Three channels, one place
Part 1 was about the habit. This part is about the plumbing, and about one question that most capture setups never ask.
Start with the plumbing, because it is simple.
Most people capture a third
Conversations happen in three places, and almost everyone captures one of them.
Online meetings. Teams, Zoom, Meet. This is the easy channel. Transcription is built in, the speakers are already labelled because everyone has their own stream and login, and it costs nothing extra. Most people stop here.
Physical rooms. Client meetings, workshops, the conversation over lunch where the important thing gets said between courses. Nothing is built in. You need a recorder in your pocket and the habit from Part 1.
Phone calls. The channel almost everyone forgets entirely. Calls contain the fastest decisions and the least prepared answers, because there is no screen, no presentation, and no one performing.
The uncomfortable pattern: the easiest channel to capture is usually the least candid one. If you capture only what is convenient, your record ends up systematically biased towards the formal and away from the honest.
One place, not three
Whatever you use to capture, the output has to land in one place.
Three tools producing three separate archives is not a capture system. It is three piles, and you will search two of them and forget the third. It also breaks everything Part 3 depends on, because you cannot ask a question across material that lives in three places with three formats.
The principle is simple: it should not matter whether a conversation happened in Teams, in a conference room, or on a phone. Same destination, same processing, same chain afterwards.
That is the whole of the plumbing. Now the part that actually takes thought.
Not everything may land anywhere
Three channels landing in one place raises a question the setup itself does not answer: which place is permitted?
And that has no single answer. It depends on what the conversation was.
A voice note you record for yourself carries essentially no requirements. An internal check-in carries few. A client meeting where someone commits to something starts to be contractually relevant. A salary review, a board discussion, a conversation about someone's health -- those carry requirements that are not negotiable, and they apply from the moment recording begins.
The principle worth internalising:
The requirement is a property of the material, not a setting in your system.
You do not configure "a secure setup" once and stop thinking. Set everything to the strictest level and capture becomes so cumbersome that you quietly stop doing it, which loses you everything. Set everything to the most convenient level and eventually the wrong conversation lands in the wrong place.
Both failure modes are common. The second is the one people notice.
Two questions that get merged
There is one distinction worth getting right, because nearly everyone collapses it into one.
Where is the material stored?
Where is it processed?
A provider storing everything in Sweden answers the first question and says nothing about the second. Material can sit on a Swedish disk and be sent to another continent for the actual inference.
The reverse also holds, and it surprises people: a model of any national origin can be entirely unproblematic if the weights run on hardware you control and nothing leaves your network. What matters is where the processing happens, not where the vendor is from.
When you evaluate a tool, ask both questions separately. Most marketing answers only the first.
Decide at capture
The practical rule that makes all of this workable:
Classify when you record, not afterwards.
The person who was in the meeting knows what the meeting was. They know it clearly for about a day. Three months later the recording is one file among a thousand and nobody can reconstruct whether it was a routine catch-up or something that needed handling.
Anything you plan to sort out later will, in practice, not be sorted out.
This does not mean filling in a form before every call. Most of it can come from what you already know at the moment of recording: is anyone external in this conversation, is this a standing internal meeting, is this about a person rather than a project. What you need is a default you can state out loud, and an easy way to override it upwards when something is sensitive.
A default worth considering: material with no stated classification may be stored and searched, but does not take part in combined answers across many conversations. That way the unhandled material stays useful to you individually without quietly leaking into a summary someone else reads. Part 3 explains why that distinction matters.
What to actually do
- List your three channels and mark which you currently capture. For most people it is one.
- Pick one destination. Everything lands there regardless of channel.
- Write down your default classification in one sentence. If you cannot state it, you do not have one.
- Add one upward override you can trigger in two seconds. Something for "this one is sensitive", usable while the conversation is still happening.
- Ask any tool you evaluate both questions. Where is it stored, and where is it processed.
What comes next
You now have material arriving from everywhere, and you know what may be done with it.
A transcript, though, is not an answer. It is raw material, and the most common mistake is to ask for a summary and consider the job done.
That is Part 3.