The new era of intelligent supply chains according to Microsoft

Artificial intelligence is reshaping how companies manage their logistics operations, and Microsoft decided to go beyond theory.

One year after publishing its first mapping of AI applied to the supply chain, the Redmond giant came back with a full update — and the landscape has changed quite a bit.

We are no longer talking about basic automation or smart dashboards. The agentic era has fully arrived, with AI agents capable of reasoning, planning, and executing complex tasks end to end, with less and less human intervention.

And the most interesting part? Microsoft is not just selling the idea to customers. It is living the experiment firsthand, using its global infrastructure as a living lab to test all of this at real scale.

In this article, you will understand what has changed, what is being built right now, and where the intelligent supply chain is heading in the coming years. 🚀

What changed in one year in Microsoft’s vision

When Microsoft published its first map of artificial intelligence applied to the supply chain, the focus was on showing where AI could help: demand forecasting, inventory management, AI-powered customer service, route optimization. The original document featured two reference architectures for logistics, involving adaptive cloud and AI-enhanced experiences, along with innovations in Microsoft Dynamics 365. It was an honest snapshot of what was available and what was already working in production.

But in technology, one year is enough time to completely flip the script, and that is exactly what happened. The recently released update is not a cosmetic revision of the previous document. It is practically a new manifesto on how supply chain operations will work going forward.

The main turning point is the consolidation of AI agents as the protagonists of the process. Before, language models and predictive systems worked as tools to support human decision-making: they suggested, alerted, recommended. Now, the approach is different. Agents do not just recommend — they execute. They query systems, make decisions within defined parameters, trigger actions in other systems, and report results, all autonomously and in sequence. Platforms like Microsoft Foundry for end-to-end agent hosting and open protocols like the Model Context Protocol (MCP) have made it easier to connect agents with enterprise systems, tools, and data. This represents a structural shift in how companies will design their operational workflows.

Beyond agents, there have been significant advances in 3D simulations, robotics, and embedded intelligence. Models like NVIDIA Cosmos and the OSMO edge-to-cloud computing framework on Azure allow machines and humanoid robots to operate more effectively in the physical world, expanding automation across warehouses, distribution centers, and transportation.

Microsoft Azure appears as the backbone of this entire architecture. It is on Azure infrastructure that models run, data is processed in real time, and agents communicate with each other using protocols like MCP and A2A (Agent-to-Agent). It is no exaggeration to say that Azure is positioning itself as the invisible operating system of the new intelligent supply chain, connecting everything from the factory floor to long-term strategic planning.

Microsoft testing on its own operations

One of the most credible aspects of this entire story is that Microsoft is not just describing a possible future for customers — it is running these experiments on its own global operations. The company operates one of the most comprehensive cloud supply chains in the world, spanning more than 70 Azure regions, over 400 datacenters, and a network of more than 600,000 km of fiber. These datacenters are the backbone of Microsoft Azure, powering everything from AI infrastructure and collaboration tools to networking and security. Microsoft also manages supply chains for Surface hardware, PC accessories, and Xbox consoles.

All of these supply chains have undergone a fundamental transformation over the past decade, evolving from a reactive, manual environment to a rapidly emerging autonomous and agentic supply chain. In the past, operations were dominated by Excel-based reports, limited visibility, and siloed data. In 2018, the company consolidated more than 30 systems into a single supply chain data lake on Azure, enabling predictive analytics and the first generation of cognitive capabilities. In 2022, it began experimenting with generative AI, followed by the development of an AI platform to operationalize agents at scale.

Today, that foundation is accelerating toward fully autonomous agents, and more than 25 AI agents and applications have already been deployed. Here are three concrete examples:

• The Demand Planning Agent runs AI-driven demand simulations for non-IT rack components, improving forecast accuracy and reducing manual reconciliation.
• The Multi-Agent DC Spare-Part Space Solver uses computer vision monitoring and multi-agent reasoning to predict spare parts storage needs and proactively mitigate space or stockout risks.
• The CargoPilot Agent continuously analyzes transportation modes, routes, cost structures, carbon impact, and cycle times, providing optimized shipping recommendations that balance speed, sustainability, and efficiency.

The goal is to operate more than 100 agents by the end of 2026 and equip every employee with agentic support. The current impact is already huge: AI in logistics is saving hundreds of hours per month for teams, demonstrating how agentic operations translate directly into efficiency and business value.

This choice to use its own operations as a lab has a very positive side effect: the learnings are real, the problems encountered are genuine, and the solutions that emerge have been validated under production conditions, not in controlled environments. When Microsoft says a particular agent reduced response time to disruptions, that data comes from a real operation.

Both in its internal transformation and with leading-edge customers, Microsoft has identified that unifying the data estate is fundamental. But it is what organizations do after that truly generates value with AI. In the supply chain, real value is unlocked by driving three elements: enabling AI-powered supply chain simulations, building agentic supply chains, and integrating the first physical AI innovations.

Simulations: the digital twins of the supply chain

One of the most fascinating elements of the new mapping is the growing emphasis on simulations as a central part of the intelligent supply chain. As supply chains become larger, more interconnected, and more exposed to global volatility, simulating scenarios before they unfold is becoming a critical capability for reducing risk and increasing resilience.

Discrete event simulations (DES) within supply chains enable the development of a risk-free virtual model to test how a complex system reacts to interventions and variables before implementation. With advanced modeling tools from Microsoft, such as Azure Machine Learning and the new machine learning model in Microsoft Fabric with Power BI semantic models, logistics organizations can simulate demand patterns, shortages, or supply chain disruptions.

Partners building on this foundation

paiqo, with its prognotix platform, offers a forecasting platform with more than 70 algorithms that enable generating and optimizing highly accurate demand forecasts directly within the Azure environment. Cosmo Tech provides an AI simulation platform for advanced supply chain risk management, offering dynamic digital twins that simulate how disruptions and decisions impact performance across the entire system. InstaDeep uses Azure for high-performance computing in deep reinforcement learning and predictive analytics that optimize last-mile deliveries, inventory levels, and fleet utilization.

The next level: 3D digital twins

The next level of simulation combines multiple physical simulations in 3D environments and discrete event-based simulations to enable teams to build comprehensive digital twins of warehouses, distribution centers, production lines, and logistics networks. These virtual environments allow modeling both the physical behavior of assets and the dynamic flow of operations. By integrating these simulation methods within a digital twin and applying AI, teams can predict future outcomes, optimize performance, and prescribe actions that drive continuous operational improvements. This can help reduce capital expenditures, shorten commissioning and ramp-up phases, and improve operational key performance indicators (KPIs).

Taking warehouses as an example, customers and partners can build advanced AI-enabled 3D visualizations for four main scenarios:

• Warehouse planning — such as greenfield and brownfield designs.
• Warehouse monitoring — such as real-time monitoring and people movement heatmaps.
• Warehouse improvement — for example, trailer dwell time optimization and collision detection for safety and automation.
• Warehouse maintenance — such as real-time asset monitoring, quality issue detection, and rework reduction.

In collaboration with NVIDIA, Microsoft offers access to libraries and frameworks including NVIDIA Omniverse, NVIDIA Isaac Sim, and NVIDIA Omniverse Kit App Streaming, which allow developers to build applications and workflows to simulate and test scenarios in digital twins before building or deploying anything in the real world.

Reference architecture for digital warehouses

The reference architecture presented by Microsoft illustrates how to combine cloud and edge computing using NVIDIA Omniverse Kit App Streaming to visualize warehouse operations in real time with GPU-accelerated Kubernetes clusters deployed natively on Azure.

Inside the physical warehouse, operational data from robotic arms, conveyors, and sensors is captured at the edge using Azure IoT Operations running on Arc-enabled Kubernetes with an MQTT broker. The architecture adopts the Universal Scene Description (OpenUSD) format to ensure that 2D, 3D geometry and point clouds can be integrated into the digital twin. Microsoft Fabric ingests data into the cloud to provide a unified analytics foundation. Microsoft OneLake serves as the centralized, governed data lake. The Digital Twin Builder transforms raw IoT signals into a contextualized virtual representation. Tools like Microsoft Copilot Studio and Microsoft Foundry enable natural language interaction. Security is maintained at every stage through Azure Arc.

Real-world implementation cases

SoftServe has proven to be an excellent delivery partner for digital twin applications. Together with Microsoft, it integrated NVIDIA Omniverse-based environments into Krones beverage production simulations, enabling physically accurate digital twins that reduced cycle times from hours to under five minutes. At Toyota Material Handling Europe, SoftServe built a digital twin to simulate autonomous forklifts in virtual warehouse environments, helping reduce training times for autonomous systems by more than 30%.

TeamViewer‘s Frontline augmented reality platform offers an additional simulation angle. Wearable devices like smart glasses deliver data directly to frontline workers for hands-free guidance in picking, packing, and AI-assisted counting. At DHL Supply Chain, TeamViewer’s solution is deployed globally to support visual picking for more than 1,500 workers across 25 sites in the United States.

Agentic supply chains: the multi-agent web

The word agents has become the center of the entire conversation about the future of the supply chain, but it is important to understand what this means in practice. An AI agent, in the context of what Microsoft is building, is a system capable of perceiving the environment it operates in, reasoning about available information, planning a sequence of actions, and executing them autonomously. It is not a glorified chatbot. It is closer to a digital coworker that has access to systems, tools, and data, and knows how to use them to achieve a goal.

Agentic supply chains mark a new era of autonomous AI systems that proactively manage and optimize end-to-end operations. These systems aim to continuously improve comprehensive KPIs like operating margin or cash conversion, as well as specific KPIs like lead time or freight cost per unit, ensuring that every agentic action contributes to measurable business impact.

These supply chains are built on tasks currently performed by humans and encode the underlying decision-making logic. They include single-purpose agents like problem solvers that continuously diagnose issues and propose fixes, as well as orchestrator agents like planners or organizers that coordinate multi-step workflows.

Leading companies already generating value

Several companies are already seeing concrete results with multi-agent systems:

• CSX Transportation deployed a multi-agent system that validates customer eligibility, routes complex requests, and supports railroad operations with multi-stage coordination.
• Dow Chemical operates invoice analysis agents that review thousands of freight invoices per day, automatically detecting discrepancies and saving millions across its global shipping network.
• C.H. Robinson implemented a large fleet of generative AI agents, including quick-quote agents that deliver personalized freight quotes and automate key steps throughout the shipping cycle.
• Blue Yonder created an Inventory Ops Agent on the Microsoft Marketplace that identifies supply-demand mismatches in real time and recommends corrective actions.
• Resilinc offers an agentic supplier risk platform on Azure with pre-built agents for disruptions, tariffs, and compliance.
• o9 Solutions‘ Digital Brain platform on Azure has been enhanced with multiple AI agents that handle everything from simple tasks to full demand reviews.
• GEP added to its GEP SMART and GEP NEXXE solutions, both built natively on Azure, a portfolio of AI agents covering sourcing, negotiation, contract lifecycle, spend analysis, and market intelligence.
• Kinaxis offers its Maestro supply chain planning platform with AI agents that detect disruptions, run scenario simulations, and provide prescriptive insights through natural language.

Delivery partners accelerating implementations

Several delivery partners have also been using Microsoft tools like Foundry and Copilot Studio to build agents at high speed:

• Avanade offers ready-to-use agents, including a trade compliance agent, a digital twin agent, and lead-to-cash agents.
• SoftServe offers a catalyst with 3 custom agents in 30 days.
• NTT DATA is developing a decision support system based on simulation and agentic AI for supply chain network rebalancing.
• PwC delivers end-to-end agentic AI consulting services.
• Capgemini is building an end-to-end agentic offering using Microsoft IQ technology, with a planned launch at Hannover Messe in April 2026.

The IQ intelligence layer

Microsoft’s Work IQ, Foundry IQ, and Fabric IQ together form an intelligence layer for supply chains — from demand planning to inventory and customer service — connecting how people work, how the business operates, and what the organization knows. This gives AI agents full business context so they can reason, simulate scenarios, and act aligned with real-world constraints and KPIs like inventory turnover.

Reference architecture with Celonis

In a strategic partnership with Celonis, Microsoft developed a new reference architecture leveraging Fabric IQ and the Celonis Process Intelligence Graph to transform fragmented supply chain data into agentic workflows.

At the system of record layer, data is often siloed and does not speak the same language. Microsoft Fabric unifies this data through mirroring, streaming, or multi-cloud shortcuts, with the goal of creating a zero-copy connection and ensuring data is fresh and accessible without the overhead of traditional ETL processes. Fabric IQ provides a reasoning layer that translates raw unified data in OneLake into contextual insights. The Celonis Process Intelligence Graph sits between data and automation, using process mining to map how the supply chain actually works.

The agentic layer breaks down into three functions:

• Build — using Copilot Studio, Microsoft Foundry, and Power Automate to create custom AI agents.
• Orchestrate — using MCP and the Agent2Agent (A2A) protocol to manage how different agents work together.
• Govern — using the control plane for agents in Agent 365 (general availability expected May 2026) to monitor agent activities.

At the top layer, with the help of Microsoft Entra ID, insights and suggested actions are surfaced in the tools employees already use, such as Microsoft Teams, Microsoft 365 Copilot, Dynamics 365, Power Apps, or the Celonis interface.

Tools we use daily

Translation

Text Inspection & Clipping

Social Media

Search & Research

Recruitment

Programming

Productivity & Organization

Operations

Marketing & Sales

Finance & Investment

Education

Design & Media

Data & Analytics

Content & Writing

Automation

A major global pharmaceutical company is using this architecture to unify fragmented logistics data, enabling real-time identification of temperature-critical pharmaceutical returns and designing an agentic returns process that unlocks annual productivity gains in the millions of euros. Uniper automated material and service needs with Celonis and Microsoft, orchestrating approvals and SAP actions, and replacing manual component planning with proactive, agentic workflows.

Physical AI: from warehouse handling to last-mile delivery

Physical AI is the ultimate evolution of intelligence in the supply chain, building on simulations and agentic AI and embedding that intelligence directly into the physical world. In the near future, humanoid robots and robotic systems will physically take on an increasing number of operational tasks across supply chains and logistics: from trailer unloading and sorting, pallet handling and replenishment, to packing, labeling, and autonomous last-mile deliveries. 🤖

Microsoft is pushing the frontier of physical AI with its new Rho-alpha robotic model that combines natural language, visual perception, and tactile feedback to make robots more adaptive and autonomous. The company launched an early access research program with selected partners to advance co-training and domain adaptation, with plans to integrate the model into Microsoft Foundry in the coming months.

The reference robotics toolchain

Already today, customers and partners can use the open-source robotics toolchain reference architecture to train and deploy warehouse robotics with NVIDIA Osmo on Azure. This toolchain is a production framework that integrates Azure cloud services with NVIDIA’s physical AI stack — from simulation to training and deployment. It combines Azure Machine Learning, Azure Kubernetes Services (AKS), Microsoft Fabric, Azure Arc, and NVIDIA’s robotics and AI stack. NVIDIA Isaac Sim and Isaac Lab enable high-fidelity simulation and reinforcement learning, while NVIDIA OSMO orchestrates scalable training workflows between cloud and edge environments.

Humanoid robots and automation in action

Hexagon Robotics has begun deploying this architecture using Azure IoT Operations and Fabric Real-Time Intelligence to deliver production-ready humanoid robotic solutions. Its industrial humanoid robot, AEON, combines dexterity, locomotion, and unique spatial intelligence to tackle complex industrial use cases in warehousing and logistics, such as inspection and inventory.

Figure AI, backed by Microsoft, enables the deployment of its humanoid robots in real logistics environments using Azure’s AI infrastructure. Its latest model, Figure 03, can take on warehouse tasks like sorting packages at conveyor-belt speed and assisting in last-mile delivery with near-human-level precision.

KUKA and Microsoft jointly developed iiQWorks.Copilot, an AI-powered assistant that enables robot programming through natural language and significantly simplifies automation tasks. The solution allows users to design, test, and deploy robot workflows faster and more safely, reducing programming time for simple tasks by up to 80%.

Wandelbots‘ NOVA software layer, combined with Azure cloud services, unifies heterogeneous robots and brings adaptive automation to the operations floor. Wandelbots NOVA streamlines warehouse and fulfillment operations like palletizing by simplifying robot programming, accelerating deployment, and enabling AI-powered path planning across multiple robot brands.

Where all of this is heading

Microsoft’s mapping is not a finished product — it is a snapshot of a process in rapid motion. And what this snapshot reveals is that the intelligent supply chain is converging toward a model where artificial intelligence stops being an additional layer on top of existing systems and becomes the central infrastructure around which everything else is organized. This has profound implications for how companies will structure their teams, processes, and technology investments in the coming years.

Agents will become increasingly sophisticated, with greater long-term reasoning capabilities and better integration with legacy systems that exist in large corporations. Simulations will gain resolution and speed, enabling complex decisions to be made with more confidence and in less time. Physical AI will bring that intelligence into the real world through humanoid robots and autonomous systems that are already being tested in production scenarios. And Microsoft Azure will continue to be the environment where much of this evolution takes place, both for Microsoft and for the partners and customers building on this platform.

What is clear is that companies that start experimenting with this approach now, even on a small scale, will be in a much stronger position when the technology fully matures. Not because there is a race to be won, but because the organizational learning curve is long, and understanding how to integrate AI agents, how to design simulation workflows, how to govern autonomous systems, and how to incorporate physical AI into a real operation takes time. Those who start earlier, learn earlier. And in the world of the supply chain, learning fast can be the difference between leading the market or playing catch-up. 💡