Agentic AI in Manufacturing: Why General-Purpose LLMs Fail

Agentic AI can transform how manufacturers make decisions. But not in the way most people think. The real promise isn't chatbots on the shop floor or copilots summarizing dashboards. It's decision intelligence. AI agents that can investigate a quality failure, trace the root cause across systems, and tell an operator exactly which lever to pull right now for a better outcome. That's the vision. But here's the problem: Most agentic workflows today are built on top of general-purpose LLMs that have no understanding of cause and effect in physical processes. To build agentic workflows that can reliably reason at scale, you need a three-layer architecture: 𝟏. 𝐃𝐚𝐭𝐚 & 𝐈𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐋𝐚𝐲𝐞𝐫 This layer combines streaming data infrastructure (MQTT, Kafka, UNS) with a process knowledge graph that models how your production actually runs. 𝟐. 𝐌𝐨𝐝𝐞𝐥 𝐋𝐚𝐲𝐞𝐫 This is where causal reasoning lives. Purpose-built models that understand time series, respect safety limits, and can trace cause-and-effect chains across your process. 𝟑. 𝐀𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧 𝐋𝐚𝐲𝐞𝐫 This is where the agentic workflows operate. The agent orchestrates by pulling data from the infrastructure layer, calling the right causal model, and reading SOPs and process documentation. Synthesizing everything into a clear, operator-ready recommendation that can be trusted and acted on. When this architecture is in place, agentic AI delivers real decision intelligence. In this episode of the AI in Manufacturing podcast, I speak with Bernhard Kratzwald, Co-Founder and CTO at EthonAI, about building and scaling agentic AI workflows for optimizing cost, quality, and speed in manufacturing operations. Watch/Listen YouTube: https://lnkd.in/dYwfES9g Spotify: https://lnkd.in/dDq2dUXt Apple Podcasts: https://lnkd.in/dgfwGn5P

We’re reaching an interesting shift in AI. New model releases are no longer the main story. The real value is moving to how these models are orchestrated in business workflows. Instead of using a single LLM, companies are building systems where: one model handles coding another focuses on analytics a third processes data On top of that, you have external data (RAG), tools, and business logic. The key challenge now is coordination. This is where management platforms come in. They enable: ▪ selecting the right model for each task ▪ controlling costs at scenario level ▪ orchestrating multi-step workflows across models, data, and tools We also see a clear move toward agent-based architectures, where AI becomes a managed, scalable system embedded into operations. Platforms like MWS GPT (with access to 90+ models) are a good example of this shift — not just using AI, but designing systems around it. The takeaway is simple: 👉 competitive advantage is no longer about having the “best model” 👉 it’s about how effectively you structure and manage AI workflows And this is exactly where the next wave of value will be created.

Most companies are still thinking about AI wrong. They’re deploying models. Not systems. That’s the difference AI orchestration makes — and it’s the gap between AI that impresses in a demo and AI that transforms how work actually gets done. Here’s what orchestration actually means: A single AI model is a capable individual. AI orchestration is a coordinated team — multiple agents, tools, and data sources working together toward a shared goal, with one layer deciding who does what and when. Think of it like a conductor and an orchestra. The conductor doesn’t play every instrument. They ensure every instrument plays at the right moment, in the right sequence, to produce something none of them could alone. In practice, this looks like: → A planning agent that breaks down complex tasks → Specialized sub-agents that execute each step → Memory systems that carry context across sessions → Tool integrations that connect AI to real systems → Guardrails that keep outputs reliable and safe The result? AI that doesn’t just answer questions — it completes workflows, adapts to failures, and scales across your entire operation. We’re moving from the era of the smart chatbot to the era of the autonomous system. Orchestration is how you get there. Are you building with this in mind yet?

The role of AI in pricing decisions There is a growing conversation around AI in business. However, most applications remain superficial. AI is largely being used to generate content, automate repetitive tasks, or improve efficiency at the margins. But rarely is it embedded where it truly matters, in decision-making. This is where the real opportunity lies. Because pricing is not a static process. It is a dynamic system influenced by market volatility, supply constraints, customer behavior, and competitive actions. And yet, in many organizations, pricing decisions are still based on fragmented data, delayed insights, and human bias. This gap is precisely where AI can create disproportionate value. With the right approach, AI can move beyond automation and become a decision-support layer across the entire pricing process: • Detect pricing inconsistencies across regions, customers, and channels • Identify margin leakage at a transactional level before it compounds • Anticipate market movements by analyzing external signals, supply shifts, and demand patterns • Simulate pricing scenarios to evaluate risk, elasticity, and profitability before execution • Support faster, data-driven decisions in environments where timing is critical But technology alone is not the differentiator. The real shift is organizational. Companies that extract value from AI in pricing do three things differently: • They integrate pricing, planning, and commercial data into a single decision framework • They challenge intuition-based decisions with structured insights • They embed AI into workflows, not as a tool, but as part of how decisions are made The competitive advantage, therefore, is not in adopting AI. It is in applying it where it directly influences revenue, margin, and strategic positioning. And in doing so, transforming pricing from a reactive function into a forward-looking capability.

I keep coming across this phrase “harnessing AI.” And every time, I end up thinking… are we really? You see it in strategy decks, roadmaps, even leadership updates. Harness AI in operations Harness AI for customer experience Harness AI for transformation It sounds right. But when you look a bit closer, it’s often not very clear what’s actually happening. To me, harnessing something means taking something powerful and making it work reliably in a real world context. Now apply that to AI. Harnessing AI is not running a few models or experimenting with prompts. That’s part of the journey, but it’s not the destination. When I look at where it actually works, a few things tend to be in place. First, the data has to be reliable. If it’s fragmented or inconsistent, AI will only amplify the problem. Then it needs to sit inside a real workflow. AI creates value only when it becomes part of how work actually gets done. There’s also the orchestration layer. APIs, pipelines, event flows, system interactions This is where most of the complexity actually sits. And finally, there’s governance. Explainability, auditability, security, cost control Without this, AI doesn’t scale in regulated environments. When these pieces start aligning, that’s when AI begins to look harnessed. Until then, most organisations are still wiring the plumbing. And that’s completely fine. Every transformation goes through this phase. But calling it harnessing too early creates a false sense of maturity. Maybe the better question is: Where are we really in the journey from experimenting with AI to actually harnessing it? What are you seeing? Are we truly there yet, or still building the foundations? #ArtificialIntelligence #AI #DataArchitecture #EnterpriseAI #DigitalTransformation #DataStrategy #Innovation #TechLeadership #MachineLearning

Everyone keeps asking: “How do we get more out of AI?” Better models Better prompts Better tools That’s where all the attention is. But that’s not where things break. Things break right after. When the output is good enough to use… and no one is 100% sure what to do with it. I’ve been sitting in real workflows lately where AI is already in the system. Nothing is broken. But everything feels heavier than it should. Decisions that used to take minutes start taking hours More people get pulled in More conversations happen More “just want to double check this” moments Not because the AI is bad. Because it’s good enough to be trusted… and no one defined what happens when it is. Who owns the final call? What does “good enough” actually mean? When does something move forward… vs get escalated? Most teams don’t have clear answers to those questions. So what happens? AI speeds everything up… and decision-making slows everything down. This is the part no one is really talking about. Not the model. Not the tool. Not the workflow. The decision layer underneath it. You can have the best AI in the world. If decisions aren’t clearly defined… you don’t get speed. You get hesitation. Decision architecture involving AI is the difference between: a system that actually moves and one that just looks like it should Change my mind.

The AI challenge in 2026 is execution. This new Futurum Research brief, developed in partnership with Anastasia Ustinova, Werner Goertz, me and Dylan Bristot highlights the five top failure modes that prevent organizations from turning AI pilots into real inference ROI. From my own conversations with AI builders and founders around the world, three themes stand out to me as a professional marketer and, business leader. First, the best AI outcomes usually come from redesigning a process from the ground up, not forcing AI into existing human workflows in unnatural ways. Start with the failure points, bottlenecks, and areas where speed and precision matter most, then build a system optimized to react to new data, detect failures, review outcomes, and respond automatically. Second, drive AI adoption from the top so innovation becomes cultural, rather than optional. Organizations move faster when leadership is clear, consistent, and sets expectations for AI instead of creating an internal game of thrones over who gets to wear the Token Crown. Finally, stop limiting advanced tools to a small group of AI insiders. If you want innovation and speed, train everyone to work with AI and agents across operations, supply chain, GTM, and core workflows, with incentives that support the change you expect. Start with strategic leaders and make coding tools, accessible without IT bottlenecks or workarounds. When you’re ready to make inference your product, Nebius Token Factory will be here to deliver real value. And…read the report. #AI #Inference #EnterpriseAI #Nebius #DigitalTransformation

Most companies say: 👉 “We are building AI.” But what they actually mean is: 👉 “We built a demo that works on good days.” Here’s the difference no one talks about: Demo AI impresses. Production AI survives. We worked with a team handling thousands of support cases. They already had “AI”: • It could answer queries • It looked smart in demos • It worked… sometimes But in production? • Inconsistent responses • No control over outputs • No memory of past context • Engineers didn’t trust it So what happened? 👉 They stopped using it. That’s how most AI projects fail. Not because they don’t work. 👉 Because they can’t be trusted. We rebuilt the system from scratch. Not focusing on the model. Focusing on everything around it. What changed: • Context-aware memory (no repeated mistakes) • Validation layers (no blind outputs) • Controlled actions (AI doesn’t overstep) • Feedback loops (system improves over time) Now: 👉 AI doesn’t “suggest” 👉 It handles real workflows The result: • Faster resolutions • Reduced manual effort • Consistent, reliable outputs • Actual adoption by engineers Here’s the truth: Production-grade AI is not about: ❌ Better prompts ❌ Better models It’s about: 👉 Designing systems that don’t break in the real world At Cortex AI Lab, this is what we build. Not demos. 👉 Systems that people actually trust. If your AI still needs constant checking… It’s not production-ready. And that’s where most teams are stuck.

Once organizations start structuring their operational data, something important begins to happen. AI moves beyond answering questions. It starts orchestrating outcomes. Most customer interactions are not a single answer. They are a sequence of actions. Diagnose the issue. Identify the right path. Guide the user through the correct steps. Escalate when necessary. Without structure, AI is often guessing at how those steps should unfold. But when operational data is structured into clear procedures and decision paths, AI gains the ability to orchestrate the experience. It can guide conversations based on the right process. It can adapt responses based on decisions along the way. It can ensure the next step is always aligned with how the organization actually operates. This is where orchestration becomes the real advantage. AI is no longer just generating responses. It is coordinating the flow of work across the customer journey. That is the power of combining conversational AI or chatbots with a structured operational intelligence layer like Procedureflow. Not just better answers. But AI that orchestrates the right outcome using structured operational data.