Context is capital — Editorial Intelligence

AI compresses execution time. It does not compress the history that made the execution possible.

I’ve written already about why the work happens before the prompt — that when AI-assisted work looks fast, the speed sits at the end of a long process, and somebody could copy my prompt exactly without getting the same result.

This piece is about what follows from that, which I think is the more consequential part.

If the accumulated thinking is the real input, then it is worth asking what kind of thing it is. Not a filing system. Not an archive. Something closer to capital: work performed once that can keep producing value, and which decays if nobody maintains it.

That framing is doing more work than it might appear.

Storage was cheap. Reuse was expensive.

Organisations have been accumulating information for decades.

Research reports. Presentations. Customer interviews. Sales calls. Meeting notes. Product discussions. Support conversations. Market analysis. Event recordings. Executive thinking. Performance data.

The storage problem was largely solved.

The reuse problem wasn’t.

A company could possess five years of customer research and still begin its next planning session by asking: what should we write about?

A research project might generate a launch campaign, a report and a few articles. Then everybody moved on. A customer interview became one customer story. An event became one recap. A conversation with an expert ended when the meeting ended.

The information remained technically available.

But availability is not usability.

To genuinely reuse it, somebody had to remember it existed. Find it. Read it. Understand why it mattered. Compare it with what had happened since. Recognise a connection. Then judge whether that connection was meaningful.

At organisational scale, that is expensive work. It requires sustained attention, and attention is the scarce resource.

So organisations accumulated enormous stores of potentially useful context while extracting remarkably little future value from them.

Storage was cheap. Retrieval, comparison and synthesis were expensive.

That equation is what has changed.

AI increases the return on previous thinking

One of the least interesting ways to describe generative AI is as a machine for producing more words.

The more consequential capability may be working across material that humans do not have time to continuously reprocess.

Search these interviews. Compare these research projects. Find recurring language. Show me where this idea appeared before. Identify evidence that supports this argument, then identify evidence that challenges it. Compare what customers said last year with what they’re saying now. Find expertise buried in these conversations that was never formally documented.

None of those operations removes the need for judgement.

But they drastically reduce the friction involved in bringing previous work back into consideration — and that changes the value of having done the previous work at all.

A customer interview no longer has to produce one asset and disappear. A research project doesn’t stop generating insight when its campaign ends. An observation that seemed minor in January may matter when another signal appears in August. A framework can be challenged by evidence that didn’t exist when it was written.

Old thinking can participate in new thinking.

That is a different economic model for knowledge work.

AI doesn’t only make new work cheaper. It can make old work more valuable.

What “capital” actually requires

I don’t mean that every document is an asset.

A shared drive containing 40,000 forgotten files is not intellectual capital because it is large. A transcript isn’t automatically knowledge. A customer quote isn’t automatically evidence of a wider truth. A meeting summary isn’t organisational intelligence.

Calling context capital only makes sense if it has the capacity to produce future value, and that requires conditions.

It has to be findable. Its source has to be understood. Important claims need provenance. Connections need maintaining. Outdated material needs to be recognisable as outdated. Observations need to stay distinguishable from verified evidence.

And somebody still has to decide what matters.

So the progression isn’t more documents → more intelligence. It is closer to:

experience → captured context → signals → connections → evidence → interpretation → reusable knowledge → better decisions

The Evidence Engine is essentially infrastructure for that progression. Capture what might matter. Connect it. Interrogate it. Work out what can actually be supported. Preserve the useful part. Let future work start from there rather than resetting to zero.

The capital isn’t the archive.

The capital is the reusable understanding.

Content is inventory. Context is productive capacity.

This gives content organisations a distinction worth having.

A published article is an output. It may create traffic, awareness, reputation or sales. But if the thinking that produced it disappears afterwards, the organisation has consumed its own knowledge to manufacture one asset.

That is closer to inventory. Make the thing. Ship the thing. Move on.

Context behaves differently.

Suppose an organisation runs substantial customer research. It could turn that into six articles, which is useful. Or it could also preserve the original interviews, the recurring customer language, the strongest claims, the areas of disagreement, the unanswered questions, the connections to earlier research, the patterns, the frameworks built from those patterns, and the record of what the organisation changed its mind about.

Now the research has productive capacity beyond its first outputs.

A later event can be compared against it. A new interview can strengthen or weaken it. A salesperson can recognise its language in a live conversation. An executive can build a position from it. A future AI system can retrieve it when a new question appears. A market development can suddenly make an old observation significant.

The organisation hasn’t only produced content.

It has increased what it knows.

Context depreciates

This is where the capital metaphor earns its place, because capital is not automatically productive forever.

Neither is context.

Research dates. Markets change. Terminology shifts. A strong argument can harden into a sacred cow. An assumption repeated often enough starts to look like evidence. A customer statement loses its qualifiers as it gets summarised and re-summarised.

AI makes it easier to reproduce all of those at scale.

So accumulated context creates a second obligation: maintenance.

If organisations get better at remembering, they also have to get better at forgetting, qualifying and updating. What is still true? What has changed? What was always uncertain? What did we believe when this was written? What new evidence challenges it? What should no longer guide decisions?

Otherwise context stops behaving like capital and starts behaving like debt.

Old assumptions get baked into new workflows. Outdated evidence reinforces outdated narratives. AI retrieves yesterday’s answer faster and gives it tomorrow’s confidence.

That isn’t organisational intelligence.

It’s automated inertia.

Which is why the contradictions are worth keeping

There is a tension a good knowledge system has to hold open.

You want accumulated context, because starting from zero wastes what you know. But you don’t want accumulated context to become a prison.

So the system has to make two things possible at once: remember what you’ve learned, and remain able to change your mind.

That is why I think contradictions are among the most valuable things to preserve.

The instinct in content work is to clean them up. Choose the position. Make it consistent. Publish.

But an organisation that only keeps evidence supporting its existing narrative is building a very efficient confirmation-bias machine.

The better system keeps the inconvenient material too. The interview that didn’t fit. The finding that weakened the headline. The customer who described the problem differently. The framework that stopped explaining what was happening.

Context compounds most usefully when it can challenge previous context.

This is also why agents worry me slightly

There is a rush to deploy agents on top of organisational information.

An agent with access to more material is not automatically a better agent. It can simply become faster at retrieving stale assumptions, weak evidence and organisational folklore.

The quality of automation increasingly depends on the quality of the context underneath it.

Which is an argument for doing the unglamorous work first.

The individual version of the same problem

This isn’t a new problem, and organisations didn’t need AI to notice it. James Walsh and Gerardo Ungson described organisations as systems that acquire, retain and retrieve information from their past as far back as 1991, and the harder half of that was always retrieval, not storage.

Organisational psychology has a more specific name for the individual stakes of it: transactive memory. A team doesn’t need every person to know everything. It needs a reliable sense of who knows what, where it sits, and who to ask. A 2024 study of five new-venture teams found that transactive memory develops through shared understanding, role formalisation and knowing where expertise sits, and that the teams which built it faster coordinated faster.

That’s the individual version of context as capital. The most valuable person in a team isn’t necessarily the one who knows the most. It’s the one other people reliably route through — the one who remembers why a decision was made, which research already answered this, and who actually understands the topic. That used to be almost entirely a function of tenure. AI is starting to make it something you can build deliberately, faster, regardless of how long you’ve been in the room.

Which changes what “being good at your job” can mean. Less I produce work quickly — that’s increasingly true of anyone with the same tools, whether or not they did the compressing themselves — and more I can tell you where this fits, what already answered part of it, and who else needs to know. That’s a harder thing to automate than drafting, and a more durable thing to be known for.

The same pattern is now turning up in AI research on agents, which is a useful confirmation rather than the origin of the idea. Google Research has framed AI agents through transactive memory theory directly, asking whether the dynamics that make human teams effective also apply to human-AI teams. Microsoft Research’s PlugMem starts from an unintuitive finding: simply giving an agent more interaction history can make it less effective, because large histories fill with irrelevant material. The fix isn’t more storage. It’s turning raw interaction into structured, reusable knowledge and retrieving only what the task needs.

That’s the organisational version of the individual point above. Storage isn’t memory. Memory isn’t useful because it’s large. An Alteryx survey of 1,400 technology leaders, published in August 2026, found the gap already showing up in practice: 77% say business context is critical to accurate AI output, and 53% say their organisation struggles to get that context into the systems doing the work.

That gap is exactly where an individual who does the retrieval and connection well becomes disproportionately useful — not despite the organisation’s system being incomplete, but because of it.

This isn’t about building a giant second brain

There is an obvious failure mode here.

Capture everything. Store everything. Tag everything. Build elaborate systems for every passing thought. Spend more time maintaining the knowledge system than using it.

I’ve done enough work on systems to be wary of it. A useful context system should reduce friction, not create another administrative job.

The test isn’t how much have we saved?

It’s does what we’ve saved improve the next piece of thinking?

If the answer is no, accumulation isn’t compounding. It’s hoarding.

This is why editorial judgement belongs at the centre. AI can increase capture, ease retrieval, cluster and compare, and surface possible connections. Somebody still has to make the choices about significance.

What deserves preserving? What deserves developing? What deserves challenging? What should influence future decisions? What should be allowed to disappear?

The scarce capability isn’t remembering everything.

It’s remembering what matters without becoming trapped by what you remember.

The advantage may not be where people are looking

Most discussion of AI advantage currently focuses on speed. Who can write faster, build faster, produce more, automate more.

Those gains are real. But the same tools are becoming widely available, so production speed is unlikely to stay a durable differentiator on its own.

The deeper divide may be between organisations that repeatedly start from zero and organisations whose knowledge compounds.

Two teams may have access to the same model. One briefs it from a blank prompt. The other can draw on years of structured customer evidence, previous decisions, tested narratives, research and frameworks.

Those are not equivalent systems.

The model is the same. The context isn’t.

The prompt is a transaction. The context is what compounds.

This is why I’ve become less interested in spectacular individual AI outputs, and more interested in what happens on the next interaction, and the one after that.

Does a customer conversation become future evidence? Does research stay useful after publication? Does a failed argument improve the next argument? Can a new observation collide with something recorded six months ago? Can an AI system retrieve the history without flattening its uncertainty? Can somebody joining the organisation inherit some of the thinking rather than only the finished assets?

Does yesterday’s work change what is possible tomorrow?

If it does, the organisation isn’t simply producing.

It’s accumulating capability.

Do the work before the prompt

There’s a strange conclusion in all of this.

AI makes deep work less visible. It may also increase the return on doing it.

A strong interview matters more if its value can be recovered repeatedly. Careful research matters more if it can inform decisions years later. A framework matters more if future evidence can strengthen or challenge it. A good observation matters more if it doesn’t vanish into a notebook. Experience matters more when a system can help you reconnect it at the moment it becomes relevant.

So I don’t think the lesson of AI is that we need to think less.

Almost the reverse.

Do the difficult thinking. Talk to people. Run the research. Develop the expertise. Test the arguments. Notice the contradictions. Change your mind.

Then preserve enough of what you learned that the next question doesn’t force you to start again.

Because one of the most important things AI may do for knowledge work isn’t thinking instead of us.

It’s making sure the thinking we’ve already done doesn’t have to disappear.