The Case of the IMC in NARRATE
In an industrial world where data is multiplying but decisions still often depend on human judgment and spreadsheets, one question is gaining momentum:
Can artificial intelligence make real, useful, and explainable decisions within a supply chain?
That is the question we asked ourselves in the European project NARRATE, and one we are answering through the development of a new kind of industrial intelligence: the IMC – Intelligent Manufacturing Custodian.
The IMC is not just an algorithm. It is an intelligent platform capable of understanding a company’s manufacturing network, anticipating disruptions, proposing optimal alternatives, and communicating with decision-makers in a natural – sometimes even conversational – way.
What is the heart of this transformation? The combination of generative AI (like LLMs) with symbolic models, rule engines, and intelligent service orchestration. In other words, combining structured reasoning with the flexibility of language.
In this post, we will explain how we are doing it, focusing on one of the project’s key components:
Creating a knowledge model for smart manufacturing networks and integrating it with generative AI.
An architecture that allows the factory not only to automate but also to reason, explain, and act.
What Exactly Does the IMC Do?
- Observes: Collects real-time information about orders, suppliers, materials, risks, and production capacity.
- Interprets: Transforms data into structured knowledge using ontologies, symbolic rules, and optimization models.
- Decides: Proposes actions in complex scenarios (logistics failures, demand shifts, bottlenecks).
- Dialogs: Connects with human operators or agents using natural language, thanks to integration with generative AI models.
Why Call It a “Custodian”?
Because it does not replace the human—it supports, assists, and optimizes their work. The IMC acts as an intelligent custodian of the manufacturing network, helping to:
- Select alternative suppliers during disruptions
- Prioritize orders based on real capacity
- Anticipate logistical impacts before they occur
- Justify each decision with clear and traceable rules
- Learn from each case to improve its own decision-making over time
What Sets It Apart?
Unlike classic planning systems (MRP, APS, etc.), the IMC is designed for uncertain and dynamic environments, where planning is not enough – you must adapt in real time, reason under uncertainty, and explain every action.
The IMC is not based solely on fixed rules or opaque predictions. Its strength lies in combining symbolic logic + generative AI + service orchestration, making it robust, transparent, and flexible all at once.
Giving the IMC Reasoning Capabilities
Talking about a system that “thinks” sounds ambitious. That is why, in the NARRATE project’s Deliverable D3.3, we tested it in practice.
We created a proof of concept (PoC) showing how a manufacturing network can reason, answer questions, and make decisions in real-world scenarios.
What Did We Build?
A knowledge model implemented on a graph database (Deliverable D4.2), representing:
- Products and components
- Orders and available stock
- Suppliers and lead times
- Risks associated with each node
We added symbolic rules like:
“If a supplier fails, look for another that minimizes delivery time or cost, based on order priority.”
And yes, the system could automatically apply these rules, explaining why it made each decision.
What Questions Could It Answer?
We designed a query interface able to answer questions such as:
- “What do I need to build a cradle?”
- “How many wardrobes can I build with current stock?”
- “Is there any order at risk of not being delivered?”
- “What alternative provider can I use for product X to minimize delivery time?”
- “Who can provide 70 mattresses with the best delivery conditions?”
These are not just answered with data—they require contextual reasoning, logic, and optimization.
Why Use a Graph Database?
Because a manufacturing network is a network: complex interdependencies, dynamic relationships, decisions dependent on many factors.
The graph allowed us to model that complexity naturally and navigably—and to give the system the foundation to “think.”
What Did We Prove?
- An industrial network can reason automatically in the face of disruptions.
- It can choose between alternatives based on objectives (cost, lead time, risk).
- It can explain the rationale behind each decision.
- By connecting this reasoning to a generative language model, any operator could interact and understand it in natural language.
In Deliverable D3.3, we gave the IMC its first real cognitive capability:
A symbolic knowledge model, integrated with business logic and ready to converse with humans or intelligent agents.
Where Does Generative AI Come In?
So far, we have discussed structures, rules, and symbolic models. But how do we make that really accessible to the humans making decisions every day?
That is where generative AI—especially LLMs—comes in, as an interface, translator, and enhancer of industrial intelligence.
From Rules to Conversation
Generative AI allows users to interact with the IMC like talking to a human expert.
Where previously one needed dashboards, complex queries, or technical teams, now it is enough to ask:
- “What alternative supplier can I use for tables to minimize delivery time?”
- “What orders are at risk if supplier X fails?”
- “How much can I produce if 500 more units of component Y arrive tomorrow?”
The LLM interprets the question, translates it into the symbolic model (graph + rules), runs inference, and returns an explained response, even with recommended options.
Generative AI as Interface and Explainer
It not only acts as a translator between human and system, but also as a mechanism for explainability:
🗣️ “Why did you choose that supplier?”
💡“Because the usual supplier is delayed, and among the available options, Z offers the best cost–lead time balance for order P456.”
This type of interaction enables transparent, trustworthy, and collaborative decision-making, with the system justifying each recommendation.
Advantages of Integrating LLMs into the IMC
✅ Accessibility: Anyone on the team, regardless of technical knowledge, can query the system.
✅ Time-saving: Drastically reduces the time spent searching for and analysing information.
✅ Trust: Decisions are explained, not imposed by a “black box.”
Generative AI does not replace the symbolic model: it extends, humanizes, and operationalizes it in everyday use.
Combined with the structured reasoning from D3.3, the IMC evolves from a technical system to a context-aware, conversational intelligent agent.
From Data to Autonomous Reasoning
Many industrial solutions focus on visibility—knowing what is happening. But in NARRATE, with the IMC, we aim to go further:
Understand context, anticipate risks, make decisions, and explain them.
To do this, we designed a system that combines three key layers of intelligence:
1. Symbolic AI: structure, rules, logic
- Represents the network knowledge (products, suppliers, risks…) as a navigable graph.
- Defines declarative rules like “if X happens, do Y,” adjustable to different goals (cost, time, priority).
- Allows inference and scenario simulations (“what if”).
This layer is stable, explainable, and traceable. It is the IMC’s logical nervous system.
2. Generative AI: language, accessibility, interaction
- Translates human questions into system actions.
- Explains symbolic model decisions in natural language.
- Learns from user feedback to refine responses or improve rules.
This is the IMC’s human face, connecting operators with internal intelligence.
3. Intelligent Orchestration: action, coordination, adaptability
- Connects the IMC with modules like:
- Demand prediction
- Risk management
- Resource planning
- Digital twins
- Coordinates tasks among agents (human or artificial) based on dynamic priorities.
This layer enables real-time action and adaptability to changing situations.
🔄 What Happens When We Combine Them?
- A supplier failure does not paralyze the network—it reconfigures automatically.
- An operator can ask in natural language and get a reasoned recommendation.
- The system anticipates disruptions and proposes alternatives before it is too late.
In short: we transform a manufacturing network into a cognitive, orchestrated system, where data is not only seen—but understood and used to decide.
📈 Industry Impact: What Did We Achieve?
Technology is great—but what really matters is the practical change. With the IMC from NARRATE and the capabilities from Deliverable D3.3, industrial companies can move toward more resilient, agile, and explainable decision-making.
1. Improved Operational Resilience
When a supplier fails, the system does not wait:
- Detects the issue
- Searches for viable alternatives based on rules and context
- Proposes the optimal decision (cost, time, priority)
All this can happen in seconds—with or without human intervention.
2. Reduced Analysis and Reaction Time
Before:
Planning teams manually reviewed spreadsheets, reports, and emails to understand the disruption.
Now:
With one query (“Can we fulfil orders if supplier A does not deliver this week?”), the system reasons and responds using data, rules, and priorities.
Decision-making shifts from hours to minutes.
3. More Collaborative and Justified Decisions
Thanks to generative AI, the system not only acts—it explains. This helps align logistics, procurement, operations, and management.
It also enables teams to adjust business rules in understandable language, without coding.
4. Scalable and Adaptable to New Factories or Sectors
The model is not limited to a single plant or scenario. It can be extended to:
- New suppliers
- New product categories
- Different operational priorities (cost, sustainability, lead time…)
This allows IMC logic to spread across entire production ecosystems, interconnecting multiple industrial nodes in an intelligent network.
In Summary
We have moved from a factory that sees what is happening to one that reasons, acts, and continuously improves. A key step toward cognitive industry, where humans and machines collaborate naturally and effectively.
What is Next?
Toward a Conversational, Explainable, and Autonomous Factory
The proof of concept in Deliverable D3.3 is just the beginning. Next steps in the NARRATE project will evolve the IMC into a fully cognitive agent within the industrial network.
From Expert System to Intelligent Agent
The end goal is to build an IMC that can:
- Coordinate services (forecasting, planning, procurement, logistics)
- Learn from human feedback and its own decision outcomes
- Converse in natural language with operators, plant managers, or executives
- Make autonomous decisions in well-defined contexts—and escalate ambiguous cases to humans
An Industry That Thinks, Adapts, and Explains Itself
We are building an industry where decision-making does not rely solely on intuition, manual reports, or isolated tools.
But on shared, distributed, and explainable intelligence, improving every day with data and human dialogue.
Conclusion
The supply chain is no longer just automated. Now, it reasons. The IMC from NARRATE represents a new generation of industrial tools.
Tools that combine symbolic structure + natural language + orchestrated services to turn data into knowledge—and knowledge into impact.
And if you are wondering whether generative AI can run your supply chain, the answer is:
Yes. And it can explain it to you, too.
Authors
Jorge Capel Planells
