Stage 1: AI as tool, human-overseen
Bottlenecks still there: office politics, ossified hierarchies, resistance to change, regulation, liability, professional norms, tacit local knowledge, human preference for dealing with humans, prompt/oversight skill as gatekeeping layer.
Stage 2: AI gradually handles coordination more, humans gradually steer more
Bottlenecks gradually no longer there: office politics, most bureaucratic rigidity, professional norms around process, resistance to change at team level.
Bottlenecks still there: regulation, liability, human taste and demand, tacit knowledge about customers and markets, branding and narrative, politics.
New bottlenecks: legibility problem (can remaining humans understand what the systems do?).
Stage 3: AI-autonomous companies, human-funded
Bottlenecks no longer there: internal hierarchies, funder to ai coordination friction
Bottlenecks still there: regulation, human demand and taste (some customers are still human), physical world constraints (supply chains, permits), liability frameworks.
New bottlenecks: alignment between funder intent and AI execution, regulatory capture as primary competitive moat.
Stage 4: AI-autonomous beyond single companies
Bottlenecks no longer there: human regulatory leverage (eroding), most human-world interface frictions.
Bottlenecks still there: physics, energy, compute.
New bottlenecks: governance vacuum (who regulates what no human fully understands?), AI-to-AI coordination frictions (competing optimization targets, standards).