DESIGN, IMPLEMENTATION, AND EVALUATION OF A SERVERLESS

AI-POWERED CHATBOT FOR E-COMMERCE IT HELPDESK

SUPPORT

USING AMAZON LEX, AWS LAMBDA, AND DYNAMODB

BY

[STUDENT NAME]

[STUDENT ID]
A PROJECT SUBMITTED IN PARTIAL FULFILLMENT OF THE

REQUIREMENTS FOR THE DEGREE OF

[DEGREE NAME]

DEPARTMENT OF [DEPARTMENT NAME]

FACULTY OF [FACULTY NAME]

[UNIVERSITY NAME]

[MONTH, YEAR]
 TABLE OF CONTENTS

TOC \h \o "1-3"
INTRODUCTION

The rapid expansion of global e-commerce has fundamentally transformed how businesses

interact with customers and manage support operations. As online retail platforms grow in scale

and complexity, the volume and diversity of customer support requests have increased

exponentially. Modern e-commerce enterprises must address a wide spectrum of user inquiries

ranging from simple order tracking and product questions to complex technical issues involving

payment processing failures, account authentication problems, system errors, and platform
functionality challenges.

Traditional human-staffed helpdesk systems, while effective for delivering personalized support,

face mounting challenges related to scalability, operational costs, service availability, and

response consistency. During peak demand periods such as promotional campaigns, holiday

shopping seasons, or unexpected platform outages, support teams often become overwhelmed,

leading to extended wait times, inconsistent service quality, and diminished customer

satisfaction. Furthermore, maintaining 24/7 support coverage across multiple time zones requires

substantial staffing investments that many organizations find economically challenging.

Recent technological advances in artificial intelligence, particularly in natural language

processing and machine learning, have enabled the development of sophisticated conversational

agents capable of understanding human language, maintaining contextual awareness, and

providing intelligent responses. These AI-powered chatbots offer potential solutions to

scalability and availability challenges by providing instant, consistent, and continuous assistance

while reducing the workload on human support agents.

Concurrently, the evolution of cloud computing architectures, especially serverless computing

paradigms, has created new opportunities for deploying AI applications efficiently and cost-

effectively. Serverless platforms abstract infrastructure management complexities, enable

automatic scaling in response to demand fluctuations, and implement pay-per-use pricing models

that align costs with actual utilization. Amazon Web Services (AWS) provides a comprehensive
ecosystem of services specifically designed for building conversational AI systems, including

Amazon Lex for natural language understanding, AWS Lambda for serverless compute

execution, and Amazon DynamoDB for scalable data persistence.

The convergence of advanced AI capabilities and serverless cloud architectures presents

opportunities to develop intelligent, scalable, and cost-efficient IT helpdesk solutions tailored to

e-commerce environments. However, despite the availability and maturity of these technologies,

significant gaps exist in academic literature documenting comprehensive implementations,

rigorous evaluations, and practical deployment experiences of serverless AI-powered chatbots

specifically designed for IT helpdesk support in e-commerce contexts.

This research addresses these gaps through systematic design, implementation, and evaluation of

a serverless AI-powered chatbot for e-commerce IT helpdesk support. The study explores how

Amazon Lex, AWS Lambda, and DynamoDB can be integrated to create a production-oriented

chatbot system capable of handling complex multi-turn conversations, maintaining persistent

context, managing support ticket workflows, and escalating appropriately to human agents when

necessary.
BACKGROUND OF THE STUDY

The concept of automated conversational agents has evolved significantly since the early rule-

based systems of the 1960s. ELIZA, developed by Joseph Weizenbaum at MIT in 1966,

demonstrated that relatively simple pattern-matching algorithms could create the illusion of

understanding human language. However, these early systems were fundamentally limited by

their reliance on predefined rules and inability to learn from interactions or adapt to linguistic

variations.
The advent of statistical natural language processing in the 1990s marked a significant

advancement, enabling systems to learn language patterns from data rather than relying

exclusively on hand-crafted rules. Machine learning approaches including support vector

machines, decision trees, and probabilistic models improved intent recognition and response

selection capabilities. However, these systems still struggled with contextual understanding,

multi-turn conversations, and handling the complexity and ambiguity inherent in natural human

language.

The deep learning revolution of the 2010s transformed the landscape of conversational AI.

Neural network architectures, particularly recurrent neural networks and subsequently

transformer models, demonstrated unprecedented capabilities in language understanding and

generation. Pre-trained language models such as BERT, GPT, and their successors achieved

near-human performance on numerous natural language processing benchmarks, opening new

possibilities for sophisticated conversational systems.

In parallel with AI advances, cloud computing evolved from simple infrastructure-as-a-service

offerings to sophisticated platform services abstracting significant operational complexity. The

serverless computing paradigm, popularized by AWS Lambda's launch in 2014, introduced

function-as-a-service models where developers could deploy code without managing servers,

operating systems, or runtime environments. This abstraction proved particularly well-suited to

event-driven applications such as chatbots, where discrete user messages trigger independent

processing operations.

Amazon Lex, introduced in 2016, brought together AWS's cloud infrastructure expertise with

advanced natural language understanding capabilities initially developed for Amazon Alexa. Lex

provides managed services for speech recognition, natural language understanding, intent

classification, entity extraction, and dialogue management, enabling developers to build

conversational interfaces without implementing these complex components from scratch.

The integration of Lex with Lambda and DynamoDB creates a powerful platform for
conversational AI applications. Lambda provides flexible, scalable compute for business logic

execution, while DynamoDB offers low-latency, highly available data storage for conversation

state, user profiles, and historical interactions. This combination addresses key technical

challenges in chatbot development including natural language understanding, business logic

execution, and state management.

Despite these technological capabilities, significant challenges remain in deploying effective IT

helpdesk chatbots for e-commerce environments. Technical support conversations often require

diagnostic questioning across multiple turns, access to customer and order data, integration with

ticketing systems, and appropriate escalation to human agents for complex issues. Furthermore,

production deployments must satisfy stringent requirements for response latency, availability,

security, and cost-effectiveness that proof-of-concept implementations often overlook.

This research builds upon the foundation of AI and cloud computing advances while addressing

practical deployment challenges often neglected in academic literature. By implementing and

evaluating a comprehensive serverless chatbot system specifically designed for e-commerce IT

helpdesk support, this study contributes empirical insights into the capabilities, limitations, and

practical considerations of deploying such systems in production environments.

CHAPTER ONE: INTRODUCTION

1.1 Background of the Study

The global expansion of e-commerce has led to a substantial increase in both the volume and

complexity of customer support requests. Modern e-commerce platforms must address diverse

user issues, including order tracking, payment failures, account authentication problems, system

errors, and platform-related challenges. Additionally, internal IT support is frequently required to

maintain the customer-facing systems and infrastructure that support online retail operations.
Traditional human-staffed helpdesk systems, while effective in many contexts, face growing

challenges related to scalability, cost, and availability. As businesses expand, support teams often

become overwhelmed during peak demand periods such as flash sales, holiday promotions, or

platform outages. These challenges result in delayed responses, inconsistent service quality,

increased operational costs, and reduced customer satisfaction (Shawar & Atwell, 2007).

Recent advances in artificial intelligence (AI), particularly in natural language processing (NLP),

have enabled the development of conversational agents capable of understanding and responding

to human language in real-time. These chatbots can provide instant, consistent, and continuous

assistance, reducing the workload on human agents while improving customer experience

(Adamopoulou & Moussiades, 2020).

Cloud computing has further enhanced the feasibility of implementing such systems. Serverless

computing, in particular, allows developers to deploy applications without managing servers,

offering automatic scaling, high availability, and pay-per-use pricing. Amazon Web Services

(AWS) provides a comprehensive ecosystem for building serverless conversational systems,

notably through Amazon Lex for natural language understanding, AWS Lambda for backend

logic execution, and Amazon DynamoDB for scalable NoSQL data storage (Amazon Web

Services, 2024).
The integration of these services enables the development of intelligent, scalable, and cost-

efficient AI-powered chatbots capable of handling complex IT helpdesk workflows in e-

commerce environments. Such systems can manage multi-turn conversations, retain user context,

log interactions, and integrate with backend workflows such as ticketing systems. However,

despite the availability of these technologies, there remains limited academic literature

documenting comprehensive implementations and evaluations of serverless chatbot-based IT

helpdesk systems specifically designed for e-commerce contexts.

1.2 Statement of the Problem

Many e-commerce organizations face difficulties in delivering timely, reliable, and scalable IT

helpdesk support, particularly outside standard business hours. Human-operated helpdesks are

constrained by staffing limitations, operational costs, and the inability to scale instantly during

demand spikes. These constraints often lead to delayed ticket resolution, inconsistent responses,

and customer dissatisfaction.

Furthermore, technical support queries are often more complex than simple frequently asked

questions. They may require diagnostic questioning, contextual understanding across multiple

interaction turns, access to historical user data, and escalation to human agents when automated

resolution proves insufficient.

Although cloud-based AI tools such as Amazon Lex, AWS Lambda, and DynamoDB are

available, there is a shortage of comprehensive, production-oriented studies that demonstrate:

persistent context and conversation history management; integration with IT helpdesk ticketing

workflows; effective escalation mechanisms to human agents; performance evaluation under

realistic workloads; and cost and scalability analysis in serverless environments.

This gap highlights the need for systematic design, implementation, and evaluation of a

serverless AI-powered IT helpdesk chatbot tailored specifically to the e-commerce domain.

1.3 Aim and Objectives of the Study

Aim

The primary aim of this study is to design, implement, and evaluate a serverless AI-powered

chatbot for e-commerce IT helpdesk support using Amazon Lex, AWS Lambda, and

DynamoDB.

Objectives

The specific objectives of the study are to:

1. Design a conversational chatbot capable of handling multi-turn IT support dialogues using

Amazon Lex

2. Implement backend business logic using AWS Lambda for request processing and system

integration

3. Develop a persistent data storage mechanism using DynamoDB to maintain conversation

history and ticket records

4. Implement escalation mechanisms to support human-agent handover where automated

resolution fails

5. Evaluate the system based on technical metrics such as latency, scalability, throughput, and

cost

6. Assess user experience factors such as satisfaction, perceived usefulness, trust, and resolution

success

1.4 Significance of the Study

This study holds significance for both practitioners and researchers. For e-commerce

organizations, it provides a practical framework for deploying scalable, cost-efficient AI-driven

IT helpdesk solutions capable of operating continuously and handling peak workloads. The
system has the potential to reduce operational costs, improve response times, and enhance

overall customer satisfaction.

From an academic perspective, the study contributes to the literature by addressing a gap in

existing research—specifically, the lack of comprehensive evaluations of serverless chatbot

architectures applied to complex helpdesk workflows. By combining conversational AI,

serverless computing, persistent data storage, and performance evaluation, this research offers

empirical insights into the viability and limitations of such systems.

Additionally, the study provides valuable insights for organizations considering the adoption of
AI-powered support systems, offering evidence-based guidance on implementation strategies,

performance expectations, and potential challenges. The findings may inform policy decisions

regarding automation in customer service and contribute to ongoing discussions about the role of

AI in enhancing service delivery while maintaining quality standards.

1.5 Scope and Limitations of the Study

This study focuses specifically on IT helpdesk support within e-commerce environments,

utilizing AWS serverless technologies including Amazon Lex, AWS Lambda, and DynamoDB.

The scope encompasses the design, implementation, and evaluation of a conversational AI

system capable of handling common technical support queries, maintaining conversation context,

and escalating complex issues to human agents.

The study is limited to text-based interactions and does not explore voice-based or multimodal

interfaces. While the system architecture is designed for scalability, testing is conducted within

defined parameters that simulate realistic but controlled workload conditions. The evaluation

focuses on technical performance metrics and user experience factors but does not extend to

long-term organizational impact or return on investment analysis.

Furthermore, the study acknowledges that chatbot effectiveness may vary across different

organizational contexts, user demographics, and problem domains. The findings are primarily
applicable to e-commerce IT helpdesk scenarios and may require adaptation for other industries

or support contexts.

1.6 Definition of Terms

Chatbot: A software application that uses artificial intelligence and natural language processing

to simulate human-like conversation.

Natural Language Processing (NLP): A subfield of AI focused on enabling computers to

understand and process human language.

Amazon Lex: An AWS service that provides natural language understanding and speech

recognition for building conversational interfaces.

AWS Lambda: A serverless compute service that executes code in response to events without

requiring server provisioning or management.

Amazon DynamoDB: A fully managed, scalable NoSQL database service provided by AWS.

Serverless Architecture: A cloud computing model where infrastructure management is

abstracted from developers.

IT Helpdesk: A support service that provides technical assistance related to systems,

applications, or infrastructure.

Intent Recognition: The process by which a chatbot identifies the user's purpose or goal from

their input.

Conversation Context: The accumulated information from previous exchanges that informs

current responses.

Escalation: The process of transferring a support request from automated systems to human

agents when automated resolution is insufficient.

Chapter One References

Adamopoulou, E., & Moussiades, L. (2020). An overview of chatbot technology. In Artificial

Intelligence Applications and Innovations (pp. 373-383).

Amazon Web Services. (2024). Amazon Lex Developer Guide.

Shawar, B. A., & Atwell, E. (2007). Chatbots: Are they really useful? LDV Forum, 22(1), 29-49.
CHAPTER TWO: LITERATURE REVIEW

2.1 Introduction

This chapter provides a comprehensive review of existing literature on artificial intelligence-

powered chatbots, serverless computing architectures, and their application to IT helpdesk

support in e-commerce environments. The review is structured to examine the theoretical

foundations, technological implementations, evaluation methodologies, and practical challenges

that characterize current research and practice in this domain.

The literature review serves multiple purposes within the context of this study. First, it

establishes the theoretical and empirical foundations upon which the research is built, drawing

from disciplines including artificial intelligence, software engineering, human-computer

interaction, and service management. Second, it identifies gaps in existing knowledge and

practice that justify the present investigation. Third, it provides a framework for understanding

the design decisions, implementation strategies, and evaluation approaches employed in

subsequent chapters.

The organization of this chapter reflects the multidisciplinary nature of the research problem.

Section 2.2 explores the role of chatbots in customer service and e-commerce, examining how

conversational AI has evolved to address service delivery challenges. Section 2.3 investigates

serverless and cloud-based architectures, with particular emphasis on AWS services and their

application to chatbot development. Section 2.4 reviews domain-specific implementations and

case studies that provide practical insights into real-world deployments. Section 2.5 examines the

fundamental principles of serverless computing and their implications for system design. Section

2.6 synthesizes findings from systematic reviews and surveys that provide broader perspectives

on the field. Section 2.7 identifies challenges and limitations evident in current literature. Finally,

Section 2.8 positions the present study within the broader research landscape, articulating how

this work addresses identified gaps and contributes to advancing knowledge in the field.
Throughout the review, particular attention is paid to the intersection of technical capabilities and

user experience, recognizing that successful chatbot implementations must satisfy both

performance requirements and human factors considerations. The review adopts a critical

perspective, evaluating not only what has been accomplished but also what remains inadequately

addressed in current research and practice.

2.2 Chatbots in Customer Service and E-commerce

2.2.1 Evolution and Adoption of Chatbot Technology

The adoption of chatbot technology in customer service has accelerated dramatically over the

past decade, driven by advances in natural language processing, machine learning, and cloud

computing infrastructure. Research on chatbots in customer service consistently highlights their

ability to deliver instant responses, improve scalability, and reduce workload on human agents.

In e-commerce contexts, chatbots are commonly deployed for order tracking, product inquiries,

payment assistance, and frequently asked questions (Brandtzaeg & Følstad, 2017).

Early chatbot systems relied primarily on rule-based approaches, using pattern matching,

keyword recognition, and decision trees to map user inputs to predefined responses. While these

systems provided consistent answers to well-defined queries, they struggled with linguistic

variation, contextual understanding, and novel questions. The limitations of rule-based systems

became increasingly apparent as user expectations evolved and query complexity increased.

The advent of statistical natural language processing and machine learning techniques marked a

significant evolution in chatbot capabilities. Systems leveraging these approaches could learn

from data rather than relying exclusively on hand-crafted rules, enabling more flexible responses

and better handling of linguistic variation. However, these systems still faced challenges in

maintaining conversational context, understanding user intent across multiple turns, and

providing consistently accurate information.

Contemporary chatbot systems increasingly leverage deep learning architectures, including

recurrent neural networks, transformers, and large language models. These advanced approaches

have dramatically improved performance on tasks such as intent classification, entity extraction,

and response generation. Nevertheless, challenges remain in ensuring factual accuracy,

maintaining appropriate behavior, and integrating with backend systems effectively.

2.2.2 Consumer Perspectives and Acceptance Factors

Cheng et al. (2024) conducted a comprehensive survey of 299 respondents to explore consumer
responses to text-based chatbots in e-commerce environments. Their findings revealed that

perceived usefulness and ease of use significantly influence user satisfaction, particularly for

moderately complex tasks. The study demonstrated that chatbots excel when handling

straightforward, well-defined queries but struggle with ambiguous or multi-faceted problems

requiring nuanced understanding.

The research identified task complexity as a critical moderating factor in chatbot acceptance. For

simple, transactional queries such as order status checks or basic product information retrieval,

users generally found chatbots efficient and satisfactory. However, as task complexity increased

—requiring diagnostic questioning, personalized recommendations, or resolution of unusual

problems—user satisfaction declined markedly. This finding has important implications for

system design, suggesting that chatbots should be deployed selectively for tasks matching their

capabilities while ensuring seamless escalation mechanisms for complex issues.

Transparency regarding chatbot identity emerged as another significant factor influencing user

acceptance. Cheng et al. (2024) found that explicit disclosure of chatbot status moderated user

expectations and responses. Users who understood they were interacting with an automated

system adjusted their communication style and expectations accordingly, whereas those who

believed they were communicating with humans experienced greater frustration when responses

failed to meet human-level expectations. These findings underscore the importance of clear

chatbot identification and appropriate expectation management.

The study also highlighted the role of anthropomorphic design elements in shaping user

perceptions and behavior. While some degree of human-like qualities enhanced engagement and

perceived friendliness, excessive anthropomorphism created unrealistic expectations and

ultimately undermined trust when the system's limitations became apparent. The optimal design

approach appears to balance approachability with honest representation of capabilities.

2.2.3 Trust, Privacy, and Security Considerations

Sharma et al. (2023) investigated consumer adoption of chatbots in the Indian e-commerce sector
through a study encompassing diverse demographic groups and shopping behaviors. Their

research identified trust, convenience, and perceived efficiency as primary drivers of chatbot

adoption. Users who had positive initial experiences with chatbots demonstrated increased

willingness to engage with automated systems for subsequent interactions, suggesting that early

experiences critically shape long-term adoption patterns.

However, the study also revealed significant concerns about data privacy, security, and the

impersonal nature of automated interactions. Respondents expressed apprehension about sharing

sensitive information such as payment details, personal identification, and purchase history with

chatbot systems. These concerns were particularly pronounced among users who lacked

understanding of data handling practices or had experienced security incidents in other contexts.

The research emphasized that addressing privacy concerns requires both technical safeguards

and transparent communication. Users need clear information about what data is collected, how

it is used, who has access to it, and what protections are in place. Furthermore, chatbot systems

must implement robust security measures including encryption, access controls, and audit trails

to protect sensitive information. The intersection of technical security and user perception

highlights the importance of security-conscious design that is also comprehensible to non-

technical users.

Sharma et al. (2023) also noted cultural variations in chatbot acceptance and privacy

expectations. Trust in automated systems varies across cultural contexts, influenced by factors
including technological literacy, institutional trust, and cultural norms regarding personal

information sharing. These variations suggest that chatbot implementations may require

localization not only in language but also in interaction patterns, disclosure practices, and

privacy controls.

2.2.4 Impact on Business Performance and Customer Satisfaction

Rasheed et al. (2025) examined the impact of AI-powered chatbots on customer satisfaction and

business performance in e-commerce through a mixed-methods study combining quantitative

metrics analysis and qualitative case studies. Their research demonstrated that chatbot

implementation positively correlates with several key performance indicators including reduced

response times, improved service consistency, enhanced customer satisfaction scores, and

decreased operational costs.

The study quantified the operational benefits of chatbot deployment, finding that organizations

implementing well-designed chatbot systems achieved average response time reductions of 60-

80% compared to human-only helpdesk operations. This improvement stemmed from the

chatbot's ability to handle multiple concurrent interactions without degradation in response

quality, effectively eliminating queue wait times for queries within the system's capability range.

Service consistency emerged as another significant benefit. Unlike human agents whose

performance may vary due to factors including fatigue, mood, knowledge gaps, or training

inconsistencies, chatbots provide uniform responses to equivalent queries. This consistency

strengthens brand perception and ensures that all customers receive accurate information

regardless of when they initiate contact or which support channel they use.

However, Rasheed et al. (2025) noted that success depends heavily on several critical factors.

Proper integration with existing systems is essential; chatbots that cannot access order data,

inventory information, or customer history provide limited value and may frustrate users by

requiring information they believe the organization should already possess. Continuous training
and updating of AI models based on user interactions is also crucial, as static systems quickly

become outdated and less effective as business offerings, policies, and customer needs evolve.

The research identified organizational factors that mediate chatbot effectiveness. Companies that

treat chatbot deployment as an ongoing process requiring continuous monitoring, evaluation, and

refinement achieve superior results compared to those adopting a "set and forget" approach.

Regular analysis of conversation logs, identification of common failure patterns, and systematic

expansion of the chatbot's knowledge base contribute to sustained performance improvement.

Furthermore, the study highlighted the importance of human-chatbot collaboration rather than
complete automation. Organizations that position chatbots as first-line support while maintaining

accessible human agent escalation paths achieve higher overall satisfaction than those attempting

to replace human support entirely. This finding reinforces the perspective that chatbots should

augment rather than replace human capabilities, handling routine queries efficiently while

preserving human judgment for complex or sensitive situations.

2.2.5 User Compliance and Behavioral Response

Adam et al. (2020) explored user compliance with AI-based chatbots through controlled

experiments examining how design characteristics influence user behavior and engagement.

Their research discovered that perceived anthropomorphism and social presence significantly

influence user engagement, compliance with requests, and overall interaction quality.

The concept of social presence refers to the degree to which users perceive the chatbot as a social

entity capable of understanding and responding to social cues. Higher social presence correlates

with increased user engagement, more natural communication patterns, and greater willingness

to follow chatbot recommendations. However, the relationship is complex and non-linear;

excessive social presence can create uncanny valley effects or unrealistic expectations that

ultimately undermine effectiveness.

The research examined specific design elements contributing to anthropomorphism, including

conversational style, response timing, emotional expression, and visual representation. Natural,
conversational language increased perceived humanity and engagement, while overly formal or

stilted responses diminished social presence. Appropriate response timing—neither

instantaneous nor excessively delayed—contributed to more natural interaction patterns.

However, the study cautioned against misleading users through excessive anthropomorphism,

emphasizing the importance of transparency about the system's automated nature.

Adam et al. (2020) also investigated how anthropomorphic design affects user compliance with

chatbot recommendations and requests. They found that users demonstrate greater compliance

when they perceive the chatbot as credible, competent, and trustworthy. These perceptions are
influenced by factors including response accuracy, appropriate confidence expression, and

acknowledgment of limitations. Chatbots that confidently provide incorrect information or claim

capabilities beyond their actual competence rapidly lose user trust and compliance.

The research suggested that optimal chatbot design balances human-like qualities with clear

identification as automated systems. Users appreciate natural, friendly interaction styles but also

value honesty about system capabilities and limitations. This balance allows chatbots to leverage

social presence benefits while maintaining realistic user expectations and preserving trust when

limitations become apparent.

2.2.6 Conversational Failure, User Frustration, and Recovery Strategies

A critical yet historically underexplored aspect of chatbot deployment is conversational failure

and the strategies systems employ to recover from such failures. Conversational failure occurs

when a chatbot repeatedly fails to understand user input, provides irrelevant or incorrect

responses, enters circular conversation loops, or prematurely escalates without meaningful

assistance. In e-commerce IT helpdesk contexts, such failures are particularly costly because

users often seek support under time pressure or emotional stress, such as during payment

failures, service outages, or technical problems preventing transaction completion.

Research indicates that repeated fallback responses—generic statements like "I didn't understand

that" or "Can you rephrase?"—significantly increase abandonment rates and user frustration
(Brandtzaeg & Følstad, 2017). The cumulative effect of multiple failures compounds user

frustration exponentially rather than linearly; a user might tolerate one misunderstanding but

becomes significantly more frustrated after two or three consecutive failures.

Studies suggest that recovery strategies can partially mitigate user frustration when implemented

thoughtfully. Effective recovery approaches include explicit clarification requests that specify

what information the chatbot needs, rephrasing prompts that guide users toward language the

system can process, offering structured options that constrain responses to recognizable patterns,

and progressive narrowing of scope to identify the user's intent through elimination.
However, few academic implementations rigorously evaluate recovery mechanisms in technical

support conversations, leaving a gap between theoretical design principles and operational

resilience. Most published research focuses on ideal-case performance—demonstrating what

systems can do when functioning properly—rather than examining degraded performance and

recovery capabilities. This gap is particularly problematic for IT helpdesk applications where

graceful degradation and effective recovery are essential to maintaining service quality.

The literature emphasizes that effective recovery strategies must be context-aware and avoid

repetitive patterns that signal system incompetence. Repeatedly asking "Can you rephrase that?"

without providing guidance on what constitutes an acceptable rephrasing frustrates users and

signals fundamental system limitations. More effective approaches include specific guidance

such as "I can help you with password resets, account issues, or payment problems. Which of

these best describes your situation?"

Progressive escalation, where the system gradually transitions from automated assistance to

human intervention, appears more effective than abrupt handoffs. This approach allows the

chatbot to gather preliminary information, attempt automated resolution, provide self-service

resources, and finally escalate to human agents with full context if automated approaches prove

insufficient. However, empirical validation of these strategies in production environments

remains limited, and additional research is needed to establish best practices.

2.2.7 Chatbots and Customer Experience Management

Beyond immediate issue resolution, chatbots increasingly influence overall customer experience

management (CEM) and long-term customer relationships. Rasheed et al. (2025) argue that

chatbots function as continuous touchpoints that shape perceptions of organizational reliability,

professionalism, and customer commitment. In e-commerce contexts, IT helpdesk interactions

may indirectly affect purchasing decisions, renewal rates, brand loyalty, and customer lifetime

value.

This broader perspective on chatbot impact underscores the need for alignment between chatbot
behavior and organizational service standards. Inconsistency between chatbot responses and

human-agent practices can erode trust and create confusion. For example, if chatbots promise

capabilities or timeframes that human agents cannot fulfill, or if they provide information

contradicting official policies, users experience cognitive dissonance that undermines confidence

in the organization as a whole.

The literature suggests that maintaining consistency requires integrated knowledge management

systems shared between chatbots and human agents. Both automated and human support

channels should draw from common knowledge bases, policy repositories, and procedural

guidelines. This integration ensures that users receive consistent information regardless of

interaction channel and facilitates seamless transitions when escalation occurs.

Furthermore, chatbot interactions contribute to customer perception formation through both

explicit content and implicit signals. Response speed, language tone, demonstrated competence,

and problem resolution effectiveness all shape user impressions. Organizations must recognize

that chatbots represent their brand in customer interactions and design systems that reflect

desired brand attributes including professionalism, helpfulness, competence, and customer

respect.
2.2.8 Emotional Intelligence and Empathy in Automated Support

Recent research has explored the role of emotional intelligence and empathy in chatbot

interactions, particularly in support contexts where users may be frustrated, anxious, or upset.

While chatbots lack genuine emotional understanding, design strategies can create perceptions of

empathy and emotional responsiveness that improve user experience.

Emotional recognition capabilities allow chatbots to detect sentiment in user messages and adjust

responses accordingly. When users express frustration or distress, emotionally-aware chatbots

can acknowledge these feelings, express understanding, and adjust their approach. For example,
a chatbot detecting frustration might prioritize immediate escalation to human agents rather than

attempting extended troubleshooting sequences that could exacerbate negative emotions.

However, the literature cautions against superficial or manipulative approaches to emotional

expression. Users can detect inauthentic empathy, and chatbots that claim to "understand how

frustrating this must be" while providing unhelpful responses undermine rather than build trust.

Effective emotional design requires substantive responses that address user concerns rather than

mere emotional vocabulary.

The challenge of emotional intelligence in chatbots becomes particularly acute in IT helpdesk

contexts where technical problems may have serious consequences for users. A payment

processing failure, for instance, may prevent a customer from completing an urgent transaction,

generating significant stress. Chatbots handling such situations must balance efficiency in

problem resolution with appropriate acknowledgment of user circumstances.

2.3 Serverless and Cloud-Based Chatbot Architectures

2.3.1 Foundations of Serverless Chatbot Development

Serverless architectures have gained substantial prominence in chatbot development due to their

inherent scalability, operational efficiency, and cost-effectiveness. The serverless paradigm

abstracts infrastructure management from developers, allowing them to focus on application

logic while the cloud provider handles server provisioning, scaling, patching, and availability.

This model proves particularly advantageous for chatbot applications characterized by variable
workload patterns, intermittent usage spikes, and the need for rapid deployment and iteration.

Yan et al. (2016) pioneered the exploration of serverless chatbot architectures, demonstrating the

fundamental feasibility of building conversational systems using function-as-a-service (FaaS)

platforms. Their seminal work established foundational patterns for event-driven chatbot design,

where user messages trigger function executions, process requests, generate responses, and

terminate without maintaining persistent server processes. This approach aligns naturally with

conversational interaction patterns, where discrete user inputs require independent processing

and response generation.

The event-driven nature of serverless architectures maps effectively to chatbot requirements.

Each user message constitutes an event that triggers backend processing, intent classification,

business logic execution, and response formulation. The stateless design of serverless functions

encourages clean separation of concerns, with distinct functions handling different aspects of

conversation management, business logic, data access, and external system integration.

However, Yan et al. (2016) also identified early challenges that subsequent research has worked

to address. State management across conversation turns proved non-trivial in stateless function

environments, requiring external state stores and careful session management. Cold start latency

—the delay incurred when functions execute for the first time or after periods of inactivity—

created inconsistent response times that could degrade user experience. Integration with external
services and databases required careful design to manage connection pooling and resource

utilization within function execution time constraints.

2.3.2 Production Deployment Experiences and Lessons Learned

Lehvä et al. (2018) presented an invaluable case study of building a serverless messenger

chatbot, providing rare insights into the practical challenges of production deployment beyond

proof-of-concept demonstrations. Their research revealed that while serverless architectures

reduce certain operational complexities, they introduce new challenges requiring careful
architectural consideration and operational expertise.

Cold-start latency emerged as a significant concern affecting user experience. When Lambda

functions had not been invoked recently, initialization overhead could add hundreds of

milliseconds or even seconds to response times. For interactive conversational applications

where users expect near-instantaneous responses, such delays create noticeably degraded

experiences. Lehvä et al. explored mitigation strategies including function warming (periodic

invocations to keep functions initialized), minimizing deployment package size to reduce

initialization time, and using provisioned concurrency for critical functions.

Debugging complexity in serverless environments presented another substantial challenge.

Traditional debugging approaches relying on local development environments, breakpoints, and

step-through execution do not transfer directly to distributed serverless architectures. The

ephemeral nature of function executions complicates issue reproduction and analysis. Lehvä et

al. emphasized the importance of comprehensive logging, structured error handling, and

distributed tracing to maintain observability in production environments.

Database connection management and throttling proved particularly problematic. Traditional

database connection pooling strategies designed for long-running server processes do not work

effectively with short-lived Lambda functions. Each function invocation potentially creates new

database connections, risking connection pool exhaustion under concurrent load. Furthermore,
DynamoDB's provisioned throughput model required careful capacity planning to avoid

throttling during usage spikes while managing costs during low-usage periods.

The study highlighted the need for careful capacity planning and performance tuning that

extends beyond simply deploying code to serverless platforms. Successful production

deployments require understanding platform-specific constraints, implementing appropriate

monitoring and alerting, designing for graceful degradation, and establishing operational

procedures for incident response and system evolution.

Lehvä et al. (2018) also noted organizational and cultural challenges in adopting serverless
architectures. Development teams accustomed to traditional application architectures faced

learning curves in understanding serverless programming models, distributed systems patterns,

and cloud-native operational practices. The gap between proof-of-concept demonstrations and

production-ready systems proved substantial, requiring investment in tooling, processes, and

expertise development.

2.3.3 AWS Services for Conversational AI

Amazon Web Services provides a comprehensive ecosystem specifically designed for building

serverless conversational systems. The integration of Amazon Lex, AWS Lambda, and Amazon

DynamoDB creates a powerful platform for developing intelligent chatbots with natural

language understanding, flexible business logic execution, and scalable data persistence.

Saini and Tajammul (2022) developed a conversational AI-powered chatbot leveraging Amazon

Lex and AWS Lambda, demonstrating the practical application of these services in creating

functional conversational systems. Amazon Lex provides natural language understanding

capabilities including intent recognition, entity extraction, and dialogue management. The

service handles the complexity of parsing user inputs, identifying user intentions, and

maintaining conversation context across multiple interaction turns.

Their implementation showcased the flexibility of serverless architectures for rapid prototyping

and iterative development. The ability to modify Lambda functions and redeploy changes
without infrastructure reconfiguration enabled quick iteration cycles and responsive adaptation to

evolving requirements. Integration with external systems and databases through Lambda

functions demonstrated the extensibility of the architecture beyond simple conversational flows.

However, the evaluation in Saini and Tajammul (2022) lacked comprehensive performance

assessment under realistic workloads. While the implementation proved functionally adequate

for demonstration purposes, questions remained regarding scalability, cost-effectiveness, and

reliability under production conditions. The study did not quantify response latency distributions,

cold start frequencies, concurrent user capacity, or operational costs, leaving practitioners with
limited guidance for production deployment decisions.

Aranjo et al. (2021) explored dynamic conversational chatbots using AWS services with

particular focus on context retention and multi-turn dialogue management. Their research

highlighted the importance of session management and conversation history for maintaining

coherent, contextual interactions. Effective chatbots must remember previous exchanges within a

conversation, reference earlier topics, and maintain consistency across multiple interaction turns.

The study examined strategies for persisting conversation state between Lambda function

invocations, including session attributes in Lex, DynamoDB storage of conversation history, and

contextual information passing through function execution chains. They demonstrated that

thoughtful state management enables chatbots to handle complex, multi-step processes such as

troubleshooting workflows requiring progressive information gathering and diagnostic

questioning.

However, Aranjo et al. (2021) acknowledged limitations in their evaluation methodology.

Assessment was primarily qualitative, based on developer observations and limited user testing

rather than quantitative performance metrics or controlled user studies. The implementation did

not explore advanced features such as human agent escalation, integration with ticketing

systems, or comprehensive error handling and recovery strategies.

2.3.4 Natural Language Understanding and Intent Recognition

Hnatushenko et al. (2024) developed and researched a chatbot using Amazon Lex V2, focusing

specifically on linguistic capabilities, intent recognition accuracy, and multilingual support.

Their work provides valuable insights into the natural language understanding component that

forms the foundation of effective conversational AI systems.

The research demonstrated measurable improvements in intent recognition accuracy and entity

extraction compared to earlier versions of Amazon Lex. Lex V2 introduced enhanced machine

learning models, improved confidence scoring, and better handling of linguistic variation
including misspellings, informal language, and regional dialects. These improvements directly

impact chatbot effectiveness by reducing misunderstandings and improving response relevance.

However, performance varied significantly across languages and dialects, highlighting

challenges in achieving consistent quality for global deployments. While Lex V2 demonstrated

strong performance for well-supported languages such as English, performance degraded for

languages with less training data or linguistic characteristics diverging substantially from

English. This finding has important implications for organizations serving diverse, multilingual

customer bases.

The study also examined entity extraction capabilities—the ability to identify specific

information within user inputs such as dates, numbers, product names, or account identifiers.

Accurate entity extraction is essential for IT helpdesk chatbots that must capture specific details

including order numbers, error codes, or system identifiers. Hnatushenko et al. found that entity

extraction accuracy depends heavily on entity definition quality, training data adequacy, and the

presence of contextual cues that disambiguate entity boundaries.

The research identified several factors influencing intent recognition accuracy including training

data quality and quantity, utterance diversity in training examples, and the distinctiveness of

different intents. Intents with clear, unambiguous language patterns achieved higher recognition

accuracy than those expressed through varied or ambiguous language. The study recommended
maintaining diverse training utterances that reflect realistic user language variation, including

common misspellings, abbreviations, and informal expressions.

2.3.5 Integration Patterns and System Architecture

Paratala Rajagopal (2020) explored cloud methodologies for seamless integrated chatbots,

proposing architectural patterns for connecting chatbot interfaces with backend enterprise

systems, databases, and external services. The research emphasized the importance of API

gateway patterns, microservices integration, and data synchronization strategies in creating

production-grade chatbot systems.

The proposed architecture employed API Gateway as a central integration point, providing

unified access to backend services while handling concerns including authentication, rate

limiting, request transformation, and response aggregation. This pattern decouples chatbot logic

from specific backend implementations, enabling independent evolution of conversation

handling and business systems while maintaining well-defined integration contracts.

Microservices integration patterns allow chatbots to interact with specialized services responsible

for distinct business capabilities such as order management, payment processing, inventory

checking, or customer data access. This architectural approach promotes modularity,

independent scalability of different system components, and technology diversity where different

services may employ different programming languages, frameworks, or data stores optimized for

their specific requirements.

Data synchronization strategies address challenges in maintaining consistency between chatbot

conversations and authoritative business system data. The research explored patterns including

synchronous API calls for real-time data access, caching strategies to reduce backend load and

improve response latency, and event-driven updates where chatbots subscribe to data change

notifications to maintain current information without constant polling.

While theoretically sound, Paratala Rajagopal (2020) lacked empirical validation of performance

and reliability under production conditions. The proposed patterns were illustrated through
conceptual diagrams and simplified examples rather than comprehensive implementations tested

under realistic loads. Questions remained regarding the practical performance implications of

different integration patterns, optimal caching strategies for various data types, and resilience

approaches when backend systems experience failures or degraded performance.

2.3.6 Observability, Logging, and Debugging in Serverless Chatbots

While scalability receives substantial attention in serverless literature, observability remains a

critical operational challenge that often receives insufficient consideration in academic research.
Serverless architectures distribute application logic across numerous ephemeral functions,

complicating debugging, root-cause analysis, and performance optimization. The stateless,

event-driven nature of serverless systems creates distributed traces spanning multiple functions,

services, and data stores, making it difficult to understand system behavior holistically.

CloudWatch and similar centralized logging mechanisms partially address observability

challenges by aggregating logs from distributed functions into searchable repositories. However,

effective observability requires more than basic log aggregation. Structured logging with

consistent formats, correlation IDs linking related log entries across function invocations,

contextual information identifying conversation sessions and user interactions, and appropriate

log levels distinguishing routine operations from exceptional conditions all contribute to

maintainable observability.

In IT helpdesk scenarios, audit trails are essential for multiple purposes beyond debugging.

Compliance requirements may mandate retaining records of support interactions, particularly

those involving sensitive data or security-related incidents. Quality assurance processes rely on

conversation logs to identify common issues, evaluate chatbot performance, and guide

continuous improvement. Troubleshooting customer complaints requires reconstructing

interaction histories to understand what occurred and why.

The lack of structured logging and comprehensive observability can severely hinder post-

incident analysis and system evolution. When users report problems or express dissatisfaction,
support teams need capabilities to review conversation transcripts, examine function execution

logs, identify where failures occurred, and understand why particular responses were generated.

Without adequate observability, organizations resort to speculation rather than data-driven

problem analysis.

Best practices emerging from industry include treating observability as a first-class design

requirement rather than an operational afterthought. This perspective advocates for designing

logging and monitoring capabilities concurrently with functional code, establishing observability

standards and patterns, implementing distributed tracing to follow requests across system
boundaries, and creating real-time monitoring dashboards tracking conversation quality metrics,

error rates, and performance indicators.

However, academic literature rarely evaluates the effectiveness of various observability

approaches in conversational systems. Research tends to focus on functional capabilities and

performance metrics while neglecting operational concerns such as debuggability,

maintainability, and the practical effort required to diagnose and resolve production issues. This

gap leaves practitioners without evidence-based guidance for observability strategy selection and

implementation.

2.3.7 Vendor Lock-In and Architectural Portability

Vendor lock-in represents a recurring concern in serverless computing research and practice. AP

Rajan et al. (2020) highlight that deep integration with proprietary services such as Amazon Lex

creates dependencies that limit portability across cloud providers. Applications leveraging AWS-

specific APIs, data formats, and service integrations cannot easily migrate to alternative

platforms without substantial re-engineering.

The lock-in concern manifests at multiple levels. At the lowest level, proprietary APIs and

service-specific configuration create technical dependencies requiring code changes for

migration. At a higher level, operational practices, deployment pipelines, and monitoring

infrastructures built around specific cloud platforms resist portability. At the organizational level,
accumulated expertise, training, and institutional knowledge focused on particular platforms

represent investments that do not transfer to alternative environments.

The serverless computing literature remains divided on whether abstraction layers can effectively

mitigate lock-in without introducing unacceptable complexity and performance overhead.

Abstraction frameworks attempting to provide cloud-agnostic serverless programming models

face challenges balancing portability with full utilization of platform-specific capabilities.

Applications limiting themselves to lowest-common-denominator features may sacrifice

performance, cost-efficiency, or functionality available in specific platforms.

This study acknowledges the lock-in trade-off while prioritizing architectural coherence and

operational maturity. The decision to leverage AWS-native services including Lex, Lambda, and

DynamoDB reflects a pragmatic balance between development velocity, operational capabilities,

and long-term portability considerations. For many organizations, the benefits of rapid

development, mature services, and comprehensive tooling outweigh theoretical portability

advantages that may never be exercised.

Furthermore, portable application architectures often prove more complex to develop, test, and

maintain than platform-specific implementations. The abstraction layers required for portability

introduce additional code, dependencies, and potential failure modes. Unless organizations have

concrete multi-cloud requirements or regulatory constraints demanding cloud independence, the

overhead of portable architectures may not justify the benefits.

The literature suggests that practical portability strategies focus on isolating platform-specific

dependencies within well-defined boundaries, maintaining clear architectural layers separating

business logic from infrastructure concerns, and documenting cloud platform dependencies

explicitly. These approaches facilitate migration if necessary while avoiding premature

abstraction that complicates development without delivering concrete value.

2.4 Domain-Specific Implementations and Advanced Approaches

2.4.1 Knowledge-Enhanced Conversational Systems

Patel (2025) proposed a graph-enhanced retrieval-augmented question answering framework

specifically designed for e-commerce customer support, combining structured knowledge graphs

with retrieval-augmented generation techniques. This approach represents a significant

advancement in chatbot architectures by explicitly modeling domain knowledge and leveraging

it to improve response accuracy and factual consistency.

The framework constructs knowledge graphs representing products, policies, procedures,

common issues, and their relationships. These structured representations enable sophisticated

reasoning about customer queries, identification of relevant information sources, and generation

of accurate, contextually appropriate responses. The knowledge graph serves as an authoritative

reference constraining generated responses to factually correct information aligned with

organizational knowledge.

Retrieval-augmented generation combines the flexibility of generative language models with

explicit retrieval of relevant information from knowledge bases. Rather than relying solely on

knowledge encoded in model parameters during training, the system retrieves pertinent

information dynamically during conversation and uses it to inform response generation. This

approach addresses knowledge currency challenges inherent in static models while maintaining

natural language generation capabilities.

Patel (2025) demonstrated approximately 23% improvement in factual accuracy compared to

baseline approaches lacking knowledge graph integration. User satisfaction scores reached

approximately 89% in test scenarios, suggesting that knowledge-enhanced approaches deliver

tangible user experience improvements. The framework particularly excelled at handling queries

requiring precise factual information about products, policies, or procedures where accuracy is

critical.
However, the study focused primarily on question answering rather than full conversational

dialogue management. The system demonstrated strong performance on information retrieval

tasks but was not evaluated on complex, multi-turn interactions requiring diagnostic questioning,

progressive information gathering, or contextual understanding spanning multiple exchanges.

Furthermore, the research did not address integration with backend services, support ticket

creation, escalation to human agents, or other operational requirements essential for production

helpdesk systems.

The knowledge graph construction and maintenance effort represents another practical
consideration not fully explored in the research. Building comprehensive, accurate knowledge

graphs requires significant domain expertise and ongoing curation to reflect evolving products,

policies, and business processes. The scalability and sustainability of knowledge graph

approaches depend on efficient workflows for knowledge capture, validation, and updating.

2.4.2 Retrieval-Based Approaches for Production Chatbots

Swanson et al. (2019) addressed production readiness concerns for retrieval-based chatbots by

proposing a dual encoder architecture optimized for rapid similarity scoring and response

selection. Their work provides valuable insights into practical system design for organizations

prioritizing deployment reliability and predictable performance over open-ended generative

capabilities.

The dual encoder architecture separately encodes user queries and candidate responses into dense

vector representations, then computes similarity scores to identify the most appropriate response.

This approach enables precomputing response encodings offline, making online inference

computationally efficient. The architecture supports real-time response selection even for large

response repositories, addressing latency requirements for interactive conversations.

Swanson et al. evaluated their system on a large proprietary helpdesk chat dataset containing

thousands of real customer support conversations. The evaluation demonstrated near-human

performance in selecting appropriate responses, with selection accuracy approaching that of

human support agents reviewing the same conversations. Furthermore, response latency

remained consistently low, meeting requirements for production deployment in interactive

support contexts.

However, retrieval-based approaches inherently limit flexibility compared to generative systems.

The chatbot can only select from pre-defined responses, making it unable to generate novel

answers to queries not represented in the response repository. This constraint proves acceptable

for many helpdesk scenarios where a comprehensive response repository covers common

questions and issues. However, it limits effectiveness for unusual queries, creative explanations,
or highly personalized responses requiring synthesis of information from multiple sources.

The research also highlighted the importance of response repository quality and coverage.

System performance depends critically on having high-quality responses addressing the full

range of user queries. Gaps in the response repository directly translate to system limitations,

while low-quality responses selected by the system reflect poorly on service quality. Maintaining

a comprehensive, high-quality response repository requires ongoing curation, quality assurance,

and expansion efforts.

2.4.3 Custom Language Models and Open-Source Approaches

Pandya and Holia (2023) explored automating customer service using LangChain to build

custom open-source GPT chatbots tailored to organizational needs. Their approach emphasized

customization capabilities and integration with proprietary knowledge bases, demonstrating

potential for organization-specific chatbot solutions that leverage large language model

capabilities while incorporating enterprise-specific knowledge.

LangChain provides a framework for building applications powered by large language models,

offering abstractions for prompt engineering, chain-of-thought reasoning, external knowledge

integration, and tool usage. The framework enables developers to combine language model

inference with database queries, API calls, document retrieval, and other capabilities, creating
sophisticated conversational systems that extend beyond basic language understanding and

generation.

Pandya and Holia demonstrated integration with organizational knowledge bases including

documentation repositories, policy databases, and historical support tickets. This integration

enables chatbots to provide accurate, organization-specific information rather than generic

responses based solely on language model training. The approach particularly suits organizations

with substantial proprietary knowledge not represented in publicly available training data.

However, their implementation relied heavily on external API services for language model
inference, raising several practical concerns for production deployment. API-based inference

introduces latency, potentially exceeding acceptable thresholds for interactive conversation.

External API dependencies create availability risks; if the API service experiences outages or

performance degradation, the chatbot becomes unavailable or unresponsive. Furthermore,

sending customer queries and conversation history to external services raises data privacy and

security concerns, particularly for sensitive support interactions.

Cost represents another significant consideration for API-based approaches. Language model

inference APIs typically charge per token processed, and conversational applications involving

extended dialogues with substantial context can accumulate significant costs. Organizations must

carefully evaluate whether API costs remain acceptable under realistic usage patterns,

particularly during peak demand periods.

The study did not comprehensively address these practical deployment challenges, focusing

instead on demonstrating technical feasibility and capability. Questions remain regarding optimal

deployment strategies balancing capability, cost, latency, and security considerations for

different organizational contexts and requirements.

2.4.4 Ethical Considerations and Responsible Automation

As chatbots assume greater responsibility in customer-facing support functions, ethical

considerations gain increasing prominence in research and practice. Automated systems may
inadvertently reinforce biases present in training data, disadvantage users with non-standard

language patterns or accessibility needs, or make inappropriate decisions lacking human

judgment and contextual understanding.

Research in fairness and bias mitigation emphasizes the importance of diverse training datasets

representing varied user populations, dialects, and communication styles. Language models

trained primarily on standard English from educated speakers may perform poorly for users with

different linguistic backgrounds, potentially creating discriminatory outcomes. Regular auditing

of chatbot responses across demographic groups helps identify and address systematic biases.
In IT helpdesk environments, ethical risks extend beyond language bias to include improper

handling of sensitive data, over-reliance on automation for decisions requiring human judgment,

and inadequate transparency about system capabilities and limitations. Users may share sensitive

information including passwords, payment details, or personal circumstances during support

interactions. Chatbot systems must implement robust data protection measures and clear policies

governing data retention, access, and usage.

Over-reliance on automation represents another ethical concern. While chatbots excel at routine

tasks, complex or sensitive situations often require human judgment, empathy, and

accountability. Organizations must establish clear boundaries defining appropriate automation

scope while preserving human agency for consequential decisions. For example, account

suspension decisions or dispute resolutions typically warrant human review even when chatbots

could technically execute them.

Transparency about system capabilities and limitations constitutes a fundamental ethical

requirement. Users deserve accurate information about whether they are interacting with

automated systems or humans, what the system can and cannot do, and how their data will be

handled. Misleading users about system capabilities or identity undermines informed consent and

autonomous decision-making.

The literature increasingly advocates for human-in-the-loop designs that preserve accountability

and user agency, particularly during escalation scenarios. Rather than positioning automation and
human agents as separate, sequential stages, effective designs integrate them collaboratively.

Human agents can monitor chatbot interactions, intervene when necessary, and provide oversight

ensuring appropriate system behavior.

2.5 Serverless Computing Fundamentals and Performance Characteristics

Understanding the underlying principles of serverless architecture is essential for effective

chatbot development and deployment. AP Rajan et al. (2020) conducted a comprehensive review

of serverless architectures and function-as-a-service platforms, synthesizing research and

practice to identify core characteristics, benefits, and challenges of serverless computing

paradigms. Serverless computing offers automatic scaling, cost-effective pay-per-use pricing,

and simplified operations. However, challenges include execution time limits, debugging
complexity, cold-start latency, state persistence requirements, and potential vendor lock-in.

These characteristics significantly influence chatbot system design, deployment strategies, and

operational practices. Research has also explored integration of serverless computing with AI

and machine learning workloads, finding that serverless approaches can reduce cost and

operational overhead for many AI inference tasks, though with constraints on model size,

inference time, and latency characteristics.

2.6 Systematic Reviews, Surveys, and Meta-Analyses

Mashaabi et al. (2022) conducted a systematic review of natural language processing in customer

service contexts, covering research from 2015 to 2022. They found that chatbots and question-

answering systems are deployed across diverse domains including e-commerce, banking, and

telecommunications. However, most studies employed custom-built datasets limiting

reproducibility, and evaluation typically focused on technical accuracy metrics rather than user-

centric measures. Kushwah et al. (2023) surveyed e-commerce chatbot technologies, reviewing

rule-based, machine learning, and deep learning approaches. They outlined benefits including

24/7 availability and scalability, alongside limitations such as context awareness challenges and

emotional intelligence gaps. A dominant limitation identified across reviews is overreliance on

accuracy-based metrics for chatbot evaluation. Resolution success, time-to-resolution, escalation

quality, and user trust retention are more meaningful indicators for production systems than
intent classification accuracy alone. The literature increasingly calls for multi-dimensional

evaluation frameworks combining technical, economic, and experiential metrics.

2.7 Challenges, Limitations, and Research Gaps

Across reviewed studies, several critical limitations are evident. Most implementations are small-

scale and lack realistic workload testing simulating production conditions with concurrent users,

varied query patterns, and peak demand scenarios. Persistent conversation history and context

management are often absent or inadequately addressed. Integration with IT helpdesk ticketing
workflows is rarely comprehensively covered, with most studies focusing on information

retrieval rather than operational support processes. User experience challenges in complex

technical support conversations remain underexplored. Serverless performance concerns

including cold-start latency and cost behavior under load are acknowledged but seldom

empirically evaluated with distributions, percentiles, or real deployment data. As Lehvä et al.

(2018) noted, while serverless architectures prove feasible, production deployment reveals

challenges including database throttling, debugging difficulties, and capacity tuning

requirements not addressed in proof-of-concept implementations. The expanded literature reveals

that technical design decisions directly influence human outcomes—latency exacerbates

frustration, poor persistence undermines trust, and inadequate escalation damages perceived

competence. These interdependencies reinforce that chatbot systems must be evaluated

holistically.

2.8 Positioning of the Present Study

Based on comprehensive literature review, this study positions itself as a systematic investigation

addressing multiple identified gaps. Unlike prior work emphasizing isolated capabilities, this

research adopts a holistic perspective encompassing design, implementation, deployment, and

evaluation. Persistent data storage and conversation history management are treated as first-class

requirements. Multi-turn technical support dialogues with contextual understanding are

prioritized. Backend integration with ticketing workflows is essential functionality rather than

future work. Escalation mechanisms for human-agent handover are core features designed

throughout the conversation flow. Holistic evaluation combines technical performance metrics

(latency, throughput, cost, scalability) with user experience factors (satisfaction, trust, perceived

usefulness, resolution success). By combining systematic implementation with rigorous

evaluation, this research aims to bridge the gap between proof-of-concept chatbot systems and

production-ready helpdesk solutions, addressing limitations identified in existing literature and

contributing empirical evidence to guide future implementations. This comprehensive literature

review has demonstrated that while chatbot technologies and serverless architectures have

matured significantly, their application to e-commerce IT helpdesk systems remains fragmented.

The identified gaps—particularly in persistence, escalation, observability, comprehensive

evaluation, and operational readiness—directly inform the methodological choices presented in

Chapter Three.
Chapter Two References

Adamopoulou, E., & Moussiades, L. (2020). An overview of chatbot technology. In Artificial

Intelligence Applications and Innovations (pp. 373-383). Springer.

Adam, M., Wessel, M., & Benlian, A. (2020). AI-based chatbots in customer service and their

effects on user compliance. Electronic Markets, 31, 427-445.

Amazon Web Services. (2024). Amazon Lex Developer Guide. Retrieved from

[Link]
AP Rajan, A., Ramachandran, M., & Alqarni, M. A. (2020). A review on serverless architectures

—Function as a service. TELKOMNIKA (Telecommunication Computing Electronics and

Control), 18(5), 2530-2539.

Aranjo, T., Fernandes, S., Gomes, J., & Pereira, R. (2021). Dynamic conversational chatbot

using Amazon Web Services. International Journal of Engineering Research & Technology,

10(7).

Brandtzaeg, P. B., & Følstad, A. (2017). Why people use chatbots. In Internet Science: 4th

International Conference, INSCI 2017 (pp. 377-392). Springer.

Cheng, X., Zhang, X., Cohen, J., & Mou, J. (2024). Exploring consumers response to text-based

chatbots in e-commerce: The moderating role of task complexity and chatbot disclosure. Internet

Research, 34(1), 183-206.

Hnatushenko, V., Ostrovska, K., & Nosov, P. (2024). Development and research of a chatbot

using the linguistic core of Amazon Lex V2. CEUR Workshop Proceedings, 3688.

Kushwah, N., Singh, R., & Sharma, S. (2023). A survey on e-commerce support using chatbot.

International Journal of Psychosocial Rehabilitation, 7(1), 45-58.

Lehvä, J., Mäkitalo, N., & Mikkonen, T. (2018). Case study: Building a serverless messenger

chatbot. In Software Architecture for Big Data and the Cloud (pp. 349-362). Elsevier.

Mashaabi, M., Al-Fuqaha, A., Khreishah, A., & Qaraqe, M. (2022). Natural language processing

in customer service: A systematic review. arXiv preprint arXiv:2212.09523.

Pandya, M., & Holia, M. (2023). Automating customer service using LangChain: Building

custom open-source GPT chatbot for organizations. arXiv preprint arXiv:2310.05421.

Paratala Rajagopal, A. (2020). Cloud methodologies for a seamless integrated chatbot.

International Journal of Engineering Research & Technology, 9(8), 234-241.

Patel, R. (2025). Graph-enhanced retrieval-augmented question answering for e-commerce

customer support. arXiv preprint arXiv:2509.14267.

Rasheed, A., Sami, U., & Tabassam, M. (2025). The impact of AI-powered chatbots on customer

satisfaction and business performance in e-commerce. Policy Journal of Management Sciences,

5(1), 12-28.

Riyan, M. (2024). The effect of artificial intelligence-based chatbots on customer satisfaction on

e-commerce platforms. Journal of Wireless Information Technology, 1(1), 78-92.

Saini, P., & Tajammul, M. (2022). Conversational AI powered chatbot using Lex and AWS.

International Journal of Trend in Scientific Research and Development, 6(3), 456-463.

Sharma, A., Satija, N., & Yadav, S. (2023). A study of consumer adoption of chatbot in e-

commerce sector in India. VEETHIKA: An International Interdisciplinary Research Journal,

9(2), 34-47.

Shawar, B. A., & Atwell, E. (2007). Chatbots: Are they really useful? LDV Forum, 22(1), 29-49.

Swanson, K., Yu, L., Fox, C., Wohlwend, J., & Lei, T. (2019). Building a production model for

retrieval-based chatbots. arXiv preprint arXiv:1906.03209.

Yan, M., Castro, P., Cheng, P., & Ishakian, V. (2016). Building a chatbot with serverless

computing. In Proceedings of the 1st International Workshop on Mashups of Things and APIs

(pp. 1-4). ACM.