sound almost identical.
- Who’s on the hook. People talk over each other, speaker labels swap, and the task ends up assigned to whoever spoke last instead of the person who actually said “I’ll do it.”
- Numbers and names. Fifteen, fifty. A client’s surname spelled three different ways in one doc.
- Stuff nobody said out loud. If the decision happened via a nod, or in the Zoom chat, it simply isn’t in the audio.
Funny thing is, the summary almost always gets the topic right. You’ll know what was discussed. What it’s shaky on is what got decided, by whom, and with what strings attached, which is, well, the only part anyone acts on.
Checking It Without Re-Listening to an Hour
Here’s what I’d do. Quit treating the summary like it’s the minutes. It’s more like the index at the back of a textbook: handy for finding stuff, useless as the actual story.
For a one-hour meeting, there aren’t that many lines that matter. Action items, anything with a number or a date, anything that sounds like a decision. Five to ten lines, usually. Check just those against the transcript. Ignore the rest.
Whether that’s painless or a slog comes down to the app. If the summary lives in one place and the audio in another, you end up dragging a scrubber back and forth trying to find one sentence, and you’ll give up by the second time. Built off a searchable transcript with timestamps? Then you type “Friday”, hit enter, done.
Quick test for whatever tool you use: can you get from a line in the summary to the exact second of audio behind it in under a minute? No? Then trust that summary a bit less than its nice bullet points suggest.
Garbage In, Garbage Out
Put the phone in the middle of the table. Not next to the laptop fan, not under the AC vent. Get people to say their names at the start, it gives the transcript something to hang speaker labels on.
The best trick costs half a minute. Right before everyone leaves, one person says the decisions out loud, something like: “OK, so Maria sends the budget Friday, assuming finance signs off on headcount.” That one clean sentence is gold for the summarizer, conditions and all. And it gives the room a last shot at “wait, that’s not what I said,” which no software is going to do for you.
Also, tell people you’re recording. Some US states require everyone’s consent for private conversations, and anywhere else it’s just basic courtesy. Worth knowing too whether your app processes audio on the phone or ships it off to a server, since meeting recordings are full of exactly the stuff companies like to keep in-house.
So, Are Summaries Worth It?
Yes. I’d still use them. Fast, searchable, and they pick up side stuff a half-asleep note-taker would totally miss. If you skipped the meeting, a summary beats nothing by a mile.
Where it goes wrong is when the summary sort of quietly becomes the official story. You can check a transcript. A summary is a model’s take on a transcript that might already be off, and it reads smoother than both. Keep the audio around, check the handful of lines people will act on, and Maria’s Friday budget mess mostly stops happening.
Related: AI Meeting Notes: Why Perfect Transcripts Still Fail Teams
