Artificial Intelligence Applications for Smart and Sustainable Mobility as a Service Concept: A Systematic Literature Review

Abstract

Over recent years, driven by intertwined economic, social, environmental, and technological factors, urbanization has accelerated at an unprecedented pace, posing complex challenges to metropolitan transport systems. This has intensified the demand for innovative mobility solutions, notably Mobility as a Service (MaaS), which promotes a paradigm shift from private vehicle ownership to mobility consumed as a service. With rapid advances in digital technologies, MaaS has gained substantial momentum, attracting significant scholarly attention for its potential to enable intelligent and sustainable transportation systems. This study aims to provide a comprehensive conceptual foundation of MaaS and its components, and to systematically examine how artificial intelligence (AI), machine learning (ML), and big data techniques are applied in this domain. Following PRISMA guidelines, a bibliometric and systematic review was conducted on peer-reviewed articles published between 2020 and 2024 and indexed in the Scopus and Web of Science databases. The analysis classifies AI applications across four MaaS integration levels: basic, intermediate, advanced, and full integration. The results show that machine learning and basic optimization dominate at the basic level; blockchain and big data are most prominent at the advanced and full levels; and deep learning is applied across all levels, with a particularly strong presence at the advanced stage for real-time, personalized mobility solutions. The findings also indicate that while most implementations focus on developed countries, there is substantial potential for adaptation in emerging markets. The paper concludes by discussing key challenges in regulatory compliance, inclusivity, and the protection of sensitive user data, and outlines future research avenues for building socially equitable, intelligent, and sustainable MaaS ecosystems.

1. Introduction

Improved mobility is seen as an engine of economic growth for countries. Public transport is designed for customers with diverse profiles and offers a variety of services to satisfy the widest possible audience. Nevertheless, this diversity has led to a variety of challenges, including concerns about ecology, comfort of transportation modes, and overuse of public transport, among others [1].

Global megalopolises typically boast a variety of transportation options, including buses, trams, metros, trains, buses with high levels of service (BHLS), taxis, and private transport. In navigating these complex networks, travelers often rely on their instincts and past experiences, particularly when confronted with unposted or unpredictable timetables. This seemingly ad hoc approach is frequently adopted due to the absence of comprehensive navigation tools, a challenge particularly prevalent in emerging cities of developing countries [2]. This approach of relying on instinctive choices can lead to several dysfunctions, such as extended waiting times at stops, overcrowding on certain transport lines despite the availability of alternatives, and the propensity for non-captive travelers and commuters to opt for private vehicles as soon as possible, which comes with high marginal social costs and generates negative externalities [3].

In the literature, a range of alternatives has been proposed and tested to mitigate the external costs associated with urban mobility. While not the sole or exclusive choice, Mobility as a Service (MaaS) is among the solutions proposed to face the aforementioned issues. MaaS aims to provide travelers with personalized sustainable transportation options offering easy access to information and the many available modes of transportation [4], while also offering efficient time management and optimizing the direct, indirect, and induced costs of travel [5].

With the emergence of this new concept, academic literature has rapidly developed and several research and review articles dealing with Mobility as a Service have been published. For example, Jittrapirom et al. [6] in their state of the art provided a comprehensive overview of the term, while Kamargianni et al. [7] focused on testing and implementation requirements, and Mulley [8] compiled a brief summary of key topics and challenges. However, only a limited number of studies have conducted systematic analyses of the literature. For instance, Utriainen and Pöllänen [9] analyzed 31 documents, focusing on the roles of various transport modes within MaaS, considering publications available up to 2018, whereas Maas et al. [10] extends the included literature into the year 2021, Durand et al. [11] examined 14 papers, focusing on users preferences and travel behavior; Wittstock and Teuteberg [12] analyzed 95 documents with the aim of identifying the fundamental components of MaaS. Kriswardhana and Esztergár-Kiss [13] provided a literature review that explores social factors influencing MaaS adoption and acceptance.

As the field of Mobility as a Service continues to evolve, researchers have increasingly turned their attention to the integration of Artificial Intelligence (AI) to further enhance its capabilities. By leveraging AI technologies, MaaS platforms can intelligently analyze vast amounts of data, predict user behavior, optimize route planning, and enhance overall system performance.

While these reviews have provided valuable insights, they share some limitations. First, most are narrow in scope, focusing on transport modes, user preferences, or social adoption, but not on the role of advanced technologies. Second, none has explicitly mapped Artificial Intelligence (AI) techniques to MaaS integration levels, leaving a gap in understanding how emerging computational methods shape the progression of MaaS from basic service coordination to full integration with societal goals.

The present paper addresses these gaps by offering the systematic literature review dedicated to AI applications in MaaS Systems, conducted under PRISMA guidelines, our study not only synthesizes the existing knowledge but also provides a framework that links AI techniques, such as machine learning, deep learning, reinforcement learning, and optimization, with specific MaaS applications, ranging from route planning and payment systems to sustainability and governance. These contributions offer new significance by clarifying how AI drives the evolution of MaaS and highlighting directions for future research.

In this context, the following research questions have been defined to guide our review study:

• RQ1: What is the relationship between MaaS level of integration, and the methods of AI used?
• RQ2: What specific AI algorithms and technologies are commonly employed in MaaS systems?
• RQ3: What practical applications do AI techniques have within the Mobility as a Service (MaaS) ecosystem?

To systematically address these research questions, we followed the review process illustrated in Figure 1. The process began by identifying relevant papers through the PRISMA procedure, followed by an initial screening to capture the main research hotspots in the MaaS domain. Each study was then classified according to its level of MaaS integration (basic, intermediate, advanced, or full), which allowed us to investigate the relationship between integration stages and the adoption of AI methods (RQ1). In the next step, we analyzed in depth the AI techniques employed, including machine learning, deep learning, optimization approaches, and other emerging computational strategies, thereby addressing RQ2. These techniques were subsequently mapped onto the key MaaS application domains—such as route planning, shared mobility, user preference modeling, and electric mobility—providing insights into the practical applications of AI in MaaS (RQ3). Finally, the outcomes of this analysis were synthesized in the discussion section, highlighting patterns, identifying trends and gaps, while future directions and recommendations are presented together in a dedicated section.

2. Search Method and a First Glance at Publications

2.1. Search Method

We conducted a literature review following the guidelines outlined in the PRISMA Statement [14], PRISMA is a widely recognized framework for ensuring clarity and comprehensiveness in reporting systematic reviews and meta-analyses. This included electronic search strategy, screening and selection of studies, and reliable methods for handling and combining the results (see Supporting Materials).

The study selection process (Figure 2) was conducted on 18 April 2024, and papers were collated from two electronic databases: Web of Science and Scopus. The search strategy (

Table 1. Database queries for both Scopus and Web of Science.

Database	Search Query	Filters Applied
Scopus	TITLE-ABS-KEY ((“Mobility as a Service” OR “MaaS” OR “Mobility Service”) AND (“machine learning” OR “neural network” OR “artificial intelligence” OR “big data” OR “deep learning”)) AND (LIMIT-TO (PUBYEAR, 2024) OR LIMIT-TO (PUBYEAR, 2023) OR LIMIT-TO (PUBYEAR, 2022) OR LIMIT-TO (PUBYEAR, 2021) OR LIMIT-TO (PUBYEAR, 2020)) AND (LIMIT-TO (DOCTYPE, “ar”) OR LIMIT-TO (DOCTYPE, “re”))	Years: 2020–2024; Document Type: Article, Review
Web of Science	TS = ((“Mobility as a Service” OR “MaaS” OR “Mobility Service”) AND (“machine learning” OR “neural network” OR “artificial intelligence” OR “big data” OR “deep learning”)) AND PY = (2020 OR 2021 OR 2022 OR 2023 OR 2024) AND (DT = “Article” OR DT = “Review”)	Years: 2020–2024; Document Type: Article, Review

) includes relevant keywords and controlled vocabulary terms related to Mobility as a Service and Artificial Intelligence techniques.

We included only articles published within the last five years (2020–2024), written in English, and addressing the application of one or more AI techniques within a MaaS context. The 5-year period was chosen to ensure that the review reflects the most recent developments in this rapidly evolving research domain, given that both MaaS and its integration with AI methods are relatively new and have gained significant attention in recent years. Upon the initial retrieval of records from both the Web of Science and Scopus databases, a total of 275 articles were identified. This comprised 179 records from the Web of Science search and 96 records from Scopus. To eliminate redundancy and ensure the inclusion of only unique articles, duplicates and workshop records were removed, resulting in a final pool of 227 unique articles for screening. During screening, 133 papers were excluded because their titles indicated they were not relevant to transportation. We were unable to retrieve 2 records, leaving us with 92 articles. After analyzing the abstracts, we excluded 43 more studies, leaving us with 49 selected articles for data extraction. Title and abstract screening were conducted as a single-reviewer process, following predefined inclusion and exclusion criteria to ensure consistency and reduce potential bias.

Table 2. Selected primary articles for review.

No.	Paper	No.	Paper
A1	Anthony Jnr et al. (2020) [15]	A26	Malik et al. (2022) [16]
A2	Cohen and Jones (2020) [17]	A27	Panichpapiboon and Khunsri (2022) [18]
A3	Longhi and Nanni (2020) [19]	A28	Sun and Song (2022) [20]
A4	Tempelmeier et al. (2020) [21]	A29	Tu et al. (2022) [22]
A5	Yoon and Chow (2020) [23]	A30	Vitetta (2022) [24]
A6	Alsaleh and Farooq (2021) [25]	A31	Yang et al. (2022) [26]
A7	Hüttel et al. (2021) [27]	A32	You et al. (2022) [28]
A8	Kong et al. (2021) [29]	A33	Cao et al. (2023) [30]
A9	Li et al. (2021) [31]	A34	Chu et al. (2023) [32]
A10	Luo et al. (2021) [33]	A35	Ding et al. (2023) [34]
A11	Miyake and Nishida (2021) [35]	A36	Esztergár-Kiss (2023) [36]
A12	Park and Hwang (2021) [37]	A37	Fabiani et al. (2023) [38]
A13	Phithakkitnukooon et al. (2021) [39]	A38	Jian et al. (2023) [40]
A14	R. Wang et al. (2021) [41]	A39	Liang et al. (2023) [42]
A15	X. Wang et al. (2021) [43]	A40	Martin et al. (2023) [44]
A16	Aman and Smith-Colin (2022) [45]	A41	Rajabi et al. (2023) [46]
A17	Boonjubut and Hasegawa (2022) [47]	A42	Servou et al. (2023) [48]
A18	Cheng et al. (2022) [49]	A43	Turno and Yatskiv (2023) [50]
A19	Choi et al. (2022) [51]	A44	Zhang et al. (2023) [52]
A20	Cokyasar et al. (2022) [53]	A45	Zou et al. (2023) [54]
A21	Cuomo et al. (2022) [55]	A46	Chu and Guo (2024) [56]
A22	Duan et al. (2022) [57]	A47	Ibrahim et al. (2024) [58]
A23	Lee and Kim (2022) [59]	A48	Jianing et al. (2024) [60]
A24	Lee et al. (2022) [61]	A49	Sun et al. (2024) [62]
A25	Li et al. (2022) [63]

presents the results in alphabetical order, organized by the corresponding author’s name.

2.2. First Glance at Publications

In this section, we provide an analysis and description of the publications using bibliometric methods. Bibliometric analysis employs quantitative techniques to systematically examine large datasets, identifying knowledge gaps and positioning researchers’ contributions within their field. This approach offers valuable insights into emerging research areas [64].

We utilized the package Bibliometrix of R 4.3 for this analysis [65]. Bibliometrix is an open-source toolset designed for quantitative research in scientometrics and bibliometrics. By following the methodology of Echchakoui [66], we merged data from Scopus and Web of Science databases to perform this comprehensive analysis.

Table 3. Main information about the study data.

Description	Results
Timespan	2020–2024
Documents	49
Average citation per doc	7.02
References	392
Document contents
Keywords Plus (ID)	452
Author’s Keywords (DE)	228
Authors
Authors	176
Authors of single-authored docs	2
Author collaborations
Single-authored docs	2
Co-Authors per Doc	3.92

summarizes the key details of the analysis, including: document type, document content, and author collaboration. The review covers the period from 2020 to 2024, resulting in a total of 49 documents, 38 sources, 176 authors, with an average of 7.02 citations per document, and a total of 392 references. It also shows that a total of 452 keywords were used across all the publications in this study sample.

To forecast the future of Artificial Intelligence applications in the Mobility as a Service concept, a co-word analysis was conducted (

Table 4. Most Frequently Used Words in the MaaS and AI Literature.

Words	Occurrence	Words in Titles	Occurrence	Words in Abstracts	Occurrence
deep learning	10	demand forecasting	4	shared mobility services	6
mobility service	10	deep reinforcement	3	passenger transport hub	5
mobility	6	mobility service	3	hybrid forecasting model	4
artificial intelligence	5	mobility services	3	artificial intelligence	3
big data	5	reinforcement learning	3	artificial neural network	3

). This analysis revealed 423 keywords within the study sample. The most frequently used keywords included “deep learning” (10 occurrences), “mobility service” (10 occurrences), “mobility” (6 occurrences), “artificial intelligence” “big data” “forecasting” and “machine learning” (each with 5 occurrences). Additionally, a distinct search through titles and abstracts identified further keywords such as “shared mobility services” (6 occurrences), “demand forecasting” (4 occurrences) and “deep reinforcement” (3 occurrences).

Using VOSviewer 1.6.20 to analyze the co-occurrence of indexed keywords, we generated the keyword co-occurrence map shown in Figure 3, which highlights the main research hotspots in the MaaS domain. Each node represents a keyword, where circle size is proportional to frequency, and distance between nodes reflects co-occurrence affinity. The color clusters indicate thematic groupings, enabling us to identify both mature and emerging streams of research.

The central positioning of “mobility service” and “mobility” underlines their role as the conceptual anchors of MaaS research. These keywords form the operational core of the map, around which technical, behavioral, and policy-oriented dimensions revolve. Their prominence reflects the fact that MaaS is studied not only as a digital innovation but also as a service-oriented paradigm designed to improve user experience and urban mobility efficiency.

A first major technical driver is “deep learning”, which emerges as one of the most recurrent terms, clustered with related concepts such as “machine learning”, “learning systems” and “reinforcement learning”. This cluster highlights the growing use of advanced AI techniques to solve complex MaaS challenges, including real-time demand forecasting, multimodal route optimization, autonomous decision-making and feedback-driven systems that allow MaaS platforms to self-optimize in dynamic, uncertain conditions.

A second key theme is “big data”, positioned at the intersection of multiple clusters. Its strong connections with “forecasting”, “optimization”, and “transport” confirm the reliance of MaaS applications on large-scale data collection and processing. This indicates that data availability and analytics are as critical as algorithms themselves, forming the foundation upon which predictive models and integration frameworks operate.

Another significant cluster revolves around “urban transport”, “traffic congestion”, and “route planning”, directly connecting AI techniques to real-world operational challenges. These linkages show how machine learning and optimization methods are being deployed to address pressing issues such as congestion reduction, efficient scheduling, and demand–supply balancing. The association with “behavioral research” suggests that MaaS is increasingly framed as a socio-technical system, requiring alignment with user preferences, accessibility concerns, and equity.

Peripheral clusters, including “electric vehicles”, “battery swapping”, and “sustainable mobility”, indicate the expansion of MaaS research toward environmental and climate-oriented goals. While these keywords appear less frequently, their linkages with AI clusters demonstrate a shift in focus: AI is not only enabling efficiency but also supporting decarbonization and sustainability agendas through electrification and energy-aware mobility systems.

Finally, the presence of “design”, “decision-making”, and “transportation policy” highlights that MaaS research extends beyond technical experimentation. These terms illustrate the embedding of AI-enabled MaaS into institutional, governance, and planning frameworks, where technological feasibility must align with regulation, equity, and long-term adoption strategies.

3. Classification of Artificial Intelligence Applications by MaaS Level of Integration

Mobility as a Service (MaaS) is an innovative concept that aims to induce significant changes in current transport practices. The European Alliance defines MaaS as an effective integration of various forms of transport services into a single, on-demand mobility chain to meet customers’ needs [67]. A classification of this concept was proposed by Sochor et al. [68] based on the levels of integration, which can help to compare the MaaS solutions. This classification can be summarized in the five distinct dimensions presented in Figure 4.

This section aims to provide a comprehensive and structured overview of the artificial intelligence (AI) techniques applied across different MaaS integration levels, offering a transparent classification that serves as the basis for understanding their role in MaaS development.

To achieve this, we applied a structured coding protocol to classify each article consistently. Specifically, papers focusing primarily on information provision (e.g., schedules, routing, shared mobility deployment) were assigned to Level 1; contributions describing booking strategies and pricing optimization were mapped to Level 2; studies discussing personalized mobility solutions and user preference integration were assigned to Level 3; and contributions explicitly linking MaaS with broader policy or sustainability objectives were classified under Level 4. Level 0 (no integration) was retained conceptually to represent the baseline but was not directly represented in the reviewed studies.

The next step in the analysis consists of exploring the key branches of artificial intelligence. Given the highly diverse and fragmented nature of this field, with branches often overlapping in terms of methods and applications, this analysis will rely on accessible taxonomies provided in the literature [69,70,71].

In fact, Artificial Intelligence encompasses multiple interconnected branches (Figure 5) that collectively define its scope and capabilities. In the realm of reasoning, AI systems employ knowledge representation, automated reasoning, and common-sense reasoning to process and understand information logically. The planning domain incorporates sophisticated approaches for scheduling, searching, and various optimization methods, including nature-inspired, market-based, and game theory-based techniques. Learning represents a fundamental aspect of AI, primarily manifested through machine learning, which encompasses supervised, unsupervised, and reinforcement learning approaches. Communication capabilities are demonstrated through natural language processing, enabling tasks such as text extraction, translation, and question answering. The perception domain includes both visual and auditory processing, with computer vision handling image recognition and machine vision tasks, while audio processing manages speech recognition and conversion between speech and text. Integration and interaction capabilities are exemplified through multi-agent systems, robotics, and automated vehicles. Expert systems form another crucial domain, implementing rule-based, fuzzy-based, frame-based, and hybrid approaches to problem-solving.

The rest of this section provides a classification of the selected articles based on the application of Artificial Intelligence techniques across different MaaS levels of integration.

3.1. Basic Integration Level

Articles at the basic level (

Table 5. Summary of AI Applications in MaaS at the Basic Level of Integration.

Paper Reference	Paper Aims	AI Techniques	Key Applications
Tempelmeier et al. [21]	To Study the spatial and temporal impact of large-scale events on urban road traffic and assist in route planning management.	Machine Learning: Support Vector Regression (SVR), k-Nearest Neighbors (KNN), and Ridge Regression	Route planning
Luo et al. [33]	To optimize the deployment of shared electric mobility (e-mobility) systems across urban areas.	Deep Learning: Graph Convolutional Networks (GCN) Multi-agent deep reinforcement learning (MARL)	Electric Mobility, Shared Mobility
R. Wang et al. [41]	To propose a two-phase matching model for integrating taxi-sharing and subways into a sustainable multimodal transit system	Optimization	Route planning, Multimodal Transit, Shared mobility
X. Wang at al. [43]	To develop a regression model with spatially varying coefficients to understand how factors like land-use, social demographics, and transportation infrastructure affect bike-sharing demand at individual stations.	Machine Learning: Spatially Varying Coefficients (SVC), Random Forests, Support Vector Machines (SVM)	Shared mobility
Choi et al. [51]	To analyze shared mobility services in South Korea and evaluate the social opportunity cost of the ban on these services	Machine Learning: Multinomial logit	Shared mobility
Lee et al. [59]	Enhancing first and last-mile transportation in urban areas, particularly in Seoul by optimizing the placement of shared PMs, such as bicycles and scooters,	Machine Learning: Random Forest (RF), Extreme Gradient Boosting (XGB)	Shared mobility
Panichpapiboon and Khunsri [18]	Analyzing mobility patterns and evaluating the efficiency of taxi services	Statistical models: lognormal, gamma, and Weibull distributions	On-Demand Services
Yang et al. [26]	Optimize the relocation strategies of competing carsharing companies.	Optimization: Bilevel optimization	Shared mobility
Liang et al. [42]	To develop a supply and demand matching model for transportation modes in a Mobility-as-a-Service (MaaS)	Deep learning	Route planning
Zhang et al. [52]	Optimizing route planning within ride-hailing services	Deep Learning: Graph Attention Network	On-Demand Services, Route planning
Zou et al. [54]	Explores whether ride-hailing can compensate for the lack of public transportation services (PTS) in suburban new towns in China	Machine Learning: Geographically Weighted Regression (GWR)	On-Demand Services

) of integration focus on applying AI methods to solve challenges related to shared mobility, route planning, or on-demand services problems. Tempelmeier et al. [21] investigated the spatial and temporal impacts of large-scale urban events on road traffic using machine learning models such as Support Vector Regression (SVR) and k-Nearest Neighbors (KNN). Their findings assist in improving route planning and managing traffic disruptions effectively. Luo et al. [33] focused on optimizing shared electric mobility deployments with multi-agent deep reinforcement learning and graph convolutional networks, demonstrating significant improvements in service coverage and profitability. In multimodal transit systems, R. Wang et al. [41] proposed a two-phase optimization model integrating taxi-sharing with subway systems, reducing taxi mileage and CO2 emissions while promoting sustainable shared mobility. Similarly, X. Wang et al. [43] developed regression models with spatially varying coefficients to identify factors affecting bike-sharing demand at specific stations, providing actionable insights for station-level optimization. Zhang et al. [52] introduced a divide-and-conquer framework enhanced with a Graph Attention Network (GAT) and an insertion transformer, significantly improving route planning by reducing mismatches between planned and actual routes, particularly in ride-hailing services.

Policy and operational challenges have also been addressed at this level. Choi et al. [51] analyzed shared mobility bans in South Korea, using Bayesian mixed multinomial logit models to evaluate social opportunity costs and proposing electrification incentives to mitigate unmet demand. Lee et al. [61] applied Random Forest (RF) and Extreme Gradient Boosting (XGB) to optimize personal mobility (PM) placement, enhancing first-mile/last-mile connectivity in urban areas like Seoul. Panichpapiboon and Khunsri [18] used statistical models to evaluate taxi service efficiency, offering insights for on-demand mobility improvements. Meanwhile, Yang et al. [26] used bilevel optimization to refine pricing and relocation strategies for carsharing services, addressing competition and operational sustainability. Zou et al. [54] investigated the role of ride-hailing in suburban new towns in China using Geographic Weighted Regression (GWR) and spatial autocorrelation. They demonstrated that ride-hailing effectively complements public transportation, particularly in high-tech areas and for younger demographics, enhancing transportation accessibility in underserved suburban areas. Finally, Liang et al. [42] utilized deep learning to develop supply-demand matching models for transportation modes within a Mobility-as-a-Service (MaaS) framework, optimizing resource allocation and route planning.

Despite these contributions, research at the basic level remains largely focused on operational efficiency and short-term optimization, with limited attention to societal concerns such as equity, inclusiveness, and governance. Furthermore, the majority of studies rely on isolated case analyses or simulation models, restricting the generalizability of their findings across different MaaS contexts. From an AI perspective, most contributions emphasize traditional or narrowly applied techniques without fully exploring advanced methods such as deep or reinforcement learning. This indicates a gap in leveraging AI’s full potential for systemic and long-term MaaS integration.

3.2. Intermediate Integration Level

At the intermediate level of integration, most studies are about demand prediction, booking strategies, and pricing optimization for shared mobility and on-demand services, while fewer papers still explore route planning problems (

Table 6. Summary of AI Applications in MaaS at the Intermediate Level of Integration.

Paper Reference	Paper Aims	AI Techniques	Key Applications
Yoon et al. [23]	To develop a revenue management strategy for bike-share systems	Machine Learning: Bootstrap, Multinomial Logit (MNL)	Shared mobility
Alsaleh and Farooq [25]	To develop trip production and distribution models for On-Demand Transit (ODT) services.	Machine Learning: Random Forests Ensemble learning: Baging Deep Learning with Explainability (SHAP)	On-Demand Services
Li et al. [63]	to predict short-term bicycle usage in bike-sharing systems.	Deep Learning: Spatial-Temporal Memory Network (STMN), Convolutional Long Short-Term Memory (Conv-LSTM),	Shared mobility
Park and Hwang [37]	To predict micro mobility demand and make decisions on bike relocations to improve service availability;	Deep Learning: Long Short-Term Memory (LSTM), Nature-inspired: Genetic Algorithm (GA)	Shared mobility
Phithakkitnukooon et al. [39]	Predicting Spatiotemporal Demand of Dockless E-Scooter Sharing Service	Deep Learning: Masked Fully Convolutional Network (MFCN)	Electric Mobility, Shared mobility
Boonjubut and Hasegawa [47]	To enhance bike distribution and availability in a bike-sharing network.	Time Series: ARIMA Deep Learning: Long Short-Term Memory (LSTM), and Recurrent Neural Networks (RNN)	Shared mobility
Cheng et al. [49]	To estimate the latent car-sharing demand at stations.	Time Series: Spatio-Temporal Autoregressive Moving Average (STARMA)) Machine Learning: Random Forests Ensemble learning: XGBoost	Shared mobility
Lee and Kim [59]	Enhance the operational efficiency of bike-sharing services by optimizing bike-sharing demand forecasts using data-driven approaches	Deep Learning: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), Convolutional LSTM (Conv-LSTM) and Graph Convolutional Network (GCN).	Shared mobility
Sun and Song [20]	Simulate individual traffic agents (vehicles) and optimize their timetables through a reservation system.	Simulation Agent-Based Models (ABM)	Route planning
Fabiani et al. [38]	handle competition among firms and reduce traffic congestion	Game Theory: Nash equilibrium seeking algorithm	On-Demand Services

). Yoon et al. [23] developed a revenue management strategy for bike-share systems using public data to optimize pricing plans for unlimited-ride passes, resulting in a 5.5% increase in revenue and improved consumer surplus. Alsaleh and Farooq [25] proposed trip production and distribution models for On-Demand Transit (ODT) services, the authors used Random Forest, ensemble learning, and explainability methods (SHAP) to identify key land-use and demographic factors that affect demand. Li et al. [63] utilized a Spatial-Temporal Memory Network (STMN) with Convolutional Long Short-Term Memory (Conv-LSTM) to predict short-term bicycle usage in bike-sharing systems and for forecasting demand patterns across both station-based and dockless systems. Park and Hwang [37] designed a hybrid model combining Long Short-Term Memory (LSTM) and Genetic Algorithms (GA) to predict micro-mobility demand and improve bike relocation strategies, enhancing first- and last-mile service availability. Phithakkitnukooon et al. [39] introduced a Masked Fully Convolutional Network (MFCN) for spatiotemporal demand prediction of dockless e-scooter sharing services. The results indicate that the proposed algorithms outperform other baseline models in accuracy for both short-term and next-day forecasting. Boonjubut and Hasegawa [47] addressed bike-sharing rebalancing challenges using time-series models like ARIMA and Long Short-Term Memory (LSTM). Booking optimization was tacked in consideration by ensuring that relocation efforts matched anticipated demand spikes. Cheng et al. [49] applied a Spatio-Temporal Autoregressive Moving Average (STARMA) model and machine learning techniques like Random Forests and XGBoost to estimate latent car-sharing demand at stations and to provide policy recommendations for optimizing vehicle deployment. Lee and Kim [59] evaluated deep learning models such as Convolutional LSTM and Graph Convolutional Network (GCN) for travel demand forecasting, improving operational efficiency in bike-sharing systems by addressing spatial and temporal correlations. Sun and Song [20] used agent-based simulation to optimize traffic flow through a Pareto-improving reservation system, which incorporated booking strategies to improve system reliability and reduce congestion. Fabiani et al. [38] proposed a semi-decentralized Nash equilibrium algorithm for managing competition among firms in ride-hailing services, the algorithm helps in introducing personalized incentives and booking coordination to reduce traffic congestion and to promote cooperative behavior.

In summary, research at the intermediate level still shows several limitations. First, the majority of demand forecasting and booking optimization models are designed for specific modes (e.g., bike-sharing or e-scooters), which restricts their generalizability to fully multimodal MaaS platforms. Second, while many studies achieve improvements in predictive accuracy using advanced AI techniques such as Conv-LSTMs or GCNs, they often lack interpretability, making it difficult for policymakers or operators to trust and apply these models in real-world decision-making. Third, most models optimize economic or operational efficiency (e.g., revenue, relocation costs, congestion reduction) but give limited consideration to broader sustainability, inclusiveness, or equity goals.

3.3. Advanced Integration Level

At this level of integration (

Table 7. Summary of AI Applications in MaaS at the Advanced Level of Integration.

Paper Reference	Paper Aims	AI Techniques	Key Applications
Anthony et al. [15]	To propose a multi-tier big data architecture to support the usage of Electric Mobility as a Service (eMaaS) within smart cities.	-Big Data tools: Spark and Hadoop	Electric Mobility
Longhi and Nanni [19]	To explore how data from car telematics can be used to develop advanced mobility services.	-Machine Learning: Random Forests -Ensemble learning: XGBoost -Deep Learning: Individual Mobility Networks (IMNs))	Car insurance, Electric mobility, Shared mobility
Huüttel et al. [27]	To model latent demand for shared mobility services, such as bike-sharing and electric vehicles (EVs).	-Deep Learning: LSTM, CQNN, Multi-CQNN	Electric Mobility, Shared mobility
Aman and Smith-Colin [45]	To explore the factors that influence user satisfaction with advanced MaaS platforms	Machine Learning: Latent Dirichlet Allocation (LDA), Ordinal Logistic Regression (OLR)	Users Experience
Cokyasar et al. [53]	to propose a data-driven approach to improve the traveler experience within integrated mobility services	-Big Data Analytics -Recommender System	User Experience
Duan et al. [57]	to develop and test models to predict MaaS use for different trip categories and predict the likelihood of users adopting MaaS for different trip types (social, general, and work trips)	Deep Learning:	User Preference (Experience)
Malik et al. [16]	Select safer routes for bicycle riders based on real-time traffic, environmental conditions, and rider-specific characteristics.	Deep Learning:	Shared mobility, User Preference
Tu et al. [22]	improve the efficiency of railway passenger transport hubs by offering dynamic dispatching and real-time scheduling of transport services, offering more personalized and efficient mobility solutions for passengers at railway hubs.	-Fuzzy logic -Deep Learning -Optimization: dynamic weight scheduling algorithm	User Preference (Experience), Multimodal Transit
Chu et al. [32]	balance both passenger satisfaction and transport operator profit, ensuring proportional fairness across multiple transportation providers.	Reinforcement Learning	User Preference, Multimodal Transit
Ding, et al. [34]	dynamic pricing and resource allocation mechanisms that optimize services based on real-time user demand, preferences aspects.	Optimization: Vickrey-Clarke-Groves model, column-generation algorithm	User Preference, Shared mobility
Martin et al. [44]	identify user groups with similar mobility behaviors.	Machine Learning: Clustering Algorithms	User Preference
Rajabi et al. [46]	provide personalized mobility services in a MaaS context	Knowledge-based system	User Preference
Ibrahim et al. [58]	focuses on integrating autonomous vehicle systems with real-time urban transport solutions to promote sustainability within smart cities	Fuzzy logic	Autonomous transportation

), we observe that research focuses on applications of real-time data and AI to provide highly personalized mobility solutions that can be dynamically adjusted based on user preferences and real-time conditions. Anthony et al. [15] proposed a multi-tier big data architecture using Spark and Hadoop to support Electric Mobility as a Service (eMaaS) in smart cities. The proposed approach helps in real-time interoperability and scalable deployment of electric vehicles. Longhi and Nanni [19] used car telematics data within machine learning, including deep learning techniques to provide adaptive mobility services. The results helped to provide custom car insurance plans, smarter ways to use electric vehicles, and shared mobility options designed around users’ driving habits and preferences. Hüttel et al. [27] extended these approaches by developing deep learning models, including LSTM and Multi-CQNN, to estimate latent demand for shared mobility systems like bike-sharing and EVs, enabling dynamic adjustments to resource allocation based on demand fluctuations.

Studies like Aman and Smith-Colin [45] have explored how advanced Mobility-as-a-Service (MaaS) platforms use AI to better understand and meet individual user needs. The authors identified key factors influencing user satisfaction with MaaS platforms, using machine learning techniques like Latent Dirichlet Allocation (LDA) and Ordinal Logistic Regression (OLR) to emphasize the importance of personalization and seamless integration in mobility solutions. Cokyasar et al. [53] developed a data-driven recommender system to dynamically enhance traveler experiences and to offer personalized travel recommendations that evolve with user preferences. Similarly, Duan et al. [57] applied Artificial Neural Networks (ANN) to predict the likelihood of MaaS adoption for specific trip types (e.g., social, work, or general trips). This helps in enabling tailored service offerings based on trip contexts. Malik et al. [16] incorporated ANN into cycling systems to provide safer routes for riders by combining real-time traffic and environmental data with user-specific characteristics, exemplifying how MaaS platforms can adjust services dynamically to enhance safety and satisfaction.

The role of real-time optimization in ensuring efficient operations and user satisfaction was studied in Tu et al. [22], the paper introduced a fuzzy neural network with a dynamic weight scheduling algorithm to improve the efficiency of passenger transport hubs, providing real-time scheduling and service adjustments tailored to commuter needs. Chu et al. [32] advanced this approach with deep reinforcement learning (DRL) and Markov Decision Processes (MDP) to balance user satisfaction and operator profitability. Ding et al. [34] proposed dynamic pricing and resource allocation strategies using the Vickrey-Clarke-Groves model and column-generation optimization algorithm, enabling shared mobility services to flexibly adapt to real-time user demand and preferences. Martin et al. [44] employed clustering algorithms to identify user groups with similar mobility behaviors, providing insights for designing subscription-based or on-demand services tailored to specific user segments. Rajabi et al. [46] developed a knowledge-based AI framework for MaaS platforms, integrating booking, payment, and contextual data to offer seamless, real-time personalized recommendations for users. Ibrahim et al. [58] applied a fuzzy decision-making model to integrate autonomous vehicles into urban transport systems, emphasizing sustainability and the ability to dynamically adjust services based on real-time urban mobility demands.

While these contributions demonstrate how advanced AI techniques—such as deep learning, reinforcement learning, and fuzzy optimization—enable dynamic personalization and real-time responsiveness, several shortcomings remain:

• Overemphasis on technical performance: Many studies validate models on accuracy or efficiency but pay limited attention to long-term user adoption or cross-platform interoperability.
• Data dependency: Most approaches rely on high-quality, large-scale datasets, yet few address challenges of data sparsity, bias, or privacy protection, which are critical for scalable MaaS deployment.
• Limited integration with governance: Although AI models demonstrate real-time adaptability, relatively little work connects these systems to policy, equity, or regulatory considerations, which are essential for sustainable adoption.
• Generalizability: AI models trained on specific cities or services often lack transferability to different contexts, raising concerns about broader applicability.

Thus, although advanced integration showcases the potential of AI to deliver highly responsive and personalized MaaS solutions, further research is needed to bridge technical innovation with policy frameworks, inclusivity, and data governance, ensuring that AI-driven MaaS systems are not only efficient but also socially sustainable.

3.4. Full Integration Level

At the full integration level, MaaS systems achieve seamless aggregation of multimodal transport modes, integrating advanced AI-driven frameworks, privacy-preserving mechanisms, and equitable transportation solutions to address the complexities of multimodal transport while ensuring accessibility, sustainability, and personalized services for users. For instance, Kong et al. [29] developed a blockchain-based monitoring system employing Proof-of-Stake consensus, Paillier cryptosystem, and Bloom filters to securely aggregate and share driver performance data among independent MaaS operators. This approach enhances trust and data integrity while maintaining user privacy, which is crucial for shared mobility services. Miyake and Nishida [35] integrated AI-driven solutions within buses and taxis to improve urban transportation efficiency and enhance user experience. The system utilized multivariate autoregressive models and ensemble learning (XGBoost) to optimize vehicle allocation and routing dynamically.

Chu and Guo [56] advanced MaaS integration by implementing federated deep reinforcement learning and Markov Decision Processes (MDP) to optimize multimodal transit systems. This framework balances passenger satisfaction and long-term platform profitability while safeguarding user data privacy. Jian et al. [40] proposed a Federated Personal Mobility Service (FPMS) architecture combining Convolutional Neural Networks (CNN) and Federated Learning (FL) to deliver seamless privacy-preserving travel recommendations. Jianing et al. [60] addressed equity challenges in MaaS by utilizing neural network models and Shapley Additive Explanations (SHAP) to predict mode usage across diverse demographics. Their study highlighted how pre-MaaS travel patterns influence post-MaaS behaviors and demonstrated the impact of equity constraints on system emissions, profitability, and overall efficiency (

Table 8. Summary of AI Applications in MaaS at the Full Level of Integration.

Paper Reference	Paper Aims	AI Techniques	Key Applications
Kong et al. [29]	To ensure secure data aggregation and verifiable sharing of driver performance records among mutually independent MaaS operators while protecting user privacy.	Blockchain and cryptographic techniques: Proof-of-Stake (PoS) consensus, Paillier cryptosystem, and Bloom filters	monitoring system, Shared mobility
Miyake and Nishida [35]	To integrate AI-driven mobility solutions, such as AI Taxi and AI Bus, to improve urban transportation efficiency and enhance user experience.	Time Series: Multivariate autoregressive models Deep learning Ensemble Learning: XGBoost	On-Demand Services
Chu and Guo [56]	describes a highly integrated system that optimizes passenger satisfaction and MaaS long-term profit while ensuring data privacy, preventing information leakage and inference in a centralized MaaS platform	-Federated Learning -Reinforcement Learning	User Preference, Multimodal Transit
Jian et al. [40]	provides seamless and personalized travel options using real-time data and advanced algorithms, combining multiple transit modes while protecting user privacy.	Deep Learning: Convolutional Neural Network (CNN) -Federated Learning	User Preference, data privacy, Autonomous transportation
Jianing et al. (2024) [60]	predict travel mode usage across various demographics and assesses the effects of equity constraints on system efficiency, emissions, and MaaS platform profitability. The study also investigates the influence of pre-MaaS travel patterns on post-MaaS behaviors.	Deep Learning: Neural Networks (NN) with Shapley Additive Explanations (SHAP)	User Preference, travel equity

).

While these studies demonstrate significant progress toward holistic MaaS ecosystems, several limitations remain. Full integration efforts often require high computational and infrastructural resources, which hinder scalability. Equity considerations, though increasingly recognized, are still insufficiently explored in large-scale deployments. Likewise, privacy-preserving AI models, despite their promise, raise unresolved questions about interoperability and efficiency when applied in heterogeneous real-world contexts. A promising direction for future research is to combine AI with advanced computing paradigms—such as edge computing and quantum computing—to enable scalable, real-time decision-making and adaptive services. Such approaches, together with a stronger focus on inclusivity, fairness, and trust, will be critical in advancing MaaS platforms toward sustainable, equitable, and fully integrated mobility solutions.

4. Discussion

In this section, we provide an analysis of the relationship between AI categories and their applications across various levels of integration in Mobility as a Service (MaaS). We examine how specific AI techniques are strategically aligned with practical use cases within the MaaS ecosystem and what their roles are in innovation and enhancing operational MaaS efficiency.

4.1. General Insights

Figure 6 shows the relationship between AI categories and the number of their applications across various levels of MaaS integration. The results indicate that machine learning and basic optimization techniques are mostly applied at the basic level of integration. At the advanced and fully integrated levels, blockchain, big data, and time series analysis are more prominent, which reflect the focus on full connectivity and data privacy objectives at these stages. For the deep learning techniques, we remark that they are utilized across various levels of integration, with a significant presence at the advanced level. This can be explained by their ability to leverage real-time data and to provide personalized mobility solutions that take into consideration user preferences and real-time conditions.

To examine how artificial intelligence (AI) methods are applied within the Mobility as a Service (MaaS) context, we used a Sankey diagram (Figure 7). The figure illustrates the links between MaaS integration levels, the AI techniques applied, and their main areas of use. The thickness of the flow shows how often certain methods appear in the reviewed literature, making it possible to see where different techniques are concentrated.

At the Basic and Intermediate levels, most connections cluster around optimization and machine learning. These approaches remain the backbone for solving operational problems such as route planning, multimodal scheduling, and the design of shared mobility systems [20,26,33]. This suggests that at early integration stages, MaaS efforts mainly target service efficiency and reliability.

By contrast, the Advanced level is more closely associated with deep learning, which supports complex tasks such as real-time personalization, demand forecasting, and autonomous transport systems [27,40]. This reflects a shift from classical optimization methods toward more data-driven models that can handle uncertainty and user diversity.

At the Full integration level, we observe the introduction of technologies such as blockchain and big data analytics. These are tied to broader societal goals, including privacy, transparency, and equity [29]. Their presence indicates that AI is not only used to optimize services but also to address governance and trust.

The diagram also shows that some methods cut across all levels. For example, deep learning appears from the Basic level (supporting route planning) up to the Full level (supporting equity analysis). This suggests it acts as a transversal tool, adapting to increasingly complex needs. By contrast, blockchain is more specialized, appearing only at the Full level where issues of security and fairness are critical.

Overall, the Sankey diagram reveals a progression: traditional techniques dominate in the early stages of MaaS, while more advanced and emerging methods become central at higher integration levels. This evolution shows both how MaaS research is advancing and how AI is being aligned with wider societal objectives.

4.2. Machine Learning Applications

The Parallel Coordinates Plot in Figure 8 visualizes the relationship between various machine learning techniques and their association with integration levels. Each line in the plot represents a specific integration level, while the vertical axes correspond to different machine learning techniques. The trajectories of these lines across the techniques provide insights into the prominence and distribution of these methods.

Figure 8 reveals that techniques such as Random Forests (RF) and Logistic Regression (LR) are used across various levels of MaaS integration, with RF showing a slight peak at intermediate levels. In contrast, Support Vector Machines (SVM) and K-Nearest Neighbors (KNN) are mostly applied at the basic level of integration. Ensemble Learning (EL) techniques are more prominent at intermediate levels of integration.

Logistic Regression (LR) has been used to study social opportunity costs and develop revenue management strategies for shared mobility systems [23,51]. Additionally, it has been employed in [45] to examine factors influencing user satisfaction within advanced MaaS systems. The use of Random Forests (RF) often aligns with Ensemble Learning (EL). For instance, Lee et al. [61] used RF and Extreme Gradient Boosting (XGB) to predict land-use variables and socio-economic factors that are important for determining optimal locations for personal shared mobility areas. Similarly, Alsaleh and Farooq [25] applied random forests and a bagging algorithm to model on-demand transit trip production and distribution levels. Their results indicated that RF excelled in predicting daily trip distribution, while the bagging model performed better in forecasting trip production. Longhi [19] used random forests on vehicle telematics data for crash risk prediction and XGB for traffic forecasting. Furthermore, Latent Dirichlet Allocation (LDA) was employed in spatio-temporal semantic analysis to uncover micro-mobility motivations [63], and Support Vector Machines (SVMs) were applied to determine factors influencing bike-sharing demand [43].

4.3. Deep Learning Applications

The application of deep learning techniques in MaaS-related studies shows significant advancements in addressing complex challenges across various levels of integration. Long Short-Term Memory (LSTM) models emerge as a dominant choice for handling time-series data. Studies like those by Boonjubut and Hasegawa [47] and Phithakkitnukooon et al. [39] effectively utilized LSTM to address challenges in bike distribution and spatiotemporal demand prediction. Zhang et al. [52] also employed LSTM to optimize route planning in ride-hailing services, demonstrating the versatility of these models for improving on-demand service efficiency. These studies underscore the suitability of LSTM for capturing temporal dependencies, making it a preferred choice for tasks requiring precise demand forecasting and resource management.

Convolutional Neural Networks (CNNs) also play a critical role, particularly in spatial data analysis. When combined with LSTM, as seen in [39] CNNs enhance the ability to process multi-modal data, addressing the intricate demands of spatiotemporal analysis. Graph Neural Networks (GNNs) stand out for their capability to model both spatial and temporal dynamics simultaneously. Li et al. [63] used a Spatio-Temporal GNN to predict short-term bicycle usage, showcasing their strength in managing interconnected systems within shared mobility.

Artificial Neural Networks (ANNs) have also demonstrated their significance in MaaS applications. Jian et al. [40] employed ANN to provide seamless and personalized travel optimization while ensuring data privacy within full-level MaaS integration. Similarly, Jianing et al. [60] utilized ANNs to predict travel mode usage across diverse demographics, focusing on promoting travel equity and user preference modeling. Malik et al. [16] extended these efforts by using deep learning to select safer routes for bicycle riders, prioritizing environmental and traffic data to enhance shared mobility and user preference in advanced MaaS systems.

Additionally, the study by Lee and Kim [59] takes a comprehensive and comparative approach by evaluating six deep learning models to enhance operational efficiency in bike-sharing systems. This work provides a benchmark for selecting the most effective models and demonstrates how varying architectures can address specific operational challenges, such as balancing bike availability across stations and optimizing fleet management. The findings offer a nuanced understanding of how deep learning can be tailored to maximize performance in MaaS systems, emphasizing its role in advancing shared mobility solutions.

4.4. Other Techniques

Diverse other techniques have been applied in MaaS-related studies to address a wide range of challenges. Among these, optimization techniques are widely utilized for enhancing resource allocation, system efficiency, and strategic planning. For instance, Wang et al. [43] introduced a two-phase matching model to optimize route planning for multimodal transit and shared mobility, demonstrating its effectiveness in improving operational performance. Similarly, Yang et al. [26] employed bilevel optimization to refine bike-sharing relocation strategies, providing a structured approach to address logistical complexities in shared mobility systems. Tu et al. [22] further contributed to this field by applying a dynamic weight scheduling algorithm, which significantly improved the efficiency of railway passenger transit operations.

Expanding beyond traditional optimization approaches, recent contributions have demonstrated the potential of reinforcement learning in MaaS systems. Chu et al. [32] applied deep reinforcement learning (DRL) to model passenger behavior in multimodal journey planning, formulating passenger experiences as a Markov decision process (MDP). Their approach dynamically adjusts utility weights to balance passenger satisfaction and operator profit, with experiments showing up to a 2.3× profit increase and higher retention rates. Building on this, Chu and Guo [56] proposed a privacy-preserving federated DRL framework (FDDPG) for MaaS, which addresses information leakage risks inherent in centralized training. By distributing training across client devices and employing secure aggregation, their model achieved comparable solution quality while enhancing privacy, improving MaaS profit by ~90% and passenger satisfaction by ~15%. In parallel, Luo et al. [33] explored multi-agent deep reinforcement learning for deployment optimization of shared e-mobility systems. Using a high-fidelity simulator and a hierarchical controller, their method outperformed heuristic approaches, improving both service coverage and net revenue in dynamic deployment scenarios

Complementing the strategic and operational focus of optimization techniques, time-series models provide crucial predictive capabilities for managing temporal dynamics in MaaS systems. These models are integral to demand forecasting and mobility management, particularly in systems where understanding and adapting to temporal patterns is essential. For example, Boonjubut and Hasegawa [47] utilized Spatio-Temporal Autoregressive (STAR) models to predict bike distribution, enabling more efficient resource allocation in shared mobility. Similarly, Cheng et al. [49] applied ARIMA models to estimate latent car-sharing demand, highlighting the value of time-series analysis in addressing fluctuating user needs. Miyake and Nishida [35] employed multivariate autoregressive models to integrate AI-driven mobility solutions, demonstrating how time-series models can work in tandem with optimization techniques to enhance operational efficiency.

Additionally, Big data analytics has emerged as a foundational tool for managing the complexity and scale of modern mobility networks. For instance, Anthony et al. [15] proposed a multi-tier big data architecture integrating tools like Spark and Hadoop to streamline operations in electric mobility systems, enabling efficient processing of large datasets. Cokyasar et al. [53] utilized data-driven techniques to enhance user experiences in shared mobility, demonstrating how advanced analytics can identify trends, optimize resource use, and improve overall system performance. While data-driven techniques enhance scalability, blockchain and cryptographic technologies ensure the security and integrity of these systems. Kong et al. [29] leveraged blockchain-based solutions for secure data aggregation and verifiability in MaaS monitoring systems. These technologies foster transparency and accountability, crucial for user trust and system reliability, especially in advanced and full-level MaaS integrations.

Finally, fuzzy logic and knowledge-based systems contribute to personalization and adaptive decision-making. Ding et al. [34] and Rajabi et al. [46] utilized fuzzy logic for dynamic pricing and autonomous vehicle integration, demonstrating its role in real-time decision-making. This paper also explored knowledge-based systems to deliver personalized mobility services, emphasizing the importance of user-centric design in MaaS solutions.

Taken together, the reviewed literature demonstrates a progressive shift in MaaS research from exploratory applications of classical statistical models toward increasingly sophisticated AI-driven solutions. Early approaches, such as Logistic Regression and basic machine learning, provided initial insights into socio-economic factors and adoption behaviors, but they were limited in handling the complexity of dynamic, multimodal systems. As the field advanced, ensemble methods and Random Forests significantly improved predictive accuracy in demand forecasting and trip distribution, marking a transition from descriptive modeling to more reliable decision-support tools. The adoption of deep learning models, particularly LSTM, CNN, and GNN, represents a breakthrough, enabling the capture of non-linear spatio-temporal dependencies, real-time traffic dynamics, and user preference evolution—capabilities essential for adaptive MaaS ecosystems. Beyond prediction, reinforcement learning, optimization, and big data analytics have propelled the field toward proactive and autonomous decision-making, where platforms not only forecast demand but also allocate resources, set dynamic pricing, and adjust operations in real time. Parallel to these computational advances, blockchain and federated learning address the growing concerns of data privacy and system trust, showing how MaaS research is broadening beyond technical optimization to include governance, equity, and ethical dimensions. Despite this progress, gaps persist: many deep learning models are still too computationally demanding to be easily deployed on a large urban scale, which limits their practical use. At the same time, interpretability often lags behind predictive accuracy, making it difficult for decision-makers to fully trust or explain model outputs. Moreover, although equity issues are increasingly mentioned, they are still much less studied compared to efficiency and performance goals. Taken together, the evolution of MaaS research shows a gradual shift from simple, single-mode analyses toward more integrated and intelligent platforms. The challenge ahead is to ensure that future AI-enabled MaaS systems not only achieve technical efficiency but also embed fairness, inclusivity, and long-term sustainability as core principles, with fairness, inclusivity, and long-term sustainability.

5. Future Directions and Recommendations

One significant limitation of this review is the underrepresentation of studies from developing regions, mainly due to the limited adoption of MaaS systems in these countries. Most existing studies originate from high-income countries where MaaS initiatives are already established [72]. However, recent contributions from emerging economies, such as research in Chile [3] and the Mobility Marketplace project in Morocco [2], suggest a gradual expansion of MaaS research beyond developed contexts.

A notable contribution comes from Rizzi and De La Maza [3], who examined Santiago’s large-scale public transport reforms, focusing on the transition from a fragmented bus system toward an integrated network. Their analysis highlighted the role of AI-powered data integration and simulation models in evaluating fare unification, travel demand, and service reliability. In particular, machine learning techniques were employed to analyze smart-card transactions and passenger flows, which informed the design of integrated ticketing systems and supported institutional coordination. More recently, Aritenang [73] provided evidence from Indonesia, where informal and low-capacity modes such as motorcycle taxis are being integrated into MaaS platforms. Models such as neural networks for demand forecasting and heuristic-based routing for fleet allocation are tested to improve first- and last-mile connectivity in dense urban areas.

Together, these case studies illustrate both the opportunities and unique challenges of deploying MaaS in developing contexts, where AI must account for informal mobility, affordability issues, and infrastructural constraints. This highlights, in our view, an important direction for future research: leveraging Artificial Intelligence (AI) to adapt MaaS solutions to the specific socio-economic and infrastructural conditions of developing regions, thereby achieving higher levels of integration while supporting socially sustainable goals. Unlike in developed countries, where MaaS often focuses on reducing private car use, in many African and Southeast Asian cities, “private vehicles” include motorcycles, which dominate the urban mobility landscape [73,74]. Incorporating such informal and sensor-limited modes into MaaS ecosystems remains a major challenge for AI models. Beyond mode integration, AI must also tackle critical issues such as congestion, environmental degradation, safety risks, and noise pollution that are prevalent in these regions [75].

Equally important, the integration of AI into MaaS is shaped by regulatory frameworks, particularly in data protection and user consent. Variations in these policies across countries constrain scalability and cross-border interoperability, raising barriers to wide adoption. Addressing data privacy concerns and designing universally adaptable yet locally compliant AI models will therefore be essential for building trustworthy MaaS ecosystems [76]. Another difficulty lies in the inherent complexity and unpredictability of urban mobility, influenced by fluctuating demand, evolving infrastructure, and shifting policy priorities. This calls for continuous refinement of AI methods to remain adaptive and reliable in real-world contexts. Emerging paradigms such as Explainable AI (XAI) offer promising pathways, ensuring transparency, interpretability, and stakeholder trust while maintaining predictive accuracy [77,78].

From a technological standpoint, coupling AI with advanced computing paradigms such as edge computing and quantum computing [79,80] provides a compelling research direction. These technologies promise scalable, low-latency, and real-time decision-making capabilities that are critical for the responsiveness of dynamic MaaS ecosystems. Yet, they also present challenges in terms of cost, infrastructural requirements, and the need for cross-disciplinary expertise.

Looking forward, we believe that the trajectory of MaaS research should center on designing AI-driven platforms that are efficient, scalable, equitable, transparent, and context-sensitive. The next generation of MaaS systems must not only optimize technical performance but also embed fairness and inclusivity as core principles. Key challenges to achieving these goals include managing the computational complexity of advanced AI models while ensuring their deployability in real-world, city-wide systems; reconciling regulatory heterogeneity across countries with the need for universally accessible and interoperable services; and ensuring that social sustainability and inclusivity are not overshadowed by efficiency-driven objectives. In our view, the future direction of MaaS research lies in balancing these trade-offs through explainable AI, privacy-preserving learning frameworks, and integration with advanced computing paradigms such as edge and quantum computing. Ultimately, the next stage of MaaS development must deliver solutions that are both technologically advanced and socially resilient, shaping adaptive mobility ecosystems capable of addressing the diverse realities and evolving needs of cities worldwide

6. Conclusions

Mobility as a Service (MaaS) is a pioneering concept designed to transform existing transportation practices by integrating multiple public and private transport services into a single, intuitive, and on-demand platform.

The systematic review conducted in this study confirms that the integration of artificial intelligence (AI) into Mobility as a Service (MaaS) has grown considerably in recent years, reflecting advances in digital technologies and the increasing demand for intelligent and sustainable mobility solutions. The review addressed three main objectives: providing a conceptual framework for MaaS and its integration levels, classifying the AI techniques applied within each integration stage, and identifying geographic and thematic gaps in current research.

The classification of AI applications reveals a clear distribution pattern across integration levels. At the basic stage, machine learning and basic optimization techniques are predominant, enabling core functions such as demand prediction and route planning. In contrast, advanced and fully integrated MaaS platforms increasingly incorporate blockchain, big data analytics, and time series analysis, reflecting their role in achieving interoperability, connectivity, and improved data protection. Deep learning techniques are present across all levels but are most prominent at the advanced stage, where their capacity to process large-scale, real-time data supports personalized mobility services that adapt dynamically to user needs and contextual conditions.

Geographically, the analysis shows a strong concentration of research in developed economies, primarily in Europe and North America, while studies from emerging markets remain scarce. This imbalance highlights the need for research tailored to the specific mobility patterns, infrastructure limitations, and informal transport modes prevalent in developing regions. Addressing these gaps could significantly increase MaaS adaptability and relevance in low-resource contexts.

Persistent challenges also emerged from the review, notably in regulatory compliance, inclusivity, and the protection of sensitive user data. Collaboration networks are still limited, with relatively few cross-country research initiatives, which constrain the exchange of best practices and slow the pace of standardization. Emerging technologies—such as explainable AI, edge computing, and quantum computing—offer promising opportunities to address these issues by enhancing transparency, scalability, and computational efficiency.

Based on these findings, the review proposes several key recommendations and priority directions for future research and practice. First, AI applications should be prioritized that enable real-time personalization, adaptive service delivery, and user-centric design to enhance responsiveness and overall service quality. Second, governance and regulatory frameworks must be developed to balance technological innovation with stringent data privacy, security, and ethical safeguards. Third, research efforts should be broadened to address emerging markets by integrating informal and low-technology mobility modes, thereby ensuring the global applicability of MaaS solutions. Fourth, stronger international collaboration networks should be fostered to promote knowledge sharing, standardization, and cross-border interoperability. Finally, the integration of next-generation technologies—such as explainable AI, edge computing, and quantum computing—should be explored to improve transparency, computational efficiency, and scalability within MaaS ecosystems.

Overall, this review contributes to advancing the understanding of how AI is applied within MaaS, clarifies its distribution across different integration levels, and highlights critical research gaps. These insights can inform policymakers, industry stakeholders, and researchers in steering the development of socially equitable, intelligent, and sustainable MaaS systems.