Artificial intelligence often fails when applied to real-world mathematical optimization problems, performing well on textbook examples but breaking down with actual business data. This gap exists because existing tools assume problem descriptions are complete, unambiguous, and small enough to fit into a single prompt, conditions rarely met in industrial settings. To address this, a new open-source AI agent named ORPilot has been developed to handle the complexities of production-grade optimization. Operations Research has long powered critical business decisions like supply chain design and facility scheduling, but translating vague business needs into mathematical models requires significant human expertise. While Large Language Models show promise in generating code for well-defined linear or mixed-integer programming problems, they struggle with real-world ambiguity, massive datasets, and the need for data transformation. Current tools often hallucinate when faced with large data volumes or generate code that cannot be easily ported or reproduced when data updates occur. ORPilot distinguishes itself by prioritizing understanding over speed, mimicking the workflow of a skilled human consultant. It operates through a strictly sequential five-stage pipeline to prevent cascading errors. First, an Interview Agent engages the user in a structured dialogue to clarify vague objectives, identify missing constraints, and define decision variables before any modeling begins. This ensures the AI solves the correct problem rather than relying on dangerous assumptions. Second, the Data Collection Agent treats data as separate external files rather than embedding them in prompts. It defines the necessary CSV file structures and validates that the user provides the correct data, handling scenarios where raw inputs like GPS coordinates must be transformed into distance matrices. Third, the Parameter Computation Agent automatically generates Python scripts to derive necessary model parameters from raw data, such as calculating aggregate demand or Big-M values for constraints. This step significantly reduces code generation failures by ensuring the modeling phase receives clean, model-ready data. Fourth, the Code Generation Agent produces Python scripts for popular solvers like Gurobi, CPLEX, and PuLP. It executes the code in a sandboxed environment and utilizes an automatic retry loop for syntax or runtime errors, leveraging the clarity established in previous stages to fix issues quickly. Finally, a Reporter Agent translates the numerical solution into plain English, explaining facility openings, routes, and production quantities in terms understandable to business stakeholders. ORPilot has demonstrated its capability on large-scale problems, successfully managing a supply chain network design with over 9.7 million decision variables and nearly one million constraints. The tool is now available as open-source software, supporting various large language model providers and solver backends. Future developments will focus on an intermediate representation format to ensure results remain portable and reproducible across different computational environments without re-engaging the AI.