How a Generative AI‑Powered Normative Analysis Framework Boosted Decision Efficiency 20× in a Conversational BI Product

In the era of data‑driven and intelligent decision making, enterprises need fast, accurate solutions for complex business scenarios. Traditional decision processes rely on manual expertise or single mathematical models, which struggle to capture time‑series data, resource constraints, and business objectives while being labor‑intensive. This paper proposes a normative analysis framework that integrates operations research models, statistical learning methods, and generative large language models (LLMs) to provide automated, explainable, and efficient decision support.

Technical Background and Motivation

The framework builds on advances in generative AI and autonomous agents, extending predictive analytics (machine learning, statistical models) with optimization algorithms and simulation techniques. It aims to automatically generate and solve mathematical models under business constraints, delivering actionable recommendations such as "what will happen" and "what should be done".

Framework Architecture

The core of the framework is a "perception‑prediction‑optimization‑execution‑learning" loop, implemented as a Manus‑style decision‑analysis agent enhanced by GenAI capabilities. It provides:

Natural‑language understanding to extract decision variables, objectives, and constraints.

Automatic model formulation (auto‑formulation) using LLMs.

Optimization via linear programming, integer programming, or custom solvers.

Result interpretation and report generation.

Continuous learning from outcomes to refine models.

Proof‑of‑Concept in ChatBI

The framework was evaluated in the ChatBI conversational business intelligence product through two representative scenarios:

Product Optimal Combination – a fully structured linear programming problem with clear objective (profit maximization) and constraints (resource limits, minimum order quantities). The POC compared a traditional analyst‑built model with the ChatBI agent’s auto‑generated model.

Product Dynamic Pricing – a semi‑structured problem requiring competitor analysis, customer segmentation, and price testing, without a unique optimal solution.

Scenario 1: Product Optimal Combination

Data: two workstation models (DL1, DL2) with known costs, profit margins, and processing times. Goal: determine monthly production quantities to maximize profit under resource constraints.

Traditional workflow involved manual variable definition, objective formulation, constraint specification, model building, and solving, taking ~30 minutes.

ChatBI workflow used a natural‑language query, the agent automatically formulated the linear program, invoked scipy.optimize, and produced the optimal solution (x=133.33, y=66.67, profit ≈ $78,333) in 90 seconds.

Results showed identical model structure and numerical solution, confirming correctness, while decision time improved by ~20×.

Scenario 2: Product Dynamic Pricing

Data: retail pricing dataset from Kaggle. Goal: devise a pricing strategy that maximizes profit, incorporating competitor pricing, customer segmentation, and price testing.

The agent decomposed the vague objective into three sub‑tasks: competitor analysis, customer segmentation, and price testing. It then applied clustering, elasticity estimation, and experimental design techniques, finally outputting a differentiated pricing plan with clear implementation steps.

End‑to‑end execution took 134 seconds, demonstrating the framework’s ability to handle semi‑structured decision problems.

Evaluation Metrics

Model construction correctness – 100% match with manually built models.

Numerical solution accuracy – identical optimal values.

Decision efficiency – 20× speedup for structured LP, ~10× for dynamic pricing.

Discussion

The study confirms the technical feasibility of the normative analysis framework for linear programming tasks and highlights its potential to shift decision‑making from expert‑driven to agent‑driven automation. Limitations include the current focus on well‑structured LP problems; performance on integer, non‑convex, or real‑time streaming scenarios remains to be explored.

Future work should address closed‑loop learning, domain‑specific knowledge integration, and real‑time data handling to broaden applicability across supply‑chain, marketing, and inventory management domains.

Conclusion

The normative analysis framework, powered by GenAI, delivers substantial efficiency gains and lowers the technical barrier for complex business decisions. Its successful integration into ChatBI illustrates a promising path toward enterprise‑wide intelligent decision platforms.