AI Gym is a simple idea: the practice of staying lucid under cognitive acceleration.
A few months ago, I said goodbye to Power Platform for serious product development. It is still excellent for prototyping and simple client work. But for complex products, it hit the ceiling. FinModeler, Training App and a few other secret projects now run on Azure. And they’re better, more powerful than ever.
The visible change was, of course, an impressive technical (and personal) achievement. The deeper change was cognitive. As my ideas gained the strength of a 1,000 horsepower machine, my nervous system stayed behind, looking for coffee. There’s just no way I can keep up with agents.
The practice of staying lucid under cognitive acceleration
This experience made me reflect deeper on my experimental AI Gym concept. By the end of this post I’ll show you five exercises to keep you fit.
AI can make a one-person business feel, briefly, like a small studio with tireless colleagues. But, it cannot decide what matters, and it cannot know why I care about doing the things I want to do. That part is still mine.
This is why AI Gym is not another productivity method. It’s not about doing more with less, but about help you use AI without losing judgment, taste, responsibility or the pleasure of creating.
A concrete story
A recent project made this very real. An accounting firm asked us to redesign how client documents arrive, get classified, stored, shared and controlled. The old process was familiar: emails, folders, manual checks, unclear ownership and fragile memory. Nothing dramatic. Just the usual sediment of years of workarounds.
The new system is more structured: client portals, document flows, Dataverse records, SharePoint storage, permissions, audit trails and, of course, AI-assisted classification.
This is exactly the kind of system I used to imagine but hesitate to build. Too many moving parts, too many decisions and too much plumbing.
With agents, the system became buildable. It was even a learning experience for all of us involved and the final product was even better than what we envisioned!
But when the client asked detailed questions about some process decisions, I felt a strange gap. Some of the detailed implementations had been expanded, tested and shaped with AI support.
The system and the responsibility are mine, but not every line of reasoning had passed through my head in the old slow way. I haven’t put my own hands everywhere, and I couldn’t answer all the client’s questions anymore!
That is a new (and somewhat embarrassing) tension. AI lets me debottleneck exploration and first execution. But it cannot sustain judgment or accountability.
The new rhythm: delegate, inspect, simplify, explain
The workflow is changing. You no longer need to try to understand everything before moving. That would kill the leverage. But you also cannot let AI create a fog of plausible complexity.
So, here’s a new rhythm for you:
Delegate. Inspect. Simplify. Explain.
Delegate means you let the agent carry work you no longer need to perform manually.
Inspect means you check the risk, not every comma.
Simplify means you refuse complexity that exists only because the machine can generate it.
Explain means you force yourself to describe the decision in plain language.
That last step is brutal and useful. If I cannot explain a system to a client, I do not understand it well enough. If I cannot explain why a design choice exists, I should not hide behind architecture. If I cannot say where the risk lives, I have not inspected properly.
Outsourcing ideas without outsourcing judgment
I am becoming more comfortable with dependence. I am dependent on AI, and I admit it. Professional pride used to mean knowing how to do the whole thing myself. Build the model, write the formula, design the dashboard, create the flow, fix the bug. And explain the result.
Now the skill is different. The skill is knowing what not to do yourself. In my recent experience, I no longer touch the code myself. Agents write it faster and with fewer errors. I don’t aim to resist that, but to stay responsible for what the code does.
I am also becoming comfortable with outsourcing part of my intellectual energy. Especially, the early expansion of an idea.
My ideas leave my head earlier now. They come back with structure, objections, variations and sometimes better names. That is not a small change. For a long time, ideas had to wait for my energy, my concentration and my technical ability. Now they can move before I am fully ready.
That is where the joy comes from.
Cognitive leverage becomes cognitive load
AI gives me cognitive leverage because I can explore more options, test more different architectures, draft more arguments, compare more approaches and build more ambitious products.
But leverage has a cost. Each generated option must be judged, each branch must be closed or continued, each prototype creates emotional attachment, each new possibility competes for attention.
So, the bottleneck moves. Before AI, the bottleneck was often execution. Now the bottleneck is selection.
In practical terms, this means you now need to think about:
- What should you build next?
- What should you ignore?
- What deserves another hour of your precious time?
- What should die now, before it becomes a project?
Recent research points in this direction. A Microsoft Research and Carnegie Mellon study surveyed 319 knowledge workers and collected 936 examples of GenAI use at work. It found that higher confidence in GenAI was associated with less critical thinking, while higher self-confidence in the task was associated with more critical thinking. It also found that GenAI shifts critical thinking from direct execution to verification, integration and stewardship.
Not every productivity gain should become more work
Let’s direct the productivity gain to leisure and reflection. For years, gaining time meant filling it. Thus, if a tool made you faster, you raised the target.
AI makes that temptation much worse. Because the gains are not small, they change the scale of what one can attempt. But capturing 100% of the productivity gain as extra work is operational debt with a human face.
You enjoy the upside now and pay later: in fatigue, worse judgment, shallow attention, impatient decisions and a nervous system that can no longer distinguish signal from noise.
So, here’s something for you to try that feels almost unnatural: do not capture the full productivity gain.
Let some of it become rest. Or walking, surfing (or another passion of yours). Let some of it become doing absolutely nothing.
This is not laziness; it’s a system design.
Learning must remain partly slow
There is another part of this I do not want you to lose: some learning must remain slow.
The mind needs friction to own an idea. A summary can tell me what a book says, but it cannot give you the slow internal rearrangement that happens when you sit with a difficult argument, resist it, misunderstand it, return to it, and finally make it yours.
That is why you still need books. And conversations with people who interrupt your certainty. Clients who ask annoying questions, friends who do not autocomplete your thoughts, teachers.
This is also part of the AI Gym: knowing when not to use AI. When to think alone, when to ask a person, read slowly, stop optimizing and when to go outside.
The future may be accelerated but your brain is not. And neither is the part of you that decides what the acceleration is for.
Start building your AI Gym now
I’m still designing this gym myself, but here are the five exercises I already practice. They’re organized by cadence, not by importance. The discipline comes from the rhythm, not from any single one of them.
Two daily reflections (morning and evening). In the morning, before opening anything, I picture what a perfect day would look like: what I’d want to have accomplished by the end of it. At night, I compare that picture to what actually happened. The interesting part is not the gap, but separating what was under my control from what wasn’t. Most days, the gap is smaller than it feels. Some days, it teaches me something I would have missed otherwise.
Explain what you just built. Once a day, I talk to a client, a colleague or a friend and explain, in plain language, what I’m working on. Ten or fifteen minutes is enough. If I can’t explain it, I don’t understand it – which means I’ve been coordinating an agent rather than thinking. This is the cheapest, most uncomfortable test I know.
Two or three deliberate pauses per week. I’ve been doing this for decades, and I do it more now, not less. I stop working – properly – and go surfing, walking, or anywhere without a screen. The point is not rest in the productivity sense. It’s giving the part of me that decides what the acceleration is for a chance to catch up with the part that’s accelerating.
Monthly idea euthanasia. Since the new bottleneck is selection, I treat killing ideas as a real practice, not an accident. Once a month I look at every active project, prototype and direction, and ask which ones still deserve me. The ones that don’t: too much risk, too little upside, or simply no longer mine. I close kindly. Leverage without pruning becomes noise.
Quarterly client audit. Acceleration makes the wrong clients more expensive, not cheaper. The ones who drain energy without valuing judgment will quietly consume the very capacity AI just gave me. Once a quarter, aligned with the billing cycle, I review the list and let go of the relationships that no longer fit. Gently, but without negotiating with myself.
This is a gym in progress. Yours will look different: different exercises, different cadence, different reasons. The point is not to copy the routine, but to have one.
If these ideas resonate with you, you may also enjoy AI Gym: Training Cognitive Capacity, where we further explore cognition, mental bandwidth and human performance in the age of AI.
