For most people, AI still means a chatbot in a browser or a recommender that pushes videos and ads. But in the past year, much of the interesting innovation has quietly moved into places the average person will never see: power plants, factories, logistics hubs, and gigantic automated warehouses. Agentic AI refers to systems that can sense, decide, and act in relatively autonomous ways toward a goal, within constraints humans define. Instead of a single model answering questions, you have networks of software ‘agents’ that read sensor data, simulate options, talk to each other, and then trigger actions in the physical world like adjusting a turbine, slowing a crane, rescheduling a shipment, or shutting down a line before something breaks. The best analogy is not science fiction but diplomacy itself: imagine a team of junior diplomats constantly exchanging cables, updating situation rooms, and proposing options to the ambassador. The agents never become the ambassador, but they narrow options, monitor risks, and handle 24/7 complexity humans cannot track alone. This raises three questions that will run through the rest of this piece: To understand why industry cares so much about agentic AI, consider a single smartphone battery: A decision to prioritize one batch of batteries over another, or shift production between two plants, ripples through energy grids, labor schedules, customs systems, and regional power markets. Agentic AI becomes an orchestration layer: thousands of micro-decisions that keep a system functioning under stress, while humans focus on the exceptions, the strategic calls, and the moments when values clash with efficiency. This is why the past year is often described in industry as the moment when AI moved ‘from pilot to plant‘. Manufacturing and energy companies are no longer asking if AI works; they are asking where to safely hand it the wheel and where to insist that the human stays firmly in charge.What ‘agentic AI’ really is and why diplomats should care
Why a smartphone battery is harder to make than it looks
The shift from ‘pilot to plant’ is happening globally, but the motivations, players, and governance challenges vary sharply depending on who’s deploying the technology and why.
The recently announced Siemens–NVIDIA partnership aims to build an ‘industrial AI operating system’ combining Siemens’ factory knowledge with NVIDIA’s compute and simulation environment. Digital twins, high-fidelity virtual replicas of plants and grids that ingest live sensor data, are central. Agentic AI systems ‘live’ inside these twins, testing different production schedules, energy mixes, or maintenance strategies in simulation before proposing changes to human managers.
From a diplomatic perspective, this raises three issues:
Chile offers a very different entry point into the same technology. Codelco, the state-owned copper giant and one of the world’s largest producers, is fast-tracking automation at its El Teniente mine after a deadly collapse in July 2025 killed six workers. The company is in talks with technology firms, equipment suppliers, and unions to accelerate efforts to remove workers from high-risk areas. In November 2025, Codelco signed a strategic alliance with NTT DATA covering 5G/6G connectivity, generative AI, autonomous control of robots, digital twins, and quantum computing. The goal is not just optimization but life-saving: getting human miners out of the most dangerous zones and letting autonomous trucks, drills, and monitoring agents take over.
This case matters because Codelco is state-owned and has contributed over USD 158 billion to the Chilean state since 1971. Automation decisions are matters of national economic interest, not just corporate strategy. The real-time union negotiations force difficult questions: What does ‘removing workers from high-risk areas’ mean for employment? Who decides when an AI system is safe enough to operate without human oversight underground? If an autonomous agent causes injury or environmental damage, who is liable? For labour advocates, Codelco is a live case study in how agentic AI intersects with worker safety, national sovereignty, and governance of critical extractive industries.
Xiaomi has been publicly promoting its ‘black-light factory‘ concept for smartphones and consumer electronics. This refers to plants that can run almost entirely in the dark because so much is automated. In these facilities, AI orchestrates component inventories, pick-and-place robots, and quality checks. Agentic architectures allow the system to balance cost, speed, and defect rates dynamically, especially when supply chains are disrupted by geopolitics or natural disasters.
These facilities force uncomfortable questions about workers’ bargaining power when production lines can be reconfigured overnight by software, about how regulators can audit decisions made by a swarm of agents that collectively reduced headcount without an explicit ‘lay off workers’ instruction, and about which jurisdiction has authority when agents coordinate actions across plants in different countries. Xiaomi is not just a tech brand, but a laboratory where tomorrow’s transnational labour and competition disputes will be tested in real time.
The global picture is far from uniform. Mexico has seen a 23% increase in industrial robot imports between 2017 and 2022, with 5,832 robots installed in 2023 (70% in automotive) driven by supply chain regionalization. In South Africa, AI-powered safety systems in several mines have achieved a 40% reduction in safety incidents; Enaex Africa and Stellenbosch University are developing robotics for deep underground mining where agentic systems could mean the difference between life and death.
India is emerging as a developer, not just an adopter. Addverb Technologies, backed by Reliance Industries, has built Bot-Valley, one of the world’s largest mobile robot factories. India leads globally with 48.6% of enterprises naming agentic AI as a primary future focus. This uneven map reveals that agentic AI in industry is shaped by local labour markets, regulatory environments, and geopolitical pressures. A diplomat who understands only the Siemens-NVIDIA narrative will miss the union negotiations in Santiago, the safety imperatives in Johannesburg, and the sovereignty plays in Noida.
Energy systems are perhaps the most politically sensitive arena for agentic AI. As renewables grow, grids become harder to balance: solar and wind fluctuate; demand spikes during heatwaves or cold snaps; and cross-border interconnectors turn local imbalances into regional problems.
Agentic AI is more used to predict demand at fine-grained intervals, redispatch generation in minutes instead of hours, pre-position storage and flexible loads, and detect anomalies that could indicate cyberattacks or physical faults. A team of agents might monitor transformer temperatures, weather forecasts, real-time prices, and grid stability indices, then coordinate responses that no single operator could track alone.
What’s actually happening in response? Germany’s Federal Network Agency now requires grid operators using AI-driven dispatch systems to maintain detailed logs that regulators can audit. The International Energy Agency’s 2025 working group on grid digitalization has proposed minimum standards for human oversight in cross-border power exchanges. Pilot programs in Denmark and Uruguay are testing ‘guardrail architectures’, i.e. hard limits coded into agentic systems that prevent them from taking certain actions (like disconnecting hospitals) without human approval, even in emergencies.
The implications are geopolitical: cross-border power exchanges may soon be decided by interacting AI agents under different regulatory regimes, raising questions about whose rules prevail. Cyber incidents could be amplified or mitigated depending on how resilient and transparent these agents are. Energy diplomacy will have to include the invisible algorithms that increasingly arbitrate who gets power, when, and at what price, a shift that smaller countries see as a new source of inequality.
All of this can sound like a purely technical story; Siemens, NVIDIA, Codelco, Xiaomi, grid operators, and obscure vendors arguing about optimization algorithms. But, agentic AI in industry forces exactly the kind of humanistic questions that HumAInism.ai, a Diplo initiative exploring human-centered approaches to artificial intelligence, was created to ask. It also echoes concerns raised in Diplo’s earlier piece on the fading of human agency in automated systems, where humans remain formally responsible for decisions that are, in practice, increasingly shaped by opaque, automated processes.
Several trends are converging: UN and UNESCO frameworks emphasize that AI should augment human capabilities, not replace human dignity, and that tools threatening fundamental rights should be restricted or banned. Industrial AI practitioners talk increasingly about ‘responsible autonomy‘: clear limits on what agents can do alone, audit trails, and escalation paths to humans. Governments and regulators look to EU-style rules that require explainability, safety assessments, and specific human oversight for high-risk systems, including those in critical infrastructure.
How this looks in practice: The European AI Act, which came into force in 2024, classifies many industrial AI systems as ‘high-risk’, requiring conformity assessments, human oversight mechanisms, and transparency documentation before deployment. Companies like Siemens are now publishing ‘model cards’ for their industrial AI agents. These are structured reports explaining what the system does, what data it uses, what its limitations are, and where humans must intervene. Union agreements at Codelco and other mining operations increasingly include clauses requiring advance notice and consultation before new autonomous systems are deployed, effectively treating AI adoption as a collective bargaining issue, not just a technical upgrade.
For diplomats and policy-makers, the question is not whether to allow agentic AI in industry; it is already here. The real questions, posed earlier, now look less abstract and more concrete:
A new hybrid role is emerging: the industrial diplomat, someone who understands supply chains, energy systems, and software, but also international law, worker rights, and ethics. This figure could sit in trade ministries, standardization bodies, or multilateral organizations, ensuring that the invisible agents running our factories and grids serve the public interest.
Agentic AI will never show up on a ballot, but it is already affecting jobs, emissions, trade balances, and regional stability. If the 20th century required diplomats who understood oil and shipping lanes, the mid-2020s are creating a need for diplomats who can read not just treaties, but logs and model cards.
The infrastructure is being built now, in simulation environments in Germany, copper mines in Chile, smartphone factories in China, grid control rooms in Denmark. The governance frameworks are lagging, but not absent. The question is whether we shape those frameworks with foresight and inclusion, or discover their contours only after the first major failure. The agents are already negotiating. The humans need to catch up.
Author: Slobodan Kovrlija
