The Transcript Is Not the Deliverable

A recorded call produces a transcript, and a transcript is raw material. The deliverable is a short list of actions, each with an owner and a date, written in the language of whoever has to do the work. Everything between the recording and that list is processing, and processing is the part that carries the value.

What is wrong with meeting notes?

They are a compressed transcript. Notes follow the shape of the conversation rather than the shape of the work, so the commitment that mattered sits wherever it happened to come up.

One test sorts them, and most notes fail it. A week later, can somebody who was not on the call do the work from this document alone? If they have to ask what was meant, the notes recorded a meeting rather than an outcome.

What each artefact is good for, and where it fails:

Artefact	Who reads it	Useful for	Where it fails
Recording	Almost nobody	Settling a dispute about what was agreed	Nobody rewatches the call
Auto transcript	A search box	Finding the exact wording later	Reads as speech, buries the commitments
AI summary	Whoever missed the call	Catching up without watching it	Flattens the critical and the incidental
Meeting notes	The person who wrote them	Feeling productive	Follows the meeting’s order, not the work’s
Internal action list	The account owner	Deciding what to do and what to hold	Carries judgement that is not client-facing
Shared action list	The person doing the work	Getting the work done	Useless without an owner and a date

What has to be in an action item?

A verb, a named owner, a date, and a definition of done specific enough to argue about. The owner is the field people drop most often, and an item with no name against it is the one that quietly never starts. Miss the date or the definition of done instead and the failure is slower: the work stays permanently available to start later, and whatever eventually turns up is hard to challenge, since nothing on paper ever said what finished looked like.

The same field is what catches an approval nobody has assigned. Every gate on a launch needs a name and a date against it in week one, or the launch date is a guess. Naming a gate costs one line in an action list.

One more field earns its place, which is the account and the workstream the item sits in. Across 8 client accounts and 11 brands, an action item with no address gets sorted into nothing.

How do you get from a recording to a list?

Three passes over the same material. The first pulls explicit commitments, the sentences where somebody said they would do a thing. The second pulls decisions, which are frequently made without anyone announcing that a decision has occurred. The third pulls changed context, meaning anything that is true this week and was not true at the last call. A budget that moved. A reorganisation that put somebody new above the programme.

The third pass is the one to protect. Commitments are easy to extract and mostly obvious to everyone who was on the call. Changed context is the part that alters what the next quarter should contain.

How much of this can AI do?

The extraction is worth handing over. Ask a model for the explicit commitments in a transcript and it returns them, sentence by sentence, without anybody rereading the whole call. We take that every week, on every account.

What it will not do is rank them. A summary puts the critical and the incidental in the same register. Read the output before it goes anywhere.

Output that follows the order the meeting happened in records what was said. Nobody can work a week from it.

Who should see which version?

Two documents come out of every call. The internal one carries strategy, judgement about the account, and the things worth holding back a week. The shared one is status, owners and dates, and it never makes the strategic call on the client’s behalf, because that decision is theirs.

The transcript itself goes to nobody.

Related reading: batching every call into two days, and what belongs in a client update.