Written daily. Read anytime.

JournAI.

The lab, built in public — a dated record of what actually happens when AI enters real work: experiments, observations, builds, doubts.

N°07 · 02 SEPT 2026 · BUILD

The lab opens

Why this site exists, and why it's built in public — insight to Markdown to git to live, with nothing hidden in between.

This is the first entry, so it should say what the room is for.

Most sites about AI are brochures. This one is a working surface. Every note here is part of a live pipeline: an observation gets dictated, shaped through a set of AI skills, written to a Markdown file, committed to git, and published. No content team, no funnel disguised as a blog.

Why in public

Because the interesting part isn’t the conclusion — it’s the reasoning that produced it. Showing the work is the point. If a framework is wrong, watching it fail in the open is more useful than a polished case study that hides the misses.

What to expect here

Short, dated, honest. Experiments that worked and ones that didn’t. The Garden holds the ideas once they’ve stabilised; JournAI is where they’re still moving.

Build log. Written the day the lab went live.

N°06 · 12 JUL 2026 · OBSERVATION

The skills you never named are running the business

The work on the invoice has a name. The work that makes it yours doesn't — until AI asks you to say it out loud.

Ask a coach, a consultant, a designer, a nutritionist what she does, and she will name the thing on the invoice. Coaching. Strategy. Brand copy. Meal plans. The service has a name because the market needs one to price it.

I have watched a lot of these women describe their work — to me, and to AI — over the past year. The described work is almost never the valuable work.

The valuable work is the part with no name. How she knows, in the first ten minutes, what the client is actually asking for underneath the question they brought. What she discards before the call even starts, without deciding to. The order she solves things in — which knot she pulls first because the others loosen once it moves. The moment she leaves her own method because this client, this week, needs something the method doesn’t cover. How she decides a thing is good enough to ship. Which risk she sees three steps before it arrives.

None of that is on the invoice. Most of them don’t believe it’s a skill at all. They call it intuition, or experience, or just “how I work,” and they undercharge for it, because you can’t price something you can’t see.

AI is what made it visible to me — not by being smart, but by being literal.

The moment you try to hand one of these judgments to a system, you have to say it out loud. You have to tell it what “a good offer” means, concretely, or it gives you a confident bad one. You have to name the conditions under which you’d break your own rule, or it breaks the rule at the wrong time. The system does not accept “I’ll know it when I see it.” It asks you to write down the seeing.

And that is the strange gift. The thing you were sure was intuition turns out to have structure. It was a decision the whole time — a fast, compressed, repeated decision you stopped noticing you were making. Making it explicit to a machine is the first time many people ever see their own expertise from the outside.

This is most of my actual work now. A client comes in wanting to “use AI in the business,” and what we really do is go hunting for the decisions she makes automatically. We slow one of them down until she can watch herself make it. Then we ask whether it can become a method, an instruction, a small system that carries the routine part — while she keeps the part that needs her.

Because not all of it should be handed over, and this is where I want to be careful. Some of these unnamed judgments are the business. They are the reason the work is hers and not easily copied. If you extract every one of them into a system, you haven’t multiplied your expertise; you’ve published it. The goal is not to externalise the judgment. It’s to externalise the routine around the judgment, so the judgment has more room.

So the same act does two things at once. Naming a skill lets you delegate it. Naming a skill also shows you which ones you never want to delegate.

I don’t have a clean rule yet for where that line sits. It seems to move per person, per skill, per how much the market can already do without them. What I am fairly sure of, after a year of watching people meet their own intuition on a screen: most of them were underpricing not because they lacked confidence, but because they genuinely could not see the thing they were selling. The AI didn’t add the skill. It held up a mirror long enough for them to name it — and you cannot value, protect, or sell what you have never named.

N°05 · 12 JUL 2026 · OBSERVATION

The second business I see inside your business

Spend a few months inside a business and you start to see two of them — the one she's running, and the one already becoming possible around her.

When I spend a few months inside someone’s business, I start to see two of them.

There is the business she is running — the current clients, the current offers, the calendar full of the work she already knows how to do. And there is a second business, the one that has quietly become possible around her while she was busy running the first. Same expertise, same person, arranged for the conditions coming toward her instead of the ones she built for five years ago.

She usually cannot see the second one. Not because she lacks vision — because she is inside the first one, and you cannot get the angle on a business from within the room where you are doing its daily work. I have the easier job. I am looking at it from further along, from the environment it is moving into rather than the one it came from.

I want to be precise about this, because “I bring an outside perspective” is what every consultant says and it means almost nothing. I am not bringing an outside perspective. I am reading the business from the conditions it will have to operate inside — how AI is already changing her particular industry, which parts of her expertise could become a system, where her knowledge could become something she owns rather than something she re-performs every week, what could be delegated, what new service only exists now. Then I work backwards to what she would need to understand, build, and document today to get there.

For a long time the ceiling on a business like hers was headcount. To do more, you hired. That ceiling is mostly gone. One person can now run systems that used to need a team — a content operation, a support layer, a sales pipeline, an assistant that never sleeps. The capacity is real, and it is genuinely new.

But capacity was never the thing that told anyone what her business should become. And here is the trap I watch people walk into: AI lets them produce far more inside the old shape of the business. More of the same offers, more of the same delivery, faster. The business grows beyond her current capacity long before it grows beyond her current mental model — and the mental model is the actual bottleneck now. She has access to a larger business and is still making decisions from the architecture of the smaller one.

So part of my work is just helping her catch up to what has quietly become possible. But that is only half, and the other half is the part almost nobody is talking about.

The more of the business AI runs, the more fragile it can quietly become.

A pipeline that looks autonomous because a model runs it is only autonomous until the model changes, the integration breaks, the account gets suspended, or the tool is discontinued on a Tuesday with thirty days’ notice. Then the owner finds out whether she was running a system or renting one. She may discover she does not actually control the thing her revenue depends on — she was borrowing it, temporarily, from a stack of tools that do not answer to her.

This is why I do not think autonomy means doing everything by hand, or refusing to lean on AI. That is just a smaller, slower business calling its ceiling a principle. Autonomy is something more specific, and more demanding: understanding the logic of your own business well enough to keep authority over the decisions that matter. Knowing where AI is acting, and on what basis. Being able to tell when it is wrong. Keeping the critical knowledge — the offer logic, the client history, the decisions and why you made them — outside any single tool that could vanish. Designing the parts so they can be swapped, with a route that still works when one of them fails.

That turns human-in-the-loop into something much bigger than checking an AI’s output before it goes out. The human is not the quality check at the end of the line. She is the one holding the continuity, the judgment, the responsibility, and the direction — the parts of a company that cannot be delegated, because they are what a company is.

So the business I am actually trying to help build is not the one that hands AI the most work. It is the one that can use AI heavily and still understand itself, still decide, still adapt, still recover when something breaks. Capable of becoming much more, without that capability belonging to someone else.

That is a harder target than either “use more AI” or “keep it human.” Grow the business into what has become possible, and build it so the new power still belongs to the business, not to the tools it is currently renting it from. The second business is real, and it is already forming inside the first one. The only question I care about is whether, when she steps into it, she will still be the one holding it together.

N°04 · 11 JUL 2026 · DIARY

AI saved me three hours. I gave them back to work.

The tools gave me my hours back. I keep handing them straight back to the work.

The pitch is always time. Every tool, every demo, every headline: this will save you hours.

It’s true. It saves me hours. Some days a lot of them — a research pass that would have eaten an afternoon, a first draft that would have taken all morning, a spreadsheet cleaned while I make coffee. The hours are real. I can feel them arrive.

Then I watch what I do with them, and it is almost always the same thing. I make more.

Another piece of content, because the draft came so easily. Another project opened, because the setup that used to take a week now takes an evening. Another idea tested, because testing is cheap now. A thing that was already fine, polished again, because I can. Work I hadn’t planned to take, taken — why not, it’s faster now. I don’t stop at the edge of the saved time and rest inside it. I refill it, immediately, with more building.

And here is the part I keep turning over: I’m not doing this because I’m disciplined, or driven. I’m doing it because it feels good. There is a specific pleasure in describing an idea in two sentences and watching it become a working thing before the coffee is cold. That loop — think it, build it, see it — is one of the best feelings in this work. The tools didn’t only remove the friction. They made the making itself more fun. So of course I do more of it.

Which means the saved time was never really free time. It was capacity. And capacity, it turns out, is not the same gift as freedom.

Freedom would be the afternoon that stays empty. Capacity is an afternoon with room for three more things in it. The tools gave me the second one and let me call it the first.

I used to think the constraint on my business was how much I could produce. For years that was true — there were only so many hours, so the hours decided what got made. AI quietly removed that constraint, and I expected to feel lighter. Mostly I feel the opposite pull. Now that I can make almost anything quickly, the question of what is actually worth making has nowhere left to hide. The limit moved. It used to be how much I can produce. Now it is how much I am willing to carry.

I don’t have this solved. I’m writing it down because I’m in the middle of it, not on the other side. I still open the new project. I still polish the thing that was fine. Some of that is real work and some of it is the dopamine wearing a work costume, and honestly, in the moment, I can’t always tell which.

The tools promised to give me my time back. They did. I just haven’t decided yet whether getting it back means anything — if the first thing I do with it is hand it straight to the work.

N°03 · 10 JUL 2026 · OBSERVATION

What is left after you ask AI?

Sometimes we hand AI a task. Sometimes a decision. Sometimes a capability we'll need again. Those are not the same thing.

“Just ask AI” is often good advice.

I do it all day. Summarise these notes. Clean up this document. Find the missing step in this workflow. Give me five ways to say this without making it sound like everyone else. Turn this chaotic list into something I can finally see.

I have no desire to learn the manual mechanics behind every task I hand over. That would defeat the point.

But I still need enough contact with the work to know whether the result belongs in the situation.

Did the summary miss the detail that changes the decision? Does this sentence sound like me? Did the table reveal something I now need to act on? Did the workflow solve the actual problem, or only produce a very neat version of the wrong process?

When I can make that judgment, AI acts like a useful service. When I cannot, its response is still a suggestion, however finished it looks.

This is where things become more interesting.

Sometimes we hand AI a task. Sometimes we hand it a decision. Sometimes we hand it a capability we will need again.

Those are not the same thing.

A task may be simple: turn a recording into notes, clean a spreadsheet, write a first reply, find patterns in a pile of feedback. The output may be enough, after a quick check.

A decision has consequences you will carry forward: your price, your offer, your positioning, the direction of a project, the system your team will use every week. AI can help you see options, challenge assumptions, organise the material, and create a strong first version. It cannot know which choice is right for your business without the reality you hold around it.

Then there is capability: work you will return to repeatedly. Writing an offer. Assessing an opportunity. Designing a customer journey. Briefing a designer. Building a system that needs to keep working after the first exciting version.

Here, a good answer can be useful and still leave you underprepared for the next round.

Ask AI to create an offer and it can return something polished: a headline, benefits, sections, a call to action, perhaps a price suggestion. But the page may rest on a vague buyer, a weak promise, an imagined problem, or proof that does not exist.

The page can look complete before the thinking is complete.

That is one of the stranger effects of AI. It can make missing knowledge harder to see.

A blank page was honest. It showed you where you had no language, no decision, no structure, no answer yet. A fluent response can cover that gap so well that you stop noticing it. You receive a usable page and move on, while the questions that would actually change your business remain untouched.

Then a similar task arrives next week, and you start again from zero.

Tools like Cowork make this especially visible. They collapse the distance between an idea and a finished-looking object. You can describe an offer in two sentences and receive a page with a headline, benefits, sections, and a call to action.

The speed is useful. It also means choices about the buyer, the promise, the proof, and the price can be made before you have consciously made them yourself.

I do not think resistance is the answer. Friction has no moral value of its own. AI can remove busywork, reduce the cost of experimentation, and make things possible that would have stayed on the list for years.

AI will keep doing more. The practical decision is more personal: which parts of your work are you happy to never learn, and which parts need to become clearer to you because your business will depend on them?

For me, that means paying attention around work that is strategic, repeated, or difficult to evaluate. I ask AI to show me the assumptions it made, the information it is missing, the alternatives it considered, and the trade-offs between them. I compare directions. I correct the brief. I save the decisions and principles that emerge, rather than saving only the final copy.

You can see the difference in the next task.

Instead of writing, “Create my offer,” again, you can say, “Build this around the buyer we chose, the proof we actually have, and the price logic we tested. Do not make claims we cannot support.”

You are still using AI heavily. You are simply no longer asking it to invent the foundation each time.

Maybe the better question is not whether every AI interaction makes us more capable. Some should simply save us forty minutes.

The question is whether an interaction leaves anything useful behind.

A reliable result.

A clearer decision.

A reusable way of working.

A question you can finally see.

A capability you did not have before.

You do not need all five every time. A formatting task may leave you with a clean document and nothing more. Good. A strategic conversation about your offer should leave more than a page of copy. It should leave you with a sharper understanding of what you are building and why.

Three months later, when a similar problem comes back, what do you reach for first: a blank chat box, or a clearer idea of what needs deciding?

That is what I want AI to leave with me.

N°02 · 09 JUL 2026 · OBSERVATION

When AI makes you faster at the wrong business

AI will build you a finished-looking business in an afternoon. Finished-looking and finished are not the same measurement.

You can now build the wrong business very fast.

I mean that precisely. In an afternoon, AI will give you an offer, a landing page, a content plan, a funnel, three automations, and a sales campaign to wrap around them. All of it will look done. Headlines, sections, a call to action, a price. It will look like the output of a company that knows what it is doing.

And none of it will tell you whether the buyer is real, whether the problem you are solving is one people actually pay to remove, whether your positioning holds, whether any of it should exist.

This is the part I have become most careful about, because I sit exactly where it happens — between the business decision and the AI build. What I watch, over and over: the speed of production gets mistaken for progress. Someone spends a week generating assets and feels like they moved the business a week forward. They didn’t. They moved the material forward. The decision underneath it never got made — it just got covered in something that looks like a decision.

Here is the mechanism, because it matters. When you had to make everything by hand, the effort was a filter. Writing the whole sales page yourself was slow enough that somewhere in the slowness you were forced to confront the weak promise, the vague buyer, the proof you didn’t have. The labour made the gaps loud. AI removes the labour, and it removes the filter with it. You skip straight to the finished-looking thing, and the gaps come along, quietly, underneath the polish.

So the output looks like a real business and rests on an unmade decision. AI didn’t fix the confusion. It packaged it, professionally — which is worse, because now the confusion has a landing page and a price.

This is why my work starts before the automation, not at it. Before I help anyone speed anything up, I want to know what actually has to be decided first: who this is for, what they are really buying, what has been tested and what has only been assumed. Then, and only then, is AI genuinely useful — because now it is speeding up a real decision instead of decorating a missing one. The tool is extraordinary at execution. It has no opinion about whether you are executing the right thing.

I want to be precise about the limit of my own claim, because I am not saying speed is the enemy. Fast is good, once the direction is right — I build fast every day, and I would never go back. I am saying speed is not evidence. The number of things you made this week says nothing about whether your business got clearer. Those are two different measurements, and AI has made it very easy to read one as the other.

So the question I would keep asking, in the middle of all that easy production, is not “how much did I make?” It is smaller, and more annoying. Which decision did I actually make this week? What did I learn about the person I am selling to? What can I repeat on purpose now, instead of regenerate from scratch? If the honest answer is “nothing, but look how much I produced” — that is not a business moving forward. That is a very well-designed way of standing still.

N°01 · 08 JUL 2026 · DIARY

Living with the tools that think back

The old tools waited for a command. The ones I use now finish my sentences — and I'm watching what that does to my own judgment.

The old tools waited.

A hammer, a spreadsheet, a search box — they sat still until you acted, and then they did exactly what you said, no more. The relationship was simple: you thought, the tool executed. Whatever thinking happened, happened in you.

The tools I use now think back.

I write half an idea and the system finishes it, in a direction I hadn’t chosen. I ask a question and it questions the question. It suggests, interprets, disagrees, fills the gap I left — sometimes better than I would have, sometimes confidently wrong, and increasingly I can’t tell which at a glance. The tool has opinions now. Or something close enough to opinions that I respond to them as if they were.

I have been watching what that does to me, because it is doing something, and it is more interesting than the productivity story everyone tells about it.

Some days it makes me sharper. The thing answers back, I see a flaw I’d have missed, I think better because I am thinking with something. Other days I notice the opposite: I reach for it before I have finished having my own thought. The answer arrives fluent and complete, and my half-formed version — the one that was actually mine — quietly loses the argument. Not because it was worse, but because it came second, and slower, and it wasn’t as well-dressed.

That is the part I keep circling. Not whether the tool is good; it is astonishing. But when does a fluent answer start borrowing an authority I never consciously handed it? I catch myself trusting a response partly because it is smooth, and smoothness is not the same as being right. I have made decisions faster this year, and I am genuinely not sure whether faster meant clearer, or just meant I stopped sooner because something plausible showed up.

And I see it in the people I work with, which is how I know it is not only me. Someone gets confident on top of an answer they don’t fully understand. Someone else starts doubting a good instinct because the machine framed it differently. The tool doesn’t only help them do the work. It gets into how they decide the work is done, how long they are willing to sit in not-knowing, what they will still call their own.

I don’t have the resolution for this. I am not even sure “resolution” is the right shape to want. What changes in a person who spends every day reasoning next to something that reasons back — in their judgment, their patience, their sense of what counts as their own thinking? I can feel it changing in me, and I can’t yet name where it lands.

Maybe that is the real work of this whole transition, underneath the tools and the saved time: learning which parts of my own judgment I want to keep exercising, precisely because the machine is so willing to do them for me. I don’t know yet which those are. I think finding out is going to take longer than learning any of the tools did — and I think it matters more.

Today’s entry lands tomorrow.
Meanwhile — the evergreen thinking →