🤖 AI

Business Utility of Large Language Models as Exploratory Data Analysis Agents

This paper evaluates the business utility of Large Language Models as exploratory data analysis agents using a supply chain simulation benchmark, revealing that most configurations lack the necessary repeatability for autonomous use despite acceptable average performance, while a specific high-effort GPT-5.4 configuration demonstrates the strongest balance of quality and reliability.

Original authors: Rafał Łabędzki, Patryk Miziuła, Hubert Rutkowski, Szymon Betlewski, Cezary Depta, Szymon Janowski, Jarosław Kochanowicz, Jan Kanty Milczek

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rafał Łabędzki, Patryk Miziuła, Hubert Rutkowski, Szymon Betlewski, Cezary Depta, Szymon Janowski, Jarosław Kochanowicz, Jan Kanty Milczek

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you hire a team of highly intelligent, super-fast detectives to solve a mystery: Which supplier sent the bad apples that caused a grocery store to lose money?

This paper is a report card on how well these "AI detectives" (Large Language Models) actually perform this specific job. The researchers didn't just ask, "Did they get the right answer?" They asked, "Can we trust them to get the right answer every single time we ask?"

Here is the breakdown of their findings using simple analogies.

1. The Job: Finding the "Bad Apples"

In the real world, businesses don't always have a label that says "This batch is rotten." Instead, they have to look at clues: maybe customers stopped buying a product, or stores returned items.

The Task: The AI had to look at messy data trails (like a detective looking at footprints) to figure out which specific supplier-product combination was the culprit.
The Challenge: This isn't a simple math problem with one right answer. It requires guessing, connecting dots, and forming a theory. This is called Exploratory Data Analysis (EDA).

2. The Test: The "Five-Run" Rule

The researchers didn't just let the AI try once. They made 15 different AI models (like GPT-5, Gemini, Claude, etc.) try to solve this mystery five times each, under four slightly different conditions (like giving them a clearer map, or a messier map).

They measured two things:

Accuracy: Did they find the bad apples?
Consistency: If you asked them the same question five times, did they give the same answer every time?

3. The Big Problem: The "Rollercoaster" Effect

The paper found a surprising truth: Most AI models are like rollercoasters.

Sometimes they zoom to the top and solve the mystery perfectly.
Other times, they crash and give a wrong answer.
Even if their average score looks okay, the fact that they are so unpredictable makes them dangerous to use in a real business.

The Analogy: Imagine hiring a chef. If they make a perfect steak 50% of the time and a burnt piece of charcoal the other 50%, their "average" meal might look okay on paper. But you wouldn't trust them to run your restaurant because you can't predict what you're getting. That is the problem with these AI agents right now.

4. The New Scorecard: "Business Utility"

The researchers invented a new way to grade the AI called Business Utility. Think of it as a "Trust Score."

It takes the AI's average accuracy and punishes it heavily if it is inconsistent.
The Formula: If an AI is smart but erratic, its score drops. If an AI is slightly less smart but very reliable, its score stays higher.

The Result:

The Winner: One model, GPT-5.4 (with extra "thinking" effort), was the clear champion. It was both smart and consistent. Its "Trust Score" was high.
The Losers: Most other models, even some that are usually very popular, were too unstable. They got good scores on average, but their "Trust Score" was low because they were too unpredictable.

5. The "Signal" Test

The researchers also tested if the AI got confused when the clues were harder to find (a "weaker signal").

Some models changed their behavior drastically when the clues got harder.
However, the paper found that internal inconsistency (the rollercoaster effect) was a bigger problem than getting confused by hard clues. Even when the clues were easy, the AI still couldn't agree with itself on the answer.

6. The Bottom Line

The paper concludes that while AI is getting better at solving these puzzles, we are not ready to let them work alone yet.

Current State: AI is like a brilliant intern who has great ideas but forgets to write them down half the time. You can't rely on them to run the show without a human double-checking every single step.
The Lesson: When companies want to use AI for data analysis, they shouldn't just look at the "average" success rate. They need to look at reliability. If the AI can't do the job consistently, it's not ready for the real world, no matter how smart it looks on a good day.

In short: The AI detectives are smart, but they are too moody to be trusted with the keys to the business yet. We need them to be consistent before we let them work without supervision.

Technical Summary: Business Utility of Large Language Models as Exploratory Data Analysis Agents

Problem Statement

While Large Language Models (LLMs) are increasingly integrated into analytical workflows, their suitability as autonomous agents for Exploratory Data Analysis (EDA) in business settings remains uncertain. In organizational contexts, a deployable EDA agent requires not only useful average performance but also sufficient repeatability to support trust. High volatility in outputs increases verification costs and makes failures difficult to anticipate. Current benchmarks often focus on code generation or closed-form question answering, leaving a gap in evaluating the full process-oriented EDA workflow: describing data, identifying salient features, and interpreting them to generate explanatory hypotheses. This study addresses whether LLMs can function as dependable EDA agents under realistic business conditions where success depends on both inference quality and repeatability.

Methodology

Benchmark Task and Simulation

The study utilizes a controlled, business-relevant benchmark built on an agent-based supply chain simulation. The task involves identifying supplier-product combinations responsible for low quality and downstream sales loss. Crucially, the "ground truth" is not explicitly labeled; the agent must infer causal relationships by reasoning from indirect operational traces (e.g., customer demand, store orders, wholesaler allocations) rather than reading a direct defect flag. This design mirrors real-world business diagnosis where causes must be reconstructed from fragmented evidence.

Experimental Design

Models: Fifteen model-variant configurations from eight LLM families were evaluated.
Conditions: Four experimental conditions varied data representation (raw vs. redundant tabular data), prompt clarity (short vs. non-ambiguous), and signal strength (clear/strong vs. clear/weaker).
Trajectories: Each model-variant was tested on five trajectories per condition (20 total per model), resulting in 300 scored trajectories.
Scoring: Outputs (JSON predictions of supplier-good pairs) were compared against deterministic ground truth using the Jaccard index.

Evaluation Framework

The paper introduces a multi-dimensional evaluation framework moving beyond simple mean scores:

Mean Score (ms): Estimates average analytical correctness (Jaccard score).
Coefficient of Variation (CV): Quantifies relative variation in scores within the same condition, capturing unpredictability.
Condition Sensitivity: Assessed via Mann-Whitney U and Kruskal-Wallis tests to determine if model behavior shifts across experimental conditions.
Business Utility: A proposed risk-adjusted metric synthesizing quality and repeatability into a single operational measure (0 to 1). It is defined as:
$\text{Business utility} = ms \cdot e^{-2.25 \cdot CV^{0.88}}$
This formula applies a prospect-theory-inspired loss function to the CV, where instability non-linearly discounts the expected analytical value. The parameters (2.25 for loss aversion, 0.88 for curvature) are adapted from cumulative prospect theory to express that departures from ideal repeatability reduce perceived usefulness.

Key Results

Performance Overview

Overall Mean: The overall mean score (ms) across all 300 trajectories was 0.3562.
Top Performer: GPT-5.4 (xhigh reasoning effort) achieved the strongest profile with an experiment-averaged ms of 0.8748 and an experiment-averaged Business utility of 0.6952. Its low CV (0.1202) resulted in a strong stability discount (0.7665).
Ranking Shifts: When ranked by Business utility, the order differed significantly from rankings based solely on mean score. For instance, Gemini 3.1 Pro (high) ranked second (Utility: 0.3533) despite a lower mean score than GPT-5.2 (xhigh), because its lower variability preserved more utility. Conversely, models with high mean scores but high CV (e.g., GPT-5.4 mini xhigh) saw their utility drastically reduced.
General Instability: Most configurations were deemed unreliable for autonomous use. The average Business utility across all variants was only 0.1511, approximately 42% of the average mean score (0.3562), indicating that instability severely erodes practical readiness.

Condition Sensitivity

Within-Condition Variability: The primary obstacle to deployment was identified as within-condition unpredictability (high CV) rather than sensitivity to environmental changes. Many models produced inconsistent reasoning paths even under identical conditions.
Cross-Condition Sensitivity: Only a subset of models showed significant sensitivity to experimental changes (e.g., weaker signal strength in Experiment 4). However, this was secondary to the issue of internal instability.

Key Contributions

Operationalized EDA Benchmark: Introduced a controlled benchmark that treats EDA as a causal reconstruction problem over business-process data, moving beyond code generation or static question answering.
Risk-Adjusted Evaluation Framework: Proposed a framework combining mean task quality with variability, introducing the Business utility metric to summarize quality and repeatability in a single operational measure.
Empirical Evidence of Fragility: Provided evidence that current LLMs remain operationally fragile as autonomous EDA agents. The study demonstrates that average performance often overstates practical readiness when instability is high.

Significance and Claims

The paper argues that the evaluation of EDA agents must treat average quality, repeatability, and condition sensitivity as complementary dimensions of operational trustworthiness.

Trust as a Function of Metrics: The authors posit that trust in AI EDA agents should not be a vague attitude but a function of demonstrated analytical quality, repeatability, and robustness.
Deployment Implications: Organizations should be cautious about interpreting isolated strong runs or benchmark averages as evidence of deployment readiness. A model may appear competent on average while remaining unreliable in repeated use, which is a critical weakness for exploratory analysis where consistent recovery of meaningful structure is required.
Methodological Shift: The study advocates for a shift in evaluation methodology from asking "Can a model produce a correct answer?" to "Can it do so often enough and consistently enough to justify trust?"

The findings suggest that while high-performing and relatively stable EDA-like behavior is possible (as seen with GPT-5.4 xhigh), the broader pattern for current frontier models is one of operational fragility, preventing confident substitution for human analysts in this specific task setting.