I was in the meeting. I still needed AI to understand what was happening

I spent part of today in a campaign meeting. I was listening, I understood most of what people were saying, and I was trying to keep track of what had changed.

When it finished, I still did not have a particularly good mental model of the project.

That sounds like a meeting problem. I am starting to think it is more often a knowledge-work problem.

The campaign already includes several articles I have written. There is influencer activity around it that I need to understand, a roundtable that may produce useful evidence, a communications plan I need to revisit, a content calendar someone else has built, a production plan that needs updating and a separate GEO plan where some of the same activity already has a place.

None of those things is especially complicated on its own. The difficulty is understanding how they relate to each other, which version is current, what has genuinely changed and what I am actually expected to do.

Increasingly, that sensemaking feels more important than creating another piece of content.

Being in the meeting is not the same as understanding the work

There is a fairly comforting assumption behind meetings: get the right people together, talk through the work and everybody leaves with greater clarity.

Sometimes that happens. But a lot of modern knowledge work does not fit neatly inside the meeting you are currently attending.

Someone mentions an influencer programme and I am trying to remember how it connects to a campaign document I saw last week. Someone talks about publication dates and I am wondering whether they match the content calendar. Someone introduces a roundtable and I am already thinking about whether the resulting evidence could support something else entirely.

Meanwhile the meeting keeps moving.

The information arrives sequentially. The work itself is not sequential. It exists as a network of documents, people, dependencies, assumptions, old decisions and new activity.

I can therefore be present for the whole conversation and still need to reconstruct it afterwards. Not because I was not paying attention, but because attention is not enough to hold the entire system at once.

I wrote recently that AI can connect silos but you still have to learn how people work. This feels like the other side of that argument. AI helps me see connections across teams, but first I often need it to help me recover what the work currently is.

Meetings can distribute information without creating understanding

I can also see why people end up saying they have spent the whole day in meetings. The meetings may genuinely contain important information. The problem is that attending them is often treated as if the information has therefore been processed.

It has not.

I can sit through an hour-long campaign discussion, understand almost every sentence and still leave without a reliable answer to fairly basic questions: what actually changed, which date is now real, what I own, which document is current and what happens next.

Multiply that across five or six meetings and somebody can spend an entire day absorbing context without ever getting the time to turn it into an operating picture.

That may explain another familiar feature of organisational life: the next meeting. If the previous conversation distributed information but nobody converted it into durable shared state, people need to reconvene to recover the context again.

The problem is not necessarily too many meetings. It may be too little sensemaking between them.

That distinction matters because meetings are often counted as coordination. They can be coordination, but only if the decisions, changes, ownership and dependencies leave the room in a form other people can actually use. Otherwise the organisation has distributed more context without necessarily creating more understanding.

Tomorrow’s job is not writing

The useful work tomorrow is fairly mundane on the surface.

I need to go back through the meeting transcript and compare it with the existing communications plan. I need to look at the content calendar and work out where the articles are actually going. I need to understand what is happening with the influencers, what the roundtable is supposed to produce, how those pieces map into the production plan and where they fit into the GEO work already underway.

If I treated each source separately, I would get five perfectly reasonable summaries and still not know what to do.

The better questions are comparative.

What changed in the meeting? Which parts of the communications plan are still current? Does the calendar reflect the latest discussion? Which activity already satisfies something in the GEO plan? What is a dependency, what is merely context and what is genuinely mine to own?

That is not summarisation. It is state reconstruction.

I am essentially asking AI to help rebuild a current model of the project from several partial versions of reality.

Content creation is becoming downstream

This is why I am increasingly sceptical of treating AI productivity mainly as a content-generation question.

How quickly can it produce a first draft? How many articles can a team make? How much time can it save on writing?

Those are real gains, but they are increasingly not where I feel the bottleneck.

The expensive part is often deciding what should exist in the first place.

The articles I have already written may now be one component of a larger campaign. The roundtable may generate evidence that changes another piece of work. The influencer activity may satisfy part of a distribution or GEO need that would otherwise look like a gap. Something that sounds like a new request might turn out to be another description of work already underway.

If I misunderstand those relationships, faster creation does not help much. I just produce the wrong thing faster.

This is close to what I meant when I wrote that the system can be complicated while the work should not be. The complexity does not disappear. It has to be carried somewhere so the next decision can become simpler.

There is a reconstruction tax in knowledge work

A surprising amount of office work is spent rebuilding state.

You return to a project and try to remember where you left it. You search Teams, reopen a spreadsheet, find the newer spreadsheet, check the deck, remember there was another meeting, locate the transcript and discover somebody created a calendar while you were doing something else.

Then you work out whether any of it changes what you need to do.

Only after that do you start.

I think of this as a reconstruction tax. It is the cost of getting back to the point where useful judgement is possible.

Organisations increase this tax for understandable reasons. Another planning document gives visibility. Another calendar improves coordination. Another specialist workstream brings expertise. Another research platform adds evidence. Another meeting aligns stakeholders.

Each layer may be useful. Each layer also creates another relationship somebody has to understand.

The problem is not necessarily that the organisation has too much information. It is that the information is distributed across systems that each contain only part of the truth.

AI is becoming my continuity layer

This is probably the biggest change in how I use AI.

Earlier on, I thought mostly about generation. Now I think much more about continuity.

I have work happening across transcripts, spreadsheets, decks, briefs, videos, messages and conversations. The useful thing is not having one model memorise every detail forever. It is being able to bring the relevant sources together and ask what the current state appears to be.

That distinction matters.

The transcript is authoritative for what somebody actually said. The content calendar may be authoritative for publication dates. The production plan may own delivery status. A strategy document may explain why the work exists. AI should not flatten those differences and invent one clean story simply because a clean story is easier to read.

Its value is in helping me compare the sources, expose the gaps between them and preserve enough continuity that I can make a judgement.

That is why AI work needs a save point. A transcript is not a project state, and neither is a summary. The useful state includes what has changed, what remains unresolved, which source is authoritative and what should happen next.

This is not an excuse for bad meetings

There is an obvious counterargument here: perhaps the meeting was unclear, the project has too many documents and better project management would solve most of this without AI.

Sometimes that is absolutely true.

AI should not become a sophisticated excuse for organisations that refuse to decide who owns what, maintain a plan or run a coherent meeting. If five documents contradict each other because nobody updates any of them, a model can help identify the contradiction but it cannot make the operating model healthy.

But I do not think all distributed work is accidental mess.

A campaign genuinely does need different views. A publishing calendar is not the same thing as a communications strategy. A GEO plan asks a different question from a production plan. A roundtable can be both an event and an evidence source. Different specialists need different artefacts because they are solving different parts of the problem.

The mistake is expecting a person to carry the relationships between all of those artefacts perfectly in working memory.

That is where AI becomes interesting to me: not as a replacement for good operations, but as a way to make legitimate complexity more navigable.

Sensemaking may be the more valuable AI skill

As AI lowers the cost of producing words, slides, summaries and plans, the ability to generate another thing becomes less scarce.

The scarce part moves upstream.

What is actually happening? What changed? What does this evidence alter? Which plan owns the decision? Does something already exist? Is this genuinely a gap? Who needs to know? What should we not create?

Those are editorial questions as much as operational ones.

They require selection, hierarchy, provenance and judgement. They also require restraint. A model can easily turn every new signal into another recommendation, another asset and another workstream. The better outcome is often to recognise that the work already has a home and route the new context into it.

That is why I keep coming back to prompting as routing and editorial judgement as infrastructure. The interesting AI work is increasingly not producing the final artefact. It is helping the right evidence, decision and context reach the right part of the system.

The next productivity problem is legibility

I do not have a shortage of things I could create.

What I have is a finite amount of attention for understanding the relationships between all the things already happening.

Without AI, I could still do tomorrow’s sensemaking exercise. I would reread the notes, compare the documents, keep another set of notes beside them and gradually reconstruct the project in my head. I have done that kind of work for years.

The difference is that AI can compress the cost of rebuilding the model. It can hold several sources in view while I ask comparative questions, point out where things do not line up and help me write the resulting state back somewhere durable.

That is a much bigger gain for me than saving twenty minutes on a first draft.

The risk is obvious too. If AI makes creation cheaper without making the surrounding work more legible, organisations can simply generate more plans, more summaries, more content and more noise for somebody else to reconcile.

So the productivity question I am increasingly interested in is not: how much more can AI help us produce?

It is: can AI help us understand enough of the existing work to make better decisions about what deserves to be produced at all?