Developers at Anthropic say Claude is involved in ~60% of their work.
But they estimate they can fully hand off only 0–20% of their total work to Claude — meaning no meaningful human verification or intervention.
That’s the interesting gap.
AI is involved in a lot of the work.
Very little can simply be handed over.
And honestly, that’s where the real advantage is.
Not just writing code.
Deciding what’s safe to ship.
Try this 5-minute test:
Open your last 10 AI-assisted pull requests. Count how many you merged without changing.
That’s your real number.
I checked my last 10. Only 5 went through without changes.
What’s yours?