Siemens CEO Roland Busch just delivered the most candid executive assessment yet of AI's promise and limits in industrial automation. In a wide-ranging interview, the leader of the 320,000-employee manufacturing giant explained why base LLMs achieve only 60-70 percent accuracy in factory settings - and why domain-specific training is the only path to viability. But his vision of fully automated factories raises uncomfortable questions about employment, economic inequality, and the future of work in an AI-driven economy.
Siemens controls one-third of the world's manufacturing lines, automates buildings across every major city, and powers nearly half the global electrical grid. But CEO Roland Busch knows his biggest challenge isn't hardware - it's the uncomfortable truth that current AI can't replace human expertise without massive intervention.
In a rare extended interview, Busch laid out the clearest picture yet of what enterprise AI implementation actually looks like when hallucinations mean production failures and 70 percent accuracy equals catastrophic losses. The reality contradicts much of Silicon Valley's automation hype.
"If you only use an LLM or an LLM-based agent to fix a problem, the hit rate is near what we need," Busch explained. Raw models from Microsoft, OpenAI, and others hover around 60-70 percent effectiveness in factory settings. Only after extensive training on proprietary industrial data - machine logs, repair histories, sensor readings - does accuracy jump to the 95 percent-plus threshold manufacturing demands.
The gap reveals a fundamental divide in how AI works in consumer versus industrial settings. Chat about vacation plans with 70 percent accuracy and users barely notice. Apply that to a pharmaceutical production line and people die.
Siemens' solution involves what Busch calls "industrial AI models" - foundation models augmented with decades of manufacturing data from thousands of factories. The company built data-sharing alliances with German machine builders including Trumpf and DMG Mori, pooling operational information to train models that understand not just generic factory processes but specific quirks of individual equipment.
One example: optical inspection systems struggled until Siemens trained models with synthetic data generated from actual defects. Another breakthrough came when the company used Nvidia photorealistic ray tracing to simulate parts under different lighting conditions. "These little details of having a normal representation of a digital part and a really photorealistic one made the hit rate come up substantially," Busch said.
The admission carries weight. Siemens operates in sectors where failures cascade - power grids, transportation systems, medical equipment. If the world's largest industrial automation company can't make LLMs work reliably without extensive customization, the implications for rapid AI deployment across other industries are sobering.
But Busch's vision for the future raises even harder questions about employment and economic opportunity. He describes a world of "fully automated factories" where digital twins simulate every product before manufacturing begins, AI agents diagnose problems in real-time, and workers primarily serve as hands carrying out instructions delivered through augmented reality glasses.
"An AI factory fully automated uses a lot of space, uses a lot of energy, and it creates a limited number of jobs," Busch acknowledged. When pressed on who buys products if fewer people have employment, he pointed to aging demographics in Germany, Japan, Korea, and China. The argument: labor shortages in developed economies will absorb displaced manufacturing workers into service sectors.
It's a theory that sidesteps uncomfortable realities. Manufacturing historically provided middle-class incomes for workers without advanced degrees. Service jobs - healthcare aides, restaurant workers, retail clerks - typically pay less and offer fewer benefits. Busch's optimistic framing about "deploying labor where it makes a difference" glosses over what happens when blue-collar manufacturing jobs transform into lower-wage service work.
The geopolitical dimension adds another layer of complexity. Siemens operates as what Busch calls "a Chinese company, as much as a United States company and a European one," with 87 percent local content in major markets. That global integration faces pressure from rising trade barriers and nationalist industrial policy.
Busch revealed Siemens is actively "forking" technologies by region - training industrial AI applications for China on Chinese LLMs while using American systems for US operations. The company maintains 45,000 US employees, 30,000 in China, 35,000 in India, and 85,000 in Germany. Managing AI development across fractured technology ecosystems adds cost and complexity but preserves market access.
When asked directly about planning for NATO's potential dissolution, Busch demurred on scenario planning but confirmed Siemens is building regional technology resilience. "We cannot solve it if we box ourselves too small," he said, expressing hope that globalization's benefits will eventually overcome nationalist pressures.
Inside Siemens, Busch is driving what he calls the company's deepest organizational transformation in two decades. The "ONE Tech company program" aims to break down divisional silos that prevent AI and data from scaling horizontally. Under the old structure, simply determining total revenue from a customer like BMW required manually aggregating numbers across business units.
The new model creates unified "fabrics" - data, technology, and sales - as horizontal layers supporting six core business units. It's a classic functional-versus-divisional reorganization challenge, complicated by 170 years of accumulated organizational complexity and 320,000 employees across radically different businesses from trains to medical scanners to building automation.
"AI doesn't respect silos," Busch explained. "AI doesn't respect data silos, doesn't respect any kind of boundaries." The implicit admission: Siemens' traditional strength in domain expertise across industries became a weakness in the data-hungry AI era. Scale requires consolidation.
The program launched a year ago with plans to complete major structural changes by October 2026 and achieve full scaling within two years. Early results show progress, particularly in sales systems. But the human cost is real - not everyone survives such transformations.
Busch's approach involves extensive communication, gradual rollout, and bringing in outside experts with "gravitas" to drive change. "We give people an opportunity to contribute, to bring their ideas, but we have a clear idea of where we want to go," he said. It's collaborative autocracy - input welcome, but the destination is set.
For decision-making, Busch emphasizes empowerment at the lowest possible level within strategic boundaries. Major decisions like M&A follow structured processes with "P" proposals for preliminary approval and "I" proposals for final investment authorization. The system aims to balance agility with governance.
Looking ahead, Siemens is betting big on what Busch calls an "industrial AI operating system" - a platform that connects digital twins of products, machines, and entire production lines, ingesting real-time data and deploying AI agents to optimize operations autonomously. Customers like PepsiCo are already testing early versions.
The vision is compelling: seamless, optimized manufacturing with minimal waste and maximum flexibility. Products designed virtually, simulated comprehensively, then manufactured with near-perfect quality. Changes to production happen in minutes, not weeks. Yields stay high, variants proliferate, customization becomes trivial.
It's also slightly dystopian. Workers become appendages to AI systems, valued for manual dexterity rather than judgment or expertise. The satisfactions of craftsmanship, problem-solving, and mastery give way to following augmented reality instructions. Economic security depends on service jobs that may pay poverty wages.
Busch maintained optimism about technology solving global challenges - feeding 10 billion people, addressing climate change, managing healthcare for aging populations. "We cannot solve it if we box ourselves too small," he repeated, arguing that international cooperation ultimately benefits everyone.
But his own admission about tariffs not disappearing - "they help close the budget deficit, and I never saw taxes going back" - suggests the forces pushing toward fragmentation may prove stronger than globalization's pull.
For now, Siemens is hedging. Doubling US manufacturing capacity for electrical equipment. Investing in Indian operations. Maintaining Chinese production. Building redundant supply chains for critical semiconductors. The company that built its empire on free trade is quietly preparing for a world where walls matter more than bridges.
Busch's candid assessment offers a reality check for enterprise AI hype. The gap between foundation model capabilities and industrial reliability requirements is wider than most vendors admit. Success requires massive data investment, domain expertise, and organizational transformation - not just API access. For workers facing automation, the picture is equally sobering: jobs aren't disappearing overnight, but the nature of work is shifting toward lower-skill, lower-pay roles that serve AI systems rather than developing expertise. Siemens' ability to navigate this transition while maintaining global operations amid rising nationalism will test whether 20th-century industrial giants can thrive in an AI-powered, increasingly fragmented world.