A Governance and Implementation Framework for Artificial Intelligence in Swiss External Asset Management

Abstract

Artificial intelligence (AI) is creating new opportunities to improve operational efficiency, information processing and client services in financial services. Yet, for Swiss external asset managers (EAMs), translating these opportunities into sustainable business value remains challenging. Limited resources, fragmented data, legacy systems, skills gaps and regulatory expectations can constrain adoption, particularly among smaller firms. This paper examines AI adoption readiness in the Swiss EAM sector, focusing on the organisational and governance capabilities required to move from experimentation to responsible implementation.

Drawing on the existing literature on AI in financial services and sector-specific empirical findings from Swiss EAMs, the paper examines the gap between perceived AI potential and organisational readiness. It considers the implications of this gap for use-case selection, data and technology infrastructure, workforce capabilities, risk management and regulatory compliance. Particular attention is given to the Swiss context, where AI initiatives must be aligned with applicable legal and supervisory requirements, proportionate governance arrangements and clear lines of accountability.

Building on this analysis, the paper proposes a conceptual readiness framework comprising five interrelated dimensions: business and process readiness; data and technology readiness; people and organisational readiness; governance and control readiness; and value and scale readiness. The framework is intended to support use-case-level assessment and help EAMs determine whether to proceed with implementation, conduct a controlled pilot, defer investment or reject a proposed application.

The paper argues that AI readiness is not a one-off technical assessment, but an ongoing organisational capability that must evolve as use cases, technologies and risks change. Its contribution is a context-specific, practice-oriented approach to AI adoption in Swiss EAM. As the underlying evidence is exploratory and the proposed framework has not yet been empirically validated, the framework should be understood as a structured basis for managerial decision-making and future research rather than a predictive maturity model. The findings suggest that sustainable AI adoption depends less on technological ambition alone than on the alignment of business objectives, organisational capabilities and effective governance.

Keywords: artificial intelligence; external asset management; wealth management; generative AI; agentic AI; AI governance; financial services; organisational readiness; regulatory compliance.

1. Introduction

Artificial intelligence (AI) is becoming increasingly significant in financial services, where organisations must process large volumes of structured and unstructured information, operate within complex regulatory environments, and make decisions based on data that may be incomplete, uncertain or continuously changing. Its applications span a broad range of activities, including fraud detection, risk assessment, customer service, document processing, investment research and financial advice. Reviews of the financial-services literature confirm both the breadth of these applications and the diversity of ways in which AI can support operational processes, analytical capabilities and customer-facing services (Fares, Butt and Lee, 2023; Pattnaik, Ray and Raman, 2024). However, the existence of technically feasible applications does not, in itself, establish that their implementation will generate sustainable business value. The benefits realised depend on how AI is integrated into organisational processes, decision-making structures and existing control environments.

These considerations are particularly relevant to Swiss external asset managers (EAMs). Their operating models combine relationship-driven investment advisory services with substantial administrative, analytical and compliance responsibilities. Client onboarding, know-your-customer (KYC) procedures, anti-money-laundering (AML) due diligence, investment research, investment proposal preparation and client reporting all require the processing and interpretation of information from multiple sources. These activities offer potential opportunities for AI to reduce manual effort, improve information retrieval and support more consistent analysis. AI-enabled financial advisory services may also reshape how investment information is synthesised and translated into client-specific insights (Zhu, Vigren and Söderberg, 2024). Nevertheless, the sensitivity of client data, the importance of suitability and professional judgement, and the regulatory obligations associated with financial services constrain how such applications can be designed and deployed.

The potential of AI must therefore be considered alongside the organisational capabilities required to use it responsibly. Research on AI in banking identifies opportunities for integration while highlighting challenges associated with implementation in a regulated industry (Moharrak and Mogaji, 2025). Similarly, research on financial-sector AI suggests that the relationship between technological adoption, internal business processes and external regulation is important to understanding how AI may contribute to organisational performance (Cheng et al., 2025). These perspectives shift attention away from technology selection alone and towards the interaction between business objectives, data quality, staff capabilities, process design and governance. In practice, an AI application may perform well in a controlled demonstration yet deliver limited value if its outputs cannot be reliably incorporated into existing workflows, appropriately reviewed or translated into measurable improvements.

The study by Agnesens, Stüttgen and Ivanov (2026) provides a sector-specific starting point for examining these issues in Swiss external asset management. Based on a structured survey of 41 EAMs and nine semi-structured interviews, the study explores current AI adoption, perceived use cases and organisational readiness. Its findings indicate substantial interest in AI alongside uneven levels of implementation preparedness. Respondents identify potential applications across operational and advisory activities, but also point to concerns relating to data protection, output reliability, regulatory uncertainty, internal expertise and the allocation of responsibilities. The study further suggests that data readiness remains a constraint: only a minority of respondents consider their data sufficiently structured for AI. These findings should be interpreted in light of the study’s exploratory design and limited sample, rather than treated as representative estimates for the entire Swiss EAM population. Nevertheless, they provide a useful indication of the practical tensions that may arise when smaller, relationship-oriented financial-services organisations seek to adopt AI.

The distinction between identifying promising use cases and developing the capacity to implement them is consequently central. Generative AI can assist with activities such as summarising research, drafting investment proposals and extracting information from documents. More advanced, workflow-oriented applications may coordinate tasks across systems or initiate actions based on information obtained from multiple sources. Yet the operational and governance demands of these applications differ. The greater the autonomy, system access or potential impact of an AI-enabled process, the more important it becomes to establish appropriate boundaries, human oversight, testing, monitoring and accountability. For EAMs, implementation decisions must therefore account not only for potential efficiency gains, but also for the consequences of inaccurate outputs, inappropriate disclosures, weak controls or unclear responsibility for decisions.

Against this background, the central research question addressed in this paper is: Which organisational and governance capabilities enable Swiss external asset managers to translate AI potential into measurable, reliable and sustainable business value?

This paper argues that the principal challenge is not simply to identify additional AI use cases, but to establish the conditions under which selected applications can be implemented responsibly, integrated into business processes and scaled on the basis of demonstrated value. Drawing on the exploratory findings of Agnesens, Stüttgen and Ivanov (2026) and recent peer-reviewed research on AI in financial services, banking and financial advisory, the paper examines the relationship between adoption, organisational readiness and governance. It then develops a practical implementation framework tailored to the characteristics of Swiss external asset management. The framework places business-process alignment, data and technology foundations, organisational capabilities, governance and control, and value measurement at the centre of implementation. In doing so, it treats responsible AI adoption not as a discrete technology project, but as an organisational capability that must be developed and maintained over time.

2. Conceptual background and literature review

2.1 AI adoption as an organisational capability

AI adoption is often approached as a technology-selection exercise in which an organisation identifies a suitable task, evaluates available solutions and deploys a tool. While technology selection is an important component of implementation, this perspective provides an incomplete account of how AI creates organisational value. Effective adoption also depends on the availability and quality of data, the suitability of existing processes, integration with operational systems, employee competence and clearly defined accountability. AI should therefore be understood not simply as a technological asset, but as a capability whose value depends on how it is incorporated into organisational structures and working practices.

Reviews of AI in financial services support this broader perspective. Fares, Butt and Lee (2023) examine AI adoption in banking across strategic, operational and customer-related dimensions, demonstrating that its implications extend beyond individual applications or technical functions. Similarly, Pattnaik, Ray and Raman (2024) identify a diverse research landscape encompassing risk management, anti-money-laundering activities and other specialised financial-services applications. Taken together, these reviews suggest that AI adoption should be examined in relation to the business activities it supports and the organisational arrangements required to implement it effectively.

Empirical research provides further support for the relationship between AI, organisational processes and performance. Cheng et al. (2025), examining Chinese financial companies over the period 2011–2022, associate AI adoption with improved financial performance and identify mechanisms involving risk management, internal governance and internal control. Their findings indicate that the contribution of AI cannot be separated from the organisational environment in which it operates. However, the study's empirical setting differs from Swiss external asset management, and its results should not be assumed to transfer directly to EAMs with different operating models, regulatory circumstances and resource constraints.

For EAMs, the implication is that organisational readiness must be assessed more broadly than the availability of AI tools, the presence of an innovation strategy or the willingness of employees to experiment. Readiness encompasses the capacity to identify worthwhile applications, provide reliable and appropriately governed data, integrate solutions into business processes, develop staff competence, manage operational and regulatory risks, and evaluate outcomes. These capabilities are interdependent: technically capable tools may deliver limited benefits when data are fragmented, responsibilities are unclear or outputs cannot be incorporated reliably into established workflows.

AI adoption can consequently be conceptualised as an organisational capability developed through the interaction of technology, people, processes and governance. This perspective provides the foundation for analysing not only which AI applications are feasible in EAMs, but also the conditions under which their implementation is likely to generate sustained value.

2.2 AI opportunities across the EAM value chain

Agnesens, Stüttgen and Ivanov (2026) examine AI opportunities across five broad areas of the external asset management value chain: client onboarding and due diligence; investment planning, research and reporting; portfolio construction and implementation; risk management and compliance; and finance, accounting and data. Their study identifies potential applications ranging from document analysis and information extraction to investment research synthesis, proposal preparation, client reporting, data-quality improvement and reconciliation.

The breadth of these applications reflects the information-intensive nature of EAM activities. Client-facing and investment-related work requires the interpretation of financial information, client documentation and market research, while operational and compliance functions depend on accurate records, consistent procedures and appropriate evidence. AI may support these activities by reducing manual information-processing effort, improving access to relevant information and assisting employees with analytical or documentation tasks. However, the value of a particular application depends on the business problem being addressed, the quality of the available information and the extent to which the resulting output can be used reliably.

The EAM study also highlights an important distinction between content-oriented tasks and applications that require coordination across processes and systems. Summarising investment research or preparing an initial draft of a client proposal can often be approached as a relatively contained activity. By contrast, automating onboarding across multiple systems may require the reconciliation of inconsistent records, the application of business rules, the management of exceptions and the coordination of actions among different functions. The latter therefore presents broader requirements for technical integration, process ownership and control.

This distinction is consistent with the wider literature on AI-enabled financial advisory services. Reviewing 132 articles, Zhu, Vigren and Söderberg (2024) examine how AI may influence the design and delivery of financial advice and identify a research agenda that connects technological developments with financial-service implementation. Their review supports the importance of considering AI in the context of adviser–client relationships and service delivery, rather than evaluating it solely in terms of technical performance.

For EAMs, this has strategic implications. The objective should not be to maximise the number of automated tasks or to introduce AI wherever a technical solution is available. Instead, applications should be selected according to their potential to improve defined business outcomes, such as the timeliness of client reporting, the consistency of due diligence, the quality of investment research or the efficiency of administrative processes. These improvements must also be assessed against the need to preserve professional judgement, appropriate client treatment and the relationship-based nature of external asset management.

A value-chain perspective therefore provides a practical basis for prioritisation. It connects AI opportunities to identifiable business activities while making visible the dependencies, controls and organisational changes that implementation may require. It also helps distinguish applications that can be tested within a limited workflow from those that depend on broader transformation across systems and functions.

2.3 Generative AI and agentic AI

Distinguishing between generative AI and agentic AI is important when assessing implementation complexity and risk. Generative AI produces or transforms content in response to instructions. In an EAM context, potential applications include summarising investment research, drafting client correspondence, preparing meeting notes and extracting information from documents. These applications can support employees without necessarily transferring operational authority to the AI system.

Agentic AI, by contrast, refers here to systems capable of pursuing a goal through multiple steps, using connected tools or systems and potentially initiating actions with limited intervention at each stage. Illustrative EAM applications include coordinating elements of client onboarding, retrieving and reconciling information across systems, or enabling natural-language interaction with portfolio information. Agnesens, Stüttgen and Ivanov (2026) distinguish between more immediately accessible content-oriented use cases and these more advanced workflow-based applications.

The difference has consequences for both implementation and governance. A generated research summary can generally be reviewed before it informs a decision or is communicated to a client. A system that operates across customer relationship management, document management and portfolio platforms may, however, affect several stages of a business process. Errors can propagate between steps, arise from incomplete or conflicting information, or result from inappropriate tool use. Where the system can initiate consequential actions, the potential impact extends beyond the quality of its generated content to the integrity of the process it controls.

Aldasoro et al. (2025) examine the implications of AI, including generative systems and emerging AI agents, across financial intermediation, insurance, asset management and payments. Their analysis considers both the opportunities arising from these technologies and their wider implications for financial institutions, markets and financial stability. This broader perspective reinforces the importance of assessing AI not only as an individual productivity tool but also in terms of its interaction with institutional processes and the financial system.

For EAMs, the distinction implies that implementation controls should be proportionate to the capabilities and authority of the application. An assistant that drafts an internal research summary does not present the same operational exposure as an agent that changes client records, coordinates compliance actions or initiates a transaction-related workflow. As autonomy and system access increase, stronger requirements may be needed for permission management, validation, human approval, exception handling, traceability and ongoing monitoring.

The distinction should not, however, be interpreted as a simple separation between low-risk generative AI and high-risk agentic AI. A content-generation tool can still create material risks if its outputs influence investment decisions, contain misleading information or expose confidential client data. Conversely, a narrowly scoped workflow agent operating under restrictive permissions may have a manageable risk profile. The relevant assessment must therefore consider the application’s intended purpose, data access, degree of autonomy, potential consequences and position within the wider business process.

2.4 Governance, confidentiality and trust

Governance is a central consideration in financial-services AI adoption because organisations must maintain confidentiality, operational reliability, regulatory compliance and effective control over business decisions. AI may introduce additional dependencies involving external technology providers, data-processing arrangements, model behaviour, system integration and the reliability of generated outputs. These dependencies do not displace existing responsibilities; they change the risks that organisations must identify, assess and manage.

The findings of Agnesens, Stüttgen and Ivanov (2026) illustrate these challenges in Swiss EAMs. Respondents identify data protection and confidentiality as prominent barriers to AI adoption, followed by concerns about reliability and accuracy and uncertainty regarding regulatory requirements. The study also highlights limitations in organisational readiness, including the perceived adequacy of data structures, internal expertise and clearly assigned responsibilities. These findings suggest that the barriers to adoption are not exclusively technological. They also concern the ability of organisations to establish and maintain the conditions necessary for reliable and accountable use.

Moharrak and Mogaji (2025), examining generative AI integration in banking, likewise identify challenges associated with implementation in a regulated industry, including data privacy, reliability and organisational preparedness. Read alongside the EAM study, their findings support the view that successful adoption requires governance arrangements to develop alongside technical capability. However, neither study should be interpreted as establishing a single governance model that will be appropriate for every financial institution. The nature and intensity of controls must reflect the specific application, organisational context and consequences of failure.

Governance should consequently be incorporated into use-case assessment and solution design rather than introduced only after a technology has been selected. Relevant considerations include which data may be processed, where those data are stored or transmitted, which external providers are involved, how access is controlled, who is accountable for outputs, what level of human review is required and what evidence must be retained to demonstrate that controls are operating effectively. Testing should address not only technical performance, but also the reliability of outputs under realistic conditions, the handling of exceptions and the consequences of incorrect or incomplete information.

Trust is closely connected to these arrangements. Employees must understand the capabilities and limitations of the systems they use, while managers need sufficient evidence to determine whether the technology is suitable for its intended purpose. Where AI supports client-related or investment-related activities, the organisation must also be able to establish how outputs were produced, reviewed and used, to the extent appropriate to the application. Such traceability helps support accountability and provides a basis for investigating errors, responding to incidents and improving system performance.

A proportional approach is therefore essential. An internal drafting assistant operating on approved, non-sensitive information may require a relatively limited set of controls. An application that processes confidential client records, contributes to suitability assessments or executes actions across operational systems requires a more extensive assessment of data protection, reliability, permissions, oversight and accountability. Importantly, the risk classification should follow an assessment of the actual use case, information processed and potential consequences, rather than relying on the product category or the label assigned to the technology.

Overall, the literature points towards a common conclusion: AI adoption in financial services is best understood as the development of an organisational capability that combines business-process alignment, appropriate data and technology, staff competence and risk-based governance. For Swiss EAMs, this perspective connects the identification of promising applications with the practical requirements for implementing them responsibly. It also establishes the conceptual basis for the implementation framework developed in the following chapters.

3. The empirical readiness gap in Swiss external asset management

3.1 Interest does not equate to implementation maturity

The study by Agnesens, Stüttgen and Ivanov (2026) provides an empirical illustration of the distinction between interest in artificial intelligence (AI) and organisational readiness to implement it systematically. Among the 41 Swiss external asset managers (EAMs) surveyed, 49 per cent reported using AI in selected processes, 39 per cent were exploring or planning initiatives, and 12 per cent had discussed AI without developing concrete plans. No respondent reported having neither plans nor interest in AI. These findings indicate broad engagement with the technology among the participating firms, although the exploratory nature and limited size of the sample mean that the results should not be interpreted as representative of the entire Swiss EAM sector.

The prevalence of experimentation does not, however, necessarily indicate that AI has become an established organisational capability. The study describes adoption as frequently fragmented or insufficiently integrated into business processes, with initiatives often driven by individual employees rather than a coordinated firm-wide strategy (Agnesens, Stüttgen and Ivanov, 2026). This distinction is consistent with the broader financial-services literature, which suggests that the organisational and operational context is important to the value realised from AI (Fares, Butt and Lee, 2023; Cheng et al., 2025).

An organisation may use several AI tools while lacking consistent arrangements for approved use, data handling, supplier assessment, risk management and performance measurement. In such circumstances, adoption can remain dependent on individual initiative, with limited visibility into the overall risk exposure or the benefits being achieved. Conversely, an organisation with fewer deployed applications may be better positioned for sustainable adoption if it has clearly defined processes, accountable owners and proportionate controls. The number of tools in use is therefore an incomplete measure of implementation maturity.

This distinction is particularly relevant to smaller EAMs, which may have limited technology budgets and restricted access to specialist expertise. Individual experimentation can help identify practical opportunities without requiring substantial initial investment. However, when experiments develop independently, they can also produce inconsistent working practices, duplicated effort and uncertainty about which tools are approved for particular activities. The challenge is to preserve the flexibility that makes experimentation valuable while introducing sufficient coordination to manage risks and scale successful applications.

A transition from experimentation to managed implementation requires an explicit operating model. This should establish how use cases are proposed and prioritised, how technology and data arrangements are assessed, who authorises deployment, and how performance and risk are monitored. Such an approach shifts the emphasis from the mere presence of AI towards the organisation's ability to use it consistently, demonstrate its value and maintain appropriate oversight.

3.2 Data readiness and process integration

Data readiness emerges as a significant constraint in the EAM study. Only 27 per cent of respondents considered their data sufficiently structured for AI (Agnesens, Stüttgen and Ivanov, 2026). This finding is important because EAM activities often involve information distributed across portfolio management systems, customer relationship management platforms, custodians, document repositories, spreadsheets and external data sources. The usefulness of AI depends not only on the capabilities of the model, but also on whether relevant information can be accessed, interpreted and used appropriately.

Poorly structured or inconsistent data can undermine an otherwise promising application. A system may successfully extract information from know-your-customer (KYC) documents, for example, while remaining unable to resolve discrepancies between records, establish the reliability of conflicting source information or determine whether missing evidence satisfies the firm's applicable procedures. These activities may require contextual knowledge, explicit business rules and professional judgement beyond the capabilities of the extraction tool itself. Automation can therefore reduce manual effort without eliminating the need for validation, exception handling or accountable decision-making.

Data readiness should consequently be assessed in relation to the intended application rather than treated as a single organisation-wide condition. Relevant considerations include data completeness, consistency, timeliness, provenance, accessibility and permitted use. For applications involving confidential client information, access restrictions and data-protection requirements must also be incorporated into the assessment. Where AI generates summaries, recommendations or extracted information, the organisation should consider whether outputs can be traced to the underlying sources and whether users can identify omissions, inconsistencies or unsupported conclusions.

These requirements connect data quality directly to governance. Reliable data are not simply a technical prerequisite; they influence whether an application can produce sufficiently dependable outputs for its intended business purpose. The broader literature on AI implementation in financial services reinforces the importance of considering these organisational and operational dependencies alongside technical capability (Moharrak and Mogaji, 2025; Cheng et al., 2025).

Process integration presents a related challenge. An application that operates within a clearly bounded task may require relatively limited technical changes, whereas an application spanning multiple systems may depend on interfaces, workflow redesign, business rules and arrangements for handling exceptions. Even when individual components perform as intended, failures can occur if information is transferred incorrectly, tasks are not assigned to the appropriate owner or exceptions remain unresolved. The EAM study's findings suggest that the perceived potential of AI does not automatically translate into implementation where these wider operational requirements remain unmet (Agnesens, Stüttgen and Ivanov, 2026).

The practical implication is that implementation should often begin with an analysis of the business process and its information requirements, rather than with software configuration. Organisations should establish what problem needs to be addressed, how the process currently operates, which information is required, where inconsistencies arise and which decisions demand human or specialist judgement. Only then can they determine whether AI is appropriate, what supporting data and integration capabilities are necessary, and which controls must be maintained. This sequence reduces the risk of automating an inadequately understood process or introducing technology that improves an isolated task without improving the end-to-end outcome.

3.3 Skills, accountability and the operating model

In addition to data readiness, the EAM study identifies internal expertise and clearly assigned responsibilities as significant organisational concerns. In each case, 41 per cent of respondents reported the issue as a readiness concern (Agnesens, Stüttgen and Ivanov, 2026). These findings highlight the human and organisational dimensions of AI implementation. Successful adoption requires coordination among business stakeholders, technology specialists, compliance professionals, risk owners and, where relevant, external technology providers.

The required expertise extends beyond the ability to operate an AI tool. Business users need to understand the system's intended purpose, limitations and appropriate use. Technology specialists must assess integration, access controls and operational dependencies. Compliance and risk professionals need sufficient understanding of the application to evaluate its implications, while business owners must determine whether its outputs are suitable for the decisions or activities it supports. These responsibilities need not always be performed by separate individuals, but the necessary capabilities must be available collectively.

This challenge can be particularly pronounced in smaller EAMs, where employees may hold several roles and specialist resources may be limited. A governance model designed for a large financial institution may impose disproportionate overhead if reproduced without adaptation. Nevertheless, limited organisational size does not remove the need for clear accountability, appropriate challenge or effective monitoring. The relevant objective is proportionality: governance arrangements should reflect the firm's scale and complexity without leaving material responsibilities undefined.

A workable operating model should establish who sponsors each use case, who owns the underlying business process, who approves data and technology arrangements, who assesses relevant risks, who validates outputs and who monitors the solution after deployment. It should also specify who can authorise changes, how incidents are escalated and who is responsible for suspending or withdrawing an application when its performance or risk profile becomes unacceptable. For applications with material implications for clients or compliance, the arrangements should provide appropriate independent challenge rather than relying entirely on the individuals promoting or operating the solution.

Clear accountability is particularly important because AI-related decisions do not always align neatly with existing functional boundaries. A technology team may manage a model or platform without owning the business decision it supports. A business user may rely on an AI-generated output without being responsible for the underlying data or supplier relationship. Unless these boundaries are made explicit, important tasks can fall between functions, especially when the system produces unexpected results or an incident requires investigation.

The objective is not to maximise governance complexity, but to make decisions, responsibilities and escalation routes explicit. This enables EAMs to retain the flexibility needed for experimentation while establishing the minimum organisational discipline required for controlled deployment. It also provides a foundation for developing internal expertise through practical implementation, structured training and the review of outcomes.

Taken together, the empirical findings suggest that the readiness gap in Swiss EAMs is multidimensional. Interest in AI is widespread among the surveyed firms, but data limitations, uneven process integration, restricted expertise and unclear responsibilities may impede the transition from experimentation to sustainable implementation. These constraints are interrelated: weak data can undermine output reliability, limited expertise can make those weaknesses harder to recognise, and unclear accountability can prevent issues from being resolved consistently. Accordingly, readiness should be addressed as a coordinated organisational challenge rather than through isolated investments in tools, data or training. This provides the basis for the implementation approach developed in the following chapters.

4. From opportunity identification to controlled implementation

4.1 Selecting use cases by value, feasibility and risk

A disciplined approach to AI implementation begins with clearly defined business problems rather than the capabilities of available technologies. Although AI offers a broad range of potential applications in external asset management (EAM), not every technically feasible use case will generate meaningful business value or be appropriate for immediate deployment. The challenge is to identify applications that address genuine operational needs while remaining feasible to implement and proportionate to the organisation's capacity to manage their associated risks.

Agnesens, Stüttgen and Ivanov (2026) identify several potential starting points for Swiss EAMs, including investment research synthesis, investment proposal generation, client reporting, know-your-customer (KYC) and anti-money-laundering (AML) document analysis, and data reconciliation. These applications differ in their data requirements, implementation complexity and exposure to operational or regulatory consequences. A research summarisation tool, for example, may operate within a relatively contained workflow, whereas automating client onboarding may require coordination across multiple systems, the application of compliance rules and the resolution of exceptions.

Use-case prioritisation should therefore consider four interrelated dimensions: business value, implementation feasibility, data readiness and risk. Business value concerns the significance of the problem and the potential improvement in efficiency, quality or service delivery. Feasibility encompasses the availability of suitable technology, the effort required for integration and the maturity of the underlying process. Data readiness concerns whether the information required is sufficiently accessible, reliable and appropriate for the intended use. Risk relates to the consequences of inaccurate, incomplete or inappropriate outputs, including the potential effects on clients, compliance obligations and operational continuity.

This approach is consistent with the broader financial-services literature, which emphasises the diversity of AI applications and the importance of considering their organisational context (Fares, Butt and Lee, 2023; Pattnaik, Ray and Raman, 2024). It also reflects the distinction between technical potential and implementation capability identified in research on AI-enabled financial advisory services (Zhu, Vigren and Söderberg, 2024).

Use cases that are frequent, clearly defined and relatively straightforward to validate may provide suitable starting points. An investment research assistant, for example, could be evaluated against source documents for factual accuracy, coverage of material information, traceability and the time required for an analyst to produce an acceptable output. By contrast, an application that autonomously coordinates an entire onboarding process would require stronger evidence of data quality, workflow integration, exception handling and control effectiveness before deployment.

This does not imply that complex or potentially high-value applications should be excluded. Rather, their implementation should be conditional on demonstrated readiness and an appropriate level of assurance. Prioritisation should therefore be treated as a continuing decision: as data foundations, organisational expertise and governance mature, applications that were previously impractical may become viable. The objective is to sequence investment so that early deployments generate useful evidence and build capabilities that support more demanding applications over time.

4.2 Establishing a baseline and measuring value

A recurring challenge in technology transformation is the difference between anticipated benefits and demonstrable outcomes. Efficiency improvements, better service quality and reduced administrative effort are frequently cited as reasons for adopting AI, but their actual magnitude may be difficult to establish. The EAM study identifies efficiency and quality as important motivations for adoption while highlighting the broader challenges organisations face in translating interest into implementation (Agnesens, Stüttgen and Ivanov, 2026).

A credible business case should therefore begin with a documented baseline of the existing process. Depending on the use case, relevant measures may include processing time, manual effort, rework, error rates, exception-resolution time, completeness of documentation and the time employees can devote to client-facing or analytical activities. Establishing these measures before implementation makes it possible to distinguish actual improvements from subjective impressions or benefits that would have occurred independently of the technology.

Measurement should reflect the intended business outcome rather than rely exclusively on productivity indicators. An AI tool that reduces the time required to prepare a client report may deliver little net benefit if employees subsequently spend an equivalent amount of time correcting unsupported statements or reconciling inconsistent information. Similarly, faster KYC document processing cannot be considered an unqualified improvement if it increases unresolved exceptions or allows material inconsistencies to go undetected. In financial services, efficiency must be assessed alongside accuracy, completeness and the effectiveness of relevant controls.

The business case should also account for the total cost of implementation and operation. In addition to software licensing, costs may arise from data preparation, system integration, testing, validation, staff training, monitoring and vendor management. These costs can be material, particularly where an application depends on fragmented data or requires extensive changes to existing workflows. A realistic assessment should compare the resulting benefits with the full cost of maintaining the application, rather than with the cost of the technology alone.

The evaluation should further consider whether benefits are sustainable under normal operating conditions. Results from a demonstration or limited pilot may not reflect performance across different users, document types, client circumstances or exception scenarios. Operational experience should therefore be used to assess whether expected improvements persist and whether new risks or costs emerge as usage increases.

This emphasis on measurable outcomes is consistent with the argument that AI's contribution to financial performance depends partly on internal processes and governance arrangements (Cheng et al., 2025). For EAMs, the resulting evidence should inform decisions to expand, modify, restrict or discontinue a use case. Investment in AI should be treated as a hypothesis to be tested through implementation, rather than a benefit that can be assumed from the outset.

4.3 Embedding governance in the implementation lifecycle

AI governance is most effective when incorporated throughout the implementation lifecycle rather than introduced as a final approval step. This approach enables risks and control requirements to influence use-case selection, solution design, testing and operational deployment. It also reduces the likelihood that material constraints will be discovered only after resources have been committed to a particular technology or implementation approach.

Before development or configuration begins, the organisation should define the intended purpose, expected users, data requirements, permitted uses and potential consequences of failure. The initial assessment should also consider applicable legal and regulatory obligations, contractual restrictions, external providers and dependencies on other systems. These factors help determine whether the proposed application is appropriate and what level of oversight will be required.

During design and testing, the organisation should establish acceptance criteria that reflect the actual business purpose. Testing should cover routine operating conditions as well as incomplete information, conflicting records, ambiguous instructions and other cases that may expose weaknesses. It should also establish how the system behaves when information is insufficient or a request falls outside its intended scope. In such circumstances, appropriate behaviour may involve withholding an answer, identifying uncertainty or referring the matter to a qualified person rather than generating a plausible but unsupported output.

Before production deployment, the organisation should confirm that operational responsibilities, access rights, escalation routes and monitoring arrangements are established. It should also determine how errors will be investigated, how incidents will be managed and how the application can be restricted or withdrawn if its performance becomes unacceptable. Following deployment, monitoring should consider relevant performance indicators, incidents, user feedback and changes in the underlying data, technology or business process. Governance must remain active because an application's risk profile and operating conditions may change over time.

Additional safeguards are warranted where agentic AI can interact with multiple tools or systems and initiate downstream actions. In such cases, controls should reflect the scope of the system's authority, the reversibility of its actions and the criticality of the process. Depending on the use case, these controls may include restricted permissions, bounded workflows, action limits, approval requirements, detailed action records and tested mechanisms for suspending or constraining the system. The objective is to prevent an error at one stage from propagating unchecked through a wider process.

This lifecycle approach reflects the broader finding that AI implementation in regulated financial services requires attention to organisational preparedness, reliability and data protection alongside technological capability (Moharrak and Mogaji, 2025). It also recognises that governance requirements should be proportionate to the application rather than uniform across all AI use cases.

The aim is not to eliminate uncertainty, which is rarely possible in complex systems, but to establish an acceptable and observable risk position. This requires clear ownership of residual risks, evidence that relevant controls are functioning and the capacity to respond when actual performance diverges from expectations. In this sense, governance is not merely a constraint on innovation; it is part of the infrastructure that enables organisations to deploy AI with greater confidence.

4.4 Scaling from pilots to institutional capability

A successful pilot demonstrates that an application can produce useful results under particular conditions; it does not, by itself, establish that the organisation is ready to scale it. Broader deployment introduces additional considerations, including user access, technical support, data consistency, integration, operating costs, staff behaviour and the ability to detect and manage incidents across a larger user population. Scaling can also expose weaknesses that remain hidden when an application is used by a small group under close supervision.

A staged implementation approach enables evidence and organisational capability to develop progressively. Initial experimentation can establish whether a use case addresses a meaningful business problem. A controlled pilot can then test the solution against defined acceptance criteria using realistic data and operating conditions. Production deployment should follow only when performance, ownership, support arrangements and controls are sufficiently established. Wider adoption should depend on evidence that the application continues to deliver value and that its risks can be managed under normal operating conditions.

This progression should not be understood as a purely technical sequence. Each stage also requires organisational learning. Employees need to understand when the system is appropriate to use, how its outputs should be reviewed and when issues should be escalated. Process owners need evidence that the application improves the overall workflow, while management must determine whether the benefits justify the cost and whether the organisation can support the solution over time. Scaling is therefore as much a matter of operational and organisational readiness as it is of technical performance.

As the number of AI applications grows, governance must also extend beyond individual use cases. Several applications may depend on the same external provider, underlying model, data source or technical platform. These shared dependencies can create correlated risks, including service disruption, inconsistent outputs or a common weakness affecting multiple business processes. Assessing each application independently may fail to reveal the cumulative exposure arising from these relationships.

Aldasoro et al. (2025) examine the wider implications of AI adoption across financial institutions and markets, highlighting why the effects of AI should not be considered solely at the level of individual applications. For EAMs, this supports the value of maintaining an organisation-wide view of AI use, dependencies and associated risks, while recognising that the appropriate level of formality should reflect the firm's size and complexity.

A proportionate inventory of AI applications provides a practical foundation for this oversight. It should identify each application's purpose, accountable owner, principal data dependencies, relevant external providers, degree of autonomy and risk classification. Maintaining this information helps management understand where AI is being used, identify overlapping or unapproved applications, assess concentration in shared providers and determine where additional controls or investment may be required.

Ultimately, scaling should represent a transition from isolated technical success to repeatable organisational capability. This requires a consistent method for selecting use cases, establishing baselines, validating performance, managing risk and learning from operational experience. For Swiss EAMs, such an approach offers a practical means of progressing from individual experimentation towards controlled, value-driven adoption without imposing governance structures disproportionate to their operating models.

5. Regulatory and risk considerations in the Swiss context

The implementation of artificial intelligence (AI) by Swiss external asset managers (EAMs) must be considered within the legal, regulatory and supervisory framework applicable to each organisation and its activities. The relevant requirements may vary according to a firm's regulatory status, the financial services it provides, the nature of the information processed and the extent to which AI influences operational or client-related decisions. Regulatory readiness should therefore be assessed in relation to the intended use case rather than treated as a uniform requirement across all AI applications.

Agnesens, Stüttgen and Ivanov (2026) identify regulatory uncertainty, data protection and output reliability as important concerns among Swiss EAMs considering AI adoption. Their discussion of FINMA Guidance 08/2024 further highlights the relevance of governance, risk classification, data quality, testing, documentation, monitoring and accountability. These considerations reinforce the distinction between adopting AI as an individual productivity tool and incorporating it into a controlled business process. Although the appropriate supervisory expectations depend on the circumstances of each institution, EAMs need to establish how their existing obligations apply to the specific technologies, data and activities they intend to support with AI.

The wider financial-services literature provides a complementary perspective. Fares, Butt and Lee (2023) and Pattnaik, Ray and Raman (2024) demonstrate the breadth of AI applications across banking and financial services, while Moharrak and Mogaji (2025) highlight the implementation challenges associated with integrating generative AI into a regulated industry. Taken together, these studies suggest that regulatory and operational considerations cannot be separated from the adoption process itself. They must inform decisions about use-case selection, system design, implementation and ongoing operation.

A first implication is that regulatory obligations must be translated into specific operational requirements. General commitments to confidentiality, security or compliance are insufficient unless they inform decisions about which information may be processed, which systems may access it, how outputs are validated and retained, and who is accountable for their use. For example, an AI application used to summarise investment research presents different control requirements from one that processes confidential client documentation or supports know-your-customer (KYC) and anti-money-laundering (AML) procedures. In each case, the organisation should determine the applicable obligations, identify the relevant risks and establish controls proportionate to the potential consequences of failure. This approach is consistent with the broader argument that the contribution of AI depends on its integration with internal processes and governance arrangements (Cheng et al., 2025).

A second implication concerns the management of external technology providers. Outsourcing AI functionality does not eliminate the EAM's need to assess the suitability of the service for its intended purpose. Provider documentation, security assurances and contractual commitments may support this assessment, but they do not, on their own, establish that a particular implementation is appropriate. Firms should consider the provider's data-processing arrangements, access permissions, service dependencies, data retention practices and arrangements for handling incidents or service disruption. They should also determine whether the available information is sufficient to assess system limitations and investigate problematic outputs. These considerations are particularly important where AI services process sensitive client information or become embedded in business-critical workflows.

A third implication is that governance must be supported by evidence. Depending on the application and applicable obligations, relevant evidence may include documented use-case assessments, approved requirements, data and access controls, testing results, records of human review, exception-handling procedures, monitoring outcomes and evidence of corrective action. Such documentation should enable the organisation to establish how an application was assessed, who authorised its use, how its outputs were validated and how identified problems were addressed. For more complex or autonomous applications, records of system interactions and actions may also be necessary to support traceability and incident investigation. The objective is not to maximise documentation, but to ensure that the evidence retained is sufficient for the risk profile and purpose of the application.

Risk management should also extend beyond initial approval. AI systems may produce unreliable outputs, behave inconsistently across different inputs or become less suitable as data, technology and business processes change. Generative AI can introduce unsupported statements or omit material information, while systems with access to multiple tools or operational platforms may create additional risks through inappropriate actions or failures in workflow coordination. Aldasoro et al. (2025) examine the broader implications of AI for financial institutions and markets, reinforcing the importance of considering the effects of AI beyond individual model outputs. For EAMs, this means assessing not only whether a system performs adequately during testing, but also whether its use remains controlled under operational conditions.

Accordingly, the intensity of oversight should reflect the application's purpose, data sensitivity, degree of autonomy and potential impact. A narrowly scoped internal assistant may be managed through approved access, user guidance, output verification and proportionate monitoring. An application that influences client suitability, processes sensitive due diligence information or initiates actions across interconnected systems warrants more extensive validation, restricted permissions, explicit approval points and defined escalation or suspension procedures. This distinction should be based on the actual use case rather than on whether a solution is labelled generative or agentic AI. As Zhu, Vigren and Söderberg (2024) emphasise through their review of AI-enabled financial advisory services, technological implementation must be considered in the context of the financial-service activity it supports.

For Swiss EAMs, the practical priority is therefore to integrate regulatory assessment and risk management into the entire AI implementation lifecycle. Before deployment, firms should establish the intended purpose, determine the applicable requirements, assess relevant data and provider dependencies, and define the controls needed to support reliable operation. During implementation, these requirements should inform testing, validation, user responsibilities and acceptance criteria. After deployment, firms should monitor performance, investigate incidents, review material changes and reassess whether the application remains suitable for its intended use.

Regulatory considerations should not be treated solely as constraints on innovation. When translated into proportionate and operationally effective controls, they help EAMs distinguish between applications that are ready for deployment and those that require further development of data, technical or organisational capabilities. Responsible AI adoption consequently depends not on eliminating risk, but on identifying, assessing and managing it in a manner consistent with the firm's obligations and the consequences of the activities being supported. This makes regulatory and risk management an integral component of the organisational readiness required to translate AI potential into measurable, reliable and sustainable business value.

6. A proposed readiness framework for Swiss EAMs

The preceding analysis indicates that the successful adoption of artificial intelligence (AI) by Swiss external asset managers (EAMs) depends on a combination of technological, organisational and governance capabilities. The findings of Agnesens, Stüttgen and Ivanov (2026) highlight constraints relating to data readiness, internal expertise, accountability and regulatory uncertainty, suggesting that interest in AI does not necessarily translate into the capacity to implement it systematically. These findings are consistent with the broader financial-services literature, which emphasises the importance of organisational processes, internal governance and implementation conditions in determining the value derived from AI (Fares, Butt and Lee, 2023; Cheng et al., 2025).

Building on this evidence, this paper proposes a readiness framework comprising five interdependent dimensions: business and process readiness; data and technology readiness; people and organisational readiness; governance and control readiness; and value and scale readiness. Together, these dimensions provide a structured basis for evaluating whether a particular AI use case is appropriate for implementation, identifying capability gaps and determining the conditions required for controlled deployment. The framework is designed specifically to reflect the operating characteristics of EAMs, where limited specialist resources, sensitive client information, relationship-based services and regulatory responsibilities must be considered alongside the potential benefits of automation.

Business and process readiness concerns the extent to which the organisation has identified a meaningful business problem and established a sufficiently well-defined process for addressing it. Before introducing AI, an EAM should understand the existing workflow, the information required at each stage, the decisions involved and the sources of delay, inconsistency or unnecessary manual effort. Success criteria should be defined in advance so that the proposed application can be evaluated against an identifiable operational need. Processes characterised by undocumented exceptions, inconsistent practices or unclear decision rights may require clarification or redesign before AI can be integrated reliably. This emphasis on process alignment reflects the wider argument that AI creates organisational value through its integration into business activities rather than through technology adoption alone (Cheng et al., 2025).

Data and technology readiness concerns whether the information, systems and technical infrastructure required for the intended application are sufficiently reliable, accessible and secure. The EAM study identifies data structure and readiness as important constraints, indicating that organisations may struggle to realise AI's potential when information is fragmented or inconsistently maintained (Agnesens, Stüttgen and Ivanov, 2026). Accordingly, firms should assess data completeness, consistency, timeliness, provenance and permitted use in relation to the specific task. They should also understand relevant system interfaces, external dependencies, access restrictions and integration requirements. Where AI processes confidential client information, data protection and appropriate access controls must be addressed before deployment. Readiness in this dimension does not require every organisational dataset to be fully standardised; rather, the data and technical arrangements must be adequate for the proposed application and its associated risks.

People and organisational readiness concerns whether the firm possesses the collective expertise, responsibilities and working practices necessary to use AI effectively. Employees need to understand the system's intended purpose, recognise its limitations and evaluate its outputs critically. The skills required will vary according to the application: users of an internal research assistant may primarily need to assess factual accuracy and source coverage, whereas employees involved in AI-supported compliance workflows may require a stronger understanding of exceptions, evidence requirements and escalation procedures. Business ownership, technical support, risk assessment and ongoing operational responsibilities should also be clearly assigned. These responsibilities may be combined in smaller EAMs, but the necessary capabilities and accountability must remain explicit. This dimension reflects the organisational challenges identified by Agnesens, Stüttgen and Ivanov (2026) and the broader literature's emphasis on implementation preparedness in regulated financial services (Moharrak and Mogaji, 2025).

Governance and control readiness concerns the firm's ability to identify, assess and manage the risks associated with a particular AI application throughout its lifecycle. This includes determining applicable legal and regulatory obligations, defining permitted uses, establishing approval responsibilities and setting requirements for testing, validation, monitoring and documentation. Controls should reflect the sensitivity of the information processed, the potential consequences of inaccurate outputs and the degree of autonomy granted to the system. Applications that interact with multiple systems or initiate consequential actions require particular attention to permissions, traceability, human approval and exception handling. Governance should also establish how incidents are investigated and how an application can be restricted, suspended or withdrawn when its performance becomes unacceptable. The importance of these arrangements is consistent with research highlighting the interaction between AI, internal governance and organisational performance (Cheng et al., 2025), as well as research on the challenges of AI implementation in regulated financial services (Moharrak and Mogaji, 2025).

Value and scale readiness concerns whether the organisation can demonstrate that an AI application produces sufficient benefit to justify its implementation and ongoing operation. Before a pilot begins, the EAM should establish a baseline and define measurable objectives appropriate to the use case. Relevant indicators may include processing time, manual effort, error rates, rework, documentation completeness or the time required to resolve exceptions. Evaluation should account for the total cost of implementation, including integration, data preparation, training, validation, monitoring and provider management. It should also consider whether productivity improvements are achieved without compromising output quality, client service or control effectiveness. The wider financial-services literature suggests that AI's contribution to organisational performance should be considered alongside the processes and governance arrangements through which it operates (Cheng et al., 2025). Consequently, expansion beyond a pilot should depend on demonstrated benefits and evidence that operational support, risk controls and technical dependencies remain manageable at the intended scale.

The five dimensions should be assessed collectively because weaknesses in one area can undermine progress in the others. A technically capable system may deliver unreliable results when the underlying data are inconsistent. A well-defined business problem may remain unsuitable for automation if the necessary information cannot be accessed appropriately. Similarly, strong governance can limit unacceptable risks but cannot compensate for an application that addresses an insignificant business problem or generates insufficient value. Readiness is therefore not a single technological characteristic, but a combination of conditions that determine whether a particular use case can be implemented and maintained successfully.

For practical application, EAMs should use the framework at the use-case level rather than assign a single readiness rating to the organisation as a whole. Each proposed application should be assessed against the five dimensions, with material gaps documented alongside the actions required to address them. These actions may include clarifying the underlying process, improving data quality, restricting system permissions, providing targeted training or establishing more robust monitoring. The assessment should then inform a decision to proceed, conduct a limited pilot, defer implementation or reject the use case. Where significant risks remain unresolved, potential business benefits should not, by themselves, justify deployment.

The framework should also be revisited as applications move from experimentation to production and wider adoption. A use case that is manageable within a limited pilot may introduce additional requirements when more employees gain access, when new data sources are incorporated or when the system is authorised to perform further actions. Periodic reassessment enables the organisation to identify emerging dependencies, evaluate changes in performance and confirm that the original controls remain appropriate. In this respect, readiness is dynamic rather than a one-off approval condition.

Importantly, the proposed framework is a conceptual and practice-oriented decision aid derived from the literature and the exploratory evidence on Swiss EAMs. It has not been empirically validated as a predictive maturity model, and its five dimensions should not be interpreted as a statistically established measurement instrument. Its contribution is to organise the principal conditions for responsible AI adoption into a structure that firms can apply to individual implementation decisions. Future research could evaluate its practical usefulness across EAMs of different sizes, compare readiness assessments with realised implementation outcomes and examine which capability gaps most strongly influence the sustainability of AI-generated business value.

Overall, the framework translates the paper's central argument into a practical approach to implementation. Rather than prioritising AI adoption on the basis of technological novelty alone, Swiss EAMs should assess whether the business problem is sufficiently defined, the required data and technology are suitable, the necessary expertise and accountability are available, the risks can be controlled and the expected benefits can be demonstrated. Developing these capabilities in parallel provides a more credible basis for moving from isolated experimentation towards reliable, responsible and sustainable AI adoption.

7. Discussion

The findings and literature reviewed in this paper support a shift in how artificial intelligence (AI) adoption in Swiss external asset management (EAM) should be understood. The central challenge is not simply to identify promising applications or introduce new technologies, but to develop the organisational capabilities required to translate AI potential into measurable, reliable and sustainable business value. This distinction between technological adoption and organisational readiness provides the principal contribution of the paper.

The argument is consistent with the broader financial-services literature. Fares, Butt and Lee (2023) and Pattnaik, Ray and Raman (2024) demonstrate the diversity of AI applications across financial services, while Zhu, Vigren and Söderberg (2024) highlight the importance of considering AI implementation in the context of financial advisory services and client relationships. Cheng et al. (2025) further associate AI adoption with financial performance through mechanisms involving internal business processes, governance and control. Collectively, these studies support a perspective in which the value of AI depends not only on technical functionality but also on the organisational conditions under which applications are implemented and used.

Nevertheless, the literature does not establish a single implementation model that can be transferred directly to every financial institution. The studies differ in research design, empirical setting and analytical focus, and findings from banking, broader financial services or Chinese financial companies cannot automatically be generalised to Swiss EAMs. Their contribution is therefore primarily conceptual: they help explain why business-process alignment, organisational competence and governance should be considered alongside technology selection. The extent to which particular arrangements generate superior outcomes in Swiss external asset management remains an empirical question.

The exploratory study by Agnesens, Stüttgen and Ivanov (2026) provides a more directly relevant indication of the challenges facing the sector. Its findings suggest that interest in AI among participating EAMs exceeds the maturity of some of the capabilities required for systematic implementation. Data readiness, internal expertise, accountability and regulatory uncertainty emerge as important concerns. These findings reinforce the distinction between experimentation and institutional capability developed throughout this paper. However, they should be interpreted as indications of potential sector-specific challenges rather than definitive estimates of their prevalence across the entire Swiss EAM population.

Several methodological limitations constrain the conclusions that can be drawn from this evidence. The survey covers 41 organisations from a population of 1,427 authorised Swiss EAMs, limiting the extent to which the results can be generalised. The sample also contains a substantial proportion of larger firms, whose resources, technology infrastructure and governance arrangements may differ from those of smaller EAMs. Furthermore, eight of the nine interviewees also participated in the survey. Although the interviews provide additional qualitative context, the overlap limits the extent to which they independently broaden the range of organisational perspectives represented. Finally, the study is exploratory rather than longitudinal and does not establish causal relationships between readiness, AI adoption and business performance. In particular, it does not demonstrate that the reported capability gaps necessarily result in lower performance or that specific AI applications generate measurable financial returns.

These limitations have implications for the proposed readiness framework. The five dimensions—business and process readiness, data and technology readiness, people and organisational readiness, governance and control readiness, and value and scale readiness—synthesise the principal implementation considerations identified in the literature and the EAM study. They provide a structured means of assessing individual use cases and identifying gaps that may prevent successful deployment. However, the framework should be understood as a conceptual and practice-oriented contribution rather than an empirically validated model of organisational maturity or implementation success. Its practical usefulness, relative importance of individual dimensions and ability to predict realised outcomes require further investigation.

Future research could address these questions through longitudinal studies that examine how AI readiness develops as firms progress from experimentation to production deployment. Such research could investigate whether improvements in data quality, process integration and staff competence are associated with more reliable implementation and stronger business outcomes. Comparative studies across firms of different sizes could also clarify whether smaller EAMs face distinctive constraints and which governance arrangements offer effective control without imposing disproportionate administrative costs. Further investigation could examine whether differences in business models, technology infrastructure or regulatory circumstances influence adoption patterns and implementation outcomes.

Research focusing on operational deployments would be particularly valuable. Studies of AI applications in know-your-customer (KYC) and anti-money-laundering (AML) procedures, investment research and client reporting could compare expected benefits with realised outcomes, including processing time, error rates, rework, control effectiveness and total operating costs. Such evidence would help establish which implementation practices are associated with sustained improvements rather than short-term productivity gains. It could also provide a stronger empirical basis for evaluating the proposed readiness framework and refining its application to different categories of AI use case.

The implications for practitioners are more immediate. EAMs should not interpret the growing availability or use of AI tools as evidence that they are ready for large-scale adoption. Instead, they should begin with a limited number of clearly defined business problems, establish baseline performance and evaluate results under realistic operating conditions. These initial implementations should be used not only to assess individual applications but also to develop reusable organisational capabilities in data management, workflow integration, staff competence, governance and operational support. Decisions to expand should depend on demonstrated value and evidence that the associated risks remain manageable.

This approach is particularly important when considering agentic AI. Systems capable of coordinating activities across applications and initiating actions may offer opportunities to automate more complex workflows, but they also introduce additional dependencies and potential consequences when errors propagate between process stages. Aldasoro et al. (2025) examine the broader implications of AI for financial institutions and markets, reinforcing the need to consider both institution-level implementation and wider dependencies. For EAMs, progress towards more autonomous applications should therefore be conditional on clearly defined permissions, effective oversight, traceable actions, appropriate exception handling and tested arrangements for restricting or suspending system activity. Greater autonomy should be treated as a reason to strengthen the assessment of readiness, rather than as an objective in itself.

Overall, this paper contributes a context-specific interpretation of AI adoption in Swiss external asset management by connecting sector-level evidence with the organisational and governance considerations identified in the wider literature. Its central implication is that sustainable adoption depends on the alignment of business objectives, processes, data, technology, people and controls. Although the available evidence does not establish a universally applicable implementation model or prove that the proposed framework predicts success, it provides a reasoned basis for structuring implementation decisions and identifying priorities for further investigation. For Swiss EAMs, the practical challenge is therefore to develop the capabilities that make AI use repeatable, accountable and demonstrably valuable, rather than to pursue adoption as an end in itself.

8. Conclusion

Artificial intelligence offers Swiss external asset managers (EAMs) opportunities to improve operational efficiency, strengthen information processing and support selected aspects of client service and investment-related work. However, the central question is not whether AI has potential, but whether individual firms possess the capabilities to implement it effectively, responsibly and at a scale that creates demonstrable value. This paper has argued that the distance between recognising AI’s potential and realising its benefits is fundamentally an organisational and governance challenge, as well as a technological one.

The analysis brings together broader research on AI in financial services and empirical insights into the Swiss EAM sector. The literature highlights the growing range of potential applications while also identifying persistent challenges related to data, skills, process integration, oversight and regulation (Fares et al., 2023; Pattnaik et al., 2024; Zhu et al., 2024). The Swiss EAM context adds a distinct practical dimension: firms vary in their resources, operating models, digital maturity and capacity to manage implementation risk. As a result, the same AI application may be feasible for one firm but premature, disproportionately costly or unsuitable for another. Readiness must therefore be assessed in relation to the firm, the intended use case and the consequences of failure, rather than inferred from general market enthusiasm or technological availability.

Regulatory and risk considerations reinforce this conclusion. Responsible AI adoption requires more than formal compliance or a general commitment to innovation. It calls for clear accountability, proportionate risk assessment, reliable data, appropriate human oversight, adequate documentation and ongoing monitoring. These controls should be embedded throughout the AI lifecycle, from use-case selection and testing to deployment, review and, where necessary, suspension. The involvement of external technology providers does not remove the need for firms to understand and oversee the risks associated with the systems they use. In this sense, governance is not merely a constraint on adoption; it is part of the operational foundation required to establish trust, manage uncertainty and scale successful applications.

To translate these insights into practice, the paper proposes a conceptual readiness framework built around five interrelated dimensions: business and process readiness; data and technology readiness; people and organisational readiness; governance and control readiness; and value and scale readiness. The framework provides a structured way for EAMs to assess individual use cases, identify capability gaps and make more disciplined implementation decisions. It also encourages firms to distinguish between applications that are ready for deployment, those that warrant controlled experimentation, and those that should be deferred until critical prerequisites are met. By linking readiness to business objectives, risk exposure and measurable outcomes, the framework seeks to move AI assessment beyond technology selection towards a more integrated model of organisational decision-making.

The framework is deliberately presented as a practice-oriented conceptual contribution rather than an empirically validated maturity model. The available sector-specific evidence is exploratory and does not establish that particular readiness characteristics cause better AI outcomes, nor does it demonstrate that AI adoption necessarily produces measurable performance improvements. The proposed dimensions and their practical value therefore require further testing. Future research could examine how readiness evolves over time, compare firms of different sizes and operating models, and investigate whether specific governance and capability investments improve implementation outcomes. Longitudinal studies of deployed applications could also assess measurable changes in processing time, error rates, rework, control effectiveness, cost and service quality. Such evidence would help establish whether the framework reliably supports prioritisation and whether its dimensions need to be adapted to different categories of AI use.

For practitioners, the principal implication is to favour disciplined progression over indiscriminate adoption. EAMs should begin with clearly defined business problems, assess the suitability and sensitivity of the relevant data, establish a credible baseline, and test whether an AI application delivers benefits under realistic operating conditions. Successful pilots should be scaled only when their value, reliability and risk controls have been demonstrated. Where more autonomous or agentic systems are considered, firms should pay particular attention to the boundaries of delegated authority, human intervention, traceability and the ability to restrict or stop system activity. This approach allows firms to build reusable capabilities while limiting the risk of investing in applications that are technically impressive but operationally unsuitable.

Ultimately, sustainable AI adoption in Swiss EAM will depend not on how many tools a firm acquires, but on whether it can integrate them into its operating model with appropriate skills, controls and accountability. AI readiness should consequently be understood as a continuing capability: one that develops through experience, evidence, organisational learning and periodic reassessment. By offering a structured way to connect business ambition with implementation capacity and governance, this paper provides a starting point for more deliberate AI adoption in the Swiss EAM sector. Its central proposition is that AI creates durable value only when technological opportunity is matched by organisational preparedness, proportionate risk management and a clear understanding of what successful implementation should achieve.

References

Agnesens, T., Stüttgen, M. and Ivanov, A. (2026) AI Use Cases and Crypto Assets in External Asset Management (EAM). Lucerne: Lucerne University of Applied Sciences and Arts, in cooperation with Finup and funded by Innosuisse, 22 July.

Aldasoro, I., Gambacorta, L., Korinek, A., Shreeti, V. and Stein, M. (2025) ‘Intelligent financial system: How AI is transforming finance’, Journal of Financial Stability, 81, article 101472.

Cheng, X., Du, A.M., Yan, C. and Goodell, J.W. (2025) ‘Internal business process governance and external regulation: How does AI technology empower financial performance?’, International Review of Financial Analysis, 99, article 103927.

Fares, O.H., Butt, I. and Lee, S.H.M. (2023) ‘Utilization of artificial intelligence in the banking sector: A systematic literature review’, Journal of Financial Services Marketing, 28(4), pp. 835–852.

Moharrak, M. and Mogaji, E. (2025) ‘Generative AI in banking: Empirical insights on integration, challenges and opportunities in a regulated industry’, International Journal of Bank Marketing, 43(4), pp. 871–896.

Pattnaik, D., Ray, S. and Raman, R. (2024) ‘Applications of artificial intelligence and machine learning in the financial services industry: A bibliometric review’, Heliyon, 10(1), article e23492.

Zhu, H., Vigren, O. and Söderberg, I.-L. (2024) ‘Implementing artificial intelligence empowered financial advisory services: A literature review and critical research agenda’, Journal of Business Research, 174, article 114494.

Contact

Reach out via email for inquiries.

Email

Subscribe to newsletter

info@grcadvisory.ch

© 2025. All rights reserved.