Artificial Intelligence In Real-Time Payment Systems: A Bibliometric And Thematic Analysis Of Emerging Research Trends (2004–2026)

Fast or real-time payment systems (RTPS) have become a central component of digital financial infrastructure because they allow funds to be transferred and made available to recipients in real time or near real time, typically on a near-continuous basis. Their economic significance extends beyond speed. Fast-payment rails can lower transactional frictions, broaden participation in digital finance, and provide a foundation on which banks, fintech firms, merchants, governments, and households build new services (Frost et al., 2024; Cornelli et al., 2024). Yet instantaneity changes the risk architecture of payments. In slower systems, suspicious activity may be reviewed before final settlement or recovered through operational delay. In real-time systems, authorization, fraud screening, compliance checks, and customer authentication must occur within a narrow decision window. A mistaken approval can become an irrevocable loss within seconds, while an overly conservative model can block legitimate transactions and undermine trust in the payment rail.

Artificial intelligence (AI) and machine learning (ML) offer a set of computational capabilities that are well matched to this environment: non-linear classification, anomaly detection, sequence learning, graph analytics, behavioral modeling, natural-language processing, reinforcement learning, and, increasingly, generative and agentic AI. Earlier payment-fraud research demonstrated that real-world usefulness depends not only on predictive accuracy but also on latency, class imbalance, evolving fraud vectors, data quality, and the business cost of false positives (Ryman-Tubb et al., 2018). Recent central-bank research extends the problem beyond card fraud. Layered ML architectures can identify unusual high-value payment activity in real time (Desai et al., 2024), network-wide analytics can reveal coordinated financial-crime patterns that are invisible to individual institutions (Bank for International Settlements [BIS] Innovation Hub, 2025), and generative AI agents are beginning to be tested for intraday liquidity and payment-prioritization decisions (Aldasoro & Desai, 2025). These developments indicate that AI is shifting from a peripheral fraud-detection tool to a broader operational and governance layer within payment infrastructures.

The scholarly literature, however, remains fragmented across computer science, cybersecurity, fintech, blockchain, banking, fraud analytics, Internet of Things (IoT), and payment-system research. Broad bibliometric studies of AI and ML in finance identify fraud, distress, forecasting, risk, and fintech as major clusters (Goodell et al., 2021; Pattnaik et al., 2024), while bibliometric work on explainable AI emphasizes interpretability, trust, risk assessment, and regulatory relevance (Chen et al., 2023). Those reviews are valuable but do not isolate the specific combination of AI and real-time payment environments, where low-latency decisions, rapid finality, network effects, interoperability, fraud migration, and operational resilience interact. Conversely, engineering studies often examine a single model, dataset, or payment problem without locating that contribution within the larger evolution of the field.

A bibliometric and thematic approach is useful because the field is expanding quickly and is distributed across disciplines. Bibliometrics can quantify scientific production and impact, while science mapping can expose conceptual relationships that are difficult to see through narrative reading alone (Donthu et al., 2021; Zupic & ÄŒater, 2015). Co-word analysis is especially appropriate for an emerging interdisciplinary domain because repeated keyword combinations reveal how researchers connect technologies, risks, and applications over time (Callon et al., 1991; Cobo et al., 2011). The present study therefore uses the Biblioshiny report generated from a Scopus-derived corpus to examine how research on AI in real-time payments has developed, which actors and sources structure the field, which topics have matured, and which research fronts are now emerging.

The study addresses four research questions. RQ1 asks how scientific production and citation impact have evolved in AI–real-time payment research. RQ2 asks which sources, authors, institutions, countries, and collaborations contribute most strongly to the field. RQ3 asks which conceptual themes organize the knowledge structure and how those themes have changed over time. RQ4 asks which emerging themes should shape the next research agenda for intelligent payment systems. By answering these questions, the article contributes a focused map of a rapidly forming research domain, distinguishes mature motor themes from new fronts, and develops a forward-looking framework that connects technical model development with payment-system requirements for speed, security, explainability, privacy, and resilience.

2. Conceptual Background: AI as an Intelligence Layer in Real-Time Payments

The defining characteristic of an RTPS is not merely faster messaging but the compression of the complete decision cycle around a transfer. Fast-payment systems generally aim to provide immediate or near-immediate availability of funds and broad operating hours, with design choices varying across jurisdictions (Committee on Payments and Market Infrastructures [CPMI], 2016; Frost et al., 2024). This architecture increases the value of automation because large transaction volumes must be evaluated continuously. It also heightens the cost of delayed intelligence: the interval between observing a suspicious event and taking preventive action may be shorter than the time required for traditional manual review.

Three analytically distinct layers of AI application can be identified. The first is transaction intelligence. Here, models score individual transactions using historical behavior, device or channel attributes, transaction sequences, and contextual features. Supervised classifiers remain useful when reliable labels exist, but fraud detection often suffers from severe class imbalance, delayed labels, and changing attack patterns. Anomaly-detection and hybrid methods are therefore important for identifying previously unseen behaviors. Desai et al. (2024), for example, demonstrate a layered approach in which supervised learning first separates typical from unusual high-value payments and unsupervised learning then ranks anomalies among the unusual subset. The architecture illustrates a core RTPS design principle: computational resources and human attention should be concentrated on the smallest set of transactions that warrant intervention.

The second layer is network intelligence. Fraud and money laundering rarely operate as isolated transactions; they involve relationships among accounts, merchants, devices, mule networks, counterparties, and institutions. Graph neural networks (GNNs) can learn from these relational structures rather than treating each transaction as independent. A systematic review by Motie and Raahemi (2024) shows that GNN-based financial-fraud research is expanding because graph representations can capture patterns that conventional tabular models miss, while important gaps remain in unsupervised, dynamic, and multi-level anomaly detection. Project Hertha provides a payments-specific illustration: network-wide analytics applied to a large synthetic retail-payment ecosystem improved the detection of illicit accounts, with particularly strong gains for novel patterns (BIS Innovation Hub, 2025). Such evidence helps explain why GNN-related terms appear as an emerging frontier in the present bibliometric corpus.

The third layer is system intelligence. AI can support decisions about routing, liquidity, settlement priorities, operational anomalies, and capacity rather than only policing end-user transactions. Aldasoro and Desai (2025) experimentally evaluate a generative AI agent for intraday liquidity management in a wholesale payment setting and find that it can reproduce several prudential cash-management behaviors under simulated constraints. Reinforcement learning has also been applied to optimization problems involving blockchain and distributed infrastructures, including the highly cited work by Liu et al. (2019). In this layer, the AI system is no longer only a classifier; it becomes a decision-support or decision-making component inside a complex financial market infrastructure. That transition magnifies governance requirements because model errors can affect liquidity, settlement queues, or systemic resilience.

Across all three layers, four constraints recur. First, latency matters: a model that is accurate but too slow may be unusable in a real-time authorization path. Second, fraud is non-stationary: adversaries adapt to controls, producing concept drift and novel behaviors. Third, models operate under asymmetric error costs. False negatives create losses and compliance exposure; false positives impose friction on legitimate users and can suppress adoption. Fourth, high-performing black-box models may conflict with requirements for explainability, auditability, contestability, and supervisory review. Research on explainable AI in finance accordingly identifies transparency and risk assessment as persistent themes (Chen et al., 2023). These constraints imply that the most valuable AI research for RTPS will not be defined by accuracy alone but by the joint optimization of detection quality, response time, interpretability, privacy, resilience, and operational cost.

3. Methodology

This study adopts bibliometric performance analysis and science mapping to examine the evolution and conceptual structure of research on AI in real-time payment systems. Bibliometric analysis is appropriate for large, heterogeneous literatures because it provides reproducible quantitative measures of productivity, influence, collaboration, and conceptual relatedness (Donthu et al., 2021; Zupic & ÄŒater, 2015). The analysis was conducted from the supplied Biblioshiny workbook, which reports a Scopus-derived corpus and the associated Bibliometrix/Biblioshiny outputs. Bibliometrix is an R-based science-mapping environment designed for comprehensive bibliometric workflows, while Biblioshiny provides an interactive interface for comparable analyses and visualizations (Aria & Cuccurullo, 2017; Aria et al., 2026).

The verified corpus contains 545 documents published between 2004 and 2026 across 402 sources. The workbook records 55,452 references, 1,825 authors, 1,727 author keywords, and 3,136 Keywords Plus terms. The average document age is 1.95 years and the mean citation count is 9.415 per document. International co-authorship is 25.32%, and the average number of co-authors per document is 3.78. Conference papers are the largest document type (296; 54.3%), followed by articles (158; 29.0%), book chapters (51; 9.4%), reviews (23; 4.2%), books (12; 2.2%), conference reviews (3; 0.6%), one data paper, and one letter. The composition is consistent with a technically oriented field in which conference dissemination often precedes journal consolidation.

The underlying search logic targeted the intersection of two concept families: AI-related terms (for example, artificial intelligence, machine learning, deep learning, neural networks, intelligent systems, AI-driven approaches, and predictive analytics) and fast-payment terms (for example, real-time payment, instant payment, faster payment, real-time transaction, and instant payment system). The supplied workbook does not preserve the literal Scopus search-history string, extraction timestamp, or a complete screening log. To avoid creating a false audit trail, this article does not present a reconstructed query as though it were the verbatim executed syntax. For submission, the exact Scopus search-history export should be appended to the methods or supplementary material if available. This limitation does not affect the internal accuracy of the descriptive and science-mapping results reported from the verified corpus, but it does constrain full search-level reproducibility.

The analysis proceeds in four stages. First, performance analysis describes annual scientific production, annual citation patterns, leading sources, leading authors, affiliations, country-affiliation frequency, internationally collaborative links, and globally cited documents. Annual citations are interpreted with attention to citation-window effects: recent publications have had less time to accumulate citations and should not be compared mechanically with older work. Similarly, the 2026 publication count is treated as a partial-year observation.

Second, trend-topic analysis is used to identify the temporal movement of recurrent terms. Biblioshiny reports each term’s frequency and the first quartile, median, and third quartile of publication years in which it appears. These distributions make it possible to distinguish earlier infrastructure terms from more recent analytic fronts. Third, keyword co-occurrence analysis maps conceptual proximity. Network prominence is examined through betweenness, closeness, and PageRank measures. The highest PageRank values in the supplied co-word network belong to learning systems, crime, fraud detection, machine learning, and machine-learning, indicating that learning-based fraud analytics occupies the network core.

Fourth, a strategic thematic map is interpreted using Callon centrality and density (Callon et al., 1991; Cobo et al., 2011). Centrality captures the degree to which a thematic cluster connects with other clusters and therefore its importance to the overall field; density represents internal development and cohesion. In the supplied Biblioshiny configuration, 250 merged keywords were analyzed with a minimum frequency of two and Louvain clustering. Four principal clusters emerge: learning systems; artificial intelligence; graph neural networks; and behavioral research. Because lower-left themes can represent either emerging or declining topics, quadrant position alone is insufficient for temporal interpretation. This study therefore triangulates thematic-map location with trend-topic medians and quartiles. Graph neural networks and behavioral research have medians concentrated in 2026 and are thus interpreted as emerging fronts rather than legacy themes in decline.

Source: Authors’ analysis of the supplied Scopus-derived Biblioshiny report.

Figure 1. Annual scientific production in AI and real-time payment research

Rank	Source	Documents
1	Lecture Notes in Networks and Systems	21
2	Lecture Notes in Computer Science family	16
3	IEEE Access	10
4	IEEE Internet of Things Journal	7
5	Procedia Computer Science	7
6	Communications in Computer and Information Science	6
7	Lecture Notes in Computer Science	5
8	ACM International Conference Proceeding Series	4
9	Journal of Risk and Financial Management	4
10	ICSES 2024 proceedings	4

Source: Authors’ analysis of the supplied Scopus-derived Biblioshiny report.

Table 2. Most productive publication sources

Source: Authors’ analysis of the supplied Scopus-derived Biblioshiny report.

Figure 2. Country scientific-production frequency

Source: Authors’ analysis of the supplied Scopus-derived Biblioshiny report.

Figure 3. Trend topics and their temporal concentration

Source: Authors’ analysis of the supplied Scopus-derived Biblioshiny report.

Figure 4. Keyword co-occurrence network

Cluster	Centrality	Density	Freq.	Representative terms	Interpretation
Learning systems	11.06	28.72	2,156	learning systems; crime; fraud detection; machine learning; deep learning; anomaly detection	Mature motor theme
Artificial intelligence	9.60	27.17	1,426	AI; blockchain; network security; data privacy; cybersecurity; scalability; federated learning	Mature motor theme
Behavioral research	6.99	25.51	589	behavioral research; banking; big data; decision making; fintech; risk management	Emerging (recent temporal profile)
Graph neural networks	3.51	22.04	280	GNN; reinforcement learning; graph neural network; benchmarking; phishing; heterogeneous graph	Emerging frontier

Source: Authors’ analysis of the supplied Scopus-derived Biblioshiny report.

Table 4. Strategic thematic clusters and interpretation

Source: Authors’ analysis of the supplied Scopus-derived Biblioshiny report.

Figure 5. Strategic thematic map of AI in real-time payment research

7. Research, Managerial, and Policy Implications

For researchers, the main implication is methodological. Static benchmark datasets and random train–test splits are increasingly misaligned with the problem structure. RTPS studies should use chronological validation, delayed-label simulation, concept-drift analysis, cost-sensitive evaluation, and, where relevant, temporal or heterogeneous graphs. Claims of “real-time” capability should be supported by measured latency and throughput under realistic load rather than inferred from model type. Synthetic data can be valuable when real payment data cannot be shared, but realism, privacy risk, and distributional fidelity should be documented.

For payment service providers and system operators, the findings favor layered controls rather than a single monolithic model. High-speed rules and lightweight models can perform first-pass screening; specialized anomaly or graph models can examine higher-risk events; human investigators can focus on the smallest, most consequential set of alerts. System-level analytics may complement institution-level models by exposing cross-provider relationships. However, governance must specify who can act on system-wide signals, how false positives are handled, and how customer rights and confidentiality are protected.

For regulators and central banks, the emerging themes suggest that AI governance in payments will increasingly intersect with operational resilience, model risk, data protection, and financial-crime supervision. The relevant question is not whether AI is used, but where it sits in the decision chain and what happens when it fails. Supervisory frameworks should distinguish advisory models from models that directly influence authorization, settlement, liquidity, or customer access. Documentation of data lineage, performance drift, human override, incident response, and explainability will become increasingly important as AI moves closer to the payment core.

8. Limitations and Reproducibility Note

This study has four limitations. First, it relies on a Scopus-derived corpus; records unique to Web of Science, IEEE Xplore, ACM Digital Library, SSRN, working-paper repositories, or industry sources may be underrepresented. Second, the supplied Biblioshiny report does not retain the verbatim Scopus query, extraction timestamp, or full screening log. The analysis therefore reports only verified corpus-level outputs and avoids reconstructing unverified audit details. Third, 2026 is an incomplete publication year, and recent citation counts are affected by short exposure time. Fourth, co-word and thematic maps depend on author/indexer terminology. Variants such as “machine learning” and “machine-learning,” or “blockchain” and “block-chain,” can fragment frequencies despite keyword merging. Future replications should preserve the full search history, document preprocessing rules explicitly, normalize lexical variants before mapping, and compare Scopus results with at least one complementary database.

CONCLUSION

Research on artificial intelligence in real-time payment systems has moved from a small, infrastructure-oriented literature into a rapidly expanding interdisciplinary field. The 545-document corpus shows an inflection after 2023, with output rising sharply in 2024 and 2025. Its current conceptual core is organized around learning systems, crime, fraud detection, AI, blockchain, and security. These are mature motor themes with high thematic centrality and density. Graph neural networks and behavioral research are less central but temporally newer, making them credible emerging fronts rather than declining topics.

The field’s next phase will be defined by a shift from isolated transaction classification to adaptive payment intelligence. Such intelligence will need to understand temporal sequences, relational networks, user behavior, and operational context while meeting strict requirements for latency, privacy, explainability, and resilience. The most consequential research opportunities therefore lie at the boundaries between technical performance and payment-system governance: dynamic GNNs for coordinated fraud, federated analytics across institutions, explainable low-latency decisioning, behavioral models for manipulated payments, adversarially robust learning, and carefully constrained AI agents for liquidity and operational management. A mature AI–RTPS research program will be judged not only by whether a model detects more anomalies, but by whether it improves the safety, efficiency, accountability, and trustworthiness of real-time financial infrastructures.

Data availability statement. The bibliometric results reported in this article were derived from the Scopus/Biblioshiny workbook supplied for the study. The exact Scopus search-history export should be retained with the submission materials to support full replication.

Conflict of interest. The authors declare no conflict of interest.

Funding. No external funding is reported in the supplied study materials.

REFERENCES

1. Abdallah, A., Maarof, M. A., & Zainal, A. (2016). Fraud detection system: A survey. Journal of Network and Computer Applications, 68, 90–113. https://doi.org/10.1016/j.jnca.2016.04.007
2. Aldasoro, I., & Desai, A. (2025). AI agents for cash management in payment systems (BIS Working Papers No. 1310). Bank for International Settlements.
3. Allen, F., Gu, X., & Jagtiani, J. (2021). A survey of fintech research and policy discussion. Review of Corporate Finance, 1(3–4), 259–339. https://doi.org/10.1561/114.00000007
4. Aria, M., & Cuccurullo, C. (2017). bibliometrix: An R-tool for comprehensive science mapping analysis. Journal of Informetrics, 11(4), 959–975. https://doi.org/10.1016/j.joi.2017.08.007
5. Aria, M., Cuccurullo, C., D’Aniello, L., & Spano, M. (2026). Biblioshiny and the SAAS workflow: An integrated framework for transparent and reproducible science mapping—A demonstration through the replication of a study. Journal of Informetrics. https://doi.org/10.1016/j.joi.2026.101837
6. Bank for International Settlements Innovation Hub. (2025). Project Hertha: Identifying financial crime patterns in real-time retail payment systems. Bank for International Settlements.
7. Brogi, M., & Lagasio, V. (2024). New but naughty: The evolution of misconduct in FinTech. International Review of Financial Analysis, 95, 103489. https://doi.org/10.1016/j.irfa.2024.103489
8. Cai, J., Liang, W., Li, X., Li, K.-C., Gui, Z., & Khan, M. K. (2023). GTxChain: A secure IoT smart blockchain architecture based on graph neural network. IEEE Internet of Things Journal, 10(24), 21502–21514. https://doi.org/10.1109/JIOT.2023.3296469
9. Callon, M., Courtial, J.-P., & Laville, F. (1991). Co-word analysis as a tool for describing the network of interactions between basic and technological research: The case of polymer chemistry. Scientometrics, 22, 155–205. https://doi.org/10.1007/BF02019280
10. Chen, X.-Q., Ma, C.-Q., Ren, Y.-S., Lei, Y.-T., Huynh, N. Q. A., & Narayan, S. (2023). Explainable artificial intelligence in finance: A bibliometric review. Finance Research Letters, 56, 104145. https://doi.org/10.1016/j.frl.2023.104145
11. Cobo, M. J., López-Herrera, A. G., Herrera-Viedma, E., & Herrera, F. (2011). An approach for detecting, quantifying, and visualizing the evolution of a research field: A practical application to the fuzzy sets theory field. Journal of Informetrics, 5(1), 146–166. https://doi.org/10.1016/j.joi.2010.10.002
12. Committee on Payments and Market Infrastructures. (2016). Fast payments—Enhancing the speed and availability of retail payments (CPMI Papers No. 154). Bank for International Settlements.
13. Committee on Payments and Market Infrastructures. (2021). Developments in retail fast payments and implications for RTGS systems. Bank for International Settlements.
14. Cornelli, G., Frost, J., Warren, J., Yang, C., & Velásquez, C. (2024). Retail fast payment systems as a catalyst for digital finance (BIS Working Papers No. 1228; revised June 2025). Bank for International Settlements.
15. Desai, A., Kosse, A., & Sharples, J. (2024). Finding a needle in a haystack: A machine learning framework for anomaly detection in payment systems (BIS Working Papers No. 1188). Bank for International Settlements.
16. Donthu, N., Kumar, S., Mukherjee, D., Pandey, N., & Lim, W. M. (2021). How to conduct a bibliometric analysis: An overview and guidelines. Journal of Business Research, 133, 285–296. https://doi.org/10.1016/j.jbusres.2021.04.070
17. European Banking Authority, & European Central Bank. (2025). 2025 report on payment fraud. European Banking Authority and European Central Bank.
18. Fan, S., Zhang, H., Zeng, Y., & Cai, W. (2021). Hybrid blockchain-based resource trading system for federated learning in edge computing. IEEE Internet of Things Journal, 8(4), 2252–2264. https://doi.org/10.1109/JIOT.2020.3028101
19. Frost, J., Koo Wilkens, P., Kosse, A., Shreeti, V., & Velásquez, C. (2024). Fast payments: Design and adoption. BIS Quarterly Review, March.
20. Goodell, J. W., Kumar, S., Lim, W. M., & Pattnaik, D. (2021). Artificial intelligence and machine learning in finance: Identifying foundations, themes, and research clusters from bibliometric analysis. Journal of Behavioral and Experimental Finance, 32, 100577. https://doi.org/10.1016/j.jbef.2021.100577
21. Hernández, L., Baladrón, C., Aguiar, J. M., Carro, B., Sánchez-Esguevillas, A., & Lloret, J. (2014). Artificial neural networks for short-term load forecasting in microgrids environment. Energy, 75, 252–264. https://doi.org/10.1016/j.energy.2014.07.065
22. Kayikci, S., & Khoshgoftaar, T. M. (2024). Blockchain meets machine learning: A survey. Journal of Big Data, 11, 9. https://doi.org/10.1186/s40537-023-00852-y
23. Lăzăroiu, G., Bogdan, M., Geamănu, M., Hurloiu, L., Ionescu, L., & Ștefănescu, R. (2023). Artificial intelligence algorithms and cloud computing technologies in blockchain-based fintech management. Oeconomia Copernicana, 14(3), 707–730. https://doi.org/10.24136/oc.2023.021
24. Liu, M., Yu, F. R., Teng, Y., Leung, V. C. M., & Song, M. (2019). Performance optimization for blockchain-enabled industrial Internet of Things (IIoT) systems: A deep reinforcement learning approach. IEEE Transactions on Industrial Informatics, 15(6), 3559–3570. https://doi.org/10.1109/TII.2019.2897805
25. Motie, S., & Raahemi, B. (2024). Financial fraud detection using graph neural networks: A systematic review. Expert Systems with Applications, 240, 122156. https://doi.org/10.1016/j.eswa.2023.122156
26. Özyılmaz, K. R., DoÄŸan, M., & Yurdakul, A. (2018). IDMoB: IoT data marketplace on blockchain. In 2018 Crypto Valley Conference on Blockchain Technology (CVCBT) (pp. 11–19). IEEE. https://doi.org/10.1109/CVCBT.2018.00007
27. Pattnaik, D., Ray, S., & Raman, R. (2024). Applications of artificial intelligence and machine learning in the financial services industry: A bibliometric review. Heliyon, 10(1), e23492. https://doi.org/10.1016/j.heliyon.2023.e23492
28. Roy, A., Sun, J., Mahoney, R., Alonzi, L., Adams, S., & Beling, P. (2018). Deep learning detecting fraud in credit card transactions. In 2018 Systems and Information Engineering Design Symposium (SIEDS) (pp. 129–134). IEEE. https://doi.org/10.1109/SIEDS.2018.8374722
29. Ryman-Tubb, N. F., Krause, P., & Garn, W. (2018). How artificial intelligence and machine learning research impacts payment card fraud detection: A survey and industry benchmark. Engineering Applications of Artificial Intelligence, 76, 130–157. https://doi.org/10.1016/j.engappai.2018.07.008
30. Yun, J., Goh, Y., & Chung, J.-M. (2021). DQN-based optimization framework for secure sharded blockchain systems. IEEE Internet of Things Journal, 8(2), 708–722. https://doi.org/10.1109/JIOT.2020.3006896
31. Zupic, I., & ÄŒater, T. (2015). Bibliometric methods in management and organization. Organizational Research Methods, 18(3), 429–472. https://doi.org/10.1177/1094428114562629