Build something small
A course teaches you the tool; a project teaches you the job. Pick one real job, run it end to end on a real piece of work, and let it be rough
The goal of this work is the clarity of your judgment. Everything downstream — the tools, the automation — depends on it. The Fractals System is a full-day workshop I run with one team, on its real work. It exists because adoption turns on two things: a change in how people see their own role, and the ability to say what good looks like, clearly enough that someone else can produce it and you can check it. This page is the belief, the method, and the day itself.
One full day, your team, your real work. Ask me for the outline and available dates.
[email protected]A stalled AI rollout starts as a decision about tools and ends as a licence nobody opens. The pilot produces something plausible, the operator spends an afternoon fixing it by hand, and the habit of doing the work yourself wins. The technology worked; the work didn't change, and a faster tool got bolted onto a role that stayed the same.
Tom de Bruyne draws the distinction that explains most of it, between awareness (knowing what AI can do), capability (using it when you try) and habit (reaching for it without deciding to). Most programs deliver the first, some the second, and almost none the third — which is why the returns fade three weeks after the workshop. His line: you cannot close a behavioural gap with more information.
AI training for employees: what actually works, Tom de Bruyne, SUE Behavioural Design, 24 May 2026
The pressure behind that is measurable. In June 2026, Harvard Business Review published the case from Marco Argenti, the chief information officer at Goldman Sachs, on what the last year of agentic AI has actually changed about work.
of 1,320 tasks across 44 professions — agentic models performed as well as or better than human professionals in 80% of cases, up from about 50% six months earlier.
His conclusion is the one I'd underline for anyone running a small operation: many of the skills people spent careers building can now be executed by agents, and the question of how to adapt is a pressing one for the people who do that work.
A senior banker asked Argenti which 10% of his job he should hold on to, the part AI would never be able to do. The answer was to let go of that 10% entirely. Argenti's comparison: an experienced horse rider learning to drive a car. Which horse-riding skills should they keep in order to drive well? Probably none. What carries over is instinct and judgment — reflexes built over years, pointed at a different machine.
His prescription for the banker was a change in stance: from making every line of the pitch deck to giving clear instructions, setting the controls, and supervising the result. From individual contributor to supervisor and mentor.
The piece argues that the workplace question has moved from which skills to keep to how you think about your own role: don't just reskill, reimagine skills and build new habits. Argenti's line on the argument underneath it — if we don't know what good looks like, neither humans nor AIs will know how to take the right steps towards success — is the same observation this practice is built on.
To Thrive Alongside AI, Focus on Mindset—Not Skillset, Harvard Business Review, 12 June 2026
I spent thirty years running newsrooms, which turns out to have been an agent system before anyone used the word. An editor holds the judgment, briefs the reporter, and checks the result. When a draft came back to me, I read it against the standard I already carried and said exactly where it missed, and that read is the part the reporter could not do yet. It is why the person in this seat usually knows more than they can currently say.
That is the shift this program asks for, stated plainly: you already know what good looks like, and you prove it every time you look at a draft and know it's off. The machine cannot see that standard, because it has never been said out loud.
The mindset is the decision. The habits are what carry it, and they run against how most of us were taught to adopt new software: learn the tool first, find the perfect course, wait until you feel ready. Argenti's line is the whole instruction — reimagine skills and build new habits.
The behavioural research adds two mechanics that any program has to respect. The first is anxiety: the fear of looking incompetent in front of colleagues suppresses experimentation, and experimentation is the only thing that builds capability, so the room has to be safe for imperfect work. The second is the trigger — a specific moment where reaching for the machine is the decided first step, rather than a general intention to use AI more. Habit then takes repetition in context, which the research puts at four to eight weeks on real work. A one-day workshop cannot deliver that, and neither can a course. Four habits do most of the work.
None of this needs a bigger budget or a new subscription. It needs a job you care about and the willingness to say what good looks like before you hand it over. Every job we set up gets its trigger named out loud for the same reason — when this lands, the first step is this — because a behaviour with a moment attached is the one that survives a busy week.
The mindset gets you into the room. The brief is what makes the work good. In practice it is one page, written in your own words, that a person or a machine can work from. I call it the statement, and it has five fields. The fifth is the one most people skip.
Written down, that page does something a conversation cannot: it survives. It briefs the next assistant, the next hire, and the colleague covering for you on Friday, and it gives you something concrete to judge the work against. The clarity of your judgment stops living in your hands and becomes something you can point at.
The practice runs on one job at a time, in three beats. The same question — what does good look like here — gets asked in three places.
Say it is the session. We walk through your week — what you produce, what eats the time — and pick the job with the clearest rules. The statement gets written field by field, out of what you say.
Delegate it is where the machine takes the doing. You describe the outcome and the machine produces the outputs. Stay at that level: the moment you start specifying how, the division breaks and you are back to fixing drafts by hand.
Optimize it is the rhythm after that. Run it, check it, sharpen the statement. Expect the first two weeks to feel slower; that dip is the climb, and each cycle moves you a level up.
Set up one job, not an agent. One job needs the context around it, the two tools it touches, the steps, and one way to check the work. Starting is the point — a perfect agent that never ran is just a plan.
Delegating the doing means setting the machine up to take it. Four things have to be in place before it can run.
This is where the work gets honest about your own material. An assistant can only work with what it can see, so the state of your files, your versions and your notes decides how much of the job can actually move. On my own operation all four layers run on one rented server, and the architecture, the tools and what they cost are documented here.
One full day with your team, working on the jobs that actually eat the week. The morning is about seeing the shift. By early afternoon everyone is working on their own job. The weeks afterwards are where it becomes a habit, which is why the day comes with a thirty-day check attached. Nothing on the agenda is hypothetical, and nobody leaves with a prompt library.
The day runs 09:00 to 17:00 with the people who produce the work and whoever owns the outcome, in person in Singapore or over video for teams elsewhere. For a larger organisation I run it team by team and the shared standard is written once.
Everyone leaves with their own judgment written down, in a form they can use on Monday, and the team leaves with one standard it agreed on together.
Thirty days later we run the check together: what held, what drifted, what you actually needed to look at. The work runs to ninety days, because a new habit takes weeks of repetition on real work before it stops being effortful. I ask whoever leads the team to run the System on their own job first, because people copy what they watch their boss do.
The day suits a team that produces something every week and is ready to change how it gets made. If the tools are still being chosen, start with the one-to-one work at AI coaching sessions — better to specify the job with one person before asking a whole team to change.
The tools are the easy part, and most teams already have access to one. The day is about the brief: saying what good looks like, plainly enough that a person or a machine can produce it and you can check it. People leave with their own judgment written down, not a prompt library.
You need one recurring job your team produces every week and the last real version of it. Everything else happens in the room, on the AI tool you already pay for.
The people who produce the work, plus whoever owns the outcome. The day runs on your real jobs, so observers learn much less. For a larger organisation we run it team by team and write the shared standard once.
Each person leaves with one job to hand over that week and a date on it. Thirty days later we run the check: what held, what drifted, what needs sharpening. The follow-through runs to ninety days, because a new habit needs weeks of repetition in context before it stops being effortful.
Most of that training transfers information about what AI can do, which produces awareness, and awareness fades within weeks. This day is designed as behaviour change: practice on real work inside the day, a named trigger for each person, and a follow-through long enough for the habit to take.
Yes, over video for teams outside Singapore. The in-person day works better for the handover, because people see each other's briefs and the standard gets argued out in the room.
The bottleneck is never the model — it's whether you can describe, plain and specific, what done means. That's the whole skill, and what this day exists to build.
If your team produces something every week — reports, briefs, proposals, content — and AI is not carrying any of it yet, one day is enough to change how that work gets made.
Tell me what your team produces and what is eating the week; I'll tell you straight whether the day is the right fit. Individuals start with the coaching sessions.