The 80/20 Partnership

The end state of working with AI isn’t “AI does everything” and it isn’t “AI assists you with everything.” It’s a division of labor: the system runs roughly 80% of the work, and you keep the 20% it genuinely can’t do.

The split follows each side’s strengths. AI is excellent at high-frequency, structured, repeatable execution — the daily loops, the routine processing, the work that’s heavy because there’s a lot of it, not because it’s hard to judge. AI is bad at strategy, at taste, at your voice, at high-stakes one-off calls. So you stop dividing work by task and divide it by kind: everything repeatable flows to the system; what stays with you is exactly the part only a person can bring — direction, taste, honest reflection, the judgment calls about what matters.

The hard part is the oversight instinct. If a system runs while you sleep, the reflex is to want approval on everything — and that reflex quietly rebuilds the bottleneck you just removed. Good oversight doesn’t mean approving every action; it means being in a position to intervene when it matters. Anthropic’s research on its most experienced users found exactly this: as people get comfortable running agents, they approve individual actions less and step in more sharply when something warrants it. Less gatekeeping, better catches.

Here’s the irony worth sitting with: the more execution the system absorbs, the more the human part matters — because when everyone’s execution is cheap, judgment, taste, and direction are what’s left to compete on.