AI, KYC Screening and the Evidential Limits of Automated Risk Intelligence

Abstract

This paper critically evaluates LSEG World-Check One as an artificial intelligence-enabled KYC and compliance technology, examining its technological proposition against academic research on customer due diligence, sanctions screening, machine learning and adverse-media analysis. The analysis considers the platform’s capabilities in identity verification, entity matching, sanctions and PEP screening, adverse-media monitoring, ongoing monitoring, case management and AI-assisted relevance filtering. While these functions are broadly consistent with established developments in financial-crime technology, the analysis distinguishes technological plausibility from independent validation of commercial performance claims.

Particular attention is given to LSEG’s claims concerning accuracy, reduced false positives, continuous monitoring and AI-powered risk intelligence. The paper argues that such claims cannot be adequately assessed without greater methodological transparency concerning evaluation populations, ground truth, matching thresholds, false-positive and false-negative rates, data quality and independent benchmarking. Research demonstrates that improvements in screening sensitivity may also increase false positives, highlighting the need to evaluate competing forms of error rather than treating alert reduction as an inherently positive outcome.

The analysis further identifies important conceptual boundaries between identity verification, screening, risk assessment and proof of misconduct. A screening alert represents a potential match requiring investigation rather than confirmation of sanctions exposure, while adverse media represents information requiring contextual and evidential assessment rather than proof of wrongdoing. Similarly, continuous monitoring should not be equated with predictive accuracy. The paper concludes that World-Check One is best understood as decision-support and risk-intelligence infrastructure that can improve the organisation, prioritisation and timeliness of compliance information, but whose substantive value ultimately depends on data quality, contextual interpretation, human judgement and independently demonstrated performance.

1. Introduction

Know Your Customer (KYC) and customer due diligence (CDD) are foundational components of contemporary financial-crime compliance. Financial institutions and other regulated organisations are expected not merely to establish a customer’s identity, but to understand the nature and purpose of the relationship, assess relevant risk factors and apply appropriate monitoring throughout the customer lifecycle. This reflects the broader shift towards a risk-based approach to anti-money laundering (AML), in which customer information and risk indicators are used to determine the nature and intensity of compliance measures rather than applying identical controls to every customer (De Wit, 2007). Academic research likewise indicates that KYC information can provide a basis for data-driven risk assessment and the identification of patterns associated with elevated risk (Chen, 2020).

The practical difficulty is that contemporary KYC environments generate large, heterogeneous and continually changing volumes of information. Global sanctions regimes, politically exposed persons (PEPs), complex ownership structures, cross-border relationships, regulatory developments and rapidly evolving financial-crime typologies all increase the demands placed on compliance functions. These pressures have encouraged greater reliance on automated screening, machine learning, natural language processing (NLP), entity matching and data-integration technologies. The underlying rationale is straightforward: technology may allow organisations to process substantially more information, identify potential risk indicators more consistently and reduce the operational burden associated with manual review. However, the ability to process and prioritise information should not automatically be equated with the ability to determine risk accurately.

LSEG positions World-Check One as a KYC verification and customer-screening platform designed to support AML, KYC and counter-terrorist-financing compliance. Its current product materials describe screening across sanctions, PEPs, adverse media and other risk categories, together with advanced name matching, ultimate beneficial owner (UBO) verification, identity verification, enhanced due-diligence reports, ongoing monitoring, batch screening, API integration and case-management functionality (LSEG, 2026a; LSEG, 2026b). LSEG also describes World-Check as drawing on public sources, official records, government watchlists, media and regulatory information, with research and verification undertaken by its risk-intelligence specialists.

The wider World-Check proposition extends beyond the provision of sanctions data. LSEG describes its World-Check database as structured risk intelligence concerning individuals and organisations, encompassing sanctions, PEPs, regulatory and law-enforcement information, adverse media, UBO information and other risk indicators (LSEG, 2026b; LSEG, 2026c). The resulting data can be incorporated into customer onboarding, KYC, due-diligence, payment-screening and ongoing-monitoring processes. LSEG also presents adverse-media screening and real-time or API-based screening as components of a broader compliance technology ecosystem (LSEG, 2026d; LSEG, 2026e).

World-Check One should therefore be understood not simply as a sanctions-list database, but as a broader risk-screening and decision-support infrastructure. Its proposition combines data aggregation, identity resolution, name matching, risk categorisation, adverse-media processing, ongoing monitoring and workflow automation. Each of these capabilities is technically plausible and, in isolation, has some support within the academic literature. For example, research demonstrates the potential of machine learning to use KYC information for risk assessment (Chen, 2020), while recent research on sanctions screening indicates that NLP can improve the sensitivity of screening systems to potential matches. At the same time, that research also demonstrates an important trade-off: increasing sensitivity can increase false positives, meaning that improved detection of potential matches does not necessarily translate into improved overall screening performance (Kim and Yang, 2024).

This distinction is central to evaluating commercial KYC technology. Finding more potential risk is not necessarily the same as identifying risk more accurately. A screening system must distinguish genuine matches from false positives; identify relevant information without overwhelming investigators; and ensure that potentially important entities are not excluded through overly restrictive matching or filtering. The problem is particularly acute in sanctions screening, where false negatives may allow sanctioned entities to escape detection, while excessive false positives can generate substantial manual-review burdens. Kim and Yang (2024), for example, found that an NLP-based approach could increase sensitivity and detect more true positives while simultaneously increasing false positives, illustrating that improvements in one performance dimension may come at the expense of another.

The same analytical problem applies more broadly to KYC and CDD. Customer information can support risk assessment, but it does not by itself establish that a customer is engaged in financial crime. Similarly, a name match is not necessarily a verified identity, an adverse-media reference is not necessarily evidence of wrongdoing, and a risk classification is not equivalent to a regulatory or investigative finding. Research on KYC and due diligence highlights the continuing importance of customer identification, contextual information and appropriate investigative procedures, particularly where automated systems operate on incomplete, ambiguous or heterogeneous information. The effectiveness of a risk-based AML framework therefore depends not simply on the volume of data available, but on how reliably that data can be interpreted and translated into proportionate compliance action (De Wit, 2007).

This creates an important evidential question concerning LSEG’s commercial proposition. The existence of a technically sophisticated screening architecture does not, in itself, establish the magnitude of any claimed improvement in accuracy, efficiency or risk detection. Academic studies can demonstrate that machine learning, NLP, entity matching and data-driven risk assessment are viable approaches; they cannot automatically validate the performance of a particular proprietary system operating on proprietary datasets. Such validation would require evidence concerning the relevant datasets, ground truth, matching methodology, false-positive and false-negative rates, testing procedures, temporal stability and performance across different jurisdictions, languages, customer populations and risk categories.

Accordingly, this paper critically evaluates World-Check One at three interconnected levels. First, it examines whether the technological components underlying the platform—including automated screening, name matching, machine learning, adverse-media processing, risk assessment and ongoing monitoring—are consistent with established academic research. Second, it considers whether LSEG’s specific claims concerning the effectiveness, accuracy or efficiency of these capabilities are independently substantiated by publicly available evidence. Third, it examines the methodological and governance limitations that arise when automated screening outputs are used to support consequential compliance decisions.

The central analytical distinction throughout the paper is therefore between technological capability and evidential reliability. World-Check One may plausibly improve the ability of compliance teams to identify, organise and prioritise potentially relevant information. The more demanding question is whether those signals are sufficiently accurate, contextualised and validated to support reliable conclusions about customer risk. In this respect, the critical boundary is not between manual and automated compliance, but between detecting a potential risk signal and establishing what that signal actually means. The paper consequently treats automated KYC technology as a mechanism for generating and prioritising evidence rather than as an autonomous substitute for contextual due diligence and human judgement.

2. World-Check One’s Technological Proposition

LSEG presents World-Check One as an integrated screening and compliance platform rather than as a single analytical model. Its functionality encompasses advanced name matching using secondary identifiers, sanctions and politically exposed person (PEP) screening, AI-powered adverse-media relevance filtering, ultimate beneficial ownership (UBO) verification, identity verification, enhanced due-diligence reporting, ongoing monitoring, batch screening, API integration and case-management capabilities. LSEG also describes the recording and date-stamping of user actions, creating an audit trail of screening and investigative activity (LSEG, 2026a; LSEG, 2026b). The technological proposition is therefore best understood as a layered compliance architecture in which data aggregation, entity matching, risk identification, monitoring and workflow management operate together rather than as a standalone prediction system.

This architecture is broadly consistent with the direction of academic research into data-driven AML and KYC. Machine-learning approaches can use customer and transactional information to identify patterns and risk relationships that may be difficult to capture through conventional static rules. Chen (2020), for example, demonstrates the use of machine-learning techniques for bank risk assessment using KYC-related information, while also illustrating the challenges associated with heterogeneous customer data and differing behavioural patterns. The significance of such research, however, lies primarily in demonstrating the technical feasibility of data-driven risk assessment, rather than establishing that a particular commercial platform will achieve equivalent performance in operational settings. Model effectiveness remains dependent on the quality, representativeness and preparation of the underlying data, as well as on model specification, validation and the definition of the outcome being predicted.

The literature also supports a broader shift from purely rules-based compliance towards risk-based customer due diligence. De Wit (2007) argues that AML frameworks should differentiate customers and relationships according to their relative risk rather than relying exclusively on uniform controls. More recent research similarly identifies CDD as a central component of AML/CFT risk management, while highlighting persistent practical difficulties in areas such as beneficial ownership identification, PEP assessment, customer identity, name structures and source-of-wealth verification (ElYacoubi, 2020). These difficulties are important because they occur precisely at the point where automated systems must convert fragmented and potentially ambiguous information into a meaningful customer-risk assessment.

World-Check One’s use of secondary identifiers and entity-matching functionality is consequently significant. Names alone are often insufficient to distinguish individuals or organisations, particularly where customers share common names, use transliterations or operate across jurisdictions with different naming conventions. Additional identifiers can improve the ability of a screening system to distinguish potential matches from unrelated entities. Yet improved matching capability does not eliminate the underlying identification problem: a screening match is a candidate relationship, not automatically a verified identity. The reliability of the resulting risk signal depends on the accuracy, completeness and provenance of the identifiers against which the match is made.

The same principle applies to the platform’s broader data-integration model. Aggregating sanctions, PEP, adverse-media, ownership and other risk information can increase the amount of potentially relevant information available to investigators. It may also reduce the fragmentation associated with searching multiple sources manually. However, more information does not necessarily mean more reliable intelligence. Data can be duplicated, outdated, incomplete, contradictory or unevenly reliable across jurisdictions and source types. Consequently, the value of an integrated screening platform depends not only on how much information it can access, but on how effectively that information is curated, matched, contextualised and maintained.

The AI component introduces a further distinction. LSEG describes AI-powered filtering of adverse media as a means of identifying relevant information within large volumes of potentially material content (LSEG, 2026d). This is technically plausible and consistent with the broader literature on AI applications in financial-crime detection, where machine-learning and NLP techniques are increasingly used to process large information sets and identify potentially relevant patterns (Mousavian and Miah, 2025). Nevertheless, relevance filtering should not be confused with establishing the truth or significance of an allegation. An algorithm may identify an article as relevant to a customer, but determining whether the underlying information is credible, current, contextually meaningful and indicative of material risk remains a more demanding evidential task.

World-Check One should therefore be viewed as evidence-processing and decision-support infrastructure, rather than as an autonomous mechanism for determining financial-crime risk. Its architecture can plausibly improve the speed, scale and organisation of screening, while machine-learning and matching techniques may improve the identification and prioritisation of potentially relevant information. These are meaningful technological capabilities. They do not, however, by themselves establish that the resulting alerts or risk indicators are accurate, complete or sufficient to support a compliance conclusion.

The technological proposition of World-Check One is consequently plausible at the architectural level, and several of its constituent techniques are supported by academic research. The more difficult evidential question is whether those techniques, when combined within LSEG’s proprietary platform and datasets, produce demonstrably superior risk identification in practice. Establishing that proposition requires more than evidence that the underlying technologies work in principle. It requires transparent evaluation of matching accuracy, false-positive and false-negative rates, data quality, model performance, validation methodology and performance across different customer populations, jurisdictions and risk categories. This distinction between technological plausibility and demonstrated operational effectiveness provides the basis for evaluating the specific claims made for World-Check One in the sections that follow.

3. KYC and Customer Due Diligence: Identification Is Not Risk Determination

A central conceptual issue in evaluating World-Check One is the distinction between identity verification, risk identification and risk determination. Although these functions are closely connected within the KYC process, they answer fundamentally different questions and should not be treated as interchangeable.

Identity verification asks whether the available evidence is sufficient to establish that a customer is the person or organisation they claim to be. Screening addresses a different question: whether that identified subject corresponds to information associated with sanctions, politically exposed person (PEP) status, adverse media or another defined risk category. Risk assessment then goes further by considering the relevance, reliability and significance of that information within the specific customer relationship. In other words, identification establishes who the subject is; screening identifies potentially relevant information about that subject; risk determination interprets what that information means in context.

This distinction is particularly important because each stage is subject to different sources of uncertainty. Research on customer due diligence identifies persistent practical difficulties in establishing beneficial ownership, interpreting names across different linguistic and cultural contexts, assessing PEP relationships and verifying source-of-funds information (ElYacoubi, 2020). These challenges demonstrate that customer identification is not simply a matter of retrieving a matching record from a database. It can require the reconstruction of relationships, ownership structures and contextual information from incomplete or heterogeneous sources. Similarly, research examining technological approaches to CDD highlights the potential vulnerabilities that arise when customer identity, ownership or contextual information is inadequately established (Demetriades, 2016).

World-Check One seeks to address some of these difficulties through the use of secondary identifiers and additional data. LSEG states that its matching technology can use multiple secondary identifiers alongside configurable matching algorithms and filtering technologies to improve matching and reduce false positives (LSEG, 2026a). In principle, this approach is preferable to relying exclusively on names. Additional identifiers such as nationality, date of birth or other relevant attributes can provide greater discriminatory power when determining whether two records potentially refer to the same individual or organisation.

Nevertheless, additional identifiers do not eliminate the underlying epistemic problem of entity resolution. A potential database match remains a candidate relationship requiring contextual assessment. A shared name may provide only weak evidence; the addition of nationality, date of birth or corporate information may increase confidence, but the strength of the resulting inference remains dependent on the accuracy, completeness, currency and provenance of the underlying data. Where identifiers are missing, incorrectly recorded or inconsistent across sources, an apparently sophisticated matching system may still generate either a false positive or a false negative.

This creates an important distinction between matching accuracy and risk accuracy. A system may become better at determining that two records probably refer to the same person without necessarily becoming better at determining whether that person represents a material compliance risk. Establishing identity is a prerequisite for meaningful screening, but it is not equivalent to establishing risk. Likewise, identifying that a customer is associated with a PEP record, sanctions entry or adverse-media article does not, without further contextual analysis, establish the significance of that association within the particular customer relationship.

The distinction becomes even more consequential when screening outputs are used to trigger substantive compliance action. A potential match may result in additional investigation, enhanced due diligence, transaction review, account rejection or termination of a commercial relationship. The evidential threshold for such actions is therefore materially higher than the threshold required simply to generate an automated alert. An alert indicates that something warrants attention; it does not, by itself, establish what has occurred or what action should follow.

World-Check One should consequently be understood as producing risk signals rather than definitive risk determinations. Its value lies in helping compliance personnel identify potentially relevant relationships and information at scale, while the ultimate interpretation of those signals depends on the quality of the evidence and the context in which the customer operates. This is consistent with a risk-based approach to AML, in which customer information should inform differentiated and proportionate compliance measures rather than mechanically determine outcomes (De Wit, 2007).

The critical boundary can therefore be stated simply: identification is not verification, a match is not necessarily a true match, and a screening signal is not a risk determination. The effectiveness of automated KYC technology must consequently be assessed not only by how many potential matches it can identify, but by how reliably those matches correspond to the correct subjects and how accurately the resulting information supports subsequent risk assessment. This distinction provides an essential foundation for evaluating World-Check One’s claims concerning name matching, false-positive reduction and automated screening performance.

4. Sanctions Screening and the False-Positive Problem

Sanctions screening provides one of the clearest examples of both the potential value and the methodological limitations of automated compliance technology. LSEG positions World-Check One as providing sanctions screening supported by advanced name matching, secondary identifiers and configurable filtering, and explicitly associates these capabilities with reducing false-positive alerts (LSEG, 2026a; LSEG, 2026b). In principle, improved entity matching and contextual filtering can reduce the number of irrelevant matches requiring manual investigation. The more important question, however, is what is being traded off in achieving that reduction.

Peer-reviewed research demonstrates why this problem is technically and methodologically challenging. Kim and Yang (2024) examine the application of natural language processing (NLP) to financial sanctions screening and demonstrate that improvements in detection sensitivity can be accompanied by substantial increases in false-positive alerts. Their results indicate that NLP can identify additional true-positive cases and reduce false negatives, while simultaneously increasing the number of alerts that do not correspond to genuine sanctions exposure. The finding illustrates that screening performance involves competing error costs rather than a simple progression from “less accurate” to “more accurate”.

This is particularly important when evaluating commercial claims concerning false-positive reduction. Fewer false positives do not automatically mean better screening. A system can reduce the number of alerts by becoming more selective, but if that selectivity also suppresses genuine matches, the apparent efficiency gain may be achieved at the expense of detection capability. Conversely, a highly sensitive system may identify more genuine sanctions exposures while generating a larger number of alerts requiring human investigation. The appropriate balance therefore depends on the relative costs of false positives and false negatives within the particular compliance context.

The central methodological issue is consequently not alert volume in isolation, but the relationship between detection performance and unnecessary alert generation. Kim and Yang (2024) demonstrate why this relationship cannot be captured adequately by a single measure. A meaningful evaluation should consider at least precision, recall or sensitivity, specificity, false-positive rates and false-negative rates, alongside the threshold at which the system generates an alert. Where machine-learning or NLP systems are involved, the composition of the evaluation dataset and the method used to establish the ground truth are equally important. Without these details, it is difficult to determine whether a reported reduction in false positives reflects a genuine improvement in screening quality, a change in alert thresholds, a different evaluation population or a trade-off against missed sanctions matches.

The problem becomes more pronounced in sanctions screening because genuine sanctions matches represent a relatively small proportion of the much larger population of customers, transactions or entities being screened. This creates a class-imbalance problem, in which a system may appear highly accurate by correctly classifying the overwhelming majority of non-sanctioned subjects while performing less effectively on the comparatively small number of cases that constitute genuine exposure. Consequently, aggregate accuracy can provide a misleading picture of screening effectiveness. What matters in this context is not simply how many records the system classifies correctly overall, but how effectively it identifies the relatively rare positive cases without generating an excessive number of unnecessary alerts.

The trade-off between different forms of screening error is demonstrated by Kim and Yang (2024), who examine the use of natural language processing in financial sanctions screening. Their findings illustrate that increasing the sensitivity of a screening approach can improve the identification of genuine matches while also increasing false positives. This is important because it shows that screening performance cannot be evaluated through a single measure. A reduction in one type of error may be accompanied by an increase in another, meaning that claims of improved screening effectiveness require consideration of the overall error profile rather than one favourable metric (Kim and Yang, 2024).

The composition of the evaluation population is equally important. Research on customer due diligence demonstrates that obtaining and assessing reliable customer information remains operationally difficult, particularly where customer identities, beneficial ownership and other relevant information are complex or incomplete (Challenges in customer due diligence for banks in the UAE, 2020). Similarly, Demetriades (2016) highlights the continuing importance of establishing whether an individual is genuinely the person they claim to be. These issues are directly relevant to automated screening because performance observed on relatively straightforward records may not necessarily represent performance in operational environments characterised by ambiguous identities, incomplete information and complex relationships.

This places an important evidential qualification on LSEG’s claims concerning false-positive reduction. The existence of advanced matching, secondary identifiers and filtering capabilities is technically plausible and consistent with the direction of research into automated sanctions screening (Kim and Yang, 2024). However, demonstrating that these capabilities reduce false positives without materially increasing false negatives requires empirical evidence from an appropriately defined evaluation population. Such evidence would need to establish not only the number of alerts generated, but also how many genuine matches were identified, how many were missed and how performance was determined.

The appropriate question is therefore not simply whether World-Check One generates fewer alerts. It is whether the system reduces unnecessary alerts while maintaining or improving the detection of genuine sanctions exposure. This requires assessment of the full performance profile of the screening system, including precision, recall, sensitivity, specificity and false-positive and false-negative rates. Operational measures such as analyst workload and alert-resolution time are also relevant because a technically sensitive system may create an investigative burden that affects the practical effectiveness of the screening process.

These considerations also reinforce the importance of human judgement within automated screening. A screening system determines which records warrant further attention, but investigators must establish whether the apparent match is genuine and what significance it has in the particular customer or transaction context. As Demetriades (2016) illustrates in the broader context of due diligence, establishing identity remains a substantive evidential task rather than merely a technical matching exercise. The role of automation is therefore better understood as supporting the identification and prioritisation of potentially relevant cases than as independently determining sanctions exposure.

Consequently, screening efficiency should not be confused with screening accuracy, and screening accuracy should not be confused with a substantive finding of sanctions exposure. World-Check One may improve the efficiency with which potential matches are identified, filtered and prioritised, but demonstrating that it does so without sacrificing material detection capability requires transparent evidence across the competing dimensions of screening performance.

Takeaway: The objective of sanctions screening is not minimum alert volume, but an evidence-based balance between detecting genuine exposure, limiting unnecessary alerts and keeping the investigative burden manageable.

5. Name Matching and the Problem of False Identity

The effectiveness of sanctions and politically exposed person (PEP) screening depends substantially on entity resolution: determining whether the customer being screened is in fact the same individual or organisation identified in a relevant external record. This is a more demanding task than identifying textual similarity between two names. A screening system must distinguish between subjects who are genuinely the same and subjects whose names or characteristics merely appear sufficiently similar to generate an alert.

World-Check One states that its matching technology incorporates multiple secondary identifiers alongside configurable name-matching algorithms and filtering technologies (LSEG, 2026a). The wider World-Check proposition similarly emphasises structured and de-duplicated information and the linking of associated individuals and organisations to identify relationships and networks (LSEG, 2026b; LSEG, 2026c). These capabilities are designed to address a fundamental weakness of simple name-based screening: names are identifiers, but they are not necessarily unique identifiers.

The problem becomes particularly pronounced in international KYC environments. Names may vary because of transliteration between writing systems, spelling conventions, abbreviations, cultural naming structures, changes of name and differences in the way personal or corporate names are recorded across jurisdictions. The CDD literature identifies difficulties associated with names, identity and beneficial ownership as continuing operational challenges, demonstrating that establishing the correct identity can require considerably more than matching a customer’s name against a database record (ElYacoubi, 2020). Research on customer identification similarly highlights the continuing importance of conventional due-diligence procedures where identity or contextual information is insufficiently established (Demetriades, 2016).

Secondary identifiers can therefore provide important additional evidence. Information such as date of birth, nationality, location or corporate relationships may help distinguish individuals who share similar names. In principle, combining multiple attributes should allow a screening system to move beyond simple textual similarity towards a more contextual assessment of whether two records refer to the same subject. This is an important technological improvement because the central problem is not merely finding similar records, but resolving identities reliably.

Kim and Yang (2024) provide relevant empirical evidence concerning the use of more sophisticated language-processing techniques in sanctions screening. Their research indicates that NLP can improve the identification of potential true-positive cases and reduce false negatives. At the same time, greater sensitivity can generate substantially more false-positive alerts. This illustrates an important trade-off for entity-resolution systems: making a matching algorithm more sensitive may increase its ability to detect difficult genuine matches, but it can also increase the number of unrelated subjects incorrectly identified as potential matches.

The technological objective is therefore not simply to maximise the number of matches detected. It is to distinguish true identity matches from plausible but incorrect matches. This distinction is critical because the consequences of a false identity can operate in both directions. An incorrect match may subject an innocent customer to enhanced scrutiny, transaction delays or additional due diligence, while an incorrect non-match may prevent a genuinely sanctioned or otherwise high-risk subject from being identified.

A further methodological problem is that a similarity score does not constitute proof of identity. A high degree of similarity between two names may increase the probability that the records relate to the same subject, but it does not establish that conclusion independently. Conversely, a low similarity score does not necessarily establish that two records concern different subjects where names have been transliterated, abbreviated or otherwise recorded differently. The evidential strength of a match therefore depends on the interaction between the name, secondary identifiers, source quality, data completeness and contextual relationships.

This is also where the distinction between matching and verification becomes particularly important. An automated system can determine that two records share a sufficiently strong combination of characteristics to warrant investigation. It cannot necessarily establish, from that match alone, that the customer is the listed individual or organisation. Verification requires consideration of the underlying evidence and, where necessary, additional customer or documentary information. Consequently, a match is an investigative signal, not an identity determination.

World-Check One’s use of secondary identifiers and configurable matching is therefore technically appropriate to the entity-resolution problem. The existence of such functionality, however, does not by itself establish a particular level of matching accuracy or superiority over alternative systems. Demonstrating such a claim would require independent empirical evaluation against a representative screening population containing both genuine matches and challenging non-matches. The evaluation should also account for common names, transliteration differences, incomplete identifiers, corporate structures and cross-jurisdictional variation.

A credible assessment would therefore need to report not only the proportion of correct matches, but also the system’s false-positive and false-negative performance, the threshold at which alerts are generated, the composition of the test population and the method used to establish the ground truth. Without these elements, a claim that advanced matching improves screening accuracy remains technically plausible but empirically under-specified.

The broader lesson is that better matching is not synonymous with certain identification. Advanced algorithms and additional identifiers can improve the probability that the correct subject is identified, but the reliability of the final conclusion depends on the quality of the evidence surrounding the match. For consequential KYC decisions, the appropriate evidential chain is therefore: candidate match → identity verification → contextual assessment → risk determination, rather than candidate match → assumed identity → assumed risk.

6. Ongoing Monitoring and the Changing Nature of Customer Risk

World-Check One also positions ongoing monitoring and rescreening as important components of its KYC proposition. LSEG states that World-Check data are updated daily and that its research teams monitor sanctions, regulatory and enforcement lists, as well as media sources (LSEG, 2026b; LSEG, 2026c). The wider World-Check proposition supports the continued screening of existing customer relationships and integration with onboarding, KYC and third-party due-diligence workflows. This reflects an important recognition within modern compliance: customer risk is dynamic rather than fixed at the point of onboarding.

The rationale for ongoing screening is consistent with the risk-based nature of customer due diligence. A customer assessed as relatively low risk when a relationship is established may subsequently become subject to sanctions, regulatory enforcement, adverse media, changes in beneficial ownership or other events that materially alter the risk profile. A one-time KYC assessment can therefore become outdated as the circumstances surrounding a customer change. Ongoing monitoring seeks to address this temporal limitation by repeatedly comparing existing customers against newly available risk information rather than treating the original due-diligence assessment as permanently valid.

The academic literature supports the broader premise that financial-crime risk is affected by changing customer characteristics, behaviours and operating environments. Research on data-driven AML risk assessment demonstrates that customer and financial-crime patterns can differ across jurisdictions and business environments, highlighting the limitations of simplistic or static approaches to risk identification (Chen, 2020). More broadly, the risk-based approach to AML requires customer risk to be assessed in relation to changing circumstances rather than treated as a fixed characteristic (De Wit, 2007). Ongoing screening is therefore conceptually consistent with the underlying logic of risk-based CDD.

Ongoing monitoring addresses an important limitation of conventional KYC: customer risk is not static and information relevant to that risk can change after the initial customer assessment. This is consistent with the risk-based approach to AML, under which customer due diligence should respond to changes in the nature and level of identified risk rather than treating KYC as a one-off exercise (De Wit, 2007). However, the effectiveness of ongoing monitoring introduces a distinct methodological issue: timeliness. The fact that a database is updated daily does not establish that every material risk event will be detected immediately. Between an external event occurring and an actionable customer-level alert being generated, several stages intervene. Information must become available, be captured by an appropriate source, be incorporated into the provider’s data environment, be correctly attributed to the relevant individual or organisation, and then be processed through the applicable screening workflow.

This creates an important distinction between data-update frequency and detection latency. Daily database updates demonstrate a particular maintenance frequency, but they do not establish how quickly information moves through the entire evidential and technological chain. The distinction is particularly relevant in KYC because customer information can be incomplete, inconsistent or difficult to attribute, creating challenges for due diligence and customer identification (Challenges in customer due diligence for banks in the UAE, 2020; Demetriades, 2016). A sanctions designation, regulatory action or adverse-media report may therefore exist before it is incorporated into a screening environment, while information that has already been captured may still require entity resolution before it can be reliably associated with the correct customer. Consequently, database freshness should not be treated as a direct proxy for the speed or reliability of risk detection.

The operational significance of ongoing monitoring therefore depends not simply on whether new information is eventually incorporated, but on whether material changes in risk are identified within a timeframe that permits appropriate compliance action. A system that identifies a sanctions designation shortly after publication has different operational implications from one that identifies the same designation only after a substantial delay. The relevant performance concept is consequently the latency between the underlying event and the generation of a reliable, actionable customer-level signal. This is particularly important because an apparently “current” database can still contain information that is operationally delayed if the underlying source, ingestion process, matching process or alert workflow introduces a significant time lag.

A meaningful evaluation of ongoing monitoring should therefore examine the individual stages through which information becomes actionable. These include source latency, data-ingestion time, processing and verification time, entity-resolution performance, alert-generation latency and delivery to the responsible compliance function. Each stage can introduce delay or error, meaning that the performance of the overall monitoring system cannot necessarily be inferred from the update frequency of the underlying database. Research on AI-enabled financial-crime detection similarly indicates that technological capability must be assessed through empirical performance rather than assumed from the existence of automated processing alone (Mousavian and Miah, 2025). In this context, a claim that data are updated frequently is evidence of system maintenance, not independent evidence of end-to-end detection performance.

There is also an important distinction between monitoring and prediction. Ongoing screening can identify newly available information that changes the known risk profile of an existing customer, but this is fundamentally different from predicting that a risk event will occur before evidence of that event becomes available. A system may therefore be highly effective at maintaining current screening information without being a predictive model of future customer behaviour or financial-crime risk. This distinction is important when terms such as “continuous”, “ongoing” or “real-time” monitoring are used, because such terminology can describe the frequency or immediacy of information processing without demonstrating predictive capability. Continuous access to new information is not equivalent to forecasting future risk.

The distinction is particularly important given the broader development of AI-based financial-crime technologies. Academic research identifies significant potential for AI and machine learning to support financial-crime detection, but also emphasises the need for appropriate evaluation of system performance and limitations (Mousavian and Miah, 2025). Similarly, research on machine-learning approaches to KYC demonstrates that data-driven methods can be used for customer risk assessment, but evidence from a particular dataset or institutional environment does not automatically establish performance across different populations, jurisdictions or operational conditions (Chen, 2020). The same principle applies to ongoing monitoring: the existence of automated and frequently updated screening does not, by itself, establish that material risk changes will be detected consistently, rapidly or accurately across all relevant circumstances.

LSEG’s ongoing-monitoring proposition is therefore technically consistent with the dynamic nature of customer risk and with the broader risk-based approach to customer due diligence. Regularly updated risk intelligence and rescreening can address limitations inherent in one-time KYC by allowing new information to be incorporated into an existing customer assessment. Nevertheless, frequent updating should not be interpreted as evidence of comprehensive or instantaneous detection. The practical effectiveness of the system depends on the quality and coverage of the underlying sources, the speed with which information enters the system, the accuracy of entity resolution and the reliability of the subsequent alerting process.

The relevant performance question is consequently not simply how often the database is updated, but how reliably and quickly material changes in customer risk are converted into accurate and actionable screening signals. Establishing that performance would require evidence concerning source coverage, update latency, processing times, entity-resolution accuracy, alert-generation speed and the treatment of delayed, incomplete or ambiguous information. This produces an important distinction between technological capability and demonstrated operational performance: World-Check One may provide a technically credible mechanism for continuous risk monitoring, but the claim that it detects material changes rapidly and reliably requires empirical evidence across the complete monitoring chain.

Takeaway: Ongoing monitoring can make KYC more responsive to changing customer risk, but database freshness is not the same as detection speed—and continuous detection is not the same as prediction.

7. Adverse Media Detection and the Problem of Information Quality

Adverse media is a major component of the World-Check One proposition. LSEG describes adverse-media screening as a mechanism for identifying information relating to alleged criminal activity, regulatory breaches, sanctions exposure, money laundering, corruption and other potentially relevant risks. Its current solution combines structured adverse-media records with screening of unstructured media sources and describes the use of artificial intelligence, machine learning, natural language processing (NLP) and intelligent tagging to identify and prioritise potentially relevant information (LSEG, 2026d). The technological rationale is clear: automated analysis can process volumes of textual information that would be difficult for compliance personnel to review manually.

The academic literature supports the general proposition that AI and NLP can assist in processing large and heterogeneous bodies of text. Recent reviews of AI applications in financial-crime detection identify NLP and related techniques as potentially useful for extracting patterns and indicators from large information environments (Mousavian and Miah, 2025). However, adverse-media screening presents a substantially more demanding problem than simple information retrieval. The relevant question is not merely whether a system can locate text containing a customer's name, but whether it can reliably determine what the information means, whether it concerns the correct subject and how much evidential weight it should receive.

A central challenge in adverse-media screening is source reliability and evidential status. Media reporting can describe allegations, investigations, arrests, charges, regulatory proceedings, enforcement actions or criminal convictions, but these categories carry materially different evidential significance. An allegation represents an assertion that has not necessarily been substantiated; an investigation indicates that conduct is being examined; a charge constitutes a formal accusation; whereas a conviction represents a judicial determination. Treating these categories as a generic form of “adverse” information risks obscuring important differences in evidential status. An automated system may therefore correctly identify an article as relevant to a customer while remaining unable to establish whether the underlying allegation is credible, substantiated or legally resolved.

This creates a fundamental distinction between information indicating potential wrongdoing and evidence establishing wrongdoing. The purpose of adverse-media screening is consequently better understood as identifying information that may warrant further investigation rather than establishing misconduct itself. This distinction is particularly important where adverse-media results may contribute to enhanced due diligence, customer acceptance, transaction review or decisions concerning an existing customer relationship. The presence of an adverse-media result should therefore be treated as a risk signal requiring contextual assessment rather than as a substantive conclusion about the customer. This is consistent with the broader risk-based approach to AML, in which information contributes to an assessment of customer risk rather than automatically determining the appropriate compliance outcome (De Wit, 2007).

A second challenge concerns contextual attribution. The appearance of an individual's name in an article does not establish the nature of that individual's involvement in the reported event. The person may be the subject of an investigation, an executive of an organisation involved in the matter, a witness, victim, relative or simply an individual mentioned incidentally. Similar problems arise with organisations: an entity may be the subject of regulatory action, a subsidiary or associated company may be involved, or the organisation may have only a peripheral connection to the reported conduct. Name detection therefore provides only limited evidence about the substantive relationship between the customer and the reported event. This reinforces the distinction between identifying a potential match and establishing the identity and circumstances underlying that match, a problem that remains central to effective due diligence (Demetriades, 2016).

Consequently, effective adverse-media screening requires more than textual matching. The system must first establish whether the relevant individual or organisation has been correctly identified and, where possible, determine the semantic and factual relationship between that entity and the reported event. This is considerably more demanding than detecting a name or classifying an article as potentially relevant. Natural-language-processing techniques can assist with entity recognition, topic classification and information retrieval, but these capabilities should not be treated as equivalent to determining whether an allegation is accurate, whether the reported conduct is attributable to the customer, or whether the matter remains relevant at the time of review. The distinction is important because a technically accurate classification can still produce an inappropriate compliance signal if the underlying entity or context has been misunderstood.

A third challenge concerns language, jurisdiction and source heterogeneity. Adverse-media environments span multiple languages, legal systems, journalistic practices and naming conventions. LSEG itself identifies the scale and diversity of adverse-media information, matching limitations and language barriers as challenges associated with adverse-media screening (LSEG, 2026d). The use of multiple languages and large numbers of sources can increase the breadth of information available to a screening system, but broader coverage also introduces greater variation in terminology, source quality and evidential standards. The technical ability to process more information should therefore be distinguished from the ability to interpret that information consistently across different linguistic and institutional contexts.

This produces an important paradox: more information can increase observability without necessarily increasing certainty. Expanding source coverage may increase the probability that potentially relevant information will be discovered, but it can also increase the volume of duplicated, contradictory, outdated or ambiguous material requiring human assessment. Scalability therefore addresses the problem of information retrieval more directly than the problem of evidential evaluation. A system may identify substantially more potentially relevant content without becoming proportionately better at determining which information is reliable, attributable and materially significant. This is consistent with the wider literature on AI-based financial-crime detection, which recognises the potential of automated techniques while emphasising the need for appropriate evaluation of their performance and limitations (Mousavian and Miah, 2025).

Recency creates a further methodological complication. An adverse-media report may concern an event that occurred many years previously, may subsequently have been corrected, or may relate to proceedings that have been dismissed, resolved or superseded. Consequently, the existence of a media reference does not establish its continuing relevance. Effective adverse-media assessment requires consideration of source credibility, publication date, underlying event date, subsequent developments and the current status of the reported matter. Without this temporal and contextual assessment, automated screening risks treating historical or superseded information as if it represented the customer's current risk profile.

The role of AI and NLP should therefore be understood primarily as one of information discovery, classification and prioritisation. These technologies can help compliance teams process large volumes of unstructured information, identify potentially relevant material and reduce the amount of content requiring manual review. However, they do not eliminate the need to determine whether the information concerns the correct subject, whether the source is sufficiently credible, what relationship the subject has to the reported event, whether the information remains current and what evidential significance should be attached to it. The value of automation consequently lies in improving the organisation and prioritisation of evidence rather than automatically converting media content into a substantive finding of risk.

World-Check One's adverse-media proposition is therefore technically plausible as a scalable information-screening capability, particularly given the broader application of AI and NLP to financial-crime detection (Mousavian and Miah, 2025). Nevertheless, its effectiveness cannot be inferred simply from the number of sources covered, languages processed or articles identified. A meaningful evaluation would need to examine entity-resolution accuracy, relevance-classification performance, source reliability, false-positive and false-negative rates, language coverage, temporal validity and the treatment of allegations in comparison with confirmed findings. It would also need to establish how ambiguous, contradictory or subsequently superseded information is handled.

The appropriate evidential sequence is therefore not media reference → adverse finding, but media reference → source and entity verification → contextual assessment → risk evaluation. Adverse-media screening is best understood as an information-discovery and risk-identification mechanism rather than a mechanism for establishing misconduct. Human review remains important because the most consequential questions—whether the information is credible, whether it concerns the correct subject, what actually occurred and what significance it has for the customer—extend beyond the identification or classification of relevant text. In this respect, automated adverse-media screening can increase the scale and speed of information processing without removing the underlying evidential and interpretive problems of due diligence.

Takeaway: Adverse-media AI can find and prioritise potentially relevant information at scale, but identifying a negative story is not the same as establishing that the information is true, attributable, current or materially relevant to the customer.

8. AI-Powered Relevance Filtering and the Limits of Automated Interpretation

LSEG describes World-Check One as incorporating AI-powered relevance filtering for adverse media, while its broader adverse-media proposition refers to the use of artificial intelligence, machine learning, natural language processing (NLP) and intelligent tagging to identify and prioritise potentially relevant information (LSEG, 2026a; LSEG, 2026d). The underlying proposition is that automated analytical techniques can reduce the volume of information requiring manual review by distinguishing material content from information that is less likely to be relevant to a particular customer or risk category.

The academic literature provides support for the general use of AI in financial-crime detection. Mousavian and Miah (2025), in their review of AI-based applications for money-laundering detection, identify growing use of supervised and unsupervised machine learning, deep learning and related analytical techniques. Their review nevertheless emphasises the importance of evaluating AI systems before applying them to large-scale financial-crime detection. This qualification is significant because technical capability and operational reliability are not equivalent propositions. Demonstrating that an algorithm can classify, rank or detect patterns in data does not establish that its outputs will remain accurate when deployed across different populations, jurisdictions, languages and financial-crime typologies.

The distinction between information discovery and substantive risk assessment is particularly important for adverse-media relevance filtering. An AI system may be capable of determining that a document is statistically or semantically associated with a customer's name, a particular risk category or a financial-crime topic. This is fundamentally different from determining whether the information represents a genuine and material compliance risk. Relevance classification should therefore be understood as one stage within a broader analytical process rather than as a risk determination in itself. Subsequent interpretation requires consideration of the identity of the subject, the nature of the reported conduct, the reliability of the source, the status of any allegation or investigation, the recency of the information and its significance within the particular customer relationship.

This establishes an important boundary between classification and interpretation. Machine-learning systems identify patterns and statistical relationships within data; they do not independently establish the factual or legal significance of those relationships. Their outputs are consequently dependent on the characteristics of the data used for training and evaluation. Where training data are incomplete, biased, poorly labelled or insufficiently representative of the operational environment, performance may deteriorate when the system encounters cases that differ materially from those represented during development. The existence of sophisticated NLP or machine-learning techniques therefore does not eliminate the underlying dependence of automated classification on data quality and problem definition (Mousavian and Miah, 2025).

The quality of the labels and definitions used during model development is particularly important. A system designed to identify “adverse media” requires an operational definition of what constitutes adverse information. That category may encompass allegations, investigations, arrests, regulatory proceedings, enforcement actions and criminal convictions, despite the substantial differences in their evidential status. If these categories are insufficiently distinguished, a model may perform accurately according to its technical classification objective while still producing outputs that are difficult to interpret from a compliance perspective. A model can be correct according to its label while the label itself remains an imperfect representation of the underlying risk question. The issue is therefore not simply whether the model predicts its assigned category accurately, but whether the category being predicted corresponds sufficiently well to the compliance question that investigators actually need to answer.

The same methodological issue applies to KYC risk assessment more broadly. Chen (2020) demonstrates the potential for machine-learning methods to classify bank risk using KYC-related information, providing evidence that data-driven approaches can support financial-risk assessment. However, the findings arise from a particular institutional dataset and analytical context. They therefore support the technical feasibility of machine-learning-based risk assessment rather than establishing that the same variables, model architecture or performance characteristics will generalise automatically across different institutions, jurisdictions, customer populations or regulatory environments. The distinction between demonstrating that a technique can work in a defined setting and demonstrating that it remains reliable across materially different settings is central to evaluating commercial AI claims.

This problem of generalisability is particularly important for proprietary commercial systems. A vendor may operate models across large and diverse datasets, but external users may have limited visibility into the training population, feature engineering, classification thresholds, ground-truth construction or independent test data. Without sufficient methodological transparency, it becomes difficult to determine whether reported performance reflects robust generalisation across operational conditions or strong performance within a particular proprietary evaluation environment. The scale of a dataset does not, by itself, establish representativeness, and a large volume of training data cannot compensate automatically for systematic weaknesses in labels, source coverage or population composition.

There is also a critical distinction between ranking relevance and making a compliance decision. An AI system may appropriately prioritise one article above another for investigator review without being capable of determining whether the customer actually engaged in misconduct. Similarly, a document may receive a high relevance score because it contains strong semantic associations with a customer or risk category, while the underlying information may ultimately prove unreliable, outdated, incorrectly attributed or unrelated to the customer's actual conduct. Conversely, an article with weaker textual similarity may contain information that is substantively important. Ranking therefore determines the order in which information receives attention; it does not necessarily determine its ultimate evidential significance.

For this reason, the appropriate role of AI within World-Check One is better understood as analytical assistance rather than autonomous determination. AI can potentially increase the scale and consistency with which information is searched, classified and prioritised, allowing investigators to focus attention on material that is more likely to warrant review. This is consistent with the broader literature, which identifies AI and machine learning as potentially valuable tools for financial-crime detection while emphasising the importance of evaluating their performance and limitations before treating automated outputs as reliable indicators of risk (Mousavian and Miah, 2025). The technology can therefore strengthen the information-processing stages of compliance without eliminating the evidential and interpretive stages that follow.

A rigorous evaluation of AI-powered relevance filtering would consequently require more than evidence that the technology can classify or rank adverse-media information. It should examine precision and recall, false-positive and false-negative rates, classification definitions, ground-truth methodology, language and jurisdictional coverage, temporal stability and performance across different customer and media populations. Where the system is used primarily for prioritisation rather than final decision-making, evaluation should also consider whether the ranking materially improves investigator outcomes. In particular, it should establish whether increased filtering efficiency is achieved without systematically suppressing information that investigators would otherwise consider relevant.

This distinction is important because improvements in one stage of the compliance process do not necessarily translate into equivalent improvements in the final outcome. A reduction in the quantity of material presented to investigators may represent improved efficiency if irrelevant information is successfully removed. However, the same reduction could represent a loss of detection capability if relevant but less easily classified information is systematically excluded. Consequently, the performance of a relevance-filtering system should not be assessed solely by how much material it removes or how accurately it reproduces its predefined labels. Its effectiveness must also be considered in relation to the quality and completeness of the information that ultimately reaches human investigators.

LSEG's use of AI is therefore consistent with an established direction of academic research, and there is a credible technological basis for applying machine learning and NLP to large-scale financial-crime information processing (Chen, 2020; Mousavian and Miah, 2025). The more demanding proposition—that AI-powered relevance filtering produces reliable and materially improved compliance decisions—requires separate empirical validation. The ability to identify or rank information is not equivalent to the ability to interpret that information correctly, and neither should be treated as evidence that the resulting compliance decision is inherently correct.

The appropriate evidential chain is therefore data → automated classification or prioritisation → human verification → contextual interpretation → compliance decision. AI may strengthen the first two stages considerably, particularly where information volumes exceed the practical capacity for manual review. However, the reliability of the final decision remains dependent on the quality of the underlying evidence, the accuracy of entity attribution, the validity of the classification framework and the judgement applied during subsequent investigation. World-Check One should therefore be evaluated not only according to how effectively its AI can identify relevant information, but according to whether that information ultimately supports more accurate, better-evidenced and appropriately contextualised compliance decisions.

Takeaway: AI can become highly effective at deciding what deserves attention without necessarily becoming capable of deciding what the information actually means.

9. Data Quality, Provenance and the World-Check Database

The effectiveness of World-Check One ultimately depends on the quality of the information against which customers are screened. LSEG describes World-Check as a structured, aggregated and de-duplicated database containing information derived from sanctions and watchlists, government and regulatory records, law-enforcement sources and media. LSEG also states that specialist researchers apply defined research processes and quality controls in compiling and maintaining the information (LSEG, 2026b; LSEG, 2026c). These features are important because automated screening cannot compensate for fundamentally inaccurate, incomplete, outdated or incorrectly attributed source information. The quality of the screening output is necessarily constrained by the quality of the evidence entering the screening system.

The importance of data quality is consistent with the wider CDD literature. Research identifies beneficial ownership, complex naming structures, source-of-wealth verification and the reliability of customer information as persistent challenges within practical due-diligence processes (ElYacoubi, 2020). These difficulties are particularly relevant to large-scale screening databases because information originating from different jurisdictions and source types may vary substantially in structure, completeness, terminology and reliability. Aggregating such information can make it more accessible, but aggregation does not necessarily resolve the underlying uncertainties associated with the original sources.

This creates an important distinction between data aggregation and evidential validation. Bringing information from multiple sources into a structured database may improve accessibility and consistency of screening. It does not, however, automatically establish that every underlying record is complete, current, correctly classified or accurately attributed. De-duplication can reduce unnecessary repetition, but it does not by itself resolve contradictions between sources. Similarly, categorising information under sanctions, PEP, adverse media or another risk heading can improve usability without necessarily establishing the evidential significance of the underlying information.

The issue becomes particularly important because World-Check is a proprietary database. Users may be able to examine individual records and assess whether particular information appears relevant to a customer, but they generally cannot independently reproduce the provider's complete process for source selection, data collection, inclusion, categorisation, verification, updating and removal. This creates an external-verification problem. A commercial user can evaluate the output of the database without necessarily being able to independently reconstruct the processes that generated that output.

Data provenance is therefore central to the reliability of a screening result. Ideally, users should be able to establish where information originated, when it was obtained or last updated, what type of source produced it, how it was categorised and whether subsequent information has altered its significance. Provenance is especially important for adverse media. An allegation reported by a media organisation may later be withdrawn, contradicted, dismissed by a court or superseded by a subsequent regulatory or judicial outcome. Without sufficient temporal and source context, an apparently current adverse-risk record may preserve an interpretation that no longer accurately reflects the underlying facts.

This creates a further distinction between information availability and evidential currency. A record can remain technically available within a database while becoming less representative of the customer's current circumstances. Similarly, a record can be accurate as a historical description while becoming misleading if its subsequent development is not incorporated. Effective screening therefore requires not only the discovery of relevant information but also appropriate mechanisms for updating, contextualising and, where necessary, correcting or removing information.

The problem is not unique to commercial databases. Any large-scale KYC system must address incomplete information, inconsistent source quality, changing customer circumstances and differences between jurisdictions. What is distinctive in the case of a proprietary database is the limited ability of external users to independently examine the complete evidential pipeline. This makes transparency concerning source provenance, update practices, inclusion criteria, categorisation and correction mechanisms particularly important when the resulting records may influence consequential compliance decisions.

LSEG's emphasis on auditability provides an important governance mechanism. World-Check One records user actions through a date-stamped audit trail, allowing screening activity and subsequent handling to be documented and reviewed (LSEG, 2026a). This can strengthen accountability by establishing what actions were taken, when they occurred and, depending on the workflow, how a particular screening case was handled.

However, auditability of the workflow is not the same as transparency of the underlying data-generation process. An audit trail may demonstrate that an investigator received an alert, reviewed a record and took a particular action. It does not necessarily explain why the underlying record was originally included, which sources were considered, how conflicting information was resolved or whether the information remained sufficiently reliable and current at the time of the decision. In other words, an auditable process can document the handling of evidence without independently validating the evidence itself.

This distinction is particularly important when screening outputs are relied upon in decisions with material consequences for customers. If a customer is rejected, subjected to enhanced due diligence or prevented from completing a transaction because of a database record, the organisation needs not only to demonstrate that its internal process was followed but also to understand the evidential basis for the underlying alert. A well-documented decision is not necessarily a well-substantiated decision.

A mature compliance framework therefore requires both dimensions of accountability. The first is an auditable record of workflow: what was screened, what alert was generated, who reviewed it, what action was taken and when. The second is sufficient evidential transparency to establish where the relevant information came from, how it was processed, how current it is and why it supports the resulting risk signal. The two functions are complementary rather than interchangeable.

The reliability of World-Check One should consequently be assessed not simply by the size or structure of its database, but by the integrity of the evidential chain connecting source information → data processing → entity attribution → risk categorisation → screening alert → human assessment. LSEG's aggregation, structuring and quality-control processes may strengthen that chain, but the public existence of those processes does not by itself establish their accuracy or completeness. Independent evaluation would require greater visibility into data provenance, error correction, source coverage, update latency and the treatment of contradictory or subsequently superseded information.

The central point is therefore that data quality is not a background feature of KYC technology; it is part of the technology's evidential performance. Sophisticated matching, AI filtering and automated monitoring can only be as reliable as the information to which those technologies are applied. World-Check One may provide substantial value by structuring and making accessible large volumes of risk information, but the reliability of any individual screening outcome ultimately depends on the quality, provenance, currency and contextual integrity of the underlying evidence.

10. Due Diligence, Automation and Human Judgement

LSEG’s broader World-Check proposition is explicitly positioned as decision-support infrastructure rather than simply a repository of compliance data. World-Check One incorporates case management, workflow routing, risk tagging, status tracking and collaboration functions, while LSEG describes its Screening Resolution Service as providing expert analysis to support the review and resolution of potential matches. This architecture is consistent with a broader risk-based approach to AML and CDD, in which technology can assist organisations in processing large volumes of information while leaving contextual judgement to appropriately qualified personnel (De Wit, 2007).

The strongest role for automation in KYC is therefore not necessarily the replacement of compliance professionals, but the augmentation of their investigative capacity. AI and machine-learning techniques can assist with information retrieval, entity matching, classification and prioritisation, potentially reducing the amount of routine screening work that investigators must perform. Academic research supports the use of AI-based methods for financial-crime detection, particularly where large and complex datasets would be difficult to process manually (Mousavian and Miah, 2025). However, technical efficiency should not be confused with the ability to make a complete compliance judgement.

This distinction is particularly important because customer due diligence is inherently contextual. Determining whether a customer represents a material AML or sanctions risk may require consideration of beneficial ownership, source of funds or wealth, business purpose, jurisdictional exposure, corporate relationships and, where available, transaction behaviour. A screening system may identify relevant information about these dimensions, but a match or risk indicator does not necessarily establish how those factors interact or what level of risk they collectively represent. As the earlier analysis has shown, identification, screening and risk determination are distinct analytical tasks.

The appropriate conceptual hierarchy is therefore:

data → screening signal → contextual investigation → risk assessment → compliance decision.

This sequence preserves an important distinction between detecting potentially relevant information and determining its significance. A screening alert identifies something that may require attention; it does not, by itself, establish that the customer presents a particular level of risk or that a particular compliance action is justified.

The principal governance risk arises when this sequence is compressed into:

data → automated alert → assumed risk.

Such compression can create an illusion of objectivity because the resulting decision appears to originate from a technological system rather than from human judgement. Yet automation does not remove the underlying interpretive problem. It can instead relocate that problem into model design, data selection, matching thresholds, classification rules and the way investigators interpret system-generated outputs. A highly automated workflow may therefore become efficient at producing and processing alerts without necessarily improving the substantive quality of the final risk determination.

World-Check One’s workflow and case-management functionality can support the former, more defensible model by providing mechanisms through which alerts can be reviewed, investigated, documented and resolved. Nevertheless, the existence of these controls does not guarantee that organisations will use them appropriately. Their effectiveness ultimately depends on how institutions configure the system, define escalation criteria, interpret screening results and exercise human judgement. Automation can structure and accelerate due diligence, but it cannot by itself establish that the resulting compliance decision is substantively correct.

Takeaway: The value of automation in KYC lies not in replacing judgement, but in ensuring that human judgement is applied to better-organised and more relevant evidence.

11. Evaluating LSEG’s Claims of Accuracy and Fewer False Positives

LSEG repeatedly uses language concerning accuracy, reduced false positives and improved operational outcomes. The World-Check One proposition, for example, refers to “fewer false positives”, while its broader screening proposition emphasises targeted screening and filtering intended to reduce unnecessary alerts. These claims are commercially significant because false-positive reduction is closely associated with the operational burden of sanctions and KYC screening. However, they should be interpreted cautiously because LSEG’s public materials do not provide sufficient independent methodological information to establish the magnitude, consistency or conditions of the claimed improvement.

A meaningful performance evaluation would require substantially more information than a headline measure of accuracy or alert reduction. At minimum, this would include the evaluation population, the prevalence of genuine matches within that population, the definition of a true match, the definition of a false positive and false negative, the matching threshold, the treatment of ambiguous cases and an independent benchmark against which performance is measured. The temporal and geographical scope of the evaluation would also be relevant because matching performance may vary across names, languages, jurisdictions and types of customer data. Without such information, “fewer false positives” is better understood as a claimed direction of improvement than as a fully verifiable measure of screening performance.

This distinction is particularly important because screening performance involves a trade-off between different forms of error. Kim and Yang (2024) demonstrate this problem in the context of financial sanctions screening: increasing sensitivity can improve the identification of genuine matches and reduce false negatives while simultaneously increasing the number of false positives. Consequently, improving one performance measure does not necessarily improve the overall quality of the screening system. The choice of matching threshold can materially affect what the system detects and what it excludes.

The same principle applies to World-Check One. A screening system that generates fewer alerts may reduce analyst workload and improve operational efficiency, but this is beneficial only if the reduction does not result from systematically failing to identify genuine matches. Conversely, a system configured to maximise sensitivity may identify a greater proportion of potential matches but produce a volume of false-positive alerts that creates substantial investigative costs. The relevant question is therefore not simply whether the number of alerts falls, but why it falls and what happens to the cases that are no longer being alerted.

The appropriate evaluation consequently requires a multi-dimensional performance framework. This should include precision, recall, sensitivity and specificity, together with false-positive and false-negative rates. These technical measures should be assessed alongside operational indicators such as analyst workload, alert volumes, investigation time and alert-resolution time. Crucially, the evaluation should also establish the population on which these measures were calculated and the method used to determine the ground truth. A system can appear highly accurate in aggregate while performing materially differently for particular classes of names, jurisdictions, languages or customer populations.

The methodological issue can therefore be expressed as a distinction between efficiency and effectiveness. Fewer alerts may demonstrate greater efficiency, but they do not by themselves demonstrate greater screening effectiveness. To substantiate the claim that World-Check One improves screening accuracy, LSEG would need to demonstrate that reductions in false positives occur without a material deterioration in the detection of genuine matches, using transparent and independently reproducible evaluation criteria.

Takeaway: “Fewer false positives” is not sufficient evidence of better screening; the critical question is whether fewer unnecessary alerts are achieved without allowing genuine matches to disappear from the system.

12. Sanctions Screening: Detection Is Not Confirmation

A further conceptual issue is the distinction between detecting a potential sanctions match and establishing actual sanctions exposure. A screening system generates an alert when information associated with a customer sufficiently resembles information contained in a sanctions record. The alert therefore represents a potential relationship requiring investigation; it does not, by itself, establish that the customer is the sanctioned individual or entity.

This distinction is particularly important where names are common, transliterated across languages or recorded with incomplete or inconsistent identifying information. In such cases, similarity between the customer record and a sanctions-list entry may be sufficient to generate an alert without providing sufficient evidence to establish that the two records refer to the same subject. The screening process must therefore move beyond name similarity and consider additional identifiers and contextual information capable of confirming or rejecting the proposed relationship.

Kim and Yang (2024) demonstrate why this distinction matters. Their analysis shows that automated sanctions matching can improve sensitivity while also producing substantially more false positives. Greater sensitivity can therefore increase the number of potential matches detected without necessarily increasing the proportion of alerts that represent genuine sanctions exposure. This supports understanding automated screening primarily as a candidate-generation and prioritisation mechanism, rather than as a system that independently confirms sanctions exposure.

World-Check One’s use of secondary identifiers, configurable filtering and case-management functionality is designed to support the subsequent resolution of these candidate matches. These capabilities can help investigators organise relevant information, distinguish potential matches from non-matches and document the resulting decision. Nevertheless, the effectiveness of this process remains dependent on the quality and completeness of the information available to resolve each alert. Where identifying information is limited, ambiguous or contradictory, technological matching alone cannot remove the underlying uncertainty.

This creates an important boundary around the value of automated sanctions screening. Sophisticated matching can reduce the volume of irrelevant alerts, improve the prioritisation of potential matches and make investigations more manageable. It does not, however, eliminate the evidential step required to determine whether the customer is actually the listed subject. The final compliance decision therefore remains dependent on contextual investigation and appropriate human judgement.

The distinction can be expressed simply as:

screening alert → candidate match → identity resolution → contextual assessment → sanctions determination.

Collapsing these stages into screening alert → confirmed sanctions exposure would risk treating a probabilistic or similarity-based detection mechanism as though it were an evidential determination.

Takeaway: A sanctions-screening system can identify who might require investigation, but an alert is not confirmation that the customer is the sanctioned subject.

13. World-Check One and Predictive Risk Intelligence

The broader World-Check proposition increasingly extends beyond static list screening towards continuous risk intelligence. LSEG positions its solutions around ongoing monitoring, real-time or near-real-time screening, adverse-media monitoring and broader visibility of changes in customer risk. This represents an important development in KYC technology because it moves away from treating due diligence as a purely periodic exercise and towards recognising that customer risk can change over time.

Traditional KYC can be understood as a process in which customer information is collected, verified and reviewed at defined points or intervals. Continuous monitoring instead seeks to identify material developments as they occur. This can improve the timeliness of compliance information, particularly where sanctions designations, regulatory actions or relevant media developments emerge between scheduled reviews. However, continuous access to new information should not be confused with continuous predictive accuracy.

The distinction is important because detecting a new event is fundamentally different from predicting a future outcome. A system may identify a newly published sanctions designation or adverse-media article shortly after it becomes available without predicting that the customer will subsequently become a higher-risk subject. Similarly, a risk score may identify statistical associations between particular characteristics and historical cases without establishing the causal mechanism underlying those associations or demonstrating that the same relationship will hold for a particular customer.

This creates an important conceptual boundary around the term predictive risk intelligence. Prediction implies an inference about a future or otherwise unobserved state, whereas monitoring primarily concerns the detection of information that has already become available. The latter can provide an earlier basis for human investigation without necessarily possessing predictive knowledge. A system that detects a regulatory event immediately may therefore be highly responsive while remaining fundamentally reactive rather than predictive.

Research on machine-learning-based AML detection reinforces this distinction. Mousavian and Miah (2025) identify substantial potential for AI-based approaches in financial-crime detection while also emphasising the importance of empirical evaluation and awareness of methodological limitations. The existence of predictive or classification capability within a model does not, by itself, demonstrate that its outputs provide reliable forward-looking assessments of individual customer risk. Such a claim would require evidence concerning the prediction target, training and test populations, performance measures, false-positive and false-negative consequences, temporal stability and generalisability to new cases.

The appropriate interpretation of World-Check One’s automated risk intelligence is therefore more limited and more defensible: early identification, monitoring and prioritisation of potentially relevant information. These capabilities can help compliance teams recognise changes in a customer’s risk environment and investigate them more promptly. They should not automatically be interpreted as autonomous prediction of future wrongdoing or future sanctions exposure.

The distinction can be expressed as:

monitoring → detection of new information → prioritisation → human assessment

rather than:

monitoring → prediction of future wrongdoing → automated risk determination.

This distinction matters because the first sequence describes a technologically plausible decision-support function, whereas the second makes a substantially stronger claim about predictive validity. Without independent evidence demonstrating that the system can reliably forecast future risk beyond detecting already observable events, the more appropriate characterisation is continuous risk monitoring and early-warning intelligence, rather than autonomous predictive risk assessment.

Takeaway: Continuous monitoring can make compliance more responsive to emerging information, but detecting risk earlier is not the same as predicting what risk will occur next.

14. The Evidential Boundary: Risk Signal Versus Proof of Wrongdoing

The most important conceptual issue in evaluating World-Check One is the distinction between screening, risk assessment and proof of misconduct. These are related but substantively different functions. A sanctions hit is a screening result indicating a potential relationship with a listed subject. A PEP designation is a risk characteristic that may require enhanced scrutiny. An adverse-media result identifies information that may be relevant to the customer. None of these outcomes, considered in isolation, establishes that a customer has committed financial crime or is legally liable for misconduct.

The same principle is particularly important in relation to adverse media. LSEG describes its adverse-media data as encompassing individuals who have been accused, questioned, investigated, arrested, charged or convicted. These categories have materially different evidentiary significance. An allegation is not equivalent to a charge, a charge is not equivalent to a conviction, and an investigation does not necessarily establish that misconduct occurred. Treating these categories as interchangeable would risk converting the existence of negative information into an implicit conclusion about the underlying conduct.

An automated system may nevertheless group different forms of information within a common risk taxonomy in order to improve searchability, filtering and prioritisation. This can provide operational value, but it does not resolve the evidential differences between the underlying records. The compliance analyst must still establish what the source actually reports, who is implicated, the status of the matter, whether the information remains current and whether it is sufficiently relevant to the customer being assessed.

This distinction is fundamental to responsible KYC because the primary purpose of screening is generally to identify cases requiring further assessment, rather than to substitute automated classification for a legal or factual determination. A screening technology can identify potentially relevant evidence, organise it and bring it to an investigator’s attention. It cannot, merely by generating an alert or assigning a risk category, establish the truth of an allegation or the legal significance of the underlying conduct.

World-Check One should therefore be understood primarily as an evidence-discovery, screening and prioritisation technology. Its value lies in helping organisations identify and organise potentially relevant information at scale. The resulting outputs should remain subject to contextual assessment and should not be treated as conclusive determinations of criminality, sanctions liability or reputational culpability.

The critical evidential sequence is therefore:

source information → screening signal → identity and contextual assessment → evidential evaluation → compliance decision.

The further an automated output moves beyond this sequence towards an asserted conclusion about misconduct, the greater the evidential justification required.

Takeaway: A risk signal identifies information that may matter; it does not establish that wrongdoing occurred.

15. Governance, Auditability and Explainability

LSEG places significant emphasis on auditability through date-stamped records and describes World-Check One as supporting transparent and auditable compliance processes. This capability is important because KYC decisions may subsequently need to be reconstructed during internal reviews, regulatory examinations, audits or challenges concerning how a particular customer was assessed.

However, auditability is not a single property. A meaningful audit trail should allow an organisation to reconstruct several distinct stages of the screening and decision-making process. First, it should establish what was screened. Second, it should identify what information produced the alert. Third, it should show how the alert was investigated and resolved. Fourth, it should establish who made the final decision, when it was made and on what evidential basis.

World-Check One’s logging and case-management functionality can contribute substantially to the first and fourth dimensions by recording screening activity, workflow and user actions. However, these capabilities do not necessarily provide equivalent transparency concerning the second and third dimensions. Understanding why an alert was generated and why it was ultimately resolved in a particular way requires sufficient information concerning data provenance, matching logic, filtering criteria, relevant source material and the reasoning applied during human review.

This issue becomes particularly significant in AI-assisted adverse-media filtering. If an automated system classifies an article as irrelevant, or prioritises one article over another, the compliance organisation should have sufficient contextual information to understand the basis of that classification and, where necessary, challenge it. An unexplained relevance score may indicate what the system decided without adequately explaining why the underlying information was considered irrelevant or significant.

Explainability should therefore not be reduced to the availability of an algorithmic score or a simple explanation of model output. In high-stakes compliance environments, meaningful explainability concerns the evidential chain connecting source information, entity matching, automated classification, human investigation and final decision-making.

This also reinforces the distinction between workflow transparency and evidential transparency. A system may provide an excellent record of who reviewed an alert, when it was reviewed and what decision was recorded while providing limited visibility into how the underlying data entered the system or why particular information was classified in a particular way. Auditability can therefore demonstrate that a process occurred without necessarily establishing that the evidence used in that process was complete, reliable or correctly interpreted.

A robust governance framework should consequently seek to establish not only what the organisation did, but also what evidence it relied upon and how that evidence was transformed into a compliance decision. In this sense, auditability and explainability are complementary: auditability reconstructs the process, while explainability helps make the reasoning and evidential relationships within that process intelligible.

Takeaway: A traceable decision is not necessarily an explainable decision; meaningful governance requires visibility into both the workflow and the evidence behind it.

16. Overall Evaluation

The examination of World-Check One produces a qualified rather than unconditional assessment of its technological proposition. The underlying architecture is strongly compatible with established developments in financial-crime technology. Automated KYC, machine-learning-based risk assessment, sanctions name matching, adverse-media processing and ongoing monitoring are all consistent with the direction of contemporary AML and compliance technology. Research demonstrates the potential for machine-learning approaches to support financial-risk assessment using KYC information, while research on customer due diligence highlights the continuing operational difficulties associated with identity, beneficial ownership and other forms of customer information (ElYacoubi, 2020). Kim and Yang (2024) further demonstrate the potential for NLP-based techniques to improve sanctions-screening sensitivity while also creating important false-positive trade-offs.

The technological plausibility of these functions, however, should be distinguished from independent validation of LSEG’s specific commercial performance claims. LSEG’s public materials describe reduced false positives, accurate screening, AI-powered relevance filtering and improved operational efficiency. These claims may be technically plausible, but the publicly available material does not provide sufficient independent methodological evidence to establish the magnitude, consistency or generalisability of these improvements across representative customer populations.

This evidential limitation is particularly important because screening performance cannot be reduced to a single measure of accuracy. A meaningful assessment would require information concerning precision, recall, false-positive and false-negative rates, matching thresholds, evaluation populations and the method used to establish ground truth. Performance would also need to be considered across different languages, jurisdictions, name structures and customer populations. Without this information, claims concerning improved accuracy or fewer false positives remain difficult to evaluate independently.

The analysis also demonstrates that data quality is as important as algorithmic sophistication. A large or frequently updated database does not automatically constitute a more accurate source of risk intelligence. The reliability of an output depends on the quality, completeness, currency and provenance of the information entering the system, as well as the accuracy with which that information is attributed to the relevant individual or entity. A sophisticated matching algorithm cannot fully compensate for incomplete, ambiguous or unreliable source information.

A further limitation concerns the distinction between screening and decision-making. World-Check One can identify potentially relevant information and support investigators in organising and prioritising that information, but the significance of an alert remains context-dependent. A possible sanctions match is not necessarily a confirmed sanctions match; adverse media is not necessarily evidence of wrongdoing; and a risk score is not equivalent to a determination of legal liability. These distinctions place an important boundary around what automated screening technology can reasonably be expected to establish.

The strongest evaluation framework for World-Check One should therefore combine technical, evidential and operational measures. Technical evaluation should include precision, recall, sensitivity, specificity, false-positive and false-negative rates, and matching accuracy. Data evaluation should consider freshness, source reliability, provenance, geographic and linguistic coverage and the treatment of contradictory or superseded information. Operational evaluation should consider alert volumes, analyst workload, alert-resolution time and the effectiveness of human review. Finally, these measures should be subjected, where possible, to independent validation against representative datasets rather than relying solely on vendor-reported performance.

Overall, World-Check One is best understood as a decision-support and risk-intelligence infrastructure rather than an autonomous mechanism for determining customer risk. Its technological proposition is credible and consistent with the direction of academic research, but the strongest commercial claims require a higher level of independent empirical substantiation than is evident from the public materials examined. The central issue is therefore not whether World-Check One can process large quantities of compliance information—it clearly is designed to do so—but whether its outputs can be demonstrated to improve the accuracy, reliability and evidential quality of consequential compliance decisions.

Takeaway: World-Check One’s technological capabilities are credible, but technological sophistication should not be treated as independent evidence that its outputs are accurate, predictive or substantively correct.

17. Conclusion

This paper has critically evaluated LSEG World-Check One as a technology for KYC, sanctions screening and broader financial-crime risk intelligence. The analysis indicates that the platform’s underlying technological proposition is credible and consistent with established developments in compliance technology. Automated entity matching, machine-learning-assisted classification, adverse-media processing, ongoing monitoring and workflow automation are all technically plausible applications of contemporary data-driven methods and are supported to varying degrees by the academic literature.

However, technological plausibility should not be confused with independent evidence of commercial performance. LSEG presents World-Check One as capable of improving screening accuracy, reducing false positives, supporting continuous monitoring and using AI to identify relevant information more efficiently. These claims may be plausible, but the public materials examined do not provide sufficient methodological detail to establish the magnitude, consistency or generalisability of the claimed improvements. In particular, meaningful evaluation would require transparent information concerning test populations, definitions of true and false matches, matching thresholds, precision, recall, false-positive and false-negative rates, ground truth and independent benchmarking.

The analysis also demonstrates that the effectiveness of screening technology depends on more than algorithmic sophistication. Data quality, completeness, currency, provenance and accurate entity attribution are fundamental to the reliability of the resulting intelligence. A large database does not necessarily constitute a more accurate database, while frequent data updates do not necessarily mean that material events will be detected immediately. Similarly, AI-powered relevance filtering can assist with identifying and prioritising information without establishing that the underlying information is accurate, relevant or evidentially significant.

A central finding is therefore the need to maintain clear boundaries between identification, screening, risk assessment and determination. A potential sanctions match is not automatically a confirmed sanctions match. A PEP designation is a risk characteristic rather than evidence of misconduct. Adverse media can identify allegations, investigations, charges or convictions, but these categories have materially different evidentiary significance. Likewise, continuous monitoring can provide earlier access to new information without demonstrating an ability to predict future wrongdoing. These distinctions limit the extent to which automated outputs can reasonably be treated as substantive compliance conclusions.

Human judgement consequently remains an essential component of the compliance process. The appropriate model is not necessarily the replacement of investigators with automated risk decisions, but the use of technology to organise, filter and prioritise large volumes of information so that human investigators can apply contextual and evidential judgement more effectively. World-Check One’s case-management and auditability functions can support this process, but they do not guarantee that organisations will interpret alerts correctly or rely on sufficiently reliable evidence. An auditable workflow can document what happened without necessarily establishing that the underlying evidence was complete, accurate or correctly interpreted.

The overall assessment is therefore qualified rather than unconditional. World-Check One represents a sophisticated form of compliance infrastructure with credible applications in screening, monitoring and information prioritisation. Its strongest value lies in improving the accessibility, organisation and timeliness of potentially relevant risk information. However, the stronger claims made for accuracy, false-positive reduction and predictive risk intelligence require a substantially stronger evidential basis than is available from the public materials examined.

Ultimately, the effectiveness of World-Check One should be judged not by the volume of information it processes or the sophistication of its AI capabilities, but by whether those capabilities produce demonstrably better compliance decisions. This requires evaluation across technical performance, data quality, evidential reliability, operational efficiency and human decision-making. The most defensible interpretation is therefore that World-Check One can augment due diligence and risk assessment, but it does not remove the need for contextual investigation, human judgement or independent validation of the evidence on which consequential compliance decisions depend.

References

Chen, T.-H. (2020), “Do you know your customer? Bank risk assessment based on machine learning”, Applied Soft Computing, 86, 105779.

De Wit, J. (2007) ‘A risk-based approach to AML: A controversy between financial institutions and regulators’, Journal of Financial Regulation and Compliance, 15(2), pp. 156–165. doi: 10.1108/13581980710744048.

ElYacoubi, Dina (2020), ‘Challenges in customer due diligence for banks in the UAE’ (2020), Journal of Money Laundering Control, 23(2), pp. 527–539. doi: 10.1108/JMLC-08-2019-0065.

Kim, S. and Yang, S. (2024) ‘Accuracy improvement in financial sanction screening: is natural language processing the solution?’, Frontiers in Artificial Intelligence, 7, 1374323. doi: 10.3389/frai.2024.1374323.

Mousavian, S. and Miah, S.J. (2025) ‘Review of artificial intelligence-based applications for money laundering detection’, Intelligent Systems with Applications.

Demetriades, G. (2016) ‘Is the person who he claims to be? Old fashion due diligence may give the correct answer!’, Journal of Money Laundering Control, 19(1), pp. 79–91. doi: 10.1108/JMLC-11-2014-0041.

LSEG (2026a) World-Check One: KYC verification and customer screening platform. LSEG Risk Intelligence.

LSEG (2026b) World-Check – KYC Screening. LSEG Risk Intelligence. 

LSEG (2026c) World-Check Risk Data. LSEG Risk Intelligence.

LSEG (2026d) Adverse Media Screening. LSEG Risk Intelligence.

LSEG (2026e) World-Check Verify: Real-Time API Screening. LSEG Risk Intelligence.

LSEG (2026f) AML & KYC Solutions for Law Firms. LSEG Risk Intelligence. 

LSEG (2026g) Real Estate KYC & Due Diligence Solutions. LSEG Risk Intelligence.

Contact

Reach out via email for inquiries.

Email

Subscribe to newsletter

info@grcadvisory.ch

© 2025. All rights reserved.