Fifteen Years for the Shallowest Layer

In 2011 I knew it ought to be possible. I could see how much disappeared every day, in meetings and conversations, in the gap between what people actually said and what ended up in the minutes afterwards. The problem was that the technology could not hold up.

In 2022 Whisper arrived, and suddenly it could.

In 2026 essentially everyone has it. It is built into Teams. It comes with the licence. Nobody needs convincing any more that it is possible.

Fifteen years from knowing it ought to work to it being taken for granted.

And what we reached is the shallowest of the layers.

This is not a complaint

I want to be clear that this is not a grumble about a slow market. It is an observation about where we actually stand, and it is more interesting than it sounds.

A transcript is an account of one occasion. It does not know what it is about. It does not know what it belongs with. It does not know who said anything, and almost never to whom. It is a verbatim text of an event, and that is all.

It is a good first layer. It is simply not very deep.

And right now almost the entire conversation about AI and meetings reads as though that were the goal. You get transcription, you get a summary, and the project is done.

Three layers, three questions

It gets easier to see if you ask what each layer actually answers.

The transcript answers: what was said?

The structure answers: what belongs together?

The origin answers: who said it, and to whom?

Three questions, three layers, in that order. You cannot skip any of them, and you cannot start at the top.

Why the first layer is finished

There is something easy to miss about transcription: it is effectively done, and that is good news.

The quality is sufficient. Swedish works. There are a handful of vendors who do it well, it is built into the tools most people already pay for, and the price is heading towards zero. If anything that is the final sign that a problem is solved -- when it stops being a product and starts being a feature.

Which means the layer is not somewhere to build an advantage. It is not somewhere to spend your attention either. It is done, and you should be glad about it, because it took fifteen years.

The question is what you do with it.

Gathering is the easy part

The second layer is where the work actually is.

Ten transcripts in a folder are not a basis for anything. They are a pile. What turns the pile into something usable is the structure between the documents -- what is what, what belongs with what, how something developed over time, who keeps coming back and in which contexts.

Only there can you ask a question of the whole rather than of one document. And that is a different kind of question. Not "what did we say in the April meeting", but "what is it this client always comes back to".

Gathering is the easy part. Knowing what it means is the hard part.

That is also why the second layer will not become a feature in somebody's licence the way the first one did. It cannot be delivered generically, because the structure has to fit whoever is going to use it.

And then somebody promised something

The third layer is the one almost nobody talks about.

A transcript can be perfectly structured and still be unable to drive a project, for one reason: it does not know who said what to whom. "We decided the quote goes out before Friday" is a note. Someone has to have taken it on, in front of someone, for it to become a commitment.

I have written about that layer separately, because it deserves its own text and it has an uncomfortable side that needs handling properly. Here it is enough to say that it exists, that it comes last in the order, and that it is what separates a well-ordered archive from something that actually makes things happen.

Running through all three

One thing runs through the whole ladder and gets harder the further up you go: what the material is allowed to be part of.

A voice note to myself carries no particular requirements. A salary review does. The requirement is a property of the material, not a setting in the system, and it is decided when you record rather than afterwards.

On the first layer this is simple -- one meeting, one type, one requirement. The further up you go, the more material is mixed into the same answer, and the more it matters to know what may be mixed with what. That is not a reason to avoid it. It is a reason to decide early.

Where we actually stand

Fifteen years to reach the first layer. The second has been started by a few. The third is effectively unclaimed.

I find that an encouraging picture rather than a bleak one. It means that what looks like a mature category -- meetings, transcription, AI summaries -- is a category in which almost everything remains.

Most of what is said about AI right now is about models getting better. But the model has not been the hard part for a while now. The hard part is knowing what you are missing.