In Kai-Fu Lee's July book AI Future Has Arrived, we flagged a counterintuitive call: this technology shift replaces not hands but minds—what was once the most expensive resource, human intelligence, is the first to depreciate. The author stresses that "AI is already pulling itself up by its own bootstraps"—already happening, not a forecast.
The book breaks AI-human collaboration into four tiers: chatbot, tool, copilot, and Agent (capable of autonomously planning, executing, error-correcting, and closing the loop on a task). Codex, Claude Code, and Cursor all sit in the top tier. Since March, Chinese practitioners had been running threads on the "devaluation of judgment"; in July, those observations get formalized into a book.
What This Is
Not a tool book—more a survival map. Three things worth remembering.
First, the Agent battlefield is enterprise, not consumer. Consumers measure by "free or subscription"; enterprises measure by "cost of replacing a human × profit increment"—the latter ceiling is orders of magnitude higher. Anthropic's 2025 growth curve has flipped this call from hypothesis into industry consensus.
Second, organizations are being rewritten. The smallest unit shifts from manager to DRI (Directly Responsible Individual—the one truly accountable for outcomes), orchestrating a swarm of Agents in the background. One stinging line from the book: "The higher you sit, the further from the truth you get." Middle management absorbs more impact than most public reports admit.
Third, individual scarcity is shifting from "knowing how to do" to "knowing what to choose." Starting in 2026, being the person on your team most fluent in AI is the cheapest insurance you can buy. But once execution is nearly free, right-brain capabilities—judgment, taste, courage—outvalue left-brain ones.
How the Industry Sees It
Industry consensus broadly aligns, but pushback exists.
The most visible skepticism is on pace. "AI recursively self-accelerating" sounds flashy, but OpenAI and Anthropic are still wrestling the same hard problems: long-chain inference cost, hallucination (AI confidently making things up), and multi-agent coordination. Next-generation training costs have not dropped exponentially; the self-improvement flywheel has no empirical footing yet.
Next, enterprise ROI remains murky. The book itself concedes: "Many companies are using AI, but few are capturing financial returns." Anthropic's growth, however steep, is one data point—it cannot be extrapolated across every sector.
The UBI (Universal Basic Income) section is warm but unfocused. Handing out cash after material abundance arrives is a wish, not a plan; the book offers no path through the next decade-plus of dislocation.
What It Means for Regular People
For traditional enterprise IT: the under-estimated shock lands on the middle layer. What Agents can replace isn't only execution but coordination and reporting. Rewriting job descriptions takes priority over plugging in AI tools.
For individual careers: the execution-role moat is thinning—"uses AI" is no longer a bonus but baseline. Differentiation now comes from problem framing, cross-functional judgment, and decisions made in the gray.
For consumer markets: free apps won't get cheaper because of this book. But the labor cost enterprises save will flow back into cheaper consumer goods. A slow variable, not a sudden shift.