Redesigning work with AI did not begin for me when a model became more capable.
It began when I stopped asking how to use AI and started asking what my work needed to look like if I wanted to leverage the full cognitive potential of AI.
That is a very different question.
The first question treats AI as an external object. There is the work, and then there is the tool. The tool might help with a draft, a summary, a research question, or a piece of analysis, but the underlying process stays mostly intact. The work still moves through the same old stages. The same bottlenecks remain. The same habits decide what is possible.
The second question is more uncomfortable. It asks whether the work itself is still designed for the constraints it was built around.
That is where the personal lesson started for me.
For a while, I treated AI as something I wanted it to be for me. I wanted it to remember more than I could remember. I wanted it to document progress through a project, preserve context, organize the history, and compensate for the limits of my own long-term memory.
In theory, that sounded reasonable. A large language model can absorb large amounts of information, restructure messy material, summarize the past, and surface patterns. It seemed obvious that one useful role for AI would be to become a kind of project memory.
It was useful. It was also a half-baked result.
Not because the tool was bad. Because I had made the tool the project. I was asking AI to become something I was not, instead of redesigning the work so that AI became part of how the project moved.
That distinction has become central to how I think about Cognitive Leverage. Cognitive leverage is not a feature. It is a redesign of the path from thought to execution.
The Tool Was Useful, But I Was Not Yet Redesigning Work With AI
In the first project, AI sat next to the work.
It helped me document, retrieve, reconstruct, and occasionally organize. But every serious gap still came back to me. Hallucinations had to be checked. Context windows had to be managed. Missing structure had to be repaired. The system could produce plausible order, but it could not rescue a process that had not been designed clearly enough in the first place.
That was the mistake.
I had treated AI like a database with generative superpowers.
That is an incredibly tempting misuse. You can pour in notes, transcripts, messages, observations, fragments, and half-formed conclusions, and the system will try to turn them into something coherent. It feels like the burden of structure has been lifted.
But it has not.
You are not freed from providing structure because you are using AI. If anything, the need for structure becomes more visible. AI can work with ambiguity, but it cannot know which ambiguity is productive and which ambiguity is just carelessness. It can infer intent, but it cannot always know what mattered to you before the conversation became noisy. It can produce a useful answer, but it will often optimize for plausibility unless you teach it what quality means in the specific context.
The crude version is still true: poor input produces poor output.
The more important version is sharper: unclear process produces unclear leverage.
That became obvious only after I compared the first project with a second one. The second project started from a different assumption: I was not alone.
That sounds simple, but it changed the setup completely.
If I was not alone, then AI could not be an afterthought. It could not be something I reached for only when a task became annoying. It had to be built into the structure of the work from the beginning.
So I designed the project differently.
I created an AI-assisted structure before the work started. I briefed the system on what mattered, how I thought, what good looked like, what constraints existed, what the client needed, and where judgment had to remain mine. I used a structure familiar from my own consulting work: project charters, reusable logic, progress artifacts, synthesis checkpoints, and reporting structures that made it easy to explain progress without rebuilding the story every time.
AI was no longer a database. The context window was no longer a constrained because we collaborated in co-creating a context playground that was constantly maintained by our team effort.
AI became part of the operating rhythm.
That is when the work changed.
A project that would normally have required a small team became something I could do alone, in less time, in collaboration with the client. Not because AI did the project for me. Because the work was designed so that AI could extend the parts of me that were useful: structure, synthesis, critique, documentation, and disciplined follow-through.
That is the personal version of the enterprise AI problem.
The question is not whether AI is useful. It is useful.
The question is whether the surrounding work has changed enough to capture the usefulness.
Validation Became The Center Of The Workflow
The biggest word in this shift is validation.
That surprised me at first, because most people still talk about AI in terms of production. More output. Faster drafts. Better summaries. Cheaper analysis. More speed from the same person.
All of that exists. But production was not the breakthrough for me.
Validation was.
If I teach AI what matters, it can help me evaluate whether I am still hitting the target. It can compare the current output against the standard I set before I became tired, impatient, or emotionally attached to an idea. It can challenge the version of me that wants to accept something because I want the work to be done. It can notice drift between the original intent and the current artifact.
That matters because most bad work does not become bad all at once.
It decays gradually.
An idea starts clearly. Then a useful thought gets added. Then a clever but slightly unrelated point appears. Then a sentence survives because it sounded good. Then a later section quietly contradicts an earlier one. Then the whole thing begins to carry the weight of every decision that was not made sharply enough.
Before AI, an idea often had to mature inside my head for a long time before I could really inspect it. By the time it became visible, it was already carrying bias, contradictions, emotional attachment, and accidental complexity. It had become harder to separate the original thought from the compromises made along the way.
With AI inside the process, the return is almost instant.
That changes the role of judgment.
The thought becomes visible earlier. I can examine the shape before it hardens. I can reject what does not belong before it creates a domino effect that weakens the whole idea. I can ask for counterarguments while the structure is still flexible. I can test whether a direction is worth pursuing before I have invested too much pride in it.
That is why the human premium does not disappear when AI gets better.
It moves upstream into framing, validation, taste, and responsibility.
This is also why AI has lowered the cost of starting but raised the standard for finishing. Starting is easier because a rough structure can appear quickly. Exploration is easier because adjacent ideas can be tested cheaply. Critique is easier because a second perspective is available before the work becomes precious.
But finishing is harder in a different way.
When producing material becomes cheap, accepting material becomes dangerous.
The bottleneck shifts from generating options to knowing which option deserves to survive. The scarce resource is no longer the first draft. It is discernment. It is the ability to see which structure carries the idea honestly, which paragraph is pretending to be useful, which recommendation sounds good but lacks evidence, and which line is mine rather than the statistical average of a machine trained on far too much text.
AI still struggles to maintain my voice.
My mind, my patterns, my feelings, my opinions, and the things I have actually lived are still competing with a huge body of generalized language. If I do not steer the system, it becomes too broad, too smooth, too generic, or too confident in a direction I would not choose.
So the work is not to surrender to the system.
The work is to design the collaboration.
The Real Change Was Saying No To Myself Less Often
The most practical change is that I say no to myself much less often.
That may sound like a productivity detail. It is not.
Before this shift, many ideas died before they had a fair chance to become real. Not because they were bad ideas, but because the implied workload was too large. A book idea meant months of discipline before there was anything useful to react to. A strategy article meant finding uninterrupted time to think, structure, draft, and edit. A client concept meant building the scaffolding before the actual work could even begin.
The idea was often killed by the cost of starting. That changes when AI is embedded into the work process.
Now the first move is different. The thought goes from “I would like to do something” to “let me put this into a system that already knows my preferences, constraints, standards, and line of thinking.” The next five steps appear. The work becomes actionable before motivation has had time to evaporate.
That does not mean every idea should survive. It means more ideas deserve to be tested before they are dismissed.
This is a deeply practical form of cognitive leverage. It is not about becoming a machine for more output. It is about increasing the number of serious attempts that can be made under real conditions.
Serious thinking rarely happens in ideal conditions. It happens while life is happening. It happens while there is pressure. It happens while there are financial constraints, family obligations, unfinished work, interviews, deadlines, and the ordinary mental drag of trying to hold too many things at once.
Under those conditions, the value of AI is not that it creates a fantasy world where work becomes effortless. The value is that it can reduce the activation energy required to continue.
Sometimes that means turning a messy idea into an A4 page for a possible book concept. Sometimes it means creating a structure for a project that would otherwise feel too large to start. Sometimes it means asking for the next five useful steps, not because I cannot think of them, but because I need the work to become visible quickly enough that I can decide whether it deserves energy.
That is not passive.
AI did not change my process. I changed my process once I understood what AI made possible.
I stopped trying to use AI and started redesigning my work around cognitive leverage.
The Evidence Is Pointing In The Same Direction
The broader evidence is beginning to point in the same direction, although it is often discussed in less personal language.
The 2024 Microsoft and LinkedIn Work Trend Index showed how widely employees were already using generative AI at work, often faster than their organizations could formally absorb. That matters because individual adoption is not waiting for enterprise permission. People are already trying to reduce the friction in their own work.
But individual adoption alone does not solve the process problem.
The 2026 Microsoft Work Trend Index makes the next point more directly: leaders are beginning to rearchitect work itself, deciding what humans and AI should each do. That is closer to the real question. AI is no longer only an employee productivity layer. It is becoming a design question for the work.
The research on the jagged technological frontier, conducted with BCG consultants, is useful because it complicates the easy productivity narrative. AI improved performance on tasks inside its frontier, but people could perform worse when they used it for tasks outside that frontier and trusted the output too readily.
That is exactly why validation matters.
The NBER paper Generative AI at Work found meaningful productivity gains in a customer support setting, especially for less experienced workers. But the setting matters. The work had a shape. The AI assistance was attached to a specific workflow. The value did not come from abstract access to a model. It came from applying the technology inside a defined process where the output could be used, assessed, and improved.
That is the lesson I keep coming back to.
AI access is not the same as AI leverage. Usage is not the same as redesign. Activity is not the same as transformation.
The Enterprise Version Is Hiding Inside The Personal One
I think many organizations are still stuck where I was in the first project.
They are trying to use AI as an add-on to the existing process. They look for use cases. They run pilots. They collect productivity stories. They ask where a chatbot, copilot, assistant, or agent might make a step faster.
Some of that is useful. None of it is sufficient.
The better question is not “Where can we improve things with AI?”
The better question is: “What should this process become now that certain forms of cognitive work are cheaper, faster, and easier to repeat?”
That question forces a different kind of analysis.
- If synthesis becomes cheaper, should strategy work still depend on long cycles of manual consolidation?
- If drafting becomes cheaper, should review still begin only after a polished document appears?
- If critique becomes cheaper, should weak assumptions survive until a steering committee finds them?
- If research becomes cheaper, should evidence still arrive late in the process?
- If agents can execute in parallel, should teams still organize every piece of work around sequential handoffs?
These are not tool questions. They are operating questions.
That is why the old workflow is too small for AI. If AI lives in isolation inside a single step, the Interface Tax becomes too high and the return is too hard to see. The user has to move context between systems. The organization has to translate outputs back into the old process. The value leaks away through handoffs, rework, governance gaps, and unclear ownership.
The value appears when there is a through-line across the process.
- Where does AI create options?
- Where does it critique?
- Where does it validate?
- Where does it accelerate execution?
- Where does it preserve context?
- Where must the human decide?
The Human Premium Enterprise will not be the company that automates everything it can touch. It will be the company that understands which human capabilities become more valuable as ordinary cognitive output becomes cheaper.
Judgment becomes more important. Framing becomes more important. Validation becomes more important. Responsibility becomes more important.
The ability to decide what should not be done becomes more important.
That is not a defensive argument for keeping humans in the loop out of sentimentality. It is a practical argument about value creation. When AI makes more things possible, organizations need better ways to decide which possibilities deserve resources, trust, and accountability.
The Question Changed
My own way of explaining AI has changed because my own work has changed.
I no longer think the interesting question is how AI can do things for me.
The sharper question is what I want to be able to do because I am no longer limited to one linear path of thought and execution at a time.
That is where parallel execution, better critique, faster synthesis, more perspectives, cheaper experimentation, and agentic delegation all become part of the same conversation. The value is not in any one capability. The value is in knowing when and how to inject those capabilities into the work.
Delegation is a good example.
Delegating to AI agents can be transformative, but it is impossible to delegate well if you do not know what you want, which is not at all different than when delegating to humans. A vague request does not become strategic just because the recipient is artificial. If the objective is unclear, the agent can only accelerate ambiguity.
This is why redesigning work with AI starts before the prompt. It starts with intent, standards, the very structure of the work. It starts with knowing what should be explored, what should be challenged, what should be preserved, and what should be rejected.
The more capable AI becomes, the less impressive it is to say that we use it.
The meaningful question is whether we have changed the work enough to capture what it makes possible.
For me, the breakthrough was not better prompting.
It was realizing that the work itself had to change.
I had to stop treating AI as a tool I occasionally reached for and start designing the work as if I was no longer alone.
That required structure. It required delegation. It required validation. It required judgment. It required me to be clearer about what I wanted, what good looked like, and where responsibility still belonged.
That is where cognitive leverage starts.
