Operationalizing the technology acceptance model with large language models: A framework for strategic insights from user reviews

Abstract

In the competitive digital service sector, leveraging user-generated data for strategic operational improvements is a critical engineering and management challenge. This study presents a business intelligence framework that integrates Artificial Intelligence (AI) with established management theory to operationalize technology acceptance drivers from unstructured text. We develop a systematic methodology that employs ChatGPT for theory-guided keyword generation to identify and measure the core constructs of the Technology Acceptance Model (TAM)—perceived ease of use, perceived usefulness, and Behavioral intention to use—within a massive dataset of 1,694,581 user reviews from leading US food delivery apps. Through a robust data processing pipeline incorporating sentiment analysis (VADER, AFINN) and Ordinary Least Squares (OLS) regression, we validate the framework’s efficacy, demonstrating that the AI-measured constructs explain 85.4% of the variance in users’ intention to use (R² = 0.854, p < 0.001). The results indicate that user perceptions of ease of use (β = 0.29, p < 0.001) and usefulness (β = 0.51, p < 0.001) are significant predictors of adoption intention. This research provides a tangible, data-driven framework for managers and engineers to systematically diagnose user experience, prioritize feature development, and formulate product strategies. The proposed methodology offers a replicable, theory-AI integrated analytics pipeline for transforming unstructured textual data into actionable engineering and business intelligence, offering a pathway to connect large-scale data analytics with strategic management decision-making.

1. Introduction

In an era marked by the dominance of digital technology and the rapid evolution of mobile applications, understanding user sentiment and preferences has become increasingly important, a critical concern shared by both businesses and researchers. ¹ User-generated reviews have emerged as a rich source of insights, shedding light on the intricate factors that drive the acceptance and adoption of technology. ² Within this landscape, the Technology Acceptance Model (TAM), a well-established framework in the field of technology adoption, has been widely used for evaluating users’ intentions and behaviors in the context of various technological innovations. ³ With an aim to probe the extent of technology adoption by users, the TAM has garnered substantial recognition. ⁴ This model posits that individuals are naturally inclined to embrace technology when they foster positive attitudes, a mindset cultivated through their perceptions of the technology’s utility and ease of use. ³ In this study, we investigate the intricate facets of TAM through the lens of user-generated reviews using machine learning (ML) techniques, all through the lens of user-generated reviews.

The ubiquity of smartphones and the rise of app-based services have transformed the way we interact with technology.^5,6 Nowhere is this transformation more evident than in the food delivery industry, where a plethora of mobile applications promise convenience and culinary delight at the tap of a screen. ⁷ For millions of users across the United States, these food delivery apps have become an integral part of daily life, shaping the way they satisfy their culinary cravings. ⁸ Understanding what drives users to embrace these apps is not only of academic interest but also of immense practical importance for the businesses that offer these services.

The term used to describe information resulting from user interactions is “user-generated data” (UGD), as described by Saura et al., ⁹ encompasses a myriad of information produced by users in their interactions within digital marketplaces. This data spans various types, including reviews, actions, emotions, comments, and experiences, constituting a rich tapestry of user engagement. ¹⁰ Additionally, the content users generate in collaborative online spaces is recognized as “user-generated content” (UGC), as discussed by Hossain and Rahman in 2022. ¹ The significance of UGC lies in its potential to yield diverse information, prompting extensive investigation in numerous studies. Hossain and Rahman (2022) ¹ probed UGC to discern empathy behavior in potential customers, while Saura et al. (2021) ⁹ explored its application in data-driven innovation. Pashchenko et al. (2022) ² delved into UGC to investigate customers’ emotional aspects, Hossain and Rahman ¹¹ examined customers’ sentiment, and Ettrich et al. ¹² identified customer needs embedded in user-generated content. Furthermore, Xu et al. ¹³ scrutinized User Satisfaction in New Energy Vehicles, Dong et al. ¹⁴ explored the identification and evaluation of competitive products based on UGC, and Wang et al. ¹⁵ investigated the role of user-generated travel posts in shaping travel choices.

Despite the widespread exploration of UGC in detecting core components, such as perceived ease of use and perceived usefulness, influencing users’ intentions to adopt technology, the TAM has been notably absent in the analysis of users’ text reviews. To address this gap, we direct our focus to user reviews as the primary data source. These reviews, serving as authentic reflections of individuals’ experiences and perceptions, ¹⁶ become the focal point of our inquiry. Through a textual lens, we aim to unravel the impact of two pivotal TAM factors: “perceived ease of use” and “perceived usefulness,” long recognized as critical determinants shaping users’ intentions to adopt and sustain technology use. ³ Our dataset, extensive in scope, comprises a staggering 1,694,581 reviews collected from three of the most popular food delivery apps on the Google Play Store in the United States. A distinguishing feature of this study is our approach, inspired by the capabilities of ChatGPT, an advanced natural language processing model. ChatGPT has provided us with invaluable recommendations for keywords associated with the key categories of our analysis: ease of use, usefulness, and intention to use. By incorporating these keywords into our research methodology, we aim to identify insights that may not emerge from traditional frequency-based methods within the vast sea of textual data.

Our findings not only underscore the efficacy of machine learning techniques in quantifying and understanding user sentiments but also shed light on the relationship between “ease of use” and “usefulness” with users’ intentions to embrace online food delivery apps. Moreover, this study goes beyond its immediate practical implications and has broader theoretical significance. It contributes to our understanding of technology adoption, artificial intelligence, and the ever-evolving dynamics of the business landscape. As we navigate the digital age, where user acceptance and engagement are pivotal to the success of technology-driven services, ¹⁷ This research contributes to the ongoing development of user sentiment analysis towards factors that contribute to users’ attitudes and decisions regarding the adoption of the online food delivery apps. The insights presented in this study may benefit organizations seeking to extract valuable information from user reviews but also provide a foundation for enhancing user acceptance and engagement in the digital era.

2. Literature review

2.1. Advancing theoretical novelty in AI-driven technology adoption

Recent advancements in artificial intelligence (AI) have contributed to significant changes in technological adoption across diverse domains. The surge in AI-driven systems, particularly in natural language processing (NLP) and machine learning (ML), have influenced user interaction paradigms. Mariani et al. ¹⁸ emphasize the rapid evolution of AI technologies, highlighting their burgeoning influence on modern-day business operations. The maturation of AI models, such as Generative Pre-trained Transformers (GPT), has facilitated the development of conversational AI, elevating customer experiences through human-like interactions. ¹⁹ ChatGPT represents one prominent example of this innovation, an OpenAI product offering AI-generated content, showcasing its potential to transform customer service, marketing strategies, and consumer engagement. ²⁰

In this epoch of AI-driven technological proliferation, the theoretical foundations of technology adoption models have garnered renewed interest. The TAM, initially introduced by Davis, ²¹ serves as a pivotal framework for comprehending user acceptance and adoption of technology. TAM’s fundamental tenets, perceived ease of use and perceived usefulness, have evolved in relevance and applicability within the context of AI-driven systems. Scholars increasingly highlight TAM’s adaptation to elucidate user perceptions and behaviors concerning AI technologies.^22,23 This evolution extends TAM’s utility from traditional technology domains to the forefront of AI applications, illuminating user attitudes and intentions towards AI-driven solutions across various sectors.^24,25

Moreover, the integration of AI technologies into daily life has raised ethical considerations and engendered challenges related to user trust and acceptance. ²⁶ Grasping users’ emotional responses and trust dynamics concerning AI-driven systems becomes imperative for successful adoption. Hossain and Rahman (2022) ¹ stress the importance of emotional reactions and sentiment mechanisms in AI adoption, emphasizing the necessity for comprehensive insights into user psychology and interaction patterns. Explainable AI (XAI) techniques have emerged as a crucial frontier in bolstering user trust by offering transparency and interpretability in AI systems. ²⁷ These advancements in AI-driven technology adoption underscore the need for a holistic understanding of theoretical frameworks and user-centric perspectives to navigate the evolving landscape of AI adoption and its profound societal implications. ²⁸

Despite the substantial growth in AI-driven technology adoption and the theoretical advancements in models like TAM, no literature has yet reported systematic integration of AI-generated content, derived from platforms like ChatGPT, into established theoretical frameworks. While prior studies have explored user perceptions in the context of AI technologies using traditional methods and theoretical models, a gap persists in comprehensively incorporating AI-generated content as a determinant of user attitudes and intentions within these models. This study aims to address this gap by incorporating AI-generated content, specifically keywords generated by ChatGPT, into the TAM framework. By integrating AI-generated content as a variable influencing perceived ease of use, perceived usefulness, and perceived intention to use, we bridge the gap between traditional theoretical models and the evolving landscape of AI-driven technology adoption. This approach provides insights into how user-generated AI content impacts user sentiments and intentions regarding AI-driven solutions, particularly within the domain of food delivery apps. Through statistical analyses and model validations, we illuminate the influence of AI-generated content on user perceptions, thereby contributing to a more comprehensive and contemporary understanding of technology adoption frameworks in the AI era.

2.2. ChatGPT’s text generation capabilities

In an ever-advancing digital landscape, businesses are actively exploring methods to incorporate AI-driven technology into their operations as the field of artificial intelligence (AI) continues to mature and grow in sophistication. ¹⁸ Among these technologies, chatbots have gained widespread popularity among companies worldwide, providing automated systems that effectively replicate human-to-human conversations through the application of natural language processing (NLP) techniques, thereby offering clients immediate assistance and support. ¹⁹ The primary objectives of the chatbot are to generate text and offer text-based recommendations.^19,29,30 Notably, the AI realm encompasses a sophisticated family of large language models (LLMs) known as Generative Pre-trained Transformers (GPT), which have been meticulously trained on extensive textual corpora. These models find application in a range of domains, including text summarization, sentiment analysis, chatbot functionality, and question answering. ³¹ It is imperative to acknowledge the pivotal role of the American AI research group OpenAI, ²⁹ who introduced ChatGPT, an AI-based language model that generates conversational responses based on textual cues, facilitated by a sophisticated algorithm. ²⁰ The advent of ChatGPT has propelled AI Generated Content (AIGC) into the spotlight, prompting users across diverse domains. The utilization of chatbots like ChatGPT offers businesses the potential to enhance customer service significantly, with the promise of better customer engagement, personalization, improved communication techniques, and cost-effectiveness. Furthermore, it affords valuable insights into consumer behavior. ²⁰ Notably, the implementation of ChatGPT may lead to significant changes in the marketing sector, revolutionizing the ways in which customers access information, make decisions, and how companies conceptualize, develop, and deliver personalized services and experiences. ²⁵

Artificially intelligent technologies endowed with transformative capabilities, such as ChatGPT, are capable of generating intricate text that is virtually indistinguishable from human-authored content in various settings. ²⁹ Operating on the principle of anticipating the next word based on context, ChatGPT, an AI-based generative language model created by OpenAI, produces high-quality text that closely resembles human composition. ³⁰ ChatGPT has been trained on an extensive and diverse dataset compiled from publicly available online sources, including webpages, books, articles, blogs, and forums. As a result, the model now possesses the ability to generate responses on a wide range of topics. ²⁹

The core objective of this study has been to harness ChatGPT’s text-generating capabilities to construct a comprehensive set of 100 keywords relating to the core components of the TAM, namely, ease of use, usefulness, and intention to use. These keywords are reflective of phrases users might employ while composing textual reviews of food delivery applications. A key facet of this research is to identify and analyze users’ perspectives on the ease of use, usefulness, and intention to use these food delivery apps, presenting a novel approach in this domain.

2.3. Core components and versatility of the TAM

The TAM, introduced by Davis, ²¹ stands as a widely acknowledged basis essential for realizing user acceptance of information technology. Central to TAM are the foundational constructs of perceived ease of use and perceived usefulness. ²² TAM and its expanded models delve into the intricate interplay between the system, the user, and actual use, taking a holistic perspective that considers system features, capabilities, and user motivation.^21,32 Core variables within TAM, such as perceived usefulness (PU) and perceived ease of use (PEOU), directly or indirectly elucidate outcomes, shaping positive attitudes toward technology adoption based on users’ beliefs in its utility and perceived ease of use.^3,33 In TAM, behavioral intention is driven by user attitude, which is formed by perceptions of a technology’s usefulness and ease of use. ²¹

In extending the TAM framework, our proposed study builds on the premise that customers’ perceived usefulness and ease of use toward online food delivery service app reviews play pivotal roles in shaping attitudes and influencing adoption intentions. TAM, itself an adaptation of Fishbein’s Theory of Reasoned Action (TRA), introduces perceived usefulness and perceived ease of use as principal constructs influencing behavioral intention to use. ²⁴ Researchers in information management systems frequently employ the TAM model to investigate the correlation between users’ subjective perception and behavioral intention. Prior research endeavors, such as Jo and Bang’s study ³⁴ on the continuance intention of enterprise resource planning (ERP) systems and Alyoussef’s ²³ exploration of the acceptance of flipped classrooms in higher education, underscore the versatility and enduring relevance of TAM across diverse technological domains.

Furthermore, TAM’s adaptability extends to explorations by Al-Emran ³⁵ into technology’s impact on economic, environmental, and social sustainability, introducing the Technology-Environmental, Economic, and Social Sustainability Theory (T-EESST). Studies by Putri et al. ²⁴ on financial technology acceptance in peer-to-peer lending and Nurse-Clarke and Joseph ³⁶ on technology acceptance among nursing faculty transitioning to online teaching during the COVID-19 pandemic showcase TAM’s efficacy in diverse contexts. Our study, recognizing TAM’s adaptability and demonstrated effectiveness in diverse technological domains, rigorously evaluated its core components in the context of customers’ perceptions and adoption intentions regarding reviews of online food delivery service apps. Grounded in the foundational constructs of perceived ease of use (PEU) and PU, TAM emerged as a robust framework for comprehending users’ attitudes and intentions toward technology adoption, contributing valuable insights to the broader literature on technology acceptance. Notably, our study innovatively applied TAM to users’ text reviews, employing advanced machine learning approaches for a nuanced analysis, thereby emphasizing the novel integration of established theory and modern analytical techniques.

2.4. User-generated reviews, ChatGPT, and machine learning: A holistic approach

Customer reviews wield significant influence over consumer decisions across industries, prompting researchers to employ diverse methodologies in understanding their dynamics and impact on businesses. Yuhsiang and Lichung ³⁷ investigated the interplay between user review characteristics and sales throughout different product life cycle stages, utilizing the Bass model to reveal the nuanced role of consumer heterogeneity. In the realm of online-to-offline commerce, Wan et al. ³⁸ explored how businesses respond to negative customer reviews, emphasizing the significant influence of apology strategies on customer trust and purchase intentions. Wan et al. ³⁸ employed text mining techniques to uncover determinants of customer satisfaction for grocery mobile apps, highlighting the importance of online reviews. Wu et al. ³⁹ investigated managerial responses to customer reviews in the hospitality industry, highlighting factors influencing a company’s decision to respond or not. Kim et al. ⁴⁰ proposed an answering framework based on customer reviews for accurate and prompt responses in online shopping contexts. Camilleri and Filieri ⁴¹ found that review credibility, content quality, and usefulness are key to how online reviews drive customer satisfaction and loyalty. Hossain and Rahman ¹ detected potential customers’ empathy behavior towards financial services, while Hossain and Rahman ¹¹ analyzed sentiment and prediction of insurance products’ reviews using machine learning approaches. Pashchenko et al. ² investigated the interplay between emotional expressions and normative judgments in hotel and travel reviews, employing a lexicon-based unsupervised learning approach to uncover their relationships.

These studies collectively underscore the multifaceted nature of customer reviews and their substantial impact on various business aspects, providing valuable insights. Moreover, several studies^1,2,37,41,42 focused on machine learning, revealing the potential of AI technologies like ChatGPT to generate text indistinguishable from human-authored content. Notably, machine learning enables sentiment detection, emotional aspects analysis, and other investigations undertaken by prior studies. This recognition, coupled with ChatGPT’s capability to produce text, motivates our study. While traditional data mining approaches have successfully extracted keywords based on frequency statistics and co-occurrence patterns, our methodology differs fundamentally. Rather than discovering patterns post-hoc from data, we begin a priori with established theory (TAM) and employ ChatGPT for theory-guided feature engineering. This represents a paradigm shift from extractive data mining to generative theory operationalization—using AI to synthesize potential linguistic manifestations of theoretical constructs that may not be frequent enough to emerge from traditional mining but are conceptually important.

To clarify the conceptual framework guiding this study, we define the key terms and their interrelationships. TAM serves as the foundational theory that identifies the core constructs influencing technology adoption. The constructs—perceived ease of use, perceived usefulness, and Behavioral Intention—represent the theoretical variables we aim to measure from user-generated content. The keywords generated by ChatGPT operationalize these abstract constructs into observable linguistic markers that can be systematically identified in review text. Finally, the hypotheses (H1 and H2) are testable predictions derived from TAM theory, specifying the expected relationships between these constructs. This hierarchical framework ensures transparency in how we move from abstract theory to measurable variables to empirical testing. In the rapidly evolving landscape of digital consumer services, online food delivery applications have become integral to modern living. This study introduces an approach by harnessing advanced machine learning techniques, specifically leveraging ChatGPT’s capabilities, to unravel users’ sentiments towards factors that contribute to users’ attitudes and decisions regarding the adoption of a particular technology in reviews. Focusing on the core tenets of the TAM—ease of use, usefulness, and intention to use—the research delves into the linguistic nuances of user-generated content. Through analysis of keywords recommended by ChatGPT, the study aims to identify latent markers within reviews that encapsulate users’ perceptions.

2.5. Hypotheses development

2.5.1. Perceived ‘ease of use’ and ‘intention to use toward food delivery apps

Previous studies show that how easy and useful a technology feels directly affects a user’s experience and their decision to use it. Emphasizing the significance of both perceived ease and perceived usefulness as crucial antecedents, Lewis and Sauro ⁴³ suggest that these factors play pivotal roles in shaping users’ attitudes and behavioral intentions, with perceived usefulness slightly outweighing perceived ease in influencing outcomes. Calisir and Calisir ⁴⁴ complement this perspective by underscoring the joint impact of perceived usefulness and perceived ease of use on behavioral intention. Their work highlights the interconnected nature of these constructs, emphasizing that users’ perceptions of both ease and utility significantly contribute to the formation of behavioral intentions. Alyoussef ²³ further supports this notion by demonstrating that perceived ease of use has a considerable impact on behavioral attitudes.

Building on these insights, our prior hypothesis posits that when users perceive technology as easy to use, they are more likely to form favorable opinions, thereby influencing their evaluations of utility. ²³ This aligns with the established belief that perceived ease of use reflects users’ expectations of technology being free from difficulties, contributing to positive evaluations of the information system’s usability.^21,24 Importantly, Alyoussef’s work ²³ also suggests that perceived usefulness represents an individual’s belief in decision-making. In the context of our study, this implies that users’ assessments of the utility of food delivery apps are intertwined with their perceptions of ease of use. Perceived ease of use and perceived usefulness serve as crucial antecedents, directly and indirectly influencing experiential and intentional outcomes; however, perceived usefulness demonstrates a somewhat more substantial effect.^43,44 This underscores the interconnected nature of these constructs in significantly forming users’ behavioral intentions, aligning with Alyoussef’s ²³ findings on the impact of perceived ease on behavioral attitudes. Consequently, we derive Hypothesis H1.

H1: The level of perceived ‘Ease of use’ in user-generated reviews will significantly correlate with their ‘Intention to use’ food delivery apps.

This hypothesis extends the existing literature by applying the conceptualization of perceived ease and perceived usefulness to the specific domain of online food delivery service apps, providing a focused lens through which to examine users’ attitudes and intentions in this context. Basically, this aids in gaining a deeper understanding of the pivotal factors influencing delivery platform services.

2.5.2. Perceived ‘usefulness’ and ‘intention to use’ toward food delivery apps

Continuing our exploration of users’ perceptions and intentions regarding online food delivery service apps, we turn our attention to the construct of perceived usefulness. The literature, as demonstrated by Putri et al., ²⁴ emphasizes that perceived usefulness reflects an individual’s belief in making decisions. In the context of food delivery apps, this implies that users’ evaluations of the utility of these platforms are essential determinants of their decision-making processes. Alyoussef ²³ provides further support for the importance of perceived usefulness by showcasing its considerable impact on behavioral attitudes. This aligns with the broader understanding that users’ beliefs in the utility of technology significantly shape their behavioral intentions. ⁴⁴ Moreover, Lewis & Sauro’s ⁴³ emphasis on the importance of perceived usefulness, somewhat more than perceived ease, adds weight to the argument that users’ perceptions of utility play a central role in shaping their experiential and intentional outcomes. It can be concluded that perceived usefulness is a person’s belief in making decisions. ²⁴ Building on these insights, we derive Hypothesis H2.

H2: The ‘usefulness’ of food delivery apps, as identified in user reviews, will significantly influence users’ ‘Intention to use’ these platforms.

This hypothesis underscores the pivotal role of perceived usefulness in shaping users’ attitudes and intentions specifically within the context of online food delivery service apps. By focusing on the identified usefulness in user-generated reviews, our study seeks to contribute nuanced insights into the factors that drive users’ decisions and intentions in adopting these platforms for their food delivery needs.

3. Method

The methodology presented below embodies a threefold innovation—in hypothesis operationalization, testbed design, and feature engineering—which we discuss in detail in Section 5.3. On November 10, 2023, we conducted web scraping to collect user reviews of the three most popular food delivery apps in the USA from the Google Play Store. The data collection process involved the development of a Python script specifically designed for web scraping. We gathered various data points from these reviews, including the review date, the full review text, the app name, the reviewer’s name, the star rating, and the number of thumbs-up reactions. To ensure the consistency of our dataset, we exclusively collected reviews written in English. In total, we obtained a dataset of 1,694,938 reviews. After removing missing values from the text reviews, we retained a total of 1,694,581 reviews for our subsequent analysis.

To prepare the data for analysis, we performed several preprocessing steps using Python within a Jupyter notebook environment. These steps included the removal of punctuation, stopwords, and any remaining missing values from the user reviews. We leveraged several Python libraries, including pandas, NRCLex, nltk, seaborn, sklearn, string, vaderSentiment, numpy, and matplotlib for data analysis and visualization.

Sentiment classification for the reviews was determined based on the star ratings provided by users. Specifically, we categorized reviews with 1-2 stars as representing negative experiences, those with 3 stars as neutral, and those with 4-5 stars as positive. Table 1 provides a comprehensive overview of the sentiment distribution within our dataset of user reviews for food delivery apps. The table presents two key columns, with one displaying the number of reviews and the other indicating the corresponding percentages for each sentiment class. Within this context, it is evident that a significant portion, approximately 67.74% of the total reviews (1,147,956 reviews), express a positive sentiment, signifying users’ positive experiences with the food delivery apps. About 27.59% of the total reviews (467,599 reviews) convey a negative sentiment, reflecting user dissatisfaction or negative experiences. Additionally, a smaller segment, approximately 4.66% of the total reviews (79,026 reviews), falls within the neutral sentiment category, suggesting mixed or ambivalent opinions. The final row of the table signifies the overall size of the dataset, totaling 1,694,581 reviews, equivalent to 100%, which serves as a reference point for understanding the relative distribution of sentiment classes throughout the dataset. In summary, Table 1’s numerical presentation offers valuable insights into the distribution of sentiments in user reviews for food delivery apps, facilitating a comprehensive understanding of user sentiment within the dataset.

Table 1.

Number of reviews and percentage for each sentiment class.

Sentiment	Number of reviews	Percentage
Positive	1147956	67.74
Negative	467599	27.59
Neutral	79026	4.66
Total	1694581	100

Additionally, Table 2 presents a comprehensive breakdown of sentiment distribution for user reviews across three prominent food delivery apps, namely DoorDash, Grubhub, and Uber Eats. For DoorDash, there are 128,898 negative reviews, 30,135 neutral reviews, and 478,130 positive reviews, resulting in a total of 637,163 reviews. Grubhub has 53,178 negative reviews, 12,060 neutral reviews, and 126,086 positive reviews, amounting to 191,324 reviews in total. Uber Eats received 285,523 negative reviews, 36,831 neutral reviews, and 543,740 positive reviews, with a total of 866,094 reviews. In sum, these figures depict the sentiments expressed in user reviews for each of the three food delivery apps.

Table 2.

Sentiment distribution by apps.

App	Number of negative reviews	Number of neutral reviews	Number of positive reviews	Total reviews
DoorDash	128898	30135	478130	637163
Grubhub	53178	12060	126086	191324
Uber_Eats	285523	36831	543740	866094
Total	467599	79026	1147956	1694581

The main goal of our study was to investigate the core components of TAM from the perspective of user-generated content in the context of food delivery apps. To achieve this, we employed a novel approach by utilizing ChatGPT to generate keywords that users might use when writing their reviews. ChatGPT provided us with 107 keywords for perceived ease of use, 124 for perceived usefulness, and 117 for perceived intention to use. We conducted a manual screening process based on three explicit criteria: (1) relevance to the TAM construct definition (keywords must clearly reflect user perceptions of ease, utility, or behavioral intention); (2) semantic non-redundancy (exact duplicates were removed, but near-synonyms were retained); and (3) contextual appropriateness for food delivery applications (keywords clearly referring to other domains were excluded). After manual selection, we retained the top 100 keywords for each category (presented in Table 3). Subsequently, we incorporated these keywords into our dataset, creating three distinct columns: ease of use, usefulness, and intention to use. In our approach, if any of the associated keywords for these variables were found in the text reviews and the sentiment was neutral or positive, we recorded a value of 1. Conversely, if the sentiment was negative, we assigned a value of -1, as we hypothesized that the presence of these keywords in negative reviews could indicate their use in a negative context. If no associated keywords were found, we recorded a value of 0. A sensitivity analysis comparing the neutral-as-positive (1) versus neutral-as-zero (0) coding scheme revealed no material difference in the regression coefficients (0.00% change), confirming the robustness of our coding choice.

Table 3.

100 keywords for ease of use, usefulness, and intention to use in food delivery service apps’ reviews generated by ChatGPT.

Perceived ease of use keywords	Perceived usefulness keywords	Perceived intention to use keywords
‘user-friendly’, ‘intuitive’, ‘simple’, ‘straightforward’, ‘convenient',‘effortless’, ‘easy to use’, ‘user-friendly interface’, ‘user-friendly design‘, ‘smooth’, ‘efficiency’, ‘convenient’, ‘intuitive design’, ‘user-friendly app‘, ‘user-friendly interface’, ‘user-friendly experience’, ‘easy navigation‘, ‘user-friendly features’, ‘user-friendly design’, ‘easy access’, ‘straightforward usability’, ‘user satisfaction’, ‘user experience’, ‘convenient functionality’, ‘intuitive usability’, ‘smooth operation’, ‘efficient design’, ‘straightforward interaction’, ‘convenient layout’, ‘effortless navigation’, ‘user-friendly layout’, ‘intuitive operation‘, ‘simple usage’, ‘convenient features’, ‘straightforward design’, ‘effortless experience’, ‘user-friendly functionalities’, ‘convenient operation’, ‘easy handling’, ‘efficiency features’, ‘intuitive user interface’, ‘simple design’, ‘user-friendly functions’, ‘straightforward usability', ‘convenient interface’, ‘effortless accessibility’, ‘user satisfaction features’, ‘user-centric’, ‘accessible’, ‘streamlined’, ‘hassle-free’, ‘simplified’, ‘ergonomic’, ‘responsive’, ‘painless’, ‘effortless usage’, ‘friendly interface’, ‘user-driven’, ‘clear', ‘user-focused’, ‘well-designed’, ‘easy-to-navigate’, ‘practical’, ‘helpful’, ‘user-oriented’, ‘self-explanatory’, ‘quick access’, ‘uncomplicated’, ‘trouble-free’, ‘conducive’, ‘effective’, ‘pleasant’, ‘efficient interface’, ‘straightforward features’, ‘comfortable’, ‘logically organized’, ‘intuitive functionality’, ‘user-first’, ‘user-savvy’, ‘time-saving’, ‘highly usable’, ‘user-responsive’, ‘no-brainer’, ‘intuitive operation’, ‘user-empowered’, ‘easygoing’, ‘accessible design’, ‘thoughtful design’, ‘user-engaging’, ‘user-helpful’, ‘smart design’, ‘hassle-free usage’, ‘welcoming’, ‘obvious’, ‘intuitively organized’, ‘ergonomic layout’, ‘pleasing’, ‘easily navigable’, ‘user-approved’, ‘user-driven features’	‘useful’, ‘helpful’, ‘beneficial’, ‘valuable’, ‘effective’, ‘productive’, ‘informative’, ‘practical’, ‘convenient’, ‘efficient’, ‘worthwhile’, ‘functional’, ‘handy’, ‘advantageous’, ‘essential’, ‘effective tool’, ‘useful features’, ‘valuable information’, ‘convenient solution’, ‘helpful resource’, ‘practical functionality’, ‘efficient features’, ‘beneficial features’, ‘valuable features’, ‘essential features’, ‘convenient features’, ‘practical features’, ‘efficient tool’, ‘productive features’, ‘effective resource’, ‘informative features’, ‘functional tool’, ‘handy features’, ‘advantageous tool’, ‘essential resource’, ‘valuable solution’, ‘effective solution’, ‘helpful solution’, ‘beneficial tool’, ‘practical resource’, ‘efficient solution’, ‘informative tool’, ‘productive tool’, ‘valuable resource’, ‘useful solution’, ‘convenient resource’, ‘effective functionality’, ‘user value’, ‘utility’, ‘usefulness factor’, ‘effectiveness’, ‘practicality’, ‘handiness’, ‘relevance’, ‘efficacy’, ‘useful resource’, ‘valuable features’, ‘valuable tool’, ‘convenience’, ‘functionality’, ‘useful information’, ‘beneficial functionality’, ‘convenient tool’, ‘valuable solution’, ‘efficient information’, ‘informative solution’, ‘functional features’, ‘helpful features’, ‘worthwhile information’, ‘practical information’, ‘handy information’, ‘advantageous information’, ‘essential information’, ‘efficient solution’, ‘productive solution’, ‘functional solution’, ‘informative features’, ‘valuable functionality’, ‘useful functionalities’, ‘convenient features’, ‘helpful functionalities’, ‘useful functions', ‘useful addition’, ‘beneficial aspect’, ‘valuable attribute’, ‘effective attribute’, ‘helpful feature’, ‘functional qualities’, ‘informative resources’, ‘convenient solutions’, ‘efficient characteristic’, ‘productive advantages’, ‘essential elements’, ‘user-friendly utilities’, ‘user satisfaction’, ‘user experience enhancement’, ‘convenient enhancements’, ‘functional enhancements’, ‘useful enhancements’, ‘helpful utilities’	‘plan to use’, ‘intend to use’, ‘will use’, ‘want to use’, ‘intention to use’, ‘consider using’, ‘wish to use’, ‘have to use’, ‘interested in using’, ‘going to use’, ‘future use’, ‘potential use’, ‘looking to use’, ‘aim to use’, ‘will continue using’, ‘use it more’, ‘use it again’, ‘used regularly’, ‘using it frequently’, ‘intend to keep using’, ‘future intention’, ‘future adoption’, ‘desire to use’, ‘interest in usage’, ‘aim to continue using’, ‘planning to adopt’, ‘future utilization’, ‘anticipate using’, ‘willing to use’, ‘will likely use’, ‘intend to utilize’, ‘consider future use’, ‘intent to adopt’, ‘intention to continue using’, ‘future application’, ‘intent to engage’, ‘aspiration to use’, ‘goal to use’, ‘expectation to utilize’, ‘purpose to use’, ‘prospect to use’, ‘future involvement’, ‘inclination to use’, ‘inclined to utilize‘, ‘consideration to engage’, ‘intention to utilize’, ‘likelihood of usage’, ‘likelihood of adoption’, ‘intention to apply’, ‘prospect to adopt’, ‘consider future adoption’, ‘future intention to use’, ‘plan to continue using’, ‘future utilization intention’, ‘intention to apply it’, ‘prospective use’, ‘intending to use’, ‘preference for usage’, ‘wants to continue using’, ‘aspirations to use’, ‘desires to use’, ‘future usage plan’, ‘considering using in the future’, ‘aims to use it more’, ‘potential to keep using’, ‘intent to continue using’, ‘inclination to utilize more’, ‘potential to use it again’, ‘intending to use it frequently’, ‘prospects to use regularly’, ‘future intention to keep using’, ‘planning to use it more’, ‘will likely use it again’, ‘desiring to use it frequently’, ‘future intentions to keep using’, ‘future aspirations to use’, ‘intentions to continue using’, ‘expecting to use it again’, ‘intending to keep using regularly’, ‘future intent to use frequently’, ‘prospects to keep using it again’, ‘intending to use it continually’, ‘potential to use it repeatedly’, ‘intention to apply it regularly’, ‘prospective usage’, ‘intending to use it repeatedly’, ‘future usage planning’, ‘considering repeated use’, ‘wants to use it regularly’, ‘aspires to use it repeatedly’, ‘desires to use it continually’, ‘future usage intent’, ‘intentions to apply it regularly’, ‘considering it for regular use’, ‘aims to use it repeatedly’, ‘potential to use it continually’, ‘intent to apply it repeatedly’, ‘anticipations for regular usage’, ‘intentions for continued use’, ‘intent to engage regularly’, ‘aspires to keep using’

Following this, we conducted a thorough analysis to examine the correlations between ease of use and usefulness with intention to use. Our ultimate goal was to gain insights into the TAM model, specifically to understand the impact of ease of use and usefulness on users’ intentions to use food delivery apps. To achieve this, we employed Ordinary Least Squares (OLS) Regression, a robust statistical method for investigating these relationships. This unique approach enabled us to explore and understand user behavior and sentiment towards factors that contribute to users’ attitudes and decisions regarding the adoption of a particular technology in the context of food delivery apps, shedding new light on the applicability of the TAM model to this domain.

To ensure full reproducibility and transparency, we have made the complete data processing pipeline and analysis scripts publicly available on GitHub. ¹ The repository contains data preprocessing scripts for cleaning raw Google Play reviews, complete documentation of ChatGPT prompts used for TAM keyword generation, feature engineering code implementing keyword matching and sentiment analysis (VADER and AFINN), statistical analysis scripts (Jupyter notebooks) for OLS regression and figure generation, editable research framework diagram source files, and a comprehensive README.md file with environment setup and execution instructions.

The complete dataset, including raw and processed data, totals approximately 1.52 GB. Due to file size limitations, the final processed data—comprising all reviews, sentiment scores, and TAM construct indicators—has been deposited in the Mendeley Data repository and is permanently accessible at mendeley. ² The preprocessing scripts in the GitHub repository enable researchers to apply the same cleaning and feature engineering procedures to independently collected raw data if desired. This dual-repository approach ensures that both the analytical methods and final processed data are fully available for verification, replication, and future research.

4. Results/discussion

In our research, we initially assigned sentiment to user reviews based on star ratings, a conventional approach.^1,2 However, we sought to enhance the accuracy of our sentiment analysis by incorporating two machine learning lexicons: VADER and AFINN. This approach allowed us to delve deeper into the correctness of sentiment classifications based on star ratings. Additionally, we evaluated the total word count in text reviews, recognizing that it serves as an indicator of sentiment correctness, with higher word counts often associated with negative sentiment. Table 4 encapsulates VADER and AFINN Scores, and Word Count by Sentiment Category.

Table 4.

Mean VADER and AFINN scores and mean word count by sentiment class.

Sentiment	vader_compound	vader_neg	vader_neu	vader_pos	afinn_score	word_count
Negative	-0.245642	0.155454	0.788421	0.056124	-1.603883	32.706462
Neutral	0.095455	0.079600	0.758509	0.161891	0.697846	23.046440
Positive	0.475582	0.019020	0.500246	0.480734	2.667082	8.811262

The VADER Compound Score was used to gauge the overall sentiment within the text reviews. Negative sentiment reviews exhibited an average VADER Compound Score of -0.246, indicating a predominantly negative sentiment. Neutral reviews, with a score of 0.095, showed a slightly positive sentiment, while positive reviews displayed a considerably higher score of 0.476, reflecting predominantly positive sentiment. The VADER Negative, Neutral, and Positive Scores provided insights into the level of negativity, neutrality, and positivity in the reviews. Negative sentiment reviews had a relatively high VADER Negative Score of 0.155. Neutral reviews exhibited a more balanced distribution across negativity, neutrality, and positivity. In contrast, positive sentiment reviews showed a notably low VADER Negative Score of 0.019, signifying a strong lack of negativity.

The AFINN Score, another sentiment measure, indicated the sentiment of the reviews. Negative sentiment reviews had an average AFINN Score of -1.604, suggesting a strong negative sentiment. Neutral sentiment reviews showed an AFINN Score of 0.698, indicating nearly neutral sentiment. Positive reviews had a high average AFINN Score of 2.667, signifying positive sentiment.

Lastly, the word count, representing the total number of words in the reviews, provided additional insights. Negative sentiment reviews had an average word count of about 32.7 words, reflecting more extended and detailed feedback. Neutral sentiment reviews had an average word count of approximately 23.0 words, indicating moderate-length reviews. In contrast, positive sentiment reviews had the lowest word count, with an average of 8.8 words, suggesting concise and to-the-point feedback.

Our utilization of VADER and AFINN, coupled with the word count analysis, contributes to a more comprehensive understanding of sentiment in user reviews. The alignment of these additional techniques with sentiment classifications based on star ratings underscores the accuracy and validity of our sentiment assessment within the study.

Table 5 provides an overview of the average sentiment, VADER Score, AFINN Score, and Word Count for three prominent food delivery apps: DoorDash, Grubhub, and Uber Eats. The “Sentiment” column shows the average sentiment value for each app, reflecting the overall emotional tone in user reviews. DoorDash has the highest average sentiment at 0.548105, indicating a generally positive sentiment. Grubhub follows with an average sentiment of 0.381071, while Uber Eats has the lowest average sentiment at 0.298140.

Table 5.

Average sentiment, VADER score, AFINN score, and word count.

Apps	Sentiment	vader_compound	vader_neg	vader_neu	vader_pos	afinn_score	word_count
DoorDash	0.548105	0.330321	0.044074	0.587495	0.368431	1.720996	15.434515
Grubhub	0.381071	0.267257	0.057103	0.644045	0.298852	1.452724	18.138493
Uber_Eats	0.298140	0.204397	0.071363	0.583443	0.345194	1.145801	16.078032

The table also includes VADER and AFINN scores, which further highlight sentiment analysis results. Specifically, DoorDash exhibits a higher VADER Compound score (0.330321), suggesting a generally positive sentiment. Grubhub and Uber Eats have slightly lower VADER Compound scores (0.267257 and 0.204397, respectively), indicating somewhat less positive sentiments. Moreover, the AFINN Score, representing sentiment, is presented in the table. It reveals that DoorDash has the highest AFINN Score (1.720996), indicating a relatively positive sentiment. Grubhub and Uber Eats follow with AFINN Scores of 1.452724 and 1.145801, respectively. The Word Count column provides insights into the length of user reviews for each app, with DoorDash reviews having an average word count of 15.43, Grubhub reviews averaging 18.14 words, and Uber Eats reviews containing approximately 16.08 words on average.

In our research paper, Table 6 presents a comprehensive breakdown of the findings related to specific keywords associated with perceived ease of use, perceived usefulness, and perceived intention to use within user-generated reviews of food delivery apps. These results are vital in unraveling the intricate relationship between the presence of these keywords and the sentiments expressed by users.

Table 6.

Quantity of reviews featuring keywords for ease of use, usefulness, and intention to use.

Assigned values	Perceived ease of use	Perceived usefulness	Perceived intention to use
0 (keywords not found)	1121754	1148448	1225472
-1 (keywords found but sentiment negative)	467599	467599	467599
1 (keywords found and sentiment positive)	105228	78534	1510

The “Assigned values” column classifies reviews into three distinct categories, each shedding light on different facets of user sentiment. In the “0” category, the specified keywords were notably absent from the reviews. Within this category, there were 1,121,754 reviews associated with perceived ease of use, 1,148,448 pertaining to perceived usefulness, and 1,225,472 relevant to perceived intention to use. This absence of keywords does not necessarily imply that the majority of reviews are devoid of any keywords. It’s worth emphasizing that, even though these reviews do not feature the precise keywords associated with the designated variables, they might include other keywords related to different variables. For example, within the 1,121,754 reviews tied to perceived ease of use that do not contain the specified keywords, it is entirely possible that other keywords connected to perceived usefulness and perceived intention to use are present, contributing to a more intricate perspective within those reviews. The “-1” category signifies that the keywords were identified in the reviews, but the sentiment associated with them was notably negative. This classification included 467,599 reviews for perceived ease of use, 467,599 for perceived usefulness, and 467,599 for perceived intention to use, underscoring the impact of these keywords on negative sentiments. The “1” category indicates that the keywords were present in the reviews, and the sentiment linked to them was largely positive. This segment contained 105,228 reviews for perceived ease of use, 78,534 for perceived usefulness, and 1,510 for perceived intention to use, highlighting the role of these keywords in shaping positive sentiments. These numerical findings enrich our understanding of the complex dynamics between these keywords and the sentiments expressed by users in their reviews of food delivery apps. They offer a valuable perspective on how the presence of these keywords influences user sentiment in this context. These findings suggest that discerning users’ perceptions of ease of use, usefulness, and intention to use online food delivery apps through keyword analysis is viable within this context.

Furthermore, we have calculated the correlations values, Table 7 presents the results of a correlation analysis between the variables ‘perceived intention to use’, ‘perceived ease of use’, and ‘perceived usefulness’. The correlation values between these variables are all positive, indicating positive relationships. Notably, ‘perceived ease of use’ and ‘perceived usefulness’ exhibit high positive correlations, with a value of approximately 0.9225, suggesting a strong positive relationship. “Intention to use” also demonstrates strong positive correlations with both “ perceived ease of use” and “ perceived usefulness,” with values around 0.8959 and 0.9141, respectively. These findings highlight the importance of “ perceived ease of use” and “ perceived usefulness” in predicting user perceived intentions to use food delivery apps. The positive correlations suggest that as these factors increase, user intentions to use the apps tend to increase as well, underscoring their influential role in user decision-making. These findings provide support for both H1 and H2.

Table 7.

Correlation analysis.

​	Perceived intention to use	Perceived ease of use	Perceived usefulness
Perceived intention to use	1.000000	0.895901	0.914064
perceived ease of use	0.895901	1.000000	0.922520
perceived usefulness	0.914064	0.922520	1.000000

To formally assess multicollinearity, we calculated the Variance Inflation Factor (VIF) for both independent variables. The VIF values were 6.713 for both perceived ease of use and perceived usefulness. These values exceed the commonly recommended threshold of 5 but are below the more conservative threshold of 10, indicating moderate multicollinearity. This level of multicollinearity is theoretically expected given TAM’s foundational premise that perceived ease of use directly influences perceived usefulness. Despite this, both coefficients remain highly significant (p < 0.001) with narrow confidence intervals, suggesting sufficient statistical power to overcome the multicollinearity.

In our research, we recognize that while correlations offer valuable insights into the relationships between variables, they do not inherently answer questions related to causality or influence. To address this, we conducted an in-depth Ordinary Least Squares (OLS) Regression Analysis. This regression analysis allows us to explore and quantify the causal relationships between the variables of interest.

Table 8 presents the detailed results of this OLS Regression Analysis, which sheds light on the impact of the independent variables, “ perceived ease of use” (Coefficient: 0.2933) and “perceived usefulness” (Coefficient: 0.5079), on the dependent variable, “ perceived intention to use” of food delivery apps. Several key components from the table are worth noting. The high R-squared value (R-squared: 0.854) suggests that our regression model effectively explains approximately 85.4% of the variance in users’ intentions to use food delivery apps. This demonstrates the model’s robustness in capturing the variations in user intentions. Furthermore, the coefficients associated with “ perceived ease of use” and “perceived usefulness” are positive, indicating that as these factors increase, users’ intentions to use the app also increase.

Table 8.

OLS regression results.

Variable	Coef	Std err	t	P>|t|	[0.025	0.975]
const	-0.0957	0.000	-664.937	0.000	-0.096	-0.095
perceived ease of use	0.2933	0.001	465.031	0.000	0.292	0.295
perceived usefulness	0.5079	0.001	773.407	0.000	0.507	0.509

Note. R² = 0.854 (Adj. R² = 0.854), N = 1,694,581. Model diagnostics: Omnibus = 822864.667 (p < 0.001), Durbin-Watson = 1.978, JB = 6.061e+06 (p < 0.001).

As a result, both hypotheses are substantiated in this study. H1 suggests that the level of perceived ease of use in user-generated reviews significantly correlates with their perceived intention to use food delivery apps. Concurrently, H2 proposes that the ‘ perceived usefulness’ of food delivery apps, as identified in user reviews, significantly influences users’ ‘ perceived intention to use’ these platforms. The findings affirm the interconnected relationships between perceived ease of use, perceived usefulness, and users’ intentions in the context of online food delivery service apps.

Additionally, these positive coefficients (Ease of use: 0.2933, usefulness: 0.5079), supported by low p-values (p-values: 0.000), signify the statistical significance of these relationships. Moreover, our OLS regression analysis provides evidence consistent with that “perceived ease of use” and “perceived usefulness” have a significant and positive influence on users’ intentions to use food delivery apps. These findings suggest several implications for app developers and marketers, offering guidance on how to enhance user acceptance and engagement.

5. Applications

5.1. Managerial and operational applications

This research offers a framework for a data-driven operational and strategic management paradigm, moving beyond traditional, often slow, market research cycles. The engineered system translates unstructured user-generated content into quantifiable, actionable metrics for perceived ease of use, perceived usefulness, and intention to use. For senior management and product leaders, this enables a shift from reactive problem-solving to proactive strategy formulation. A direct application is in Resource Allocation and Product Roadmapping; by identifying which specific aspects of ”ease of use” (e.g., navigation, checkout process) or “usefulness” (e.g., tracking accuracy, restaurant variety) are most strongly correlated with user intention, organizations can make evidence-based decisions on where to invest development resources for maximum impact on user retention and acquisition. This systematic approach mitigates the risks associated with subjective decision-making and ensures that engineering efforts are directly aligned with customer-driven value propositions.

Furthermore, the framework serves as a continuous Quality Assurance and Operational Benchmarking tool. Instead of relying on sporadic feedback, managers can implement this analytical pipeline to monitor these key performance indicators (KPIs) in real-time, tracking them against updates, marketing campaigns, or competitor movements. For instance, a drop in the “ease of use” score following a new app interface release provides a precise, quantifiable alert, allowing for rapid iteration and correction—a core principle of agile and lean engineering management. This methodology also offers new possibilities for strategic marketing and Communication. Marketing departments can move beyond generic messaging to craft campaigns that authentically highlight the specific utility and usability features that users themselves value and vocalize, as identified by the AI-driven keyword analysis. This ensures that external communications are deeply resonant and credible, enhancing the efficiency of customer acquisition funnels and strengthening brand positioning around proven user benefits.

5.2. Theoretical and integrative applications

This study makes a theoretical contribution by demonstrating a viable pathway for the operationalization of established psychological and sociological models within an engineering and business intelligence context. It extends the TAM beyond its traditional survey-based methodology, validating its core constructs through the organic, unsolicited language of millions of users. This not only reinforces the model’s robustness but also transposes it into the digital age, demonstrating its relevance for analyzing contemporary software-driven services. The framework’s flexibility also allows for future integration of additional theoretical dimensions, including privacy protection and security resilience—factors that have been shown to significantly influence user acceptance in mobile application contexts. ⁴⁵ The research bridges the long-standing gap between qualitative theory and quantitative big data analytics, presenting a novel Integrated Theoretical Framework where machine learning acts as the interpreter between human sentiment and managerial theory. This may inform the future application of other theoretical models, such as Service Quality (SERVQUAL) or Expectation-Confirmation Theory (ECT), to user-generated content at scale, opening a new avenue for theory-driven data science in business research. Recent studies have further demonstrated the power of integrating deep learning with established marketing models, such as the AIDA framework, to predict consumer intentions from online reviews with high accuracy, reinforcing the value of theory-driven AI analytics in digital service contexts. ⁴⁶

Moreover, the study contributes to the theoretical discourse on Human-AI Collaboration in Knowledge Management. The use of ChatGPT for domain-specific keyword generation represents a novel methodology for feature engineering in natural language processing. It demonstrates how generative AI can be leveraged not as a black-box solution, but as a collaborative tool to encode human-understandable theoretical constructs into a machine-readable format, thereby enhancing the interpretability and validity of the resulting analysis. This approach illustrates one possible method for how AI can be systematically integrated into the research lifecycle to manage and extract meaning from vast knowledge repositories—a critical challenge in the modern information economy. Complementing our findings, recent systematic analyses of conversational AI in marketing have identified key thematic clusters—including consumer engagement, sentiment analysis, and technology adoption—that underscore the transformative potential of AI in shaping consumer experiences. ⁴⁷ By successfully merging TAM with advanced AI analytics, this work lays the groundwork for a more dynamic and scalable approach to testing and refining behavioral theories in real-world business environments.

5.3. Methodological and engineering applications

Returning to the threefold methodological innovation introduced in Section 3, we now elaborate on its significance. First, in hypothesis generation, we use ChatGPT to operationalize established theoretical constructs (TAM) into comprehensive keyword sets, bridging the gap between abstract theory and empirical measurement. Second, in testbed design, we integrate AI-generated keywords with multi-lexicon sentiment validation (VADER, AFINN) and robust statistical modeling into a complete, replicable analytical framework. Third, in experiment execution, we employ ChatGPT for theory-guided feature engineering—creating the independent variables (PEOU, PU) that, when validated through OLS regression, explain 85.4% of the variance in users’ intention to use. This threefold contribution distinguishes our work from standard LLM applications that focus on direct operational tasks rather than methodological innovation for theory testing.^48–50

Recent applications of large language models have demonstrated their utility in diverse operational contexts, including healthcare appointment systems, ⁴⁸ semantic database management, ⁴⁹ and clinical implementation studies. ⁵⁰ However, these applications typically employ LLMs for direct task execution rather than as methodological tools for operationalizing and testing behavioral theories. Our study differs fundamentally by leveraging ChatGPT not for an operational task, but for theory-guided feature engineering—translating abstract TAM constructs into measurable linguistic markers at scale. This positions our contribution at the intersection of AI methodology and management theory validation.

Unlike traditional keyword mining approaches that rely on frequency-based extraction, our methodology employs theory-guided generative AI to operationalize abstract constructs into measurable linguistic markers. From a methodological standpoint, this study engineers and validates a replicable Business Intelligence Pipeline for converting qualitative textual data into structured, quantitative business insights. The core innovation lies in the development of a scalable, automated process for feature identification, moving away from manual, subjective coding of text data, which is infeasible at this volume. The methodology—encompassing AI-assisted keyword generation, sentiment analysis with multiple lexicons (VADER, AFINN), and robust statistical modeling (OLS regression)—provides a comprehensive and rigorous blueprint. This end-to-end pipeline is a significant contribution to the field of Engineering Management and Decision Support Systems, offering a tangible tool for organizations to harness their own user feedback data systematically. It demonstrates how engineering principles can be applied to the management of information systems to create sustained competitive advantage.

This methodological framework is highly generalizable and can be adapted to a multitude of other domains within the engineering business landscape. For example, the same pipeline could be deployed to analyze customer support tickets to automate the identification of systemic product flaws, to scan social media for emerging competitive threats or market trends, or to evaluate employee feedback on internal enterprise software. The approach provides a clear methodology for Operationalizing User-Generated Data across various business functions, from supply chain management (e.g., analyzing supplier communications) to product development (e.g., mining idea portals). By detailing this process, the study empowers other researchers and practitioners to replicate and refine this methodology, contributing to evolving standards for data-driven, evidence-based management in the digital era.

6. Conclusion and implications for engineering management

This study has developed and tested a novel business intelligence framework that synergizes advanced artificial intelligence with established management theory to decode the drivers of technology adoption from large-scale user-generated content. By leveraging ChatGPT to operationalize the core constructs of the TAM, we developed a scalable and replicable methodology for transforming unstructured text from 1.6 million app reviews into quantifiable metrics for perceived ease of use, perceived usefulness, and intention to use. The application of robust sentiment analysis and regression modeling indicated strong, statistically significant relationships between these constructs within the highly competitive online food delivery market. This finding not only revalidates the enduring relevance of TAM in a modern digital context but, more importantly, demonstrates the efficacy of the proposed AI-driven pipeline as a powerful tool for strategic decision-making. A primary contribution of this work is its methodological engineering, offering organizations a potential blueprint to systematically audit user experience, prioritize feature development based on empirical evidence, and align operational strategies with authentic customer sentiment. For engineering managers and business strategists, this framework moves beyond traditional, lagging indicators to offer a dynamic, data-driven system for diagnosing user acceptance and predicting engagement, thereby contributing to evolving standards for evidence-based management in the digital services sector.

7. Limitations and avenues for future research

While this research presents certain advancements, several limitations suggest directions for future scholarly and practical inquiry. The study’s scope is bounded by methodological constraints, particularly in its AI-driven keyword generation process. The dependency on a single prompt for ChatGPT and the static nature of the keyword set introduce potential biases and limit the model’s adaptability to evolving language patterns. Future research should systematically investigate prompt engineering strategies to optimize keyword relevance and explore methods for creating dynamic keyword libraries that evolve with language trends and new product features. Furthermore, the cross-sectional nature of the data provides only a snapshot in time, unable to capture the dynamic evolution of user perceptions in response to app updates or market shifts, suggesting the need for longitudinal studies to understand temporal patterns in technology acceptance.

The generalizability of our findings is constrained by the exclusive reliance on English-language reviews from the Google Play Store, which limits the demographic and technological diversity of the sample. This restriction may limit the generalizability of findings across cultural contexts or platform-specific nuances in user behavior that could significantly impact technology adoption models. Future work should expand this scope by incorporating data from iOS platforms and multiple languages, enabling a cross-cultural and cross-platform comparative analysis that would enhance the robustness of the framework and provide more nuanced insights for global expansion strategies. Such expansion would also allow for the examination of how socioeconomic factors and regional market characteristics influence the perception and adoption of digital services.

From a theoretical perspective, the current framework, while powerful, focuses on a streamlined TAM structure that may not capture the full complexity of user decision-making processes. To enhance its predictive power and business relevance, future iterations of this engineered system should integrate a broader set of variables from complementary theoretical models. Incorporating constructs such as perceived risk from the Unified Theory of Acceptance and Use of Technology (UTAUT) or service quality metrics would provide a more holistic view of the user acceptance landscape. Furthermore, as highlighted in the mobile applications literature, privacy protection, security, and resilience of security mechanisms are increasingly recognized as critical determinants of user acceptance and trust. ⁴⁵ Future research should explicitly integrate these privacy and security constructs into the theoretical framework, as user concerns about data protection may significantly influence perceived usefulness and intention to use, particularly in mobile food delivery applications where sensitive payment and location data are routinely shared. Preliminary work in this direction has already demonstrated the value of integrating service quality dimensions—such as those from the SERVQUAL model—with deep learning to analyze customer sentiment in food delivery reviews, suggesting a promising pathway for extending our TAM-based framework. ⁵¹ Additionally, enriching the analysis with user demographic or behavioral data could enable sophisticated customer segmentation, allowing for highly targeted and personalized operational strategies that account for varying user preferences and needs across different market segments. Also, the current study employed a single version of ChatGPT (GPT-3.5) for keyword generation. Future research should enhance reliability by incorporating multiple LLM versions (e.g., GPT-4, GPT-4o) and employing cross-validation techniques to assess consistency across model versions and identify potential version-specific biases.

In essence, this study offers a framework that may inform future paradigms in engineering business management, where AI-driven analysis of user-generated content informs strategic decision-making. The identified limitations do not undermine the value of the current system but rather delineate a clear and exciting research agenda that bridges computer engineering, data science, and business management. By addressing these challenges, future research can further refine this methodology, expanding its precision, scope, and applicability to empower organizations in an increasingly data-driven and competitive global marketplace, ultimately leading to more responsive and user-centric digital service ecosystems.