\n<\/p>\n
A robot pours popcorn from a cooking pot into a bowl on March 8, 2017 at the Institute for Artificial Intelligence (AI) of the university of Bremen, northwestern Germany. – Scientists of the institute work among others on “AI-based control methods for robotic agents, performing human-scale everyday manipulation-task”, and investigate “computing systems that assist humans at home, at work and during leisure activities”. – Germany OUT (Photo by Ingo Wagner \/ dpa \/ AFP) \/ Germany OUT (Photo by INGO WAGNER\/dpa\/AFP via Getty Images)<\/span><\/p>\n dpa\/AFP via Getty Images<\/small><\/div>\n<\/div>\n<\/figure>\n Box CEO Aaron Levie has a name for what afflicts most C-suites deploying AI: AI psychosis<\/u><\/a>. Executives sit far enough from front-line workflows that when they test an AI tool, they see only the happy path, never the ten or twenty steps that have to happen after the prototype before any agent delivers sustainable results. The gap between a polished demo and a production-grade system is where most enterprise AI investments go to die.<\/p>\n Levie\u2019s diagnosis is backed by data. 42% of companies abandoned most AI initiatives in 2025<\/u><\/a>, up from 17% the prior year, according to S&P Global Market Intelligence. The average organization scrapped nearly half its AI proofs-of-concept before they reached production. The companies still standing in that wreckage are now where the real capital is flowing.<\/p>\n The Capital Is Watching<\/strong><\/p>\n Venture capital has not lost faith in AI agents. It has simply recalibrated where it bets. Agentic AI companies raised $2.66 billion across 44 rounds in the first four months of 2026<\/u><\/a>, compared to $1.09 billion in the same period a year earlier, per MarketsandMarkets. Average round size for agentic AI startups nearly doubled, reaching $155 million, as investors concentrate capital on companies that have crossed $100 million ARR and demonstrated production-grade reliability, not demo-grade novelty.<\/p>\n The global AI agents market was valued at $7.84 billion in 2025 and is projected to reach $52.62 billion by 2030<\/u><\/a>. But investors are no longer writing checks based on a CEO’s enthusiastic walkthrough of a prototype. The firms raising $150 million-plus rounds in 2026 are those that can demonstrate consistent task completion at scale, audit trails, and governance infrastructure. In Levie’s framing: they solved the last mile.<\/p>\n The thesis is filtering down. Enterprise AI revenue reached $37 billion in 2025, growing more than 3x year-over-year<\/u><\/a>, according to VC Cafe’s 2025 market analysis. Investors now describe a “layer cake” thesis: foundation models for infrastructure bets, vertical applications for the biggest near-term returns. The verticals winning are those where the last-mile work is well-defined, regulated, and high-stakes enough that enterprises will pay for reliability.<\/p>\n The Prototype Is Not the Product<\/strong><\/p>\n A CEO generates a contract with an AI agent and calls it a win. What the CEO did not do: verify every term before it went to a counterparty, wire up all prior contracts to maintain consistency, or build the review workflow that prevents a single hallucinated clause from becoming a liability. “Look I generated a contract,” he wrote on X<\/u><\/a>. “Yes but you didn\u2019t verify all the terms before it goes out to the counterparty and didn’t have to wire up all the past contracts to work with.” <\/p>\n The pattern holds across code; a CEO ships a prototype and is impressed. The engineers who inherit it review the code, identify edge cases, fix issues before deployment, and absorb the maintenance burden. The CEO saw the product while the engineers saw the work it takes to make it production ready.<\/p>\n This cognitive distance is a result of the organizational role. CEOs are, by design, removed from last-mile execution. The problem is that AI tools are exceptionally good at producing convincing first drafts of almost everything, and senior executives encounter AI almost exclusively in controlled demonstrations, not in the messy middle of a production workflow.<\/p>\n The research confirms this; RAND Corporation found that over 80% of AI projects fail, twice the failure rate of non-AI technology projects<\/u><\/a>. A McKinsey 2025 survey found that organizations reporting meaningful financial returns were twice as likely to have redesigned end-to-end workflows before selecting any modeling technique. <\/p>\n Use It More, Not Less<\/strong><\/p>\n Levie\u2019s prescription is that CEOs should use AI more – not pull back, but the mode matters. Using AI to generate outputs is not the same as using AI to understand systems. The executives who will make sound deployment decisions are those who use AI tools extensively enough to encounter the failures, the edge cases, and the downstream dependencies that don’t appear in a demo.<\/p>\n “The best thing you can do as a CEO is to use AI a ton to figure out the real implications of agents in the enterprise, and come out the other side with an appreciation for both the upside and the real work that goes into them,” Levie wrote<\/u><\/a>. That has commercial value as founders and enterprise buyers who understand the full stack of effort required to get an agent into production are the ones making rational bets. Everyone else is funding a potential graveyard.<\/p>\n Levie himself has been consistent on this. Speaking at TechCrunch Disrupt in October 2025<\/u><\/a>, he argued that mission-critical business processes need a “church and state” separation between deterministic systems and probabilistic AI, citing real-world examples of agent failures including data leaks and unexpected database modifications. The enthusiasm for agents is valid, but we need to be cognizant of the governance infrastructure to support them is still being built.<\/p>\n Where the Real Bets Are<\/strong><\/p>\n The market is bifurcating. On one side: organizations where executives saw a demo, approved a budget, and are now watching agents fail in production. On the other: a smaller cohort of companies that mapped the last-mile work before they wrote the deployment brief, and built the human review layers, data governance, and workflow redesign required to make agents reliable.<\/p>\n