AI made reflection part of the work

One of the most useful things AI has changed for me is not how quickly I can write. It is how cheaply I can turn a day of conversations, observations and half-formed ideas into something I can inspect, reuse and take back into the work.

I’ve noticed something slightly strange about the way I work now: sometimes I read my own posts to work out what I’m trying to do at work. That sounds backwards, because the conventional sequence is supposed to be straightforward: do the work, understand what happened, then write about it.

Increasingly, my sequence looks more like this:

do the work → talk and think through what happened → turn the thinking into a note → read the note back → understand the work differently → take that understanding back into the work

The writing sits in the middle. That has become particularly obvious while developing Editorial Intelligence alongside my day job, where a real project produces questions, a meeting changes my understanding, a piece of research exposes a gap, and I can talk the problem through with AI while the detail is still fresh.

Within minutes, the mess can become something legible enough to inspect. Sometimes it becomes a working note on this site; then a few days later I come back to the note because I’m trying to decide what to do next on the original project. The output has become an input.

I don’t think the interesting part is that AI can write a blog post quickly. We already know that. The more interesting possibility is that AI makes reflection cheap enough to become part of ordinary knowledge work.

Reflection was already work

There is nothing new about the idea that people learn by reflecting on experience. Reflective practice has a long history in professional learning because it helps people turn activity into meaning and decide what should change next.

A 2020 overview of reflection in workplace learning describes reflection, critical reflection and reflective practice as ways of turning experience into meaning and action rather than simply accumulating activity. Faller, Lundgren and Marsick describe reflection as central to workplace learning precisely because it helps people interpret experience and decide what follows.

Gillie Bolton’s work on reflective practice makes a related point: experience alone does not guarantee development. Reflection is what allows people to revisit what happened, challenge assumptions and use experience to change future practice. Her 2009 paper frames reflective writing as a way of learning from your own experience rather than simply recording it.

That is close to what I’m doing, but the difference is the friction. Traditional reflective writing is effortful: you finish the meeting, project or working day and then face another task of reconstructing what happened, deciding what mattered, organising it, writing it coherently and preserving it somewhere you might find again.

A realist evaluation of reflective writing in UK medical education found exactly this problem. Written reflection could be valuable when people genuinely needed to think more deeply about an experience, but participants also experienced it as effortful; when reflection became externally imposed, it easily collapsed into a tick-box exercise. The study is a useful reminder that the theoretical value of reflection and the practical likelihood of people doing it are not the same thing.

This is the part AI changes for me. It reduces the production cost between I think something happened here and I can see the thinking in front of me.

AI as a reflection scaffold, not a replacement thinker

There is now research beginning to look directly at this. A 2025 study in the British Journal of Educational Technology explored the use of large language models to support reflective writing and metacognition, starting from a familiar problem: reflective writing can improve metacognitive skills, but people often struggle to structure reflection well enough to get the benefit.

Their work treats the LLM as a scaffold that can help people organise and deepen reflection rather than simply generate an answer for them. Neshaei and colleagues describe the opportunity as one of supporting reflective writing at scale while adapting prompts and support to the individual.

Another 2025 study tested an LLM-mediated journaling system called PaceMind. It was built for post-exercise reflection rather than workplace thinking, but the interaction is interesting: the system generated data-driven drafts and personalised questions, then asked users to refine and personalise what came back.

The LLM-supported condition led to deeper integration of information in journal entries and stronger perceived reflection support. But the researchers also found something important: users still had to verify and personalise what the system produced. The study explicitly argues for balancing automation with meaningful cognitive engagement and user control.

That feels much closer to the useful model than “AI writes my reflection”. The AI can do some of the expensive editorial work around the thinking:

• turn a messy conversation into a provisional structure;
• surface contradictions mentioned in different places;
• remind me of earlier evidence or decisions;
• help separate what happened from what I am inferring;
• convert a stream of thought into something I can critique; and
• give me a durable artefact while the context is still fresh.

But I still have to decide whether the note is right. If the AI becomes the source of the interpretation rather than a way of externalising and testing my interpretation, the process can easily become synthetic self-understanding: coherent-looking prose that was never really mine.

The productive loop requires resistance. I need to be able to say: No, that’s not what happened. That conclusion is too strong. You’re treating these two things as connected when they aren’t. The important part was actually this. That correction is part of the thinking.

The working diary is not really a blog

This has changed how I think about publishing here. For a while I treated the site in fairly conventional terms: there were articles, frameworks and field notes, and even when the ideas were experimental the mental model was still publishing — think first, publish afterwards.

“Working diary” increasingly feels more accurate because the entries are not always the end product of understanding. Sometimes they are how I create enough understanding to continue, which makes the public/private boundary unusual: I am publishing a cleaned-up version of my working cognition, and future-me is part of the audience too.

A useful note has to survive long enough that I can return to it two weeks later and recover the reasoning that produced a decision. This is why I sometimes read an Editorial Intelligence post before going back into a live project: not because I have forgotten the post, but because the post has compressed a messy period of thinking into a more legible object.

Reading it restores the state of the argument. In knowledge work, that is surprisingly valuable because we lose enormous amounts of reasoning between meetings: decisions remain in people’s heads, useful observations disappear into chat threads, and a project changes direction until nobody can quite reconstruct why.

A working diary can preserve some of that path. It does not need to preserve every thought; it needs to preserve enough of the reasoning that the next decision starts from somewhere better than memory alone.

Sensemaking is a legitimate output

This also changes how I judge a working day. I have historically had a strong production bias: a good day feels like a day where something tangible exists at the end — an article, a deck, a page or a finished draft.

That bias is useful because it stops thinking from becoming a substitute for doing, but it can also distort knowledge work. Some days the important output is a better model of the problem, and that model may come from conversations, disagreement, research or discovering that the thing you were preparing to build already exists.

You might talk to three people and realise the workflow is different from what you assumed. You might discover that two initiatives you thought were separate are actually overlapping, or that a content gap is really an evidence gap. Nothing dramatic has been published that afternoon, but tomorrow’s production is now pointed somewhere better.

There is established research around sensemaking in professional work: people do not simply process information; they construct a workable interpretation of ambiguous situations so they can act. Recent research is also beginning to explore generative AI specifically as an aid to professional sensemaking.

One study involving genetics professionals found sensemaking to be both a major challenge in complex knowledge work and an area where experts could imagine AI supporting them, particularly when handling large amounts of information without delegating the final judgement. Mastrianni and colleagues describe AI-supported sensemaking as a human-AI interaction problem, not merely a generation task.

That distinction is central to Editorial Intelligence. A lot of valuable editorial work happens before the obvious content exists:

What are we actually seeing? What evidence do we have? Which interpretation is defensible? What is missing? Who should we ask? Which part deserves an output at all?

If AI can make that reasoning easier to preserve and revisit, then its value is not confined to faster production. It can improve the layer that decides what gets produced.

The strange loop between practice and publishing

The most useful part of my own system is the loop between real work and Editorial Intelligence. A live project creates friction or ambiguity, and I use the working diary to understand the problem rather than treating the diary as something that only starts after the work is finished.

That may produce a provisional idea: perhaps a new editorial rule, a way of structuring evidence, a question about GEO, a different way to think about video, or a clearer distinction between content production and knowledge acquisition. Then I take the useful part of that idea back into the live project and let reality get a vote.

A colleague may already be doing it. A production constraint may make the idea impractical. A stakeholder conversation may expose a flaw, the evidence may not support the neat theory, or the intervention may work surprisingly well. Then the note changes too.

That is a healthier relationship between theory and work than treating every thought as a framework that needs adoption:

work → reflection → legible note → intervention → reality test → revised understanding

The publication is therefore not simply a marketing asset for Editorial Intelligence. It is part of the operating system.

What AI makes cheaper is not thinking itself

There is an obvious danger in saying that AI makes thinking cheaper. It can make the appearance of thinking very cheap, because a model will happily produce five lessons from a meeting, a strategic implication or a beautifully ordered story about what you learned whether or not the source material supports it.

That is not reflective practice. It is retrospective content generation. The useful gain is narrower: AI makes externalising, structuring and revisiting thinking cheaper, while the human still has to supply the experience, recognise what matters, contest the interpretation and decide what changes.

The PaceMind study is useful here because participants did not simply accept machine-produced reflection. Verification and personalisation remained cognitive work; the AI reduced some friction while still requiring the person to stay engaged. That is probably the right design principle for knowledge work too.

The point is not to automate the reflection away. It is to remove enough clerical and compositional friction that reflection can happen while the experience is still useful.

From documentation to a thinking surface

This may be the bigger shift for me. Most workplace documentation is designed as a record: minutes tell us what was discussed, project trackers tell us what is happening, knowledge bases tell us what the organisation has decided, and content repositories preserve finished outputs.

A working diary does something slightly different because it preserves interpretation in motion. That matters because knowledge work is rarely a clean sequence of decisions made from complete information; understanding develops through conversations, partial evidence, corrections and repeated exposure to the problem.

AI makes it feasible to capture more of that without turning every working day into an evening writing exercise. Once the interpretation is written down, it can become active again: it can be searched, compared with later evidence, challenged, fed into a new conversation, connected with an older project or rewritten when reality disproves it.

The note becomes a thinking surface rather than an archive entry. That is probably why I keep coming back to mine.

I still want to make things

None of this has cured my desire to produce, and I don’t particularly want it to. Production pressure is useful because thinking without any pressure to make something can become an elaborate form of avoidance.

What has changed is what I count as production. A day of conversations and investigation can produce a clearer question, and AI can help turn that question into a durable working note before it disappears. The note can then improve the article, interview, video or decision that follows.

So the loop is not thinking instead of making. It is:

think → make the thinking inspectable → use it to make the next thing better

That feels like a much better description of how I want Editorial Intelligence to work, and perhaps it is one of the less discussed changes AI is making to knowledge work. We have spent years asking how much faster AI can produce the final artefact; I am increasingly interested in what happens when it makes the messy middle — reflection, sensemaking, connection and revision — cheap enough to preserve.

Because sometimes the most valuable thing AI produces for me is not the article. It is the version of my own thinking that I can finally see clearly enough to argue with.