AI Fundamentals in NPLAN

The term Artificial Intelligence has become ubiquitous in the corporate world. Language models like GPT, Gemini, and Claude impress with their ability to generate text, code, and even answer complex questions. However, when it comes to solving real supply chain planning problems, these models alone have important limitations.

This article explains why a robust AI approach in supply chain benefits from three well-defined layers, and why placing AI Agents directly on raw supply chain data tends to produce fragile results.

The table above shows how different AI techniques apply to each supply chain domain. Language models don't act as the primary calculation mechanism, but they play an important complementary role throughout the entire process. They connect the planning layers by interpreting results, explaining decisions, and enabling a much more natural interaction with the system.

This doesn't replace Optimization, Heuristics, and Machine Learning, which remain responsible for solving the mathematical and operational problems of the chain. GenAI's role is to amplify the usability of these techniques, facilitating analysis, scenario exploration, and decision-making.

That's why, at NPLAN, GenAI is not treated as a decision engine, but as a reasoning and intelligent interface layer that operates on top of the mathematical engine's results, making planning more accessible, iterative, and connected.

Despite this valuable complementary role, it is important to clearly understand where language models reach their limits when applied to supply chain problems.

Language models (LLMs) are trained to find statistical patterns in text. They excel at tasks like summarizing documents, generating reports, and conversing in natural language. However, supply chain problems involve combinatorial constraints, time horizons, finite capacity, interdependent lead times, and financial trade-offs that LLMs were not designed to solve directly.

Without proper orchestration, an LLM doesn't know that a production order must respect machine capacity, that setup depends on the previous sequence, or that safety stock needs to consider demand variability and supplier lead time simultaneously. These are mathematical problems requiring specialized algorithms.

This doesn't mean LLMs are useless in supply chain. On the contrary, they are extremely valuable when operating at the appropriate layer of the architecture, as an intelligent interface, copilot for planners, and natural language analysis engine over the engine's results.

If LLMs were not designed to solve these problems directly, a specialized calculation and decision mechanism is needed, one that can do so with precision and consistency.

The concept of a Supply Chain Engine exists to fill exactly this gap. It is a technical layer that combines four complementary pillars to solve planning problems in a deterministic and auditable way:

It's the combination of these four pillars that enables the engine to deliver precise, fast, and feasible answers. An LLM can describe the complexity of a finite scheduling algorithm, but calculating the detailed sequencing considering setup optimization, tool constraints, and synchronism with raw materials simultaneously requires specialized algorithms. The engine does this in a more robust, fast, and effective way, for thousands of SKUs at once.

Understanding what the engine does is only part of the challenge. The central question is how to organize these responsibilities into an architecture that clearly separates the role of each layer.

A robust architecture for AI in supply chain is organized into three distinct layers, each with its own responsibility:

The middle layer, the Supply Chain Engine, is the technical differentiator. It transforms raw data into viable, optimized, and consistent plans. Without this layer, AI Agents operate without visibility into real operational constraints.

This separation of concerns is fundamental. AI Agents can ask questions in natural language, but it's the Engine that calculates the answer with mathematical precision, considering all constraints and interdependencies.

Everything presented so far, from LLM limitations to the need for an engine and a 3-layer architecture, materializes in practice within the NPLAN platform.

NPLAN was built with this three-layer architecture. The platform combines a robust Supply Chain Engine, with optimization, heuristics, statistics, and machine learning, with an AI Agents layer that acts as a copilot for planners.

In practice, this means the planner can interact with the platform naturally, asking questions, exploring scenarios, and receiving recommendations, while the engine ensures all answers are mathematically consistent and respect real operational constraints.

NPLAN combines optimization, heuristics, statistics, and ML in a single decision engine, accessible through a modern, collaborative cloud experience.

Security is a fundamental pillar of the platform. NPLAN preferentially uses LLM models contracted through corporate agreements, ensuring data does not leave the country of origin and with contractual protection so it is not used for training third-party models.

Additionally, the platform features instant protection that prevents users from sending sensitive information to AI models, with rules configurable by each company's IT department. This ensures that confidential data such as prices, margins, and strategic information are never exposed.

This combination of a precise mathematical engine, governable AI Agents, and corporate security infrastructure positions NPLAN as a solid option for companies that want to implement AI in supply chain with control, traceability, and security.