HAMBURG, Aug. 5, 2026 — The factory floor is no longer just connected. It is starting to think for itself.
A new mid-year assessment from IoT Analytics, the Hamburg-based research firm that tracks enterprise IoT spending and adoption across more than 1,000 corporate partners worldwide, argues that 2026 marks the start of the "agentic and autonomous operations era" for industrial technology. The firm's analysts, who had 12 people on the ground at Hannover Messe in March, counted 29 industrial agentic AI solutions on display — and 72 percent of them were already commercially available, not pilot projects or proof-of-concept demos.
The shift is structural and it is happening fast. For years, industrial software vendors layered generative AI copilots onto existing platforms as free add-ons to drive adoption. That pilot phase is ending. Siemens now charges €2,100 per user per year for its Eigen Engineering Agent. Microsoft has moved business applications from per-seat licensing to a "seats plus consumption" model, with nearly 60 percent of service customers already buying usage-based credits. Oracle lets customers purchase "bundles of tokens" and is testing outcome-based pricing tied directly to the value delivered.
"AI was the number-one topic for CEOs as we entered 2026," said Knud Lasse Lueth, CEO of IoT Analytics, in the firm's mid-year pulse check published Aug. 3. "Discussions around agentic AI, AI agents, and physical AI continued to climb. Cisco's Chuck Robbins called 2026 a turning point and the year of agentic applications."
From advisory to orchestration
The Hannover Messe comparison tells the story in concrete terms. In 2025, IoT Analytics analysts noted that agentic AI demonstrations were largely simple automation — chat interfaces that could answer maintenance questions or generate reports. In 2026, the same team identified 29 solutions, 78 percent of them model-agnostic, with maintenance and troubleshooting standing out as the leading agentic AI use case at the fair.
One showcase combined Norway-based Cognite's Atlas AI agent, backed by its industrial knowledge graph, with US-based Tulip's frontline operations agent and AWS infrastructure. The Cognite agent identifies an issue using industrial context — equipment history, process data, maintenance records — then passes rich contextual data to the Tulip agent, which translates it into a real-time, actionable workflow for a technician on the shop floor. IoT Analytics ranked this as "agentic AI level 3" for its multi-agent coordination capability.
At Maintenance Dortmund in May, IFS Ultimo demonstrated an agentic module that independently spots safety-related content in a technician's routine report and automatically generates a compliant safety incident record — no human intervention required.
"One case we have available in the software is based on safety incidents," an IFS Ultimo representative said. "When a technician files a report, the AI agent identifies if there is a potential safety incident as part of that registration. It then raises that incident automatically instead of needing to rely on the technician to do so. Another feature being released shortly is for the maintenance manager, who will receive a daily maintenance briefing directly on his phone."
Physical AI closes the loop
Agentic AI orchestrates decisions. Physical AI puts those decisions into motion. At Hannover, Germany's Beckhoff Automation demonstrated TwinCAT CoAgent, which uses large language models connected via the Model Context Protocol to directly control real machine motion sequences within the TwinCAT automation platform. An engineer types a command; CoAgent translates it into machine commands, orchestrates path planning, generates function blocks, and performs error diagnostics.
The promise: operators set goals — "clear this field," "optimize this harvest" — and the machine handles the cognitive load of navigating rugged, unpredictable environments. This reduces the skill gap that has long constrained industrial automation adoption, allowing less-experienced workers to operate with veteran-level proficiency.
Digital twins become execution engines
The digital twin — long a simulation and visualization tool — is becoming a real-time control layer. At Hannover, Dell Technologies showed an XMPro-NVIDIA Omniverse demo where live PLC data from a brewery centrifuge digital twin at New Belgium Brewing in Fort Collins, Colorado, fed into an LLM to detect boundary violations and trigger SCADA-level adjustments within human-defined limits.
"With agentic AI, once you have identified something that is exceeding a boundary, you may allow AI to make very small changes to your production environment," a Dell representative said at the fair. "For example, you may allow Omniverse to adjust a parameter by only 10 percent. Anything more than 10 percent would require a human to be involved. Omniverse can then connect back to the SCADA system to make those small changes — nothing big, only small changes — and notify an operator that it made the change on their behalf."
PepsiCo uses a similar approach through Siemens' Digital Twin Composer, creating physics-accurate digital twins of US manufacturing and warehouse facilities — every machine, conveyor, pallet route, and operator path. AI agents continuously monitor live shop-floor data, flag anomalies, trace them to specific physical assets, and recommend corrective actions. The company reports identifying up to 90 percent of potential issues before committing any physical capital expenditures or changes.
Proprietary models replace generic ones
Industrial adopters piloting frontier models hit a wall: outputs lack the determinism, physics-based reasoning, and domain grounding that plant-floor use cases require. In response, vendors are training proprietary industrial foundation models.
Siemens' Industrial Foundation Model, developed with Microsoft, arrived in 2025. The company committed over €1 billion to industrial AI through its One Tech Company program. US-based SymphonyAI expanded its Iris Foundry platform in April with eight industrial applications built for energy operators using industry-specific failure nodes, process dynamics, and regulatory obligations. Its model-agnostic architecture enables multi-agent systems where different agents use different underlying models for predictive maintenance, anomaly detection, and process optimization.
Pricing models reshape the market
The commercialization wave is also reshaping how industrial software gets priced. Microsoft's shift to "seats plus consumption" reflects a broader pattern: customers want predictable costs but also want to pay for actual usage. Oracle's token bundles and outcome-based experiments point the same direction. Siemens' fixed €2,100 per user per year subscription for Eigen Engineering Agent was driven by customer demand for predictable costs after years of free pilots.
This matters because pricing is often the last barrier to scale. When vendors give away copilots, adoption looks high but revenue is zero. When they attach a price, the market reveals which capabilities customers actually value enough to pay for. The early data suggests agentic AI for maintenance, troubleshooting, and safety incident handling clears that bar.
Constraints remain
The IoT Analytics assessment is not unqualified optimism. Data foundation deficits remain a major bottleneck to scaling agentic AI, according to the firm's Industry 4.0 & Smart Manufacturing Market Report 2026–2030. Non-deterministic LLM behavior challenges adoption in high-precision, highly regulated industries. Human-in-the-loop remains a core requirement — a bad decision can stop production, damage equipment, and create audit issues. Multi-agent orchestration is still rare; most agents operate inside a single defined workflow.
Yet the market signal is clear. Aside from robotics, IoT Analytics believes agentic AI platforms represent the largest new market opportunity in smart manufacturing for the remainder of 2026 and beyond. The conversation has shifted from "can we do this?" to "how do we scale, govern, and charge for it?"
For the IoT sector, the implication is direct: the sensors, gateways, and connectivity layers that IoT vendors have spent a decade deploying are finally becoming the nervous system for autonomous decision-making. The hardware is in place. The intelligence layer is arriving.
The full IoT Analytics mid-2026 pulse check is available on the firm's website.
According to the State of Enterprise IoT 2026 report published in January, many enterprises have grown beyond early tech maturity phases like connectivity, platforms, and scaling and are entering the upper end of the IoT Value–Maturity Curve, where AI and stronger ecosystems enable connected operations. Agentic AI is the next big level starting this year.
Read more on industrial automation in our Robotics and Drones and Semiconductors sections.