Pejman Mazaheri 

AI-Driven Inventory Digitization in Second-Hand Retail: 
Design and Evaluation of a Multi-Agent System 

 

 

 

 

 

 
  

Vaasa 2025 

University of Vaasa 
Master’s thesis in Industrial 

Systems Analytics  
 



2 

UNIVERSITY OF VAASA 
Author:    Pejman Mazaheri 
Title of the Thesis:  Design and Implementation of a Multi-Agent AI System for  
                                                    Real-Time Inventory Digitization: Case Study of a Finnish  
                                                    Second-Hand Store 
Degree:    Master of Science 
Programme:   Industrial Systems Analytics 
Supervisor:   Emmanuel Ndzibah 
Year:    2025 Pages: 113 

ABSTRACT : 
 
Presently, the second-hand retail sector, which is one of the main pillars of the circular economy, 
is going through a "scalability-sustainability paradox" situation. Simply put, the sector is caught 
in a paradox where the high labor costs of a manual process that aims at the digitalization of a 
heterogeneous inventory lead to low-value products being unprofitable for Small and Medium-
sized Enterprises (SMEs). Micro-retailers are severely limited in terms of digital market access 
from this operational bottleneck. The master's thesis offers a solution to the problem by ques-
tioning how the multi-agent Artificial Intelligence (AI) architecture can be restructured to auto-
mate the digitization workflow, thus reducing transaction costs and facilitating scalability in re-
source-constrained environments. 
 
The research, which is based on Design Science Research (DSR) methodology, invents "AI-MAS-
on-Serverless," an innovative artifact that employs event-driven multi-agent systems and multi-
modal Large Language Models (LLMs) on a serverless cloud infrastructure. The theoretical 
framework combines the review of the three literatures identified, i.e. inventory digitization, 
multi-agent systems, and process automation, to not only develop a solution that is cost-effec-
tive and easy to access but also to tailor it to the second-hand sector's peculiar constraints. 
 
The designed instrument was put to the test through a summative naturalistic evaluation in a 
Finnish second-hand retailer wherein an AI-assisted "Human-in-the-Loop" workflow was com-
pared with the traditional manual practices across a sample of 50 diverse items. The outcomes 
indicate a statistically significant enhancement in the operational efficiency that is accompanied 
by an 88.4% reduction of the processing time (from 310s ±45s to 36s ±5s per item) and an 88% 
decrease in digitization costs (from €1.29 to €0.155). The research achieves a 94% accuracy level 
in the extraction of objective metadata and, therefore, becomes instrumental in the provision 
of an automation design theory that is democratized and shows the manner in which accessible 
AI agents can put to an end the economic barriers to entry in the digital circular economy. 
 
 
 
 
 
 
 
 
 

KEYWORDS: Multi-Agent AI, Inventory Digitization, SME Retail Digitalization, Second-Hand 
Retail, Process Automation 

  



3 

Contents 

1 Introduction 8 

1.1 Background of the Research 8 

1.2 Research gaps 11 

1.3 Research problems, questions, and objectives 12 

1.4 Definitions and scope of the study 16 

1.5 Structure of the study 19 

2 Literature Review 22 

2.1 Inventory Digitization in SME Retail 23 

2.2 AI and Multi-Agent Systems in Retail Digitization 25 

2.3 Second-Hand Retail 30 

2.4 Process Automation 34 

2.5 SME Retail Digitalization 37 

2.6 Design Science Research and Theoretical Framework 39 

2.7 Identified Research Gap 42 

3 Method 44 

3.1 Design Science Research Methodology 44 

3.1.1 Stage 1: Problem Identification and Motivation 44 

3.1.2 Stage 2: Objectives of a Solution 46 

3.1.3 Stage 3: Design and Development 49 

3.1.4 Stage 4: Demonstration 51 

3.1.5 Stage 5: Evaluation 56 

3.1.6 Stage 6: Communication 56 

3.2 Data collection 59 

3.3 Data analysis 60 

3.4 Reliability and validity 61 

4 Analysis and results 63 

4.1 Artifact Demonstration: The "AI-MAS-on-Serverless" System 63 

4.2 Phase 1: Formative-Artificial Evaluation Findings 68 



4 

4.3 Phase 2: Summative-Naturalistic Evaluation – Quantitative Findings 70 

4.3.1 Efficiency Performance (Dimension2) 70 

4.3.2 Accuracy Performance (Dimension2) 71 

4.3.3 Operational Impact Analysis (Dimension3) 73 

4.4 Phase 2: Summative-Naturalistic Evaluation – Qualitative Findings 75 

4.5 Summary of Results 77 

5 Summary and conclusions 79 

5.1 Summary of Findings 79 

5.2 Implications for Theory and Practice 82 

5.3 Limitations and Directions for Future Research 84 

5.4 Concluding Remarks 85 

References 87 

Appendices 103 

Appendix 1. Complete Data Schema Specifications 103 

Appendix 2. Interview Protocol and Evaluation Instruments 104 

Appendix 3. Full Interview Transcripts (Qualitative Evaluation) 105 

Appendix 4. A Step-by-Step User Interaction Guide for i-Stock: A Visual 

Documentation Approach 111 

Appendix 5. UI Screens of the Developed Prototype 112 

  



5 

Figures 
 
Figure 1 Mapping evaluation dimentions to research objectives ................................. 15 

Figure 2 Overview of the DSR research methodology and evaluation framework. ....... 21 

Figure 3 Conceptual Map Synthesis - Connecting Theoretical Domains to the Research 

Question and Artifact Outcome. .................................................................................. 22 

Figure 4 Integrated Theoretical Framework linking technical enablers to market impact. 

Source: Adapted from Peffers et al. (2007). ................................................................. 40 

Figure 5 Research gap in AI-driven inventory digitization. ............................................ 43 

Figure  6 Product Catalog Creation Workflow. Source: Author's own creation, adopted 

from (Firebase, 2024) .................................................................................................. 53 

Figure  7  Inventory Adjustment Workflow. Source: Author's own creation. .................. 55 

Figure  8  Final "AI-MAS-on-Serverless" Event-Driven Architecture. Source: Author's own 

creation. ..................................................................................................................... 65 

 
 

Tables 
 
Table 1 Application of the DSRM Process Model 57 

Table 2 Final Firestore Database Schema (products Collection) 67 

Table 3 Cognitive Agent Prompt Engineering Iterations and Results. (Author's own 

creation). 68 

Table 4 Comparative Analysis of Process Efficiency (Per-Item Average, N=50). (Author's 

own creation based on field data). 71 

Table 5 AI Metadata Generation Accuracy in Naturalistic Setting (N=50) 72 

Table 6 Cost-Benefit Analysis and Scalability Projection 74 

Table 7 Summary of Results Mapped to Research Objectives and Questions 77 

Table 8 Summary of Research Questions, Objectives, and Key Findings 79 

Table 9 Synthesis of Empirical, Conceptual, and Practical Implications 82 

 

Abbreviations 
 
AaaS: Agent as a Service 



6 

ACID: Atomicity, Consistency, Isolation, Durability  

AI: Artificial Intelligence  

API: Application Programming Interface  

B2C: Business-to-Consumer  

BPA: Business Process Automation  

BPM: Business Process Management  

C2C: Consumer-to-Consumer  

CAPEX: Capital Expenditure  

CE: Circular Economy  

CSS: Cascading Style Sheets  

DSC: Digital Supply Chains  

DSR: Design Science Research  

DSRM: Design Science Research Methodology  

DTs: Data-driven Technologies  

ERP: Enterprise Resource Planning  

FEDS: Framework for Evaluation in Design Science  

HITL: Human-in-the-Loop 

HTTPS: Hypertext Transfer Protocol Secure  

ICT: Information and Communication Technologies  

IoT: Internet of Things  

IPA: Intelligent Process Automation  

JSON: JavaScript Object Notation  

MAS: Multi-Agent System  

ML: Machine Learning  

MVP: Minimum Viable Product  

NoSQL: Not Only SQL  

NPD: New Product Development  

OCR: Optical Character Recognition  

REST: Representational State Transfer  

RO: Research Objective  



7 

ROI: Return on Investment  

RPA: Robotic Process Automation  

RQ: Research Question  

SaaS: Software as a Service  

SDK: Software Development Kit  

SEO: Search Engine Optimization  

SKU: Stock-Keeping Unit  

SME: Small and Medium-sized Enterprise  

SPA: Single-Page Application  

SSL: Secure Sockets Layer  

SUS: System Usability Scale  

TAM: Technology Acceptance Model  

TLS: Transport Layer Security  

TOE: Technology-Organization-Environment  

UI: User Interface 

  



8 

1 Introduction 

1.1 Background of the Research 

One of the major consequences of rapid economic growth, industrialization, and demo-

graphic changes is the increase in carbon emissions, which in turn increases the environ-

mental pressures. The industrial development and growth of per capita income are still 

the major factors that cause pollution (Balsalobre-Lorente et al., 2024). The introduction 

of a circular economy, technological innovation, and the imposition of environmental 

taxes are some of the measures that can lead to a reduction in greenhouse gas emissions. 

On the other hand, industrialization and economic instability worsen the situation of the 

environment (Abbas et al., 2025). It is estimated that the fashion industry alone emits 

more than 1.2 billion tons of greenhouse gases and produces 92 million tons of waste 

each year. In addition, it is projected that nearly 57 million tons of textile waste will be 

added to the waste cycle yearly. The linear “take–make–waste” model is not only very 

expensive but also highly destructive (Charnley et al., 2022). 

While the second-hand retail sector is widely recognized as one of the main contributors 

to the circular economy, second-hand markets would only be responsible for emission 

reductions if buyers were to replace their consumption of new products with that of 

used ones. In case that buying second-hand only leads to increased consumption (i.e., 

items that would not have been bought otherwise), then its environmental impact may 

be harmful. Platforms and policymakers should work towards achieving higher levels of 

real substitution by, for example, directing their efforts towards buyers who are likely to 

purchase new products (Mak & Heijungs, 2022). Another concern raised in the literature 

regarding second-hand markets is that waste from supply chains that are not functioning 

properly in this sector. Every year more than 24 billion pieces of used clothing are ex-

ported from rich countries to poor ones, and the market is worth over $4.9 billion. A 

large portion of these imports is not marketable and therefore ends up in landfills or is 



9 

burned. This phenomenon contributes to the pollution of soil, water, and air in the coun-

tries that receive the waste (mostly in Africa), and it is called “waste colonialism” (Brooks, 

2025). 

Digitization can fundamentally fuel the shift towards a circular economy, ultimately re-

sulting in lower carbon emissions in the recycling sector. The changes brought about in 

this way not only lessen the expenses and increase the effectiveness but also conform 

to sustainable development goals and the climate goal of net-zero emissions by 2050 

(Kurniawan et al., 2023). The second-hand retail sector, commonly referred to as "kirp-

putori" in Finnish culture, is a distinct entity within the general retail market and features 

operational characteristics that make it highly different from typical retail environments. 

Unlike regular retail which depends on standardized inventory systems, universal identi-

fication codes, and product information provided by manufacturers, second-hand sellers 

have the arduous task of presenting unique and unforeseeable products that call for an 

individual evaluation, taking photos, categorizing, and writing a description specifically 

for the product (Guiot & Roux, 2010). Such variation significantly hinders digitization ef-

forts, especially those made by small and medium-sized enterprises (SMEs) which are 

limited both technically and financially. 

The situation has become increasingly dire for these companies as the trend towards 

sustainable consumption has gained traction among consumers. The worldwide market 

for the second-hand industry was worth roughly $177 billion in 2023 and is expected to 

double its value by 2027, reaching $350 billion—the reason for this increase being mainly 

the growing environmental consciousness and the widespread adoption of circular econ-

omy principles (Ellen MacArthur Foundation, 2021). The second-hand retail market in 

Finland is also booming, with consumers increasingly requiring sustainability to be taken 

into account alongside the traditional purely economic reasons for purchasing used 

goods (Turunen & Leipämaa-Leskinen, 2015). The trend has turned into a considerable 

market potential; however, it still remains largely unnoticed by small and medium-sized 

enterprises that do not have the proper digital infrastructure to compete effectively 

against digital-first platforms like Vinted, Depop, and specialized online marketplaces. 



10 

Second-hand retail has unique features when it comes to inventory management that 

are not found in traditional retail settings (D’Adamo et al., 2022). Entities existing in this 

field encounter rapid product turnover and the issues caused by extreme heterogeneity. 

To be precise, every product is an independent whole with its own specific features, dif-

ferent conditions, and sudden supply fluctuations. On the other hand, digital inventory 

management tools of the traditional kind, for example, ERP systems offered by SAP or 

Oracle, are unreasonably costly for small and medium-sized enterprises (SMEs), and usu-

ally, the annual investments required range from 10,000€ up to 50,000€ (Greg Roughan, 

2022).  

At the same time, the reasonably priced alternatives that exist, such as mid-tier plat-

forms like Shopify Plus or WooCommerce, do not have the necessary adaptability to han-

dle diverse inventories and the smart automation capabilities that can ensure the low-

cost digitization process (montonio, 2025). It is, therefore, a paradox, that there are or-

ganizations with staff numbers from 5 to 20 people who are managing 1,000 to 5,000 

unique products and yet do not have access to proper technological solutions. At the 

same time, the fast-paced growth of digitally native competitors is gradually putting 

pressure on these entities thus a clear need exists for a cost-effective system at both the 

level of design and implementation. One competent enough to act as a platform which 

could furnish independent second-hand businesses with a competitive advantage and 

thus contribute to the circular and sustainable economy through increased product turn-

over. 

Significant innovations in artificial intelligence, in particular, agent-based systems, multi-

modal models, and context engineering, have made feasible a practically cost-free and 

almost real-time digitization of store inventories. Examples of such platforms are Google 

Gemini, OpenAI’s GPT-4 Vision, and other competing multimodal systems which have 

shown their advanced capabilities in image understanding, structured data extraction, 

and natural language generation (Team et al., 2025). Besides, when these technologies 

are put together with serverless cloud infrastructure, the whole technological stack does 

away not only with the usual investment requirements, the complexity of infrastructure 



11 

management and the need for a technical person, but it also gives a chance to the SMEs 

to go for enterprise-grade automation without hefty CAPEX. Nevertheless, there still ex-

ists a large research gap in the area of empirical and practical evaluation of these solu-

tions in terms of the feasibility, the performance, and the commercial impact in the Finn-

ish second-hand retail context. 

 

1.2 Research gaps 

While each of the individual fields of retail digitalization, multi-agent systems, and the 

circular economy have been explored in depth, a careful reading of the existing literature 

suggests that there is a substantial gap in understanding how these three fields intersect. 

The existing research on AI in retail mostly focuses on mass-market commerce, where 

products are standardized and data is well-organized (He et al., 2016; Ren et al., 2017). 

Authors in these articles generally take for granted the presence of Global Trade Item 

Numbers (GTINs) and manufacturer-provided metadata. Therefore, the present theoret-

ical frameworks are not capable of explaining the second-hand markets' level of extreme 

heterogeneity, in which each product is a distinct stock-keeping unit (SKU) that has to be 

evaluated separately (Guiot & Roux, 2010). The body of research has so far failed to com-

bine the concepts of the circular economy with those of advanced automation in such a 

way that it is compatible with the latter's unstructured data environment. 

Besides this, a split can be detected between the possible extent of technological devel-

opment and the reachable level of organizational change. The articles on Intelligent Pro-

cess Automation (IPA) and Multi-Agent Systems (MAS) often spotlight the usage scenar-

ios at the level of large enterprises which entail significant capital investment and the 

availability of a specialized technical infrastructure (Chakraborti et al., 2020). On the 

other hand, the research on SME digitalization points out the "resource poverty" situa-

tion of small firms, particularly their lack of ability in implementing complex ERP systems, 

and yet, it rarely goes further in suggesting technical architectures that can overcome 



12 

these obstacles (Ifinedo, 2011). Few researchers have addressed the question of server-

less, low-code architectures as a means to make high-level automation more accessible 

to micro-enterprises. 

Moreover, the majority of digitization-related investigations in this field consider the 

work processes as a sequence of separate technical tasks, such as image classification or 

price estimation, without recognizing them as a continuous flow of operations (Liu et al., 

2021). These studies do not integrate the separate AI functionalities into one compre-

hensive "sorting-to-listing" pipeline; thus they fail to capture the dynamics of human-in-

the-loop interaction which are crucial for the actual implementation of the operations. 

The present research overcomes these shortcomings by building a bridge between the 

complex, agentic AI on one hand and the resource-poor, diverse second-hand retail mar-

ket on the other.  

 

1.3 Research problems, questions, and objectives  

Problem Statement 

The second-hand retail sector, a major contributor to the circular economy, is struggling 

with a "scalability-sustainability paradox" in which the highly diverse nature of the in-

ventory is at odds with the operational capabilities of Small and Medium-sized Enter-

prises (SMEs). While linear retail depends on standardized stock, second-hand inventory 

is a collection of unique and unpredictable goods that need labor-intensive "valuation 

practices" to turn raw donations into products suitable for the market (Fuentes & 

Hedegård, 2025; Guiot & Roux, 2010). The manual "sorting-to-listing" process involved 

here leads to high transaction costs, which in turn generate a "digitalization paradox" 

where the cost of digitizing the low-value items is higher than the potential return (Bar-

ton et al., 2024; Gebauer et al., 2020). As a result, micro-retailers who are in "resource 

poverty" (Ifinedo, 2011) are practically shut out of the digital marketplace thus the cir-

cular economy is not able to reach the scale necessary for it to have a positive environ-

mental impact. 



13 

On the technology front, the main issue is the deployment of sophisticated automation 

in these resource-poor settings. To be sure, enterprise-level inventory systems can lead 

to optimization but at the same time, they have high entry barriers in terms of costs and 

the level of technical knowledge which are difficult for SMEs to overcome (Chopra, 2019; 

Sándor & Gubán, 2022). In addition, the focus of current research on AI in retail is mainly 

on standardized e-commerce or isolated technical tasks thereby overlooking end-to-end 

"sorting-to-listing" workflows for heterogeneous goods (He et al., 2016; Liu et al., 2021). 

There is a significant shortage of real-world frameworks for how Multi-Agent Systems 

(MAS) can be designed not only to be cognitively able to deal with unstructured inven-

tory data but also operationally friendly for the non-technical staff in the unorganized 

retail sector (Narang & Tiwari, 2024), which means that the potential for AI democrati-

zation in this field is largely unexplored. 

Research Questions 

To achieve the objectives set and also to address the gaps in the previous studies, this 

investigation is largely based on one key question: 

RQ: How can multi-agent artificial intelligence systems be designed and deployed to au-

tomate inventory digitization in second-hand retail, and what quantifiable operational 

impacts emerge from this automation relative to manual practices? 

The answer to this main question is obtained through five interconnected evaluation 

dimensions. Instead of five separate questions, the single overarching inquiry of the 

study was operationalized by examining the artifact and its performance across five an-

alytical dimensions: (1) architectural design, looking into the structure of multi-agent 

systems to meet SME requirements; (2) performance, measuring accuracy and efficiency 

metrics under naturalistic conditions; (3) operational impact, investigating economic and 

operational advantages in comparison to manual methods; (4) user acceptance, deter-

mining employees' perception and adoption of the system; and (5) transferability, con-

sidering whether the design principles and patterns can be generalized to other re-

source-constrained contexts.  



14 

Research Objectives 

Three interrelated research objectives reflect the main research question operationally, 

with each of them covering a different aspect of the research: 

(RO1) To design and develop a multi-agent AI system for automating inventory digitiza-

tion in second-hand retail. This goal includes architectural decisions, technology stack 

choices, agent specialization and coordination mechanisms, human-in-the-loop integra-

tion, and complete end-to-end implementation from the user interface at the front end 

to the cloud infrastructure at the back end. The achievement of the goal will be judged 

by the deployment of the artifact in the operating environment and demonstration of all 

the specified functional requirements. 

(RO2) To evaluate the system's accuracy, efficiency, and usability in a real-world store 

environment. Carry out a detailed assessment of the system's accuracy, efficiency, and 

usability, as implemented, through its deployment in the operations of a real second-

hand retail environment. This comprehensive evaluation covers: (1) AI agent accuracy 

indicators (classification accuracy, metadata completeness, correctness of label tran-

scription); (2) system performance indicators (processing latency, throughput, reliability); 

(3) operational efficiency indicators (per-item processing time, overall productivity im-

provement); (4) user adoption indicators (system usability scale, staff satisfaction, will-

ingness to adopt); and (5) this assessment is based on Design Science Research principles 

and uses the Framework for Evaluation in Design Science (FEDS), which combines form-

ative and summative, artificial, and naturalistic assessments. 

(RO3) To analyze the operational impacts of the system compared to manual practices 

and provide recommendations for scaling in SME retail. Operational impacts of the sys-

tem compared to manual practices, as well as recommendations for scaling in SME retail, 

are to be analyzed. Measure the economic and operational effects of AI-based inventory 

digitization as compared to manual methods, create a persuasive business case to facil-

itate technology adoption decision-making, and research scalability across different sec-

ond-hand retail contexts. This goal includes: (1) cost–benefit analysis (labor savings, sys-



15 

tem costs, return on investment); (2) measurement of operational impacts (improve-

ment in inventory accuracy, catalog completeness, changes in staff time allocation); (3) 

business case development; and (4) evaluation of generalizability and adaptability across 

different retail scenarios. 

Mapping of Evaluation Dimensions to Research Objectives 

The architectural dimension is the main factor that shows the support of RO1 by explain-

ing the creation and implementation of a multi-agent AI system within the technical and 

organizational constraints of SMEs. The performance and user-acceptance dimensions 

are the main factors that demonstrate the support of RO2. They involve measuring the 

system's accuracy, efficiency, and usability via detailed formative and summative evalu-

ations based on the FEDS framework. The operational-impact and transferability dimen-

sions are the main factors that show the support of RO3 by determining economic ben-

efits, creating business cases, and finding generalizable design patterns that facilitate the 

spread of the retail sector in different contexts. Figure 1 illustrates the dimensions of the 

main research question (RQ) and relation to addressing the research objects (ROs). 

 

Figure 1 Mapping evaluation dimentions to research objectives 



16 

1.4 Definitions and scope of the study 

This study lies at the crossroads of two worlds that are quite different from each other- 

one aspect of it is a state-of-the-art AI technology, and the other aspect is the most or-

dinary practical side of retail. Firstly, the five fundamental domains are outlined in order 

to be quite clear about the ideas, and secondly, the specific places of their application in 

the present research are disclosed. 

Multi-Agent Systems (MAS) are the main technological base of this research. According 

to computer science literature, a multi-agent system is a network of software agents that 

are loosely connected and cooperatively interact to solve problems that are beyond the 

capabilities or the knowledge of any one of them (Wooldridge, 2012). These agents are 

described as being autonomous, socially able, reactive, and pro-active, which features 

enable them to function in ever-changing surroundings without the need for continuous 

human intervention (Russell et al., 2022). Within the framework of this work, the expres-

sion is used solely to refer to the development and the actualization of "AI-MAS-on-Serv-

erless," a serverless architectural artifact consisting of five specialized agents: an Image 

Processing Agent, a Categorization Agent, a Metadata Extraction Agent, a Label Tran-

scription Agent, and an Inventory Synchronization Agent. These agents implement Large 

Language Models (LLMs) to complete the highly coordinated task of turning unstruc-

tured visual data into structured inventory records. 

Inventory Digitization is basically the process of making physical assets into digital ones 

for easier management, trading, and participation in the digital ecosystem. At the same 

time, the recording of stock levels is not the main thing, however, through the use of 

structured metadata, the physical attributes of the products are "digitized" - for example, 

classification, condition, and specifications, which allow being searched and data-driven 

decision-making (Chopra, 2019). The scope of this research takes inventory digitization 

only as far as the "sorting-to-listing" workflow. It is concentrated on the transformation 

of physical second-hand clothing and household items into well-structured JSON data 

and high-quality photos that can be used by e-commerce platforms, while the down-

stream logistics of shipping or fulfillment are left out. 



17 

Second-Hand Retail that is usually called re-commerce or the resale market, is the trade-

in of goods that have been owned before. The difference with traditional retail, which is 

based on the use of standardized stock-keeping units (SKUs) and deep inventory levels, 

is that the second-hand retail sector features extreme inventory heterogeneity where 

each piece is singular in terms of its condition, origin, and specifications (Guiot & Roux, 

2010; Turunen & Leipämaa-Leskinen, 2015). This study is confined to the operational 

context of the Finnish “kirpputori” and brick-and-mortar indie resale stores. It is about 

the enormous challenge of managing "markets of one," where the cost of digitizing one 

single, lowly-priced unique item is usually more than the potential profit margin., lowly-

priced unique item is traditionally more than the possible profit margin. 

SME Retail Digitalization is an umbrella term referring to the process when SMEs em-

brace and integrate digital technologies in order to change their business models and 

create new paths for revenue and value generation (Gartner, 2025). The academic liter-

ature shows that SMEs experience a kind of "resource poverty" that is different from 

large enterprises, as they are deficient both in the financial resources needed for enter-

prise resource planning (ERP) systems and the specialized technical skills required for 

complex infrastructure maintenance (Ifinedo, 2011). This research scope is designed to 

overcome these limitations by coming up with a solution that does not require any in-

frastructure management (serverless architecture) and is able to function within a mi-

cro-enterprise budget (approximately less than €1,000 monthly operational cost). The 

focus is on the "democratization" of highly advanced automation, i.e., making enter-

prise-level AI technology available to staff without a technical background in small retail 

businesses. 

Process Automation is the embracing and employment of tech to carry out repetitive 

jobs or processes in a company which are currently done by hand but can be replaced or 

supplemented by machines. Whereas standard Business Process Automation (BPA) is 

concentrated on rule-based tasks, Intelligent Process Automation (IPA) goes a step fur-

ther by adding AI to be able to handle unstructured data and cognitive tasks (Chakraborti 

et al., 2020; Van Der Aalst et al., 2018). As to this thesis, the scope is the cognitive part 



18 

of the product valuation and description workflow that's been automated. The research 

uses a "Human-in-the-Loop" (HITL) model whereby the system automates 90% of the 

data entry and classification tasks, leaving the human operator to verification and the 

final decision-making role, thus the retail worker is supplemented, not replaced. 

This research is concerned with the multi-agent AI system's design, Analysis, and perfor-

mance evaluation, named i-Stock (https://i-buy-stock.web.app), and the associated UI 

portal, i-Buy (https://i-buy-en.web.app). The system is made up of 5 parallel AI agents 

whose interaction and positions are shown in Figure  8 . (1) Image Processing Agent uti-

lizing computer vision for background removal and image enhancement. (2) Product Cat-

egorization Agent carrying out hierarchical classification across 150+ categories of sec-

ond-hand retail. (3) Metadata Extraction Agent producing structured product infor-

mation with category-specific schemas. (4) Label Transcription Agent getting item codes 

from product labels through optical character recognition. (5) Inventory Synchronization 

Agent implementing real-time quantity changes through transactional database opera-

tions. 

Regarding the technology stack, the design and implementation of the artifact will be 

limited to a serverless cloud stack mainly consisting of Google Firebase (Cloud Functions, 

Firestore) (Chanaka, 2023) and the Google Gemini multimodal API (Generating Content 

| Gemini API, 2025). 

(1) Integration of Firebase platform (Firestore database, Cloud Storage, Cloud 

Functions, Hosting). 

(2) Frontend of a single-page web application allowing staff interaction 

(3) Serverless backend architecture that does not require infrastructure man-

agement. 

The artifact evaluation will be formally structured using the FEDS framework. It will in-

clude quantitative validation of performance metrics (time reduction, accuracy) and 

qualitative validation of user acceptance through semi-structured interviews. The eval-

uation of the artifact will be formally organized around the FEDS framework It will entail 

quantitative validation of the performance metrics (time reduction, accuracy (Prat et al., 

https://i-buy-stock.web.app/
https://i-buy-en.web.app/


19 

n.d.)) and qualitative validation of the user acceptance through semi-structured inter-

views (George, Tegan, 2023). 

 

1.5 Structure of the study 

The thesis after this introduction comprises four other significant sections that are sys-

tematically organized to address the research questions. Chapter 2, Literature Review is 

a review of the theoretical bases in five interrelated domains. Section 2.1 is concerned 

with inventory management theory and its use in second-hand retail, thus uncovering 

the fundamental mismatch between the classical assumptions and the operational real-

ities. Section 2.2 acquaints readers with AI and multi-agent systems that can be utilized 

for the automation of retail functions, also including computer vision, natural language 

processing, multimodal AI capabilities, and retailer applications. Section 2.3 sets Design 

Science Research as the method to be used for artifact-focused research. Section 2.4 

considers decision science viewpoints on process automation and technology adoption 

in SMEs. Section 2.5 integrates the learnings from all the previous sections into one the-

oretical framework guiding artifact design and evaluation. 

In Chapter 3, Method is a detailed account of the research design and execution plan. 

The work is carried out in accordance with the Design Science Research six-stage process 

model by Peffers et al. The protocol for the artifact design specifies the rationale for 

technology selection, the architectural decisions, and the implementation details. The 

evaluation plan makes use of the FEDS framework which combines formative/summa-

tive and artificial/naturalistic evaluation. The data collection plan considers both the 

quantitative metrics and the qualitative insights. The analysis plan makes use of the ap-

propriate statistical and thematic methods. 

In Chapter 4, Results offers the comprehensive account of the empirical findings. The 

outcome of the artifact development documents the technical specifications and the im-

plementation metrics. The results of the formative evaluation detail the iteration of the 



20 

design and the validation of the performance. The summative evaluation findings pre-

sent quantitative performance metrics, accuracy assessments, and user acceptance data 

gauged against the pre-established success criteria. The business impact analysis quan-

tifies costs, benefits, and return on investment while also forecasting scalability and gen-

eralizability potential. 

Finally in Chapter 5, Conclusions is a synthesis of the research findings and a statement 

of the contributions. The findings are directly linked to the research questions. Theoret-

ical implications concern multi-agent systems, Design Science Research methodology, 

and human-AI collaboration. Practical implications provide a guide for making technol-

ogy adoption decisions. Limitations acknowledge research constraints openly. Future re-

search avenues point to extensions and related investigations made possible by this 

foundational work. 

The research is centered around five fundamental keywords that stand for both the tech-

nical and business aspects: 

(1) Multi-Agent AI: A set of specialized artificial intelligence agents that coop-

erate and coordinate to achieve complex organizational goals 

(2) Inventory Digitization: The conversion of physical inventory assets into 

structured digital representations which makes them discoverable and 

manageable 

(3) SME Retail Digitalization: The technology-driven transformation of small 

and medium enterprises in the retail sector 

(4) Second-Hand Retail: Commercial operations that are specialized in pre-

owned, heterogeneous product inventory (kirpputori) 

(5) Process Automation: Manual work is reduced by the intelligent automa-

tion of the repetitive workflow 

These keywords position the research at the point where the emerging artificial intelli-

gence capabilities, practical retail challenges, small business constraints, and sustainable 



21 

consumption models intersect. In Figure 2, The overwiew of the thesis structure and DSR 

research methodology demonstrated. 

 

Figure 2 Overview of the DSR research methodology and evaluation framework.  

 

 



22 

2 Literature Review 

The literature review synthesizes the findings of scholarly research from five interrelated 

domains and organizes them around five main keywords. Instead of detailing these five 

domains separately and independently, the review demonstrates that each keyword pro-

vides specific insights that help to fill the gaps in the existing research. This section finally 

brings together the five domains and connects them with the research problem and re-

search questions presented in the introduction. Figure 3 synthesizes these domains into 

a unified conceptual framework for this study. 

 

Figure 3 Conceptual Map Synthesis - Connecting Theoretical Domains to the Research Question 

and Artifact Outcome. 

 



23 

2.1 Inventory Digitization in SME Retail 

Inventory digitization is the fundamental change of the company’s physical assets into 

well-structured digital formats that involve not only imaging but also classification, de-

tailed description, and real-time tracking in order to make the assets easily accessible 

and manageable through the digital platforms (Chopra, 2019). Proper digitization is not 

simply about inputting data; it involves complete integration of physical and digital pro-

cesses, together with metadata creation, data management, and synchronization of as-

sets (Fisher, M. L., & Raman, A., 2010). Nevertheless, the path of this change is not always 

straightforward, especially for SMEs, which have different developmental ways than 

large corporations (Sándor & Gubán, 2022). Large enterprises are capable of handling 

complex digital ecosystems as a result of their structural maturity. On the other hand, 

SMEs frequently become a challenge due to the fact that their business processes are 

not well-defined, hence it is difficult for them to apply standard digital tools. The digiti-

zation success is dependent on the business being operationally resilient, cost-saving, 

and having the ability to cope with volatile markets (Kallmuenzer et al., 2025). 

Even though inventory management is the major-thing which should be digitized, it has 

still been the bottleneck of SMEs that are heavily dependent on manual processes and 

thus very sensitive to human errors. Conventional manual methods result in slower in-

ventory cycles and inaccuracies, which consequently have a direct impact on customer 

service and operational expenses (Tung-Hsiang Chou et al., 2024). Panigrahi et al. (2024) 

argue that efficient inventory management is the lifeblood of SMEs because these busi-

nesses operate under resource constraints and thus have a higher risk of cash flow and 

storage problems. Therefore, inefficiencies in this field, e.g., time that is lost due to man-

ual cataloging or data quality that varies because of human inconsistency, have a dispro-

portionate impact on SME performance as compared to large companies (Fisher, M. L., 

& Raman, A., 2010; Silver et al., 2021).  

The technology environment that is supposed to solve these problems, however, pre-

sents a dilemma that very often leaves SMEs, especially micro enterprises, without a 

feasible option. On one side of the continuum are ERP systems such as SAP or Oracle 



24 

that offer a complete solution for inventory optimization but need a large capital outlay 

and technical skills, thus elevating the entry barrier for those companies that have a tight 

budget (Chopra, 2019). On the other hand, typical cloud services for SMEs like Shopify 

or WooCommerce are good in terms of cost but are often inflexible for inventories that 

are heterogeneous and are the case in second-hand retail (Ifinedo, 2011). Although the 

advancements in the Internet of Things (IoT) and cheap microcontrollers have opened 

up possibilities for affordable data collection and giving updates in real-time (Barton et 

al., 2024), the hardware-dependent solution installation usually requires technical 

knowledge which is not within the average retail SME. This leads to the existence of a 

"digitalization paradox" where companies realize that they must invest in digitalization 

in order to grow. But at the same time they find it difficult to see immediate increases in 

revenue or operational efficiency because of the high effort needed to implement tech-

nologies that do not fit well (Barton et al., 2024; Gebauer et al., 2020). 

Moreover, the peculiarities of data-driven technologies (DTs) significantly determine 

their contribution to sustainability and operational efficiency. Hernández et al. (2024) 

postulate that implementation of data-gathering and data-analysis technologies leads to 

sustainable operations; however, firm size is a very important factor that changes the 

effect since bigger SMEs can leverage sophisticated analytics and AI more easily than 

micro firms. In other words, the so-called technology 'democratization' does not happen 

automatically; micro-retailers need not only cheap but also low-tech solutions to be able 

to overcome the resource gap (Hernández et al., 2024). The difficulty is also due to the 

fact that "human-centered" methods in Industry 5.0 are required, which point out that 

technological adoption has to be combined with behavioral knowledge and human skills 

if it is to result in improved performance (Panigrahi et al., 2024). 

Therefore, there is a considerable gap in research at the point where AI adoption, inven-

tory heterogeneity, and the operational constraints of micro-retailers intersect. The ex-

isting body of work has a strong focus on systems at the enterprise level (Chopra, 2019) 

and on the technical aspects in isolation such as image processing (He et al., 2016), but 

it does not consider the complete workflow of stocktaking for unique, non-standardized 



25 

items in resource-constrained environments. Researchers have not sufficiently ad-

dressed the question of how easily accessible multi-agent AI systems can be that auto-

matically handle the extremely cognitively demanding "sorting-to-listing" workflows 

while at the same time being financially sustainable for micro-enterprises (Chavez et al., 

2022). This study is the one that intends to bridge the gap by building an inventory dig-

itization system that is not only for the second-hand SME sector but also that takes into 

account the sector's unique constraints, with the features of affordability, non-technical 

staff usability, and AI integration to overcome the traditional scalability limitations. 

On the other hand, digitization is a necessary provision for the infrastructure of modern 

retail, but merely having static databases is not enough for the dynamic nature of sec-

ond-hand inventory. The reason for this is that used goods do not have standard identi-

fiers (like barcodes), and therefore, static digitization methods fail to capture their 

unique attributes efficiently. In order to surpass this limitation, the literature proposes 

going beyond passive data entry to cognitive automation. This requires revisiting Artifi-

cial Intelligence and Multi-Agent Systems (MAS), which can provide the necessary flexi-

bility to deal with unstructured data. 

 

2.2 AI and Multi-Agent Systems in Retail Digitization 

The digital transformation of retail has been the main driver of a total reconsideration of 

business ecosystems, which have put AI and multi-agent systems at the core of both 

operational innovation and customer experience optimization. The change goes far be-

yond a mere incremental enhancement of current abilities, as it is an entirely new para-

digm of how retail organizations can create and deliver value in complex and intercon-

nected market environments. 

The fundamental design of an AI-powered retail transformation is based on three inter-

connected technological aspects: advanced machine learning algorithms that can handle 

huge transactional datasets, autonomous agent systems that function as goal-directed 

entities across various decision domains, and integrated analytics frameworks that unify 



26 

various data streams into actionable intelligence (Ameya Gokhale, 2025). These instru-

ments together make it possible for retailers to go beyond their traditional reactive busi-

ness models. They use instead predictive and prescriptive models that allow them to 

foresee market changes at a level of detail that has never been achieved before. In fact, 

nowadays, retail can count on AI systems that monitor consumer behavior by looking at 

more than 500 distinct data points per customer, thus they are able to create personal-

ized recommendations and interventions that, in a measurable way, improve conversion 

rates by 15% and, at the same time, increase average order values by 20% (David Iya-

nuoluwa Ajiga et al., 2024). 

The realization of these potentialities on the operational level is, thus, the most conspic-

uous feature of advanced recommendation engines that use collaborative filtering, con-

tent-based analysis, and hybrid methods to foresee customer preferences with a preci-

sion level of over 95% (Ameya Gokhale, 2025). The studies indicate that such AI-driven 

recommendation systems are the main influencers in up to 35% of all transactions on 

major e-commerce platforms and that, in these platforms, they have been able to raise 

the conversion rate by 75% compared to the traditional targeting methods (Satish Krish-

namurthy et al., 2024). The technological sophistication which is the basis for such an 

accomplishment involves transformer-based architectures that are capable of executing 

recommendation queries within 100 milliseconds and thereby are able to ensure real-

time personalization which changes accordingly to the customer's context (Ameya Gok-

hale, 2025). 

Conversational AI has been one of the most influential developments in the customer 

service area by creating autonomous chatbots that can understand natural language and 

even exhibit emotional intelligence. The rate of deployment has increased by 85% since 

2020, and contemporary systems are now able to achieve 82% accuracy in sentiment 

detection, thus, customer satisfaction has increased by 35%, and first-contact resolution 

rates (FCR) have decreased by 28% (Fnu Imran Ahamed, 2025). All these improvements 

are the result of transformer architectures that significantly raise the context retention 

level from 58% (for rule-based systems) to 89%  (David Iyanuoluwa Ajiga et al., 2024). As 



27 

a result, agents are now able to handle 75% of the routine inquiries without intervention 

and recognize the need for escalation with 95% accuracy. The reduction in the agency of 

operations due to this autonomy has brought about a 60% cut in operational costs and 

a decrease in response times from hours to less than three seconds (Fnu Imran Ahamed, 

2025). 

AI-powered retail revolution changes the supply chain management massively by the 

usage of predictive analytics, which eliminates the drawbacks of traditional forecasting. 

Learning algorithms on their own, after analyzing the impact of different variables such 

as sales history, weather, and sentiment over results, attain now a 92% accuracy rate of 

forecasts for regular products and an 84% one for seasonal items. This way retailers can 

lower the occurrence of out-of-stock situations by 80% and that of excess inventory by 

32% (David Iyanuoluwa Ajiga et al., 2024; Satish Krishnamurthy et al., 2024). Advanced 

neural networks that are able to consider more than 50 variables at a time achieve al-

most 94% accuracy in stable product categories (Satish Krishnamurthy et al., 2024). Such 

accuracy is being extended to supply chain operations, thus resulting in a 22% saving in 

transportation costs and a 35% improvement in supplier performance. In the end, pre-

cise demand forecasting is the cause of markdown spending reduction by 30%, thus con-

tributing to overall profitability optimization (David Iyanuoluwa Ajiga et al., 2024). 

MAS show a major change in the retail field, where independent AI units that manage 

each other in different but connected areas to achieve the best shared results are used. 

Studies on the subject have been focusing on the development of such agentic architec-

tures that manifest purposeful behavior, are able to reason adaptively, and possess long-

term memory in complicated scenarios of work (Trivedi, 2025). Such systems are basi-

cally different from conventional AI ones in that they can make decisions autonomously 

without direct human intervention, they are proactive in action as they can specify sub-

goals, invoke application programming interfaces, and figure out trade-offs among com-

peting objectives (Trivedi, 2025). Examples of practical implementations are virtual shop-

ping assistants that, in an autonomous manner, can check inventory across distribution 



28 

networks, can perform purchasing transactions with the use of stored payment creden-

tials, and coordinate same-day delivery from several fulfillment centers (Trivedi, 2025). 

The multi-agent architecture sophistication extends to the composable commerce 

frameworks where microservice backends and headless frontends dynamically interact 

with individual agents thus enabling quick adjustment to market changes while at the 

same time, presenting new security issues concerning agent autonomy and cross-system 

coordination (Trivedi, 2025). 

Among the various approaches, reinforcement learning methods, especially the contex-

tual bandit and deep learning techniques, have been the most successful in the retail 

agents' training for the pricing and promotion strategies. The performance of these ap-

proaches, as compared to the static policies, is so outstanding that they may increase 

the revenue by up to 26% as is evident from the simulations with offline batch data (Xia 

et al., 2024). In particular, Proximal Policy Optimization achieves maximum revenue by 

offering lesser discounts to price-insensitive customers, although it may produce less 

than optimal results at times due to data limitations. These frameworks have been suc-

cessful in handling complex decision models that include store visit probabilities, prod-

uct choices, and pricing sensitivities, thus, they are instrumental in solving the problem 

of how to combine the generation of revenue with the giving of discounts in the most 

efficient way (Xia et al., 2024). 

Computer vision is another area of AI-enabled retail transformation with an annual 

growth rate of 65% in its implementation and the early users claiming, on average, a 40% 

increase in operational efficiency along with a 35% rise in customer satisfaction (Fnu 

Imran Ahamed, 2025). The systems attain product identification accuracy levels of up to 

96% even though the lighting conditions may vary, thus they enable the continuous mon-

itoring of shelves at 15-minute intervals which leads to a cutting down of the manual 

auditing tasks by 85% and at the same time, an increase in stock accuracy of 92% (Fnu 

Imran Ahamed, 2025). 

Undoubtedly, these massive progresses have been made; still, there remain significant 

gaps that the researchers need to bridge, if they are to fully unleash the power of AI and 



29 

multi-agent systems in retail settings. Firstly, the strong points of isolated AI applications 

within single fields do not bring forth fully fledged frameworks that integrate several AI 

technologies along the entire retail value chain (David Iyanuoluwa Ajiga et al., 2024). The 

fragmentation issue here stands in the way of understanding the synergies between dif-

ferent tech components in achieving overarching retail goals. Secondly, the predominant 

emphasis on large enterprises hides the distinct challenges and opportunities small and 

medium-sized retailers have, who, in turn, struggle with resource constraints and oper-

ational realities which may not allow them to adopt enterprise-scale solutions directly 

(David Iyanuoluwa Ajiga et al., 2024). Third, the security risks of agentic AI need a lot 

more in-depth probing especially in terms of vulnerabilities resulting from autonomous 

decision-making capabilities that blur traditional boundaries between client and server, 

user and executor (Trivedi, 2025). Finally, there is a demand for a more thorough inves-

tigation into the transfer of agent learnings from simulation to real-world applications, 

as the prevailing mechanistic models lean more towards interpretability and tunability 

rather than predictive accuracy, thus, they may have limited practical use (Xia et al., 

2024). 

The collective effect of these shortcomings is to draw attention to a research core that 

is crucial: the designing of integrated, ethically founded, and empirically validated de-

ployment frameworks for AI and multi-agent systems which can be applied in varied re-

tail contexts. The frameworks should not only deal with tech sophistication but also ac-

count for organizational readiness, governance mechanisms, and human-AI collabora-

tion models evolution. Retail digitization path is moving toward the dependence on the 

resolution of these basic issues i.e., the tension between technological capability and 

practical use, the one between human oversight and autonomous agent operation, and 

finally, the conflict between data-driven personalization and privacy which are the chal-

lenges forming the new research agenda for AI-powered retail transformation. 

While MAS architectures have been able to effectively optimize typical retail chains, they 

still pose difficulties when one tries to apply them to the circular economy, according to 



30 

the research paper. The resale market is a supply chain with high uncertainties and vari-

abilities and thus, unlike most AI literature that assumes predictable supply chains, it is 

quite different. So, in order to figure out the exact design of the AI agent that is required, 

one has to understand the real-life operations of the Second-Hand Retail sector. 

 

2.3 Second-Hand Retail 

The second-hand or the resale/re-commerce market, is a separate unit that is rapidly 

growing to become one of the most significant parts of the global retail industry. Globally, 

the market for pre-owned goods was worth close to 177 billion dollars in 2023 and is 

expected to more than double by 2027 to reach a value of $350 billion (Ellen MacArthur 

Foundation, 2021). However, the ways of operating behind this growth are very different 

from those of the traditional retail market. Compared to the traditional linear model of 

"take–make–waste", second-hand retail is seen as a primary driver for the circular econ-

omy (CE), which helps in the redistribution and reuse of the already existing resources 

(Machado et al., 2019). Nevertheless, Nascimento et al. (2019) emphasize that accom-

plishing circularity is not only dependent on the growth of the market; it also requires 

the incorporation of sophisticated technologies such as Industry 4.0 for the optimization 

of resource loops and the reduction of waste.  

The biggest difference between traditional retail or second-hand retail is the kind of 

stock. While normal retail is based on identical units of stock and a supply chain without 

surprises, the second-hand market is highly diverse in terms of products and is always 

on the "eternal hunt" for new sources (Appelgren, 2019). Businesses and stores get their 

products from various, and most of the time, unexpected avenues like donations, con-

signment, and direct purchases from consumers thus ending up with a stock where al-

most every single piece of the stock is a unique stock-keeping unit (SKU) with a certain 

past, state, and feature set (Guiot & Roux, 2010). Such openness to change undermines 

the prevailing logic of retail efficiency. According to Yrjölä et al. 2021, second-hand firms 

may be identified as "connector" platforms which only provide a means for peer-to-peer 



31 

exchange with little or no intervention, and "controller" models (branches and profes-

sionalized SMEs) in which retailers fully take care of quality assurance, inventory man-

agement, and transaction fulfillment. 

It is these "controller" models where the operational issues are most prominent. Far 

from being a simple task of passing through the logistical chain, the transformation of a 

donation or acquisition into a marketable product is a complicated socio-material pro-

cess of valuation (Fuentes & Hedegård, 2025). An item is subjected to a series of activi-

ties such as sorting, modification (cleaning or repair), pricing, and marketing before it 

can be sold. (Fuentes & Hedegård, 2025) provide the argument that in the second-hand 

market, value is not implicit but comes into existence through these labor-intensive pro-

cedures. Retailers are required to utilize particular valuing registers (evaluating brands, 

vintage status, material quality, and sustainability potential) to resell a product. This 

"production of value" consumes a significant amount of staff skills and time as employ-

ees have to work not only as curators who devalue (discard) the unsuitable items but 

also as revalorize by creating stories through which they attract potential buyers for 

other items. 

The process of valorization deeply involves marketing and pricing strategies. Hedegård 

(2024) points out that second-hand sellers are under the obligation to "frame" their used 

goods offerings so as to win over consumers' fears of the products being unsanitary or 

of low-quality. Through this framing, they turn to storytelling, photography, and visual 

merchandising to persuade consumers that the products are "environmentally friendly," 

"creatively experiential," or giving the customer "value for money." Moreover, pricing is 

very flexible and full of doubts. Gu et al. (2023) discovered that in contrast to the fixed-

price idea for new products, second-hand pricing decisions are affected by a variety of 

factors like time elapsed since the product's original release, consumer engagement with 

product features, and emotive descriptions. Frequently, sellers need to adjust prices 

based on market feedback and how long the product has been left unsold, thus they add 

another layer of complexity to the decision-making process for each single item in their 

stock. 



32 

The presence of these complicated operations becomes even stronger when SMEs at-

tempt to scale up. Hultberg (2025) reveals that scaling circular business models entails 

different strategic stages, starting from simple organizational growth ("breadth-scaling") 

and ending with influencing broader consumer habits and institutional practices 

("depth-scaling"). Meanwhile, Tritto et al. (2024) talk about how the SMEs are generally 

more inclined to adopt circular practices as they are more flexible but at the same time 

they have limited resources compared to bigger companies. Their research across the 

EU shows that the readiness to carry out CE practices is mostly dependent on the size of 

the company and its tangible resources, thus linking this with the presence of a "scala-

bility bottleneck" for smaller players. This is the point where digital tools come into play. 

Suchek et al. (2024) show that industry 4.0 technologies like AI, big data, and cloud com-

puting can be essential resources for SMEs, allowing them to carry out complex tasks 

through automation and becoming more efficient participants in global value chains. 

Nevertheless, they are exhorting that if the SME management is not done carefully, sim-

ultaneously taking on digitalization and global expansion may lead to resource con-

straints. 

The role of technology in lessening these difficulties is quite significant although the is-

sue has received little attention in the case of second-hand retail governance and oper-

ations. Dwivedi & Paul (2022) put forward a concept for "Digital Supply Chains" (DSC) in 

the circular economy, where they claim that the digital resources can break down the 

barriers connected with the shortage of skills and facilities. Also, Acosta Llano et al. (2025) 

make a point that blockchain technology can improve transparency, incentivization, and 

standardization that are the main elements to trust-building in the origin and quality of 

second-hand goods. Still, as Nascimento et al. (2019) state, practical application of these 

technologies in functional business models for recycling and reuse is quite challenging 

and, most of the time, it needs new "smart production systems" which are still at the 

early stages. 



33 

As a result, the second-hand SMEs are still experiencing a heavy operational load. The 

necessity to take a photo of, describe, attribute, price, and market each unique item in-

dividually out of thousands makes it impossible to work on the scale issue through purely 

manual processes. Large platforms may operate "connector" models which put the re-

sponsibility of this type of work on users and thus offload it from themselves (Yrjölä et 

al., 2021), however, the traditional brick-and-mortar SMEs and independent online 

sellers are fully responsible for these digitization and valuation processes. Even though 

purchasers are more and more driven by economic, critical (environmental), and recre-

ational (treasure hunting) factors (Baruönü, 2025; Machado et al., 2019), the supply side 

is still incapable of fulfilling this demand in an efficient manner because high transaction 

costs are involved in managing heterogeneous inventory. 

This marks a vital research gap in the present literature. Existing works mainly concen-

trate on consumer motivations (Baruönü, 2025; Machado et al., 2019), the sociology of 

valuation practices (Appelgren, 2019; Fuentes & Hedegård, 2025), macro-level typolo-

gies of business models (Yrjölä et al., 2021), or advanced technological structures for the 

circular economy (Acosta Llano et al., 2025; Dwivedi & Paul, 2022). Few are the pieces 

of empirical work that focus on how specific technical artifacts, such as multi-agent AI 

systems, could ease the heavy labor load associated with valuation and digitization for 

SMEs. While the theoretical intricacies of why second-hand retail is difficult being laid 

out, and the potential of digital technologies being recognized (Nascimento et al., 2019; 

Suchek et al., 2024), little is done in terms of research about intelligent automation that 

offers a practical way to streamline the "sorting-to-listing" workflow. This study intends 

to fill that void by creating and scrutinizing a system that automates these complicated 

valuation and digitization tasks, thereby resolving the tension between the need for 

high-quality product framing (Hedegård, 2024) and the resource constraints of small 

retailers.  

An examination of the second-hand industry has unveiled a significant limitation: the 

task of valuing the uniqueness of items is so labor-intensive that it often surpasses the 

profit margin. This results in a 'scalability paradox' whereby growth leads to complexity 



34 

that is difficult to manage. Its solution demands not only more advanced software but 

also a fundamental understanding of the workflow itself. Therefore, the next chapter 

deals with Process Automation theories to find ways through which human, cognitive, 

and manual tasks, such as grading a used shirt, can be turned over to automated agents 

in a systematic manner. 

 

2.4 Process Automation 

Business Process Management (BPM) along with its technological implementation by 

means of Business Process Automation (BPA) have changed from being simply a means 

of operational support to becoming the main factors of organizational competitiveness 

and value creation. Whereas traditional views consider BPA as a technology that enables 

the automation of activities or services that are part of a certain function or workflow, 

the newest research studies highlight its significance in the absolute transformation of 

the organization's value-creation process. Automation, by shortening the time taken for 

routine tasks and reducing human involvement in repetitive workflows, frees employees 

for creative and strategic work, which is a change that is especially important for SMEs 

that want to be competitive in the digitized markets (Moreira et al., 2024). The transition 

coming from this technological change is not only about moving from static record-keep-

ing to dynamic, data-driven ecosystems where real-time visibility and decision-making 

capabilities determine market responsiveness and resilience in the case of retail and sup-

ply chain management (Ali et al., 2024; Tiwari et al., 2024). 

The department of SMEs has been a target of automation technologies where the latter 

has been welcomed with a lot of hesitations and a slow pace of implementation. Sys-

tematic reviews done lately reveal that big enterprises have been able to harness the 

power of Enterprise Resource Planning (ERP) and advanced robotics while SMEs are on 

the other side of the "digital divide" which is marked by financial difficulties, the lack of 

technical skills, and the non-existence of structured implementation methods (Moreira 

et al., 2024; Neri et al., 2023). The importance of this difference lies in the fact that the 



35 

preparedness for Industry 4.0, and especially the embracing of the digital technologies 

like the Internet of Things (IoT) and AI, is gradually becoming the main factor that deter-

mines whether a business can thrive or scale (Yosephine et al., 2025). 

Automation is necessary for various reasons, one of which being that a Circular Economy 

(CE) is the second-hand retail operational context. Presently, digital technologies are 

considered the primary enablers of the circular strategies that mainly target "slowing" 

and "closing" resource loops through reuse, refurbishment, and recycling (Neri et al., 

2025; Piedra-Muñoz et al., 2025). On this platform, automation is not only a tool for 

resource efficiency through optimising logistics and inventory management but also a 

facilitator of the data necessary for product value to be retained over time (Rusch et al., 

2023). Nevertheless, the existing literature arguing that digital technologies are indeed 

theoretical enablers of such circular strategies is yet to be convinced by the firm evidence 

of their deep integration into the core business models of SMEs (Neri et al., 2023).  

To a large extent, this tech space is covered by Robotic Process Automation (RPA), which 

mechanizes rule-bound, structured tasks through simulations of human beings' interac-

tions with user interfaces. In contrast, RPA has been limited to the use of environments 

with unstructured data and high variability, where only a few situations have been in 

administrative sectors world, like finance and administration, where it has been most 

successful (Crisan et al., 2023). Typical RPA is problematically rigid due to the fact that it 

demands structured inputs and definite rules while at the same time, the characteriza-

tion of unique, pre-owned items is the process that involves the use of unstructured 

inputs like pictures and variable text descriptions. As such, academic researchers recom-

mend cognitive automation or intelligent process automation (IPA) that combines AI and 

ML to address the issue of ambiguity, learn from data patterns, and execute complex 

decision-making tasks that a conventional software robot cannot make (Crisan et al., 

2023; Moreira et al., 2024). The transition from "doing" (RPA) to "thinking" (AI) is quite 

important if we consider the case of the automation of the valuation and categorization 

of the heterogeneous goods, which is the backbone of the recommerce sector (Kumar 

et al., 2023). 



36 

Besides that, the use of cutting-edge automated technologies in supply chain activities 

sheds light on the level of informativeness required for contemporary inventory man-

agement that cannot be done without it. Digitized supply chains using real-time moni-

toring and data analytics give organizations the power to alleviate risks, efficiently man-

age stock, and be in a position to respond to market changes dynamically (Ali et al., 2024). 

The ability to reconcile inventory through digital channels becomes the chief determi-

nant of profitability in the case of recommerce and second-hand markets where supply 

is erratic and goods are one-of-a-kind (Kumar et al., 2023). However, most of the current 

studies neglect the detailed operational aspects of circular supply chains in SMEs and 

only concentrate on linear manufacturing or large-scale logistics instead (Jiang et al., 

2025).  

The result is a clear research gap that is defined by the crossing of these domains. In 

contrast to the existing literature that recognizes digital transformation as a must for 

SMEs (Moreira et al., 2024), the potential of digital technologies to promote the imple-

mentation of the circular economy (Piedra-Muñoz et al., 2025; Rusch et al., 2023), and 

the possibilities of AI to support process automation (Crisan et al., 2023), very few inte-

grated, low-cost multi-agent systems tailored for the high-variety, low-volume context of 

second-hand retail have been the subject of empirical research. Present researches ei-

ther dwell on the high-level frameworks of the circular economy while leaving the tech-

nical automation implementation part untouched or concentrate on technical automa-

tion (such as RPA) without considering the unstructured nature of the circular inventory. 

This study intends to close that gap by the design and assessment of a device that makes 

use of intelligent agents to carry out the digitization of the heterogeneous inventory, 

thus facilitating the theoretical advantages of the circular economy to resourced-con-

strained SMEs. 

 



37 

2.5 SME Retail Digitalization 

The digital transformation of SMEs has dramatically changed from being a strategic 

choice to becoming absolutely necessary for survival and differentiation in a competitive 

way in the global economy of today. At the beginning, the academic discussions were 

only about the adoption of basic Information and Communication Technologies (ICT), 

but now, the scholars focus on the use of cutting-edge Fourth Industrial Revolution (In-

dustry 4.0) technologies in particular AI to achieve both operational efficiency and value 

creation. By bibliometric analysis, Prasetyani et al. (2025) disclose that the topic of tech-

nology adoption in SMEs has been heavily researched since 2018, with a main focus on 

sustainability and the relationship between digitalization and business management. 

Nevertheless, in contrast to the theoretical consensus of digitalization benefits, the data 

on the ground reveals that SMEs (especially those in the retail sector) are quite different 

from large companies in that they have a distinct "digital divide" which is characterized 

by their unwillingness to use complex technology because they are short of resources 

and structurally tough (Narang & Tiwari, 2024). 

With the help of dynamic capability and resource-based frameworks, researchers have 

come to realize how the adoption of the technology is carried out or whether it is just a 

mere show of capability. Arroyabe et al. (2024) provide evidence that digital capabilities 

inside the company are what mainly lead to the introduction of AI in the European SMEs. 

They emphasize that the firm's ability to detect, take, and change digital resources is 

much more important than just getting support from the environment. Jalil et al. (2025) 

also echo that point by showing that the technology orientation is the mediator between 

AI adoption and digital value creation; in other words, just having the AI tools is not 

enough if there is no strategic orientation that makes these tools compatible with the 

firm’s overall digital value proposition. Besides that, Enshassi et al. (2025) implement the 

Technology-Organization-Environment (TOE) model in Malaysia to disclose that although 

being useful is what mainly pushes the adoption, the obstacles at organizational and 

environmental levels e.g., the lack of a data-driven culture and poor infrastructure, still 

hinder AI adoption in digital marketing considerably. 



38 

The pieces of literature have been consistent in pointing out a paradox where the use of 

AI has huge potential to change how SMEs operate, but on the other hand, the real im-

plementation is still very limited. Cooper, (2025) discovers that the use of AI in SMEs is 

very low even in areas with high potential such as New Product Development (NPD), and 

this unpreparedness is what has been most of the time the main cause for it. 'Readiness' 

here is defined as the presence of the business case, senior management support, and 

trust in AI outputs. Khan et al. (2025) through their study on Finnish SMEs, finds the 

same problem and thus they conclude that most of the companies they surveyed are at 

the very initial stage of AI maturity and have a hard time facing issues like lacking tech-

nical know-how and data management. In the same vein, Ulrich & Frank (2021) suggest 

German SMEs are more comfortable with the traditional, rule-based systems and hence 

less dependent on machine learning due to lack of employee skill and high implementa-

tion costs being their major obstacles. 

By going deep into the retail-specific literature, one can find that the problems are even 

worse in the unorganized or micro-retail sector. Narang & Tiwari (2024) uncover that the 

small-scale retailers in India hardly ever come across the word "Artificial Intelligence," 

and they consider modern technology merely as a tool for backend billing rather than a 

means for customer engagement or inventory optimization. The reason behind this un-

awareness is very important, as Ishengoma & John (2025) state it in the context of Tan-

zanian manufacturing, that factors like ease of use, compatibility, and observability are 

the necessary conditions for the adoption of mobile-based AI services. Because of the 

complexity of AI systems, the micro-enterprises who do not have a specialized IT depart-

ment like the big companies are often left out (U. A. Khan et al., 2025), therefore there 

arises a need for not only technologically advanced but also user-friendly and easily in-

tegrable solutions to the current workflows. 

As a result, the missing significant piece of research is the position of micro-retailers' 

operational realities intersecting with advanced AI capabilities, particularly in the circular 

economy. On one hand, there are comprehensive studies on challenges to AI adoption 

in manufacturing (A. N. Khan et al., 2024) and the factors that lead to digital marketing 



39 

adoption (Enshassi et al., 2025), but on the other hand, empirical research is very scarce 

on cheap and easy AI artifacts that are designed specifically for labor-intensive inventory 

digitization in the second-hand retail industry. Present-day literature puts a lot of em-

phasis on the strategic aspect of AI adoption (Cooper, 2025; Jalil et al., 2025) and over-

looks the technical side of multi-agent systems which can facilitate heterogeneous, sin-

gle-item inventory management to have lower entry barriers. This research aims to fill 

that void by creating and assessing a "Human-in-the-Loop" AI model that not only tackles 

the specific capability limitations pointed out by Arroyabe et al. (2024) and the usability 

requirements emphasized by Narang & Tiwari (2024) but also makes advanced inventory 

digitization accessible to the unorganized retail sector. 

 

2.6 Design Science Research and Theoretical Framework 

The study is centered around DSR, which is a difference in behavior science. Behavioral 

sciendesrtries to explain or predict phenomena, where as DSR is concerned with the 

creation and evaluation of new artifacts for a solved problem (Hevner et al., 2004). As it 

entails the design of a socio-technical solution, i.e. multi-agent AI system, that deals with 

the complex limitations of SME second-hand retail, this problem-solving paradigm is the 

most appropriate one for the context of this study. In order to maintain the standards of 

the research, the inquiry is in accordance with the DSR process model by Peffers et al. 

(2007) and the Framework for Evaluation in Design Science (FEDS) by Venable et al. 

(2016). These approaches define the study as a series of iterations of problem identifi-

cation, objective setting, design, demonstration, evaluation and communication. 

The study has come up with a theoretical framework to solve the problem of "scalability-

sustainability paradox" that has been second-hand retail by synthesizing the ideas from 

five interconnected thematic areas. Instead of considering these domains as separate 

entities, the theoretical framework combines them into a layered structure that not only 

directs the artifact's design but also its evaluation. Figure 4 presents the model, explain-

ing how the technical capabilities of a Multi-Agent AI can fulfill the operational needs of 



40 

Process Automation, that are then adjusted to the particular Organizational Constraints 

of SMEs to finally bring about Market Impact in the Circular Economy. 

 

Figure 4 Integrated Theoretical Framework linking technical enablers to market impact. Source: 

Adapted from Peffers et al. (2007). 

 

The initial stage, problem identification and motivation, delineates the inventory digiti-

zation challenges in second-hand retail SMEs and through operational inefficiencies and 

missed opportunities for data-driven decision-making, it puts forward their importance. 

Stage two, defining the objectives of a solution, converts these challenges into clear-cut 

functional and non-functional requirements for a multi-agent AI-based inventory digiti-

zation system that also comprises performance, usability, and deployment constraints. 

Artifact creation by combining domain knowledge, theoretical foundations, and tech-

nical expertise into a consistent multi-agent system architecture and its cloud-native, 

serverless implementation is the third phase, design and development. Stage four, 

demonstration, comprises the artifact instantiation and deployment in different settings 

(from controlled evaluations to authentic SME retail environments) in order to demon-

strate how the disclosed problems are being solved by it. The fifth stage, evaluation, is a 

systematic assessment of an artifact against set objectives, using appropriate methods 



41 

such as case-based evaluations, performance analyses, and user-centered assessments. 

Lastly, the sixth-stage communication, through which the artifact and the associated de-

sign knowledge are shared with academic and practitioner communities (Peffers et al., 

2007). 

The assessment of the artifact is aligned with the Framework for Evaluation in Design 

Science (FEDS) by Venable et al. (2016) which complements the conventional DSR in-

struction with detailed specification of evaluation strategies. According to FEDS, forma-

tive evaluation is the one that takes place during artifact construction and is aimed at 

guiding refinement, while summative evaluation is an assessment of whether the final 

artifact achieves its intended goals (Venable et al., 2016). Moreover, it differentiates ar-

tificial evaluation that is done in controlled, often laboratory-like settings using synthetic 

or constrained data from naturalistic evaluation conducted in real organizational con-

texts with real users, processes, and data (Venable et al., 2016). In line with FEDS, this 

research is performed using formative (artificial evaluation for rapid prototype iteration, 

formative) naturalistic evaluation for initial deployments in SME second-hand retail en-

vironments, and summative–naturalistic evaluation when the system is mature enough 

for assessment in authentic use contexts (Venable et al., 2016). 

Together, these different perspectives form a comprehensive theoretical framework that 

has been used as the basis and is the result of the DSR process in this thesis. The tech-

nical, process, organizational, research, and market perspectives together provide the 

direction for the design of the multi-agent AI system, the structuring of its workflows in 

SME second-hand retail, and the planning of its evaluation; besides, the artifact and its 

field-testing produce new design knowledge concerning AI-based, cloud-native, and 

serverless solutions for inventory digitization in resource-constrained organizational set-

tings (Hevner et al., 2004; Peffers et al., 2007; Venable et al., 2016).  



42 

2.7 Identified Research Gap 

The detailed analysis of the five theoretical domains uncovers a "Scalability-Sustainabil-

ity Paradox" which is not yet solved by contributions from universities and the real world. 

As shown in Figure 5, the literature depicts this conflict as having three components: 

(1) The Inventory Mismatch: On one hand, conventional inventory manage-
ment theories are based on standardization and deep stock levels (Chopra, 
2019). While on the other hand, the second-hand retail research portrays 
the industry as 'markets of one' with extreme heterogeneity (Turunen & 
Leipämaa-Leskinen, 2015). None of the existing digital instruments is ef-
ficiently compatible with the latter, although they are operationally suita-
ble for the former. 

(2) The Technological Barrier: Multi-Agent Systems (MAS) domain can pro-
vide the necessary cognitive flexibility to manage heterogeneous data 
(Wooldridge, 2012). Nevertheless, the SME Digitalization literature argues 
that due to their complexity and high cost, these advanced and innovative 
enterprises are out of reach for the very SMEs that need them most (Ifi-
nedo, 2011). 

(3) The Process Disconnect: Research on Process Automation (BPA/IPA) has 
effectively used AI in structured corporate workflows. However, the 
"physical-to-digital" transition which is a bridge for the resale industry, 
where the main source of labor is the extraction of structured data from 
physical objects, has not been sufficiently tackled (Barton et al., 2024). 

 

Thus, the narrowing down to the research question involves pinpointing the lack of a 

single architectural framework that leverages commoditized, multi-modal AI to auto-

mate the digitization of the heterogeneous inventory without the need for enterprise-

grade infrastructure. Scholars have optimized certain algorithms for image recognition 

or pricing in their studies but have not combined these into an accessible, end-to-end 

artifact for the non-technical SME retailer. This study is at that very juncture, going be-

yond the mere examination of adoption barriers to the actual building of a technical 

solution that makes those barriers irrelevant. 



43 

 

Figure 5 Research gap in AI-driven inventory digitization. 



44 

3 Method 

3.1 Design Science Research Methodology 

 The research outlined here has chosen Design Science Research (DSR) as its leading 

method for a theoretical framework. It follows closely the rules set up by Hevner et al. 

(2004), the six-stage process model of Peffers et al. (2007), and the Framework for Eval-

uation in Design Science (FEDS) by Venable et al. (2016). DSR is different from behavioral 

research in nature, the latter being aimed at explanation and prediction of organizational 

phenomena, while the main focus of DSR is on prescription designing and producing new 

artifacts through which goals of the organization are achieved. Such a problem-solving 

character of DSR makes it appropriate for practical challenges, the solutions of which, 

are not readily available. 

DSR approach is a good fit for the case of inventory digitization in the context of second-

hand retail where the problem domain does not consist of only technical challenges that 

can be solved by algorithms or only social phenomena that require interpretive inquiry. 

Rather, it deals with the design of integrated socio-technical systems where technology, 

human actors, processes, and organizational context are closely interconnected. The so-

cio-technical nature of the problem domain requires an approach that is not only tech-

nically rigorous but also human factors and organizational realities sensitive, and this is 

exactly what the DSR method offers. The DSR implementation has six stages. 

 

3.1.1 Stage 1: Problem Identification and Motivation 

Problem identification and motivation were at the center of attention in the first phase 

of the Design Science Research process. Following the DSR framework, the problem 

identification was based on a thorough review of existing scholarly literature. This served 

as the empirical base for spotting the research gap and designing the research problem. 



45 

The problem understood was as a result of an exhaustive systematic literature review, 

which covered five interconnected domains: inventory digitization in SME retail, artificial 

intelligence and multi-agent systems, second-hand retail operations, process automa-

tion, and SME digitalization. The review in Chapter 2 brought together the results of the 

recent research and pinpointed six critical research gaps: 

(1) the unavailability of multi-agent AI systems that could be used for handling the het-

erogeneity of second-hand retail inventory (Ren et al., 2017, 2017); (2) the lack of AI 

automation studies that tackle the resource constraints of SMEs (Ifinedo, 2011); (3) the 

deficiency in applied research on digital transformation in Finnish second-hand retail; (4) 

the scarcity of academic literature focusing on the end-to-end inventory digitization 

workflow (Liu et al., 2021); (5) the predominance of evaluation methods that are labor-

atory-based rather than those that consider real-world deployment; and (6) the insuffi-

cient usage of the Design Science Research method in addressing digitization issues in 

SME retail contexts (Hevner et al., 2004). 

Combining these six gaps led to the revelation of a compound theoretical and practical 

problem: the lack of an integrated, empirically-validated, and multi-agent AI-accessible 

method for SMEs, which would be able to automate the inventory digitization of heter-

ogeneous, non-standardized goods in operational environments with limited resources 

(Chopra, 2019; Sándor & Gubán, 2022). According to the literature, on the one hand, 

there is the technical capability for AI-driven automation to be done by separate units, 

multimodal language models, computer vision, agent-based architectures (Russell et al., 

2022; Wooldridge, 2012), however, their orchestrated application to the specific prob-

lem of second-hand retail inventory digitization remains unexplored. On the other hand, 

factors hindering the adoption by SMEs, such as inadequate technical infrastructure, lim-

ited financial budgets, and absence of in-house expertise, are among the reasons why 

the literature has been (Hernández et al., 2024; Ifinedo, 2011) but have not been sys-

tematically solved through the design of practical, cost-effective solutions (Chavez et al., 

2022). 



46 

By identifying the gaps presented, this literature-based problem formulation set the the-

oretical and empirical grounds for the research and showed that it was a complex prob-

lem that had not been solved before. The gaps identified were not only addressed by 

existing commercial solutions (that are still inaccessible to SMEs due to cost and com-

plexity) (D’Adamo et al., 2022; Greg Roughan, 2022) but also have not been sufficiently 

tackled by prior academic research (which has concentrated on large-scale implementa-

tions or isolated technical components rather than end-to-end, SME-contextualized so-

lutions) (He et al., 2016; Liu et al., 2021). The synthesis also indicated that although the 

second-hand retail sector manifests specific operational characteristics. These issues 

have not been systematically bridged to the adoption or creation of scalable, low-cost, 

and contextually-aware digital systems within the limitations of micro and small firms 

(Appelgren, 2019; Arroyabe et al., 2024).  

 

3.1.2 Stage 2: Objectives of a Solution 

The stage converted problem understanding into concrete and measurable solution ob-

jectives. Requirements elicitation have been informed by studies and market reports 

that reveal a structural problem in the retail of used goods. The second-hand goods mar-

ket in Finland has changed radically over the last several years, the main factors being 

the rapid development of online markets and consumer-to-consumer (C2C) platforms 

such as Tori.fi and Vinted. The value of Finland’s circular market, which includes second-

hand goods transactions, went up from about €895 million in 2023 to €1.4 billion in 2025, 

showing a 56% growth over two years (Finnish Commerce Federation, 2023; Saarinen, 

2025). Such growth is indicative of a fundamental shift in consumer behavior and market 

structure that has opened up both new opportunities and challenges for traditional busi-

nesses. 

Analysis of market structure suggests that most of the exchanges are done through dig-

ital platforms nowadays. The Finnish Commerce Federation states that online C2C trans-

actions constitute 61% of the circular market, while kirpputori (traditional second-hand 



47 

shops) make up only 20%, and brick-and-mortar and online stores combined have 19% 

(Linda Lisa Maria Turunen & Maike Gossen, 2024). Besides, traditional market segments 

have increased only by 6% in two years, whereas online B2C commerce has gone up by 

106% during the same period. This change shows that as digital channels become more, 

traditional businesses are very fastly losing their market share (Saarinen, 2025). 

The online platforms' competitive strengths are complex and multi-directional, making 

it difficult for traditional retailers to imitate them. Digital marketplaces allow consumers 

to have 24/7 access which is not limited by geography, a larger product selection than 

what physical stores can usually provide, and competitive pricing resulting from seller 

rivalry (Padmavathy et al., 2019). The platforms like Tori.fi and Vinted have used these 

advantages to get a significant share of the market, with Vinted getting most rapidly 

among the younger consumers in particular (Saarinen, 2025). The ease with which online 

commerce transactions can be done, coupled with the fact that there are lower over-

heads as compared to physical retail operations, results in a structural cost advantage 

which traditional retailers find it very difficult to match (Anay, 2024). 

Consumer behavior patterns sharpen the issues of traditional second-hand businesses 

even more. Demographic analysis shows that 81% of people under 30 in Finland pur-

chase used products, whereas only 66% of the total population does (Saarinen, 2025). 

This younger demographic demonstrates a pronounced preference towards online pur-

chasing which is mostly due to the fact that it is convenient, they are tech-savvy, and 

there is a general social acceptance of digital commerce (Karhunen, 2024). The embed-

ding of second-hand shopping into social media platforms and the impact of online com-

munities have made digital resale channels quite normal, especially for the generations 

that are digital natives. 

The disruption caused by e-commerce goes even further than the individual businesses 

to the whole retail economy. The move to online shopping has changed the pattern of 

jobs, with demand for traditional retail positions going down and growth in jobs like lo-

gistics, warehousing, and digital technology increasing (Bus Econ J, 2024). Retailers who 

operate physically now have to live with lower foot traffic and are experiencing shrinking 



48 

profit margins, thus, most of them are compelled to re-evaluate their business models 

and strategic positioning (Chava et al., 2024). The retail giants that previously had a phys-

ical marketplace as their fortress have been forced to implement omnichannel strategies, 

consequently, a seamless customer experience can now be delivered by integrating 

online and offline operations (Verhoef et al., 2007). 

Traditional second-hand retailers, despite being able to maintain a loyal customer base 

which values the physical shopping experience—characterized by tactile browsing, social 

interaction, and the “treasure hunt” aspect of in-store discovery—are facing huge prob-

lems when transforming digitally (Sasu-Boakye & Olsson, 2024). Many brick-and-mortar 

stores have attempted to go digital, but have failed because of their limited digital skills, 

shortages of resources, and the operational complexities of managing e-commerce 

alongside traditional retail (Sasu-Boakye & Olsson, 2024). The technology capabilities 

gap and the requirement for specialized skills in digital marketing, inventory manage-

ment systems, and online customer engagement are significant obstacles to effective 

digital adaptation. 

The platform economy has essentially changed the traditional retail structures by facili-

tating intermediary models that link buyers and sellers without the need for physical 

inventory or retail space (Parente et al., 2018). This disintermediation poses a major 

challenge to those traditional business models which heavily rely on physical infrastruc-

ture and customer access that is location-based. Studies indicate that physical retailers 

need to be digitally integrated and at the same time use their unique strengths like in-

store experiences, community engagement, and social missions to differentiate them-

selves from purely digital players (Sasu-Boakye & Olsson, 2024). 

Considering these market dynamics and competitive pressures, the digitalization of tra-

ditional second-hand businesses is more of a strategic imperative than a mere optional 

enhancement. This DSR study's goal is to create a tool that will enable traditional kirp-

putori and brick-and-mortar second-hand retailers to be competitive in a market that is 

becoming more and more digital while at the same time keeping the unique value prop-



49 

ositions that make them different from the pure online platforms. The digitalization pro-

ject is designed to help traditional retailers tackle the main challenges they face: loss of 

market share to online platforms, difficulty in attracting younger consumptions seg-

ments, limited resources hindering digital adoption, and the need to integrate physical 

and digital channels into a single coherent omnichannel strategy (Sasu-Boakye & Olsson, 

2024; Verhoef et al., 2007). Therefore, the study considers both quantitative and quali-

tative success criteria. They comprise, among others, at least a 75% reduction in pro-

cessing time, classification accuracy above 90%, a system usability scale (SUS) score of 

at least 70, and synchronization latency under 5 seconds. 

Functional requirements for this solution involved more than 25 functional capabilities 

such as multi-channel image capture, automatic background removal, hierarchical prod-

uct classification, structured metadata generation, label transcription, and real-time in-

ventory synchronization. As far as non-functional requirements are concerned, the qual-

ity attributes specified are: AI processing latency under 30 seconds, support for 500+ 

products per month, uptime greater than 99%, operability by non-technical staff, and full 

security. The research’s limits and assumptions consist of the following contextual con-

straints: monthly budget below €1,000, only non-technical staff available, MVP deploy-

ment within 1 day per store for 1,000 inventory items, and a heterogeneous, unstruc-

tured inventory. 

 

3.1.3 Stage 3: Design and Development 

The design and development phase through iteration cycles development, evaluation, 

and design processes is a kind of the design and development phase. Integrated multi-

agent AI system coordinating five specialized agents is the product scope to carry out 

inventory digitization. The implementation is a web-based software system that is on the 

production hosting infrastructure deployed. 

The choice of the frontend was judged against the twofold standards of technical appro-

priateness and accessibility of SMEs. The frontend framework React 19.1.1 was chosen 



50 

for its component-based architecture which allows fast user interface creation. Type-

Script 5.8.2 ensures type safety and lowers the number of runtime errors. The Vite 6.2.0 

build tool makes development cycles very short. Tailwind CSS offers utility-first styling, 

which makes it easier to have a consistent and responsive design. 

Google Gemini AI, the AI/ML platform, was chosen based on the following criteria: mul-

timodal features supporting both vision and language tasks; function calling allowing 

structured data generation with JSON schema validation; being economically priced at 

€0.07 per product; and having a quite good latency averaging 2–4 seconds. The backend 

infrastructure was set up on the Firebase platform, which was chosen and fully managed 

serverless infrastructure deployed. Firebase gives a NoSQL document database that sup-

ports flexible product schemas; Cloud Storage provides scalable image asset manage-

ment; Cloud Functions allow event-driven serverless computing; and Firebase Hosting 

offers global content delivery with automatic SSL. 

The product architecture is the implementation of both the high-level architecture and 

a distributed multi-agent system design with five specialized agents features. 

Agent 1: The Image Processing Agent