The AI industry is abandoning its own pitch

Goldman Sachs put a number on it this month: across the S&P 500, only 2% of companies have managed to quantify productivity gains from AI, and even those show no statistically significant earnings difference from peers that never bothered ^[1]. A companion analysis found that while 70% of management teams now mention AI on earnings calls, there is no meaningful relationship between productivity and AI adoption at the economy-wide level ^[2]. The industry spent years selling artificial intelligence as a productivity tool. It is now, quietly and from three directions at once, walking away from that pitch. Start with the money that funds startups. Northzone, a European venture firm, says it has stopped backing basic AI productivity assistants and moved several hundred million dollars into autonomous systems for science and defense, on the grounds that the productivity-tool market is overfunded and overhyped despite generating roughly $150 billion to $200 billion in revenue ^[3]. The same logic runs through n8n's $180 million round for AI-agent orchestration, led by Accel. Its chief executive put the retreat plainly.

n8n has dramatically expanded the share of work that can be automated. — Accel

Accel framed the investment not as another generative-AI bet but as a move into agentic deployment ^[4]. And when Andreessen Horowitz and Bessemer put $100 million into Neo, a security startup built for agentic software, the founder's pitch was that nearly half of enterprise software will be agentic by the end of the year ^[5]. The capital is not leaving AI. It is leaving the productivity story for a new one that has not yet been proven. Second, the companies that actually sold the productivity tools. Salesforce authorized $75 billion in buybacks — a record — borrowing $25 billion to fund part of it, while rebranding its Sales Cloud around AI agents under the name Agentforce Sales ^[6][7]. The arithmetic is stark: Agentforce reached $800 million in annual recurring revenue against $41.5 billion in total revenue, meaning AI is less than 2% of a business now marketed around it ^[6]. The stock fell 26% in a quarter as investors worried agents would cut the number of seats customers buy. Even a bullish investor framed the bet in those terms.

phenomenal quarter — Marc Benioff

The buyback does not settle that question. It props up the share price while the core issue — whether agents lift revenue per seat or erase seats altogether — stays open. Third, the labs that build the models. DeepSeek cut its V4-Pro prices by 75%, made possible because Huawei's Ascend chips let it run the model at somewhere between a twelfth and a nineteenth of what GPT-5.5 or Claude costs to serve ^[8]. Its engineers framed the cost advantage as permanent, not promotional.

It is an efficiency gain being passed through. — Sanchit Vir Gogia

The American labs responded the way price-takers respond: OpenAI cut fees on one model by 80% and Anthropic halved prices on another ^[9]. Raw intelligence is now a commodity priced toward zero, and no lab is claiming anymore that the models themselves create enterprise value. Three independent retreats in three directions, all running from the same room. Goldman's finding is not that AI failed; it is that AI never translated into utility at the scale the pitch promised. The counter-evidence sharpens rather than breaks the diagnosis. ADP outperformed the S&P 500 by embedding AI inside its own payroll and HR workflows ^[10]. Indeed worked it into its job-matching product and steered customers toward higher-priced tiers ^[11]. The Bank of England found real gains in software and IT consulting, where firms built AI into their own processes — though the total bump came to a tenth of a percentage point ^[12]. The pattern is consistent: gains arrive when a company that owns a workflow integrates AI into that workflow itself, not when a general-purpose tool is handed across the fence. That is what the industry sold, and it is what all three retreats are abandoning. The autonomous systems the VCs are funding have not been shown to work in production any better than the assistants being discarded. The buybacks create no AI revenue; they buy time while the stock price says what the earnings call will not. And a model priced at a twentieth of its rival is still a tool that has not been integrated into anyone's job. The problem was never the model's intelligence. It was the translation from intelligence to something a company can actually use. None of these three moves does that translation. Each one, in its own direction, is a decision to change the story instead.