Episode 008 · 6 min
The follow-up that writes itself
Buy Back Your Hour, part two. Maya records a meeting with Otter's free plan and lets it draft the follow-up, the summary and the action items. Sam's objection is sharp: a machine that was not in the room decided what mattered. The turn is one human check that keeps it safe, and it costs about twenty seconds.
What this episode says
- The tool can draft the follow-up, but the action items are a fluent guess about what the meeting meant, and fluent is not the same as correct.
- The one human check worth keeping: read the action items aloud before you send. Your ear catches what your eye slides past.
- Whether a summary is true and whether it was recorded fairly are two different questions. Do not let the tool choice answer both.
- Voice beats text at the moment a task changes hands. If you are handing work to someone, say it, do not just type it.
Transcript
Every word, in full. Nothing is said in the audio that is not here.
Bit: This is Buy Back Your Hour, from The Human Bit. Maya and Sam are recurring characters who bring common workplace situations and reviewed research into the conversation. Their examples are composites, not recordings of real meetings.
Sam: You're doing it again. The meeting finished ninety seconds ago and you're already sending the follow-up.
Maya: Because in ninety seconds I'll have forgotten who agreed to what.
Sam: That is either impressive or slightly frightening.
Maya: It's mostly typing. The tool did the boring part. I'm just deciding what actually leaves my outbox.
Sam: See, that's the bit I don't trust. "The tool did it." Did what, exactly?
Maya: Give me one second, and I'll show you the thing you're about to be annoyed about.
Bit: I'm Bit, and this is The Human Bit. Buy Back Your Hour, part two. Last time these two argued about turning a meeting into notes. Today it's the job right after that, the one that quietly eats your afternoon. The follow-up. The summary, the action items, the who does what by when.
Sam: The email that dies in Drafts.
Bit: For most people, yes. Maya, apparently, has a system.
Maya: It's not a system. It's one tool and one habit. The habit is the part that matters, so hold me to getting there.
Bit: Start with the tool. What are you actually using?
Maya: Otter. The free one. I press record at the start of the call, and at the end it hands me a transcript, a summary, and its best guess at the action items.
Sam: For free.
Maya: For free, with limits. Three hundred minutes a month, and thirty minutes per conversation. Which sounds generous until you have a meeting that will not end, and it stops listening halfway through somebody's monologue.
Sam: So it has abandoned you mid-meeting.
Maya: Once. In a workshop. I've got the first half beautifully transcribed, and the second half is just my own increasingly worried face. So now I know its edge, and I plan around it.
Bit: That's the first honest thing about this. The free tool has a limit, and you found it the annoying way.
Maya: Everyone finds it the annoying way. But for a normal thirty minute meeting, it's genuinely good. It writes the draft follow-up in about the time it takes me to refill my coffee.
Sam: And then you send it.
Maya: No. Then the actual work starts. And that's where you come in, because you're about to make my point for me.
Sam: Fine. Here's my problem. A machine that was not in the meeting, that does not know anyone in that room, has decided what mattered enough to become an action item. Who approved that?
Maya: Nobody. That's exactly right.
Sam: That was quick.
Maya: Because it's the whole point. It didn't approve anything. It guessed. It's a very fluent guess about what your meeting was about, and fluent is not the same as correct.
Bit: There's a number under this, if it helps. A 2026 benchmark puts hallucination rates anywhere from about zero point seven per cent on grounded summarising, which is roughly this job, up to eighty eight per cent on much harder work like legal questions.
Sam: So the summary is the safe end.
Bit: The safe end. But zero point seven per cent is not zero. And when it's wrong, it won't be wrong in a harmless place. It'll be wrong about the one number, or it'll put a name next to a task that person never agreed to.
Maya: It once had me delivering something that was actually Sam's.
Sam: When?
Maya: It's a made up example.
Sam: It did not sound made up.
Sam: Alright, second problem, and this one isn't about accuracy. If it's recording, it's recording everyone. Half these tools now argue about exactly this. Some send a little bot that joins the call so the whole room can see it. Others, like Granola, just listen on your laptop and nobody in the room knows.
Bit: That split is honestly the argument of the year in this corner of the market. A bot in the call, or no bot at all.
Sam: And "nobody knows you're recording" is not a feature to me. It's a consent problem wearing a nice interface.
Maya: You're right, and I'm not going to pretend the tool settles that. Whether people know they're being recorded is a decision I make out loud before the call. It's not something I let a checkbox make for me.
Bit: So the tool choice, bot or no bot, is a privacy posture. It's not the thing that makes the follow-up safe.
Maya: It isn't. You can pick the most polite, most visible, fully consented tool on the market, and still send an action list that quietly handed the launch to the wrong person. Whether it recorded fairly and whether the summary is true are two different questions. People keep merging them.
Bit: So what's the habit? You keep saying the habit is the part that matters.
Maya: It's embarrassingly low tech. Before I send the follow-up, I read the action items out loud. Actually out loud, at my desk, like a slightly unhinged person.
Sam: I've heard you. I thought you were on a call.
Maya: I know. But something happens when you say "Sam owns the client update by Thursday" out loud that does not happen when you skim it. Your ear catches what your eye slides past. If it's wrong, you hear it. If I never actually agreed to Thursday, my own voice tells on me.
Bit: That's better than "check the output", isn't it. "Check the output" is advice that means nothing. Reading it aloud is a specific thing a person can do in twenty seconds.
Maya: And it moves the check to the one moment that counts. Right before a task changes hands. Because the second I hit send, someone else believes they've been told what to do.
Sam: That's the part I actually care about. Not the transcript. The moment it becomes somebody's job.
Maya: Voice beats text there. Every time. If a task is changing hands, say it, don't just type it.
Bit: So walk me through the send.
Maya: Read the action items aloud. Fix the one it got wrong, there's usually one. Delete the two it invented. Add the thing everyone agreed but nobody said clearly. Then send.
Sam: How long does all that take?
Maya: Two minutes. The tool saved me twenty. I spend two of them making sure the twenty were worth it.
Sam: That's annoyingly reasonable.
Maya: Melbourne's still in Victoria, too.
Sam: You had to bring it up.
Bit: Here's where we leave it. The tool can write the follow-up. It cannot decide what's true, and it cannot be the one who's answerable when Thursday comes and the thing isn't done. That stays yours. So pick your own version of Maya's habit. Read it aloud, or say it on the call, or make one phone call before the email. Whatever makes you hear the task before you hand it over.
Bit: This is The Human Bit. The tool does the dull part. You keep the judgement.