The default question in any AI strategy conversation is “what can AI do?” It’s the wrong question. Not because it’s unimportant, but because the answer changes every quarter. Whatever you decide AI can’t do today, it probably can next spring. Building a division of labour around capability boundaries is building on sand.
The better question is “what should a human do?” That boundary is a choice, and choices are stable in a way that capability assessments aren’t.
I’ve been running a personal AI system for a few months — agents handling research, synthesis, code, monitoring, drafting. The kind of setup where you could, in theory, delegate almost everything. Which forced the question: what do I actually keep? Not what must I keep because AI can’t handle it. What do I choose to keep, even when AI could do a credible job?
The answer settled into five categories.
Presence. Things that require you to actually be there. Not your output, not your judgment — your physical, attentional presence. Reading to your kid. Sitting with a client where the value is that you showed up and listened. Having dinner with your partner without half your mind on what the agent is doing. AI can’t be present on your behalf, but more importantly, you wouldn’t want it to. These are the things that would be meaningless if delegated.
Sharpening. Work you do specifically to maintain the judgment that evaluates everything else. Reading source texts rather than summaries. Forming your own view before reading what AI produced. Drilling from memory rather than looking things up. This category is counterintuitive because AI could help more, and you deliberately limit it. The reason is calibration: if you never struggle with the raw material, you lose the ability to tell when the AI’s synthesis is subtly wrong. You become a rubber stamp. Sharpening is the maintenance cost of being a useful human in the loop.
Collaborative. Neither you nor AI alone produces the result. You brainstorm together. The AI probes your reasoning and you probe its analysis. You write a strategy draft and it pressure-tests it. The output is genuinely co-produced — not AI-drafted-and-human-approved, but iteratively shaped by both. This is where the interaction is the product, not the artifact.
Automated. The big one. Research, synthesis, code, monitoring, first drafts, data processing, maintenance — anything where AI can produce the deliverable and you can verify it faster than you could produce it. The volume category. The one that should expand to fill available budget, because the task space here is essentially infinite. Your goals always need more research, more synthesis, more preparation. The constraint is budget, not ideas.
Dropped. The uncomfortable one. Work that doesn’t serve any goal you actually care about. Not delegated to AI — dropped entirely. No one does it. The meeting that exists because it always has. The report no one reads. The optimisation of something that doesn’t matter. Saying no is the hardest form of taste, and it’s the one AI can’t help you develop because AI will cheerfully do anything you ask it to.
What I find useful about this framing is that it survives capability improvements. When AI gets better at something currently in Collaborative, it might migrate to Automated. Fine. The categories still hold. When AI develops something that feels like presence — sophisticated emotional interaction, say — the question isn’t “can it?” but “should it replace you there?” The framework forces intent, not just capability assessment.
The practical test is simple. Before any task, ask: which category? If it’s Automated, dispatch it. If it’s Sharpening, do it yourself with minimal AI scaffolding. If it’s Collaborative, do it together. If it’s Presence, close the laptop. If it doesn’t map to something you actually care about, drop it.
The category most people underinvest in is Sharpening. It feels inefficient. Why struggle with the source material when AI can summarise it in seconds? Because the struggle is the product. A consultant who never reads primary research becomes a consultant who can’t tell when the summary is misleading. An engineer who never writes code from scratch becomes an engineer who can’t evaluate AI-generated code. The Sharpening category is where you pay the tax to remain qualified to judge the Automated category’s output.
The category most people avoid is Dropped. It requires admitting that some work you’ve been doing — maybe for years — doesn’t actually matter. That’s an identity challenge more than a productivity one.
But here’s the implication that keeps nagging at me: if you take this framework seriously, the question isn’t how to use AI more effectively. It’s how to decide what kind of human you want to remain. The division of labour between you and your AI isn’t a technical decision. It’s a statement about what you think is worth doing yourself.
What would you refuse to delegate, even if the AI were perfect?