On June 18, 2026, Microsoft laid out a blueprint for bringing agentic AI to process manufacturing floors—and it looks nothing like a chatbot. During a technical showcase anchored at Yara International’s Porsgrunn fertilizer plant in Norway, the company demonstrated a governed, human-supervised AI layer designed to assist operators with complex decisions, not replace them with autonomous agents. The setup, built on Kongsberg Digital’s Azu platform, underscores a deliberate pivot away from conversational interfaces toward closed-loop advisory systems that keep people firmly in control.
Microsoft’s pitch arrives as heavy industry confronts a deluge of sensor data, aging workforces, and rising sustainability mandates. Process manufacturing—chemicals, oil and gas, pulp and paper—has been notoriously slow to adopt AI beyond predictive maintenance. The new approach positions agentic AI as a decision-support engine that ingests real-time streams from digital twins, simulates multiple what-if scenarios, and recommends ranked actions, all while logging every step for audit and regulatory compliance.
What Agentic AI Means on the Plant Floor
In Microsoft’s framing, agentic AI is not a single model but a constellation of specialized agents that coordinate under human supervision. At Porsgrunn, these agents monitor ammonia synthesis loops, steam reformation temperatures, and catalyst health in parallel. When a disturbance—such as a feed composition upset or a utility fluctuation—ripples through the process, the agents collectively reason about the best mitigation, presenting operators with an explainable recommendation rather than executing it automatically.
“Think of it as a copilot that has deep domain expertise, not a chat window,” a Microsoft engineer told attendees at the closed-door event. The system runs on Azure infrastructure and hooks into Kongsberg Digital’s Azu, a real-time digital twin environment that mirrors the physical plant with sub-second latency. That tight coupling lets agents run forward simulations across minutes or hours in milliseconds, giving operators a preview of how the process will behave if they follow a suggested action.
Inside the Yara Porsgrunn Pilot
Yara’s Porsgrunn facility is an ideal testbed. The 100-year-old site produces roughly 500,000 metric tons of nitrate-based fertilizers annually, making it one of Europe’s largest. Its process units are instrumented with thousands of IoT sensors, feeding pressure, flow, temperature, and spectroscopic data into Azu. Microsoft’s agentic layer taps that data to build contextual awareness—learning normal operating envelopes, spotting nascent anomalies, and proposing preemptive adjustments.
In one demonstration, the agents detected a slow degradation in the primary reformer’s heat-transfer efficiency—a subtle shift that would typically take shift engineers hours to triangulate. The system cross-referenced the reformer’s digital twin with weather data, utility pricing, and ammonia inventory levels before suggesting a minor load reduction and an earlier than planned catalyst regeneration window. The recommendation was presented alongside its confidence score, downstream impact on urea production, and estimated CO₂ savings.
Yara’s operations manager, speaking on background, described the tool as “an additional expert in the room, one that never gets tired.” The company emphasized that no control actions were taken without explicit operator approval, aligning with its stringent safety management systems. Microsoft and Yara are evaluating a phased rollout to additional units at Porsgrunn throughout 2027.
Governed Decision Support, Not a Chatbot
The distinction between agentic AI and conventional generative AI drew sharp focus at the event. While chatbots excel at retrieving procedures or summarizing documentation, they lack the real-time physical context and closed-loop safety constraints demanded by process industries. Microsoft’s agentic layer operates on a “propose but not dispose” principle: it can suggest, but every actuator command must be ratified by a human.
A governance framework underpins this approach. Each agent’s reasoning chain is immutable, timestamped, and stored for audit. If an operator overrides a recommendation, the override is logged with a rationale field. Over time, this creates a traceable corpus that site leaders can review during management-of-change discussions or regulatory inspections. “This isn’t a black box; it’s a glass box,” the engineering lead noted.
Microsoft’s Azure AI Foundry provides the orchestration backbone, enforcing role-based access controls and aligning with IEC 62443 cybersecurity standards. The company is also working with Kongsberg Digital to integrate the agentic layer into Azu’s existing process historian and alarm management workflows, ensuring it slots into control-room culture rather than disrupting it.
The Digital Twin Connection
Kongsberg Digital’s Azu platform is the silent enabler. Azu, which builds on years of maritime and energy expertise, creates hyperaccurate digital replicas of complex facilities. For Porsgrunn, the twin includes every heat exchanger, reactor, and piping run, calibrated against both design specs and live operational data. Microsoft’s agents treat this twin as a sandbox, testing thousands of permutations before forming a recommendation.
This marriage of agentic AI and digital twins could become a template for the broader process industry. Microsoft hinted at plans to publish reference architectures and joint solution accelerators with Kongsberg Digital, targeting chemical parks, refineries, and LNG terminals. The go-to-market strategy leans on system integrators who already serve these sectors, rather than a direct sales push.
Benefits Beyond Productivity
Beyond keeping plants running, the agentic approach addresses three pressures gripping manufacturers: decarbonization, workforce transition, and regulatory compliance. By optimizing setpoints in real time, the system can shave several percent off energy consumption—a meaningful figure for plants that burn millions of gigajoules monthly. At Porsgrunn, initial simulations suggested a 3–4% reduction in natural gas use during ammonia production, translating to annual savings of up to €8 million and 30,000 tonnes of CO₂.
For workforce transition, the agents serve as a knowledge-capture mechanism. As veteran operators retire, their corrective actions—paired with the AI’s reasoning—create a learning library that less experienced staff can query. Microsoft is exploring a “what learned” module that auto-generates training scenarios from historical incidents.
Regulatory compliance benefits from the glass-box architecture. The system’s detailed logs can accelerate root-cause analyses and provide auditable evidence for ISO 50001 energy management or SEVESO III safety directives, reducing the administrative burden on plant leadership.
Challenges and Guardrails
No deployment of this ambition is without hurdles. The first is infrastructure: agentic AI demands low-latency edge compute and robust connectivity. At Porsgrunn, Microsoft addressed this with Azure Stack Edge devices placed on-premises, but not all plants have the same IT maturity. Kongsberg Digital’s Azu is also a proprietary environment, raising questions about interoperability with other digital-twin vendors.
Trust remains the larger obstacle. Operators accustomed to making gut-level calls may resist a system that recommends counterintuitive actions. Microsoft’s human-centric design—explainable recommendations, confidence scores, and a clear override mechanism—aims to earn that trust gradually. Early feedback from Porsgrunn operators was cautiously positive, though some asked for tighter integration with the Distributed Control System (DCS) to reduce the number of screens they monitor.
Safety integrity is non-negotiable. The agents currently run at a lower Safety Integrity Level (SIL) than the DCS and Emergency Shutdown systems, meaning they cannot directly trip alarms or close valves. Microsoft acknowledged that for high-SIL environments, additional certification under IEC 61508 would be required before agents could assume advisory roles closer to the control boundary.
Industry Reaction and Broader Strategy
The Porsgrunn showcase generated buzz across the industrial AI community. Analysts note that while startups have pitched autonomous plant operations, Microsoft’s emphasis on governance and partnership with a tier-one digital-twin provider sets it apart. “They’re not trying to beat Emerson or Honeywell at DCS; they’re building a cognitive layer above it,” said one control-system integration specialist who viewed the demos remotely.
For Microsoft, agentic AI for process manufacturing is an extension of its enterprise AI stack—Copilot, Azure AI Foundry, and Azure Digital Twins—into a domain-vertical play. By working with Kongsberg Digital, it gains domain credibility without having to build industry-specific IP from scratch. The Yara pilot also aligns with Microsoft’s carbon-negative goals, as fertilizer production is among the hardest sectors to decarbonize.
Competitors are watching closely. AWS has its own industrial IoT and twin services, while Google Cloud partners with Seeq and Cognite for similar workflows. Microsoft’s early lead in governed agentic AI, however, could give it an edge in regulated sectors where traceability is paramount.
What’s Next
Microsoft and Kongsberg Digital plan to release a technical white paper detailing the Porsgrunn architecture by September 2026. A private preview for select chemical and energy companies is slated for Q4 2026, with general availability through the Azure Marketplace expected in mid-2027. Pricing will likely follow a throughput-based model aligned with the volume of agent interactions.
As the pilot expands, the two firms are exploring additional agent specializations—environmental compliance, supply-chain optimization, and even autonomous inspection via drones. The long-term vision is a suite of governed agents that collaborate across the entire value chain, from feedstock sourcing to finished product logistics, all orchestrated under human command.
The Porsgrunn experiment demonstrates that agentic AI for heavy industry is viable today, but only when wrapped in a rigorous governance framework. For plant managers weary of AI hype, Microsoft’s message is clear: this is not a chatbot that hallucinates a maintenance procedure; it is a disciplined, auditable decision-support system designed to make the safest and most sustainable choice—every time, with a human in the loop.