Decision Intelligence Agent: Assisting Business Decisions with Causal Reasoning and Monte Carlo Simulation

This is a decision intelligence prototype system that combines large language models (LLMs) with deterministic analysis components, aiming to provide interpretable and auditable intelligent decision support for enterprises. Its core features include:

• Defining organizational causal models via YAML specifications;
• Using Monte Carlo simulation to evaluate decisions under uncertainty;
• Integrating machine learning, optimization algorithms, and LLM orchestration capabilities;
• Bridging the gap of traditional BI which only performs descriptive analysis, and outputting normative decision recommendations on "what should be done".

Traditional business intelligence (BI) systems are limited to descriptive analysis (reports, dashboards) and rely on humans to make decisions based on insights. However, in complex environments (nonlinear interactions of variables, future uncertainty), human intuition is hard to capture the optimal path.

Decision Intelligence, as an emerging discipline, integrates achievements from fields such as operations research and causal inference to build systems that output decision recommendations (different from predictive analysis which answers "what will happen", it answers "what should be done"). This project is a technical implementation of this concept, combining the language capabilities of LLMs with traditional analysis tools.

Core design philosophy: Clearly distinguish between "orchestration" and "computation".

• LLM is responsible for orchestration: Understanding queries, selecting tools, extracting parameters, and converting results into natural language;
• Deterministic components handle computation: Numerical operations (optimization, simulation, causal propagation) are performed by Python components to ensure reproducibility and auditability.

The system is defined via spec/organizational_model.yaml: decision variables, causal relationships, simulation configurations, and optimization objectives. The causal directed acyclic graph (DAG) serves as the core data structure, supporting topological sorting execution, modular expansion, and result traceability.

1. Monte Carlo Simulation Engine: Samples probability distributions of input variables, runs causal graph evaluations multiple times, and outputs probability distributions of results (e.g., "80% probability that profit is between X and Y");
2. Optimization Solver: Supports discrete optimization and multi-objective optimization with reproducible results;
3. LLM-as-a-Judge Mechanism: An independent LLM checks the grounding (based on tool outputs), responsiveness (answering user questions), and numerical consistency of results to improve reliability.

User query processing workflow:

1. Planning: LLM analyzes intent, selects tools (simulation/optimization/retrieval) and extracts parameters;
2. Execution: Calls deterministic tools to complete computation;
3. Synthesis: LLM converts raw results into natural language explanations;
4. Judgment: An independent LLM checks result quality and decides whether to return it;
5. Output: Presents the final answer.

Workflow states are persisted to SQLite, supporting multi-turn conversation context tracking.

Applicable scenarios include:

• Pricing Strategy Optimization: Simulate profit distributions under different price/marketing combinations to find the maximization strategy;
• Supply Chain Decisions: Simulate stockout risks and holding costs of inventory strategies, and recommend optimal ordering plans;
• Resource Allocation: Balance return on investment and risk under budget constraints to find Pareto optimal allocations;
• Strategic Planning: Evaluate the impact of macro assumptions on business metrics to support scenario analysis.

Helps enterprises quantify risks and make data-driven decisions.

Advantages:

• Auditability: Every step of computation is traceable, meeting compliance requirements;
• Domain-agnostic: Adapts to industries such as retail, finance, and manufacturing via YAML configuration;
• Human-machine collaboration: Enhances human decision-making capabilities while retaining final decision-making authority.

Limitations: Currently focuses on discrete decision spaces and simple optimization objectives.

Future Directions: Expand continuous optimization algorithms, multi-agent game scenarios, dynamic programming, and real-time data integration.