🤖 AI

What Makes a Sale? Rethinking End-to-End Seller--Buyer Retail Dynamics with LLM Agents

The paper introduces RetailSim, an end-to-end retail simulation framework powered by LLM agents that unifies seller persuasion, buyer-seller interaction, and purchase decisions to accurately model cross-stage dependencies and evaluate retail strategies through behavioral fidelity and economic regularities.

Original authors: Jeonghwan Choi, Jibin Hwang, Gyeonghun Sun, Minjeong Ban, Taewon Yun, Hyeonjae Cheon, Hwanjun Song

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jeonghwan Choi, Jibin Hwang, Gyeonghun Sun, Minjeong Ban, Taewon Yun, Hyeonjae Cheon, Hwanjun Song

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 are a store owner trying to figure out the perfect way to sell a new product. You want to know: If I change my sales pitch, will people buy more? If I lower the price, will I make more money? If my delivery is late, will customers get angry?

In the real world, testing these ideas is risky and expensive. If you try a bad strategy, you lose money. If you try a good one, you might miss out on making even more money because you didn't test it enough.

This paper introduces RetailSim, a "virtual reality" for shopping. Think of it not as a video game, but as a massive, digital flight simulator for retail stores. Just as pilots use simulators to practice flying in storms without crashing a real plane, business owners can use RetailSim to practice selling without losing a single dollar.

Here is how it works, broken down into simple concepts:

1. The Cast of Characters (The Actors)

In a real store, you have real people with real personalities. In RetailSim, the "actors" are AI agents (smart computer programs) that act like humans.

The Sellers: These aren't just robots reading a script. They have "personalities." Some are assertive (pushy and direct), some are friendly (warm and chatty), and some are rational (focused on facts and numbers).
The Buyers: These agents also have personalities. Some are picky (they check every detail), some are price-sensitive (they only buy if it's cheap), and some are emotional (they buy because they feel like it).

The magic of RetailSim is that it lets these two groups interact naturally. A "picky" buyer might argue with an "assertive" seller, while an "emotional" buyer might fall in love with a "friendly" seller's pitch.

2. The Full Story (The Movie, Not Just a Scene)

Previous computer simulations were like watching a single scene of a movie. They might simulate a customer asking a question, or a customer returning a broken item, but they didn't connect the dots.

RetailSim simulates the entire movie, from start to finish:

Scene 1 (The Pitch): The seller creates a sales pitch based on their personality and the product.
Scene 2 (The Chat): The buyer asks questions ("Is it durable?" "How much is shipping?"). The seller answers.
Scene 3 (The Decision): The buyer decides to buy or walk away.
Scene 4 (The Aftermath): If they bought it, what happens if it arrives late or broken? The buyer contacts support, and the seller tries to fix it.
Scene 5 (The Review): Finally, the buyer leaves a review based on the whole experience, not just the product.

This is crucial because a great sales pitch can be ruined by a rude customer service agent later. RetailSim sees the whole chain.

3. The "Reality Check" (Did it work?)

The authors were worried: Are these AI actors just making things up, or do they actually act like real humans?

To prove it works, they ran "tests" to see if the simulation followed the laws of economics:

The Price Test: They lowered the price in the simulation. Did more people buy? Yes. (Just like in real life).
The Gender Test: They tested products marketed to men vs. women. Did men buy the "men's" products more? Yes.
The "Picky" Test: They made some buyers very sensitive to price. Did those buyers react more strongly to price changes? Yes.

Because the simulation followed these real-world rules, the authors are confident it's a reliable testbed.

4. Why Should You Care? (The Superpower)

RetailSim acts like a crystal ball for business decisions. Here are three ways it helps:

Reading Minds (Persona Inference): Imagine you have a chat log with a customer, but you don't know if they are "picky" or "easygoing." RetailSim can analyze the chat and guess their personality traits, helping you tailor your future approach.
The Perfect Match (Revenue Impact): The paper found that not all sellers work with all buyers. A "rational" seller might make more money selling to a "rational" buyer, but a "friendly" seller might do better with an "emotional" buyer. RetailSim can find the perfect pairings to maximize profit.
Strategy Testing: You can ask, "What if I change my sales script to be more urgent?" The simulation runs this scenario thousands of times instantly and tells you, "This strategy will likely increase sales by 10%," before you ever spend a dime on a real ad campaign.

The Bottom Line

RetailSim is a safe, digital playground where businesses can experiment with sales strategies, customer personalities, and pricing models. It uses advanced AI to mimic the messy, complex, and human nature of shopping, allowing companies to learn from their mistakes in a virtual world so they can succeed in the real one.

It's like having a time machine that lets you try out a million different ways to run a store, so you only have to launch the best one once.

1. Problem Statement

Designing effective retail strategies is inherently difficult because outcomes are determined by a complex, multi-stage pipeline involving interdependent seller–buyer interactions. Existing research and simulation frameworks suffer from two primary limitations:

Fragmented Scope: Prior studies focus on isolated stages (e.g., only negotiation, only customer support, or only product recommendation) rather than the full end-to-end journey from seller persuasion to post-purchase feedback.
Lack of Economic Consistency: While existing simulators may generate realistic dialogue, they often fail to reproduce fundamental real-world economic regularities (e.g., the inverse relationship between price and demand, or heterogeneous price elasticity). This makes it difficult to verify if simulation outcomes are reliable for strategic decision-making.

The core challenge is creating a unified simulation environment that captures cross-stage dependencies (how early decisions propagate to downstream outcomes) and adheres to real-world economic principles.

2. Methodology: The RetailSim Framework

The authors introduce RetailSim, a multi-agent simulation framework designed to model the entire retail pipeline in a unified environment.

A. Unified Multi-Stage Pipeline

RetailSim structures the retail process into five causally linked stages, where the output of one stage serves as the context for the next:

Seller-Side Persuasion: The seller agent generates a structured sales pitch based on four strategic dimensions: Target Expansion, Value Proposition, Contextual Urgency, and Objection Handling.
Pre-Purchase Inquiry: The buyer agent engages in a multi-turn dialogue to clarify product specs, pricing, shipping, or comparisons before deciding.
Purchase Decision: The buyer decides whether to purchase and the quantity, based on the accumulated context of the pitch and inquiry.
Post-Purchase Support: If a purchase occurs, the buyer may encounter issues (e.g., damaged items, delays). A multi-turn dialogue resolves these via Delivered, Refunded, or Exchanged outcomes.
Review Generation: The buyer generates a four-dimensional review covering the sales pitch, pre-purchase support, post-purchase resolution, and product satisfaction.

B. Key Components for Fidelity

To ensure the simulation reflects real-world complexity, RetailSim incorporates:

Diverse Product Space: Utilizing metadata from the Amazon Reviews '23 dataset, covering 34 categories grouped into Food, Fashion, Home, and Electronics.
Persona-Driven Agents: Both sellers and buyers are assigned specific personas via prompt injection.
- Seller Traits: Assertiveness (Assertive vs. Passive), Friendliness (Friendly vs. Reserved), and Rationality (Rational vs. Emotional).
- Buyer Traits: Pickiness (Picky vs. Easygoing), Price Consciousness (Sensitive vs. Indifferent), and Rationality (Rational vs. Emotional).
Multi-Turn Interactions: Simulates realistic, stochastic dialogues where responses are conditioned on the full conversation history.

C. Evaluation Protocol

The authors propose a dual-level evaluation protocol:

Stage-Level Fidelity (Human Evaluation): Human annotators evaluate task fidelity (adherence to stage guidelines) and persona fidelity (whether agents exhibit intended traits) using Likert scales and A/B testing.
System-Level Meta-Evaluation (Economic Consistency): The simulation is tested against established economic principles to ensure it reproduces real-world patterns, specifically:
- Demographic purchasing patterns (gender-based differences).
- The price–demand relationship.
- Heterogeneous price elasticity (sensitivity of different buyer types).

3. Key Contributions

End-to-End Simulation Framework: RetailSim is the first framework to model the complete retail pipeline (persuasion $\to$ interaction $\to$ decision $\to$ support $\to$ review) within a unified, causally structured environment.
Economic Consistency Validation: The paper introduces a "meta-evaluation" paradigm, demonstrating that LLM agents can reproduce fundamental economic laws (e.g., demand decreases as price increases) when guided by realistic personas, moving beyond superficial behavioral mimicry.
Persona-Driven Decision Support: The framework enables the inference of latent personas from interaction data and analyzes how specific persona pairings (e.g., an assertive seller vs. a picky buyer) impact revenue and conversion.
Strategic Testbed: It serves as a controlled environment to evaluate sales strategies (e.g., the impact of human-guided sales scripts vs. naive LLM generation) before real-world deployment.

4. Experimental Results

The framework was evaluated using eight different Large Language Models (LLMs), including Qwen3, GPT-5 variants, Gemini, and DeepSeek.

Behavioral Fidelity: RetailSim achieved high scores in human evaluation.
- Task Fidelity: Sales scripts and reviews scored >4.0/5.0 across models.
- Persona Fidelity: Agents consistently reflected their assigned traits with >95% accuracy in A/B comparisons.
Economic Consistency: The simulation successfully reproduced real-world economic patterns:
- Demographics: Male buyers showed higher purchase rates for men-oriented products, while female buyers preferred women-oriented products.
- Price–Demand: Purchase rates decreased monotonically as prices increased (from 64.8% at -10% price to 56.6% at +10% price).
- Price Elasticity: "Price-sensitive" buyers exhibited significantly higher price elasticity (2.58) compared to "price-indifferent" buyers (0.76), a statistically significant difference ( $p < 0.001$ ).
Application Use Cases:
- Persona Inference: Classifiers could accurately infer latent buyer/seller traits from interaction traces (85.5%–98.5% accuracy).
- Revenue Analysis: Revealed that "assertive and friendly" sellers generated higher sales, but "emotional" sellers had higher refund rates. It also showed that revenue is not just about the seller model but the pairing of seller and buyer personas.
- Strategy Evaluation: Demonstrated that human-guided sales strategies consistently outperformed naive LLM-only scripts, and that larger models closed the gap only when fully guided.

5. Significance and Impact

Scientific Testbed: RetailSim provides a rigorous, controlled environment for studying LLMs as economic agents, allowing researchers to measure latent behavioral traits and their interaction effects on economic outcomes.
Practical Utility: It offers a low-risk, high-fidelity testbed for retailers to evaluate sales scripts, train customer service agents, and optimize pricing strategies without the cost and irreversibility of real-world A/B testing.
Theoretical Advancement: By bridging the gap between behavioral simulation (LLM dialogue) and economic theory (market regularities), the paper establishes a new standard for evaluating AI agents in socio-economic domains, proving that LLMs can be aligned with fundamental economic principles through structured persona and pipeline design.