AI Agents, Platform Power and the Future of Financial Intermediation

Abstract

The emergence of agentic artificial intelligence (AI) may fundamentally alter the structure of financial intermediation. Unlike conventional digital tools, agentic AI can interpret objectives, evaluate alternatives, make recommendations and, subject to delegated authority, initiate actions on behalf of consumers. This paper examines the implications of these capabilities for the competitive position of banks and the governance of financial decision-making. Drawing on the literature on financial intermediation, platform power, algorithmic decision-making, AI-based financial advice and automation bias, the paper develops the concept of interface intermediation to describe the mediation of access, information and choice between consumers and financial institutions. It argues that agentic AI could become a financial gatekeeper without becoming a bank: its power may derive from controlling discovery, recommendation and execution rather than ownership of financial assets or balance-sheet capacity.

The paper further develops the concept of delegated financial agency, through which consumers transfer increasing degrees of objective-setting, interpretation, provider selection and execution to AI systems. This creates significant potential benefits, including lower search costs, greater personalisation, reduced administrative burdens and potentially reduced forms of human bias. However, efficiency does not necessarily imply consumer welfare. Automation bias, commercial conflicts of interest, ambiguous optimisation objectives and fragmented accountability may weaken meaningful consumer autonomy. The paper therefore proposes transaction-chain governance as a regulatory perspective that considers the entire process connecting consumer intention to financial action rather than focusing exclusively on the AI model.

The analysis does not assume that AI platforms will inevitably replace banks. Consumer preferences, regulatory constraints, interoperability and incumbent adaptation may limit platform dominance. Instead, the paper argues that agentic AI creates the possibility of a new locus of financial gatekeeping in which control over the customer decision interface becomes economically significant. The future of banking may consequently depend not only on who provides financial products, but on who controls the processes through which consumers discover, evaluate and act upon those products.

Keywords: artificial intelligence; agentic AI; banking; platformisation; financial intermediation; Big Tech; consumer autonomy; algorithmic decision-making; financial advice; gatekeeping

1. Introduction

Financial intermediation has traditionally been organised around institutions that combine capital, information, distribution and trust. Banks do not simply hold deposits, extend credit or facilitate payments; they also provide an important interface through which customers discover, evaluate and acquire financial products. The digital transformation of banking has progressively loosened this relationship between the provision of financial services and the institution through which customers access them. Online banking, fintech platforms, comparison services and embedded finance have increasingly separated product provision from the traditional branch and, in some cases, from the bank's own digital channels. Vives (2019) characterises this process as a fundamental digital disruption of banking, in which technology changes both the competitive structure of financial services and the relationship between incumbent banks, fintech firms and large technology platforms.

The emergence of generative and increasingly agentic artificial intelligence may represent a further stage in this transformation. Earlier forms of financial technology largely required customers to navigate digital interfaces themselves: they searched for products, compared alternatives and initiated transactions. Agentic systems potentially alter this arrangement by allowing customers to delegate elements of these activities to an artificial intelligence system capable of maintaining context, interpreting objectives, evaluating alternatives and, where appropriately authorised, interacting with external services. The significance of this development therefore extends beyond the automation of individual banking processes. It raises the possibility that the interface through which financial choices are formed could become institutionally separate from the financial institution that ultimately provides and executes the service.

Kamm's (2026) discussion of OpenAI's "Dots" provides a useful starting point for examining this possibility. He proposes that an AI-based "digital concierge" could become an intermediary between customers and financial institutions, with the bank continuing to provide the underlying regulated financial service while the AI system increasingly mediates the customer's interaction with the market. In such a configuration, the customer might no longer begin a financial journey by entering the digital environment of a particular bank. Instead, the customer could state an objective to an AI agent, which would interpret that objective, identify potentially suitable providers, compare alternatives and potentially initiate the resulting transaction. The bank would remain responsible for the regulated financial product, but the AI platform could increasingly influence how the customer discovers, evaluates and selects that product (Kamm, 2026).

This possibility requires a reconsideration of what is meant by financial disintermediation. Conventional accounts of disintermediation imply that an intermediary is removed from a transaction, allowing customers and providers to interact more directly. Agentic AI suggests a potentially different trajectory. Rather than eliminating intermediation, it may relocate it. The bank could remain indispensable as a balance-sheet institution, regulated entity and provider of financial infrastructure while another organisation assumes increasing control over the interface through which customers formulate financial decisions. The resulting structure would not necessarily be one in which banks disappear from the customer relationship, but one in which their relationship with the customer becomes increasingly mediated by an external technological intermediary.

This distinction is particularly important because the strategic value of a customer relationship does not reside exclusively in the execution of a transaction. It also resides in the processes that precede execution: identifying a need, framing the relevant financial problem, determining which alternatives are considered, evaluating trade-offs and selecting a course of action. Digital platforms have demonstrated that control over these processes can constitute a significant source of economic power. Qi (2022), for example, conceptualises platform power in terms of the ability of large technology platforms to structure interactions between market participants and to occupy strategically important positions within economic networks. Research on the platformisation of financial infrastructure similarly illustrates that technological transformation does not necessarily eliminate incumbent intermediaries; rather, it can redistribute power between different layers of the financial system. Robinson, Dörry and Derudder (2024), in their analysis of SWIFT, demonstrate how an incumbent infrastructure provider can remain an "obligatory passage point" even as financial technologies and platform architectures evolve.

Agentic AI therefore introduces the possibility of what may be termed interface intermediation. The central intermediary may no longer be the organisation that provides the financial product, but the technological system that mediates access to the products available in the market. This distinction shifts attention from the ownership of financial assets and execution infrastructure towards the control of customer attention, product discovery and decision formation. In this context, the strategically significant question is not simply whether an AI platform can perform a banking transaction, but whether it can become the system through which customers decide which banking transaction to perform.

This paper consequently examines whether agentic AI platforms could become gatekeepers of financial choice, and what such a development would mean for banks, consumers and the organisation of financial intermediation. The concept of the "gatekeeper" is used deliberately. An AI agent would not need to control every financial product or execute every transaction to exercise significant influence over the market. Its power could arise earlier in the decision process, through determining which providers are surfaced, how their products are compared and how their relative suitability is communicated to the customer. In this respect, the relevant form of power is not necessarily control over financial assets but control over the decision environment in which financial choices are made.

The emerging literature on AI-based financial advice provides grounds for taking this possibility seriously while also cautioning against technological determinism. Baeckström and Matkovskyy (2026), in comparing human and AI financial advice behaviour, examine the extent to which financial decision-making may differ when advice is generated by artificial intelligence rather than human advisers. Brüggen et al. (2025) similarly demonstrate that the development of AI-based financial advice raises questions extending beyond technical performance to autonomy, responsibility, transparency and the ethical relationship between consumers and advisory systems. These concerns are particularly relevant when AI moves beyond providing information towards interpreting a customer's objectives and recommending actions on their behalf.

The potential benefits are nevertheless substantial. AI systems may reduce the cognitive and informational costs associated with financial decision-making, particularly in markets characterised by complex products and significant information asymmetries. Jaffri, Jha and Butt (2026) find evidence of household interest in using artificial intelligence for financial advice, suggesting that consumers may value AI where it can reduce the time and effort required to navigate financial decisions. Similarly, Chen et al. (2025) demonstrate that FinTech-based decision systems can reduce certain forms of human bias in lending. These findings complicate a purely dystopian interpretation of AI-mediated finance: delegating aspects of financial decision-making to algorithms may in some circumstances improve consistency, accessibility and decision quality.

The same delegation, however, creates a new set of risks. The relevant issue is not simply whether an AI system produces a better recommendation than a human being, but who determines the objectives according to which "better" is defined. When an individual asks an AI system to identify the most suitable mortgage, investment or savings product, the system must translate a relatively open-ended human objective into a set of criteria against which alternatives can be evaluated. That translation is itself a form of decision-making. The more capable the system becomes at performing this interpretive function, the greater the potential for the customer to delegate not merely the execution of a decision but part of the reasoning through which that decision is constructed.

This creates a tension between convenience and autonomy. Sargeant (2023) highlights the economic and normative consequences of algorithmic decision-making in financial services, including questions concerning autonomy, discrimination and the distribution of responsibility. Brüggen et al. (2025) similarly emphasise that AI-based financial advice cannot be assessed solely through the technical accuracy of its outputs because the relationship between the consumer and the system raises broader questions about agency and ethical responsibility. The problem is therefore not that consumers necessarily lose autonomy when they use AI; rather, autonomy may increasingly depend upon whether delegated decision-making remains transparent, contestable and subject to meaningful human control.

This concern is reinforced by the literature on automation bias. Skitka, Mosier and Burdick (1999) demonstrate that individuals can become overly reliant on automated decision aids, including in circumstances where the automated recommendation is inconsistent with other available information. In financial services, where AI systems may communicate recommendations through highly fluent and personalised natural-language interactions, the risk of over-reliance may be particularly significant. Trust in the system may become difficult to distinguish from trust in the underlying financial recommendation. Recent research by Sungkarungsri and Kiattisin (2026) further highlights the importance of trust and perceived credibility in the adoption of AI-based financial advisory services. The more trusted an AI intermediary becomes, therefore, the greater its potential influence over financial choice—and the greater the importance of understanding how that trust is constructed and governed.

The emergence of agentic AI also raises questions about the organisational distribution of authority. A system that merely provides financial information occupies a different position from one that recommends a particular provider, and both differ fundamentally from an agent authorised to execute a transaction. The distinction becomes particularly significant when an AI system acts across multiple financial institutions. A customer's instruction to "manage my finances" may require the agent to coordinate actions across several providers, each of which may have only partial visibility of the customer's wider instructions and constraints. The problem is therefore not simply one of authentication. It is a question of how delegated authority is represented, interpreted, monitored and revoked across institutional boundaries (Kamm, 2026).

This creates the possibility of a new layer of financial intermediation situated between the customer and the regulated financial institution. Whereas conventional financial intermediaries transform and allocate capital, an AI intermediary could increasingly transform and allocate customer attention and decision-making. Its economic significance would derive from its position within the sequence through which financial intentions become financial actions. This distinction is central to the argument advanced in this paper.

The paper therefore develops the proposition that agentic AI may generate a new form of financial intermediation in which decision-forming power becomes strategically significant alongside traditional control over financial assets, infrastructure and execution. The argument is not that AI platforms will inevitably replace banks, nor that customers will necessarily abandon direct relationships with financial institutions. Rather, the proposition is that the competitive boundary of banking may shift towards the layer at which customers formulate financial choices. If a small number of AI platforms become the principal systems through which customers ask financial questions, compare providers and delegate actions, those platforms could acquire a gatekeeping position even while banks retain responsibility for the underlying regulated services.

Such a development would represent a significant evolution of the platformisation of finance identified in the existing literature. Vives (2019) describes digital disruption as a force reshaping the structure of banking, while Qi (2022) demonstrates how Big Tech platforms can acquire economic power through control of strategically important interfaces. Robinson, Dörry and Derudder (2024) show that financial platformisation can simultaneously disrupt and reinforce existing institutional arrangements. Agentic AI potentially extends these dynamics into the domain of financial decision-making itself: the platform may become not merely the place through which a financial service is accessed, but the system through which the customer determines which financial service should be accessed in the first place.

The implications extend beyond competitive strategy. For banks, the emergence of AI gatekeepers could challenge the traditional assumption that owning the account equates to owning the customer relationship. For consumers, AI-mediated finance could reduce search costs and improve access to personalised financial guidance while simultaneously increasing dependence on systems whose objectives and commercial incentives may not be fully visible. For regulators, the challenge is to determine how responsibility should be allocated when financial decisions emerge from a chain involving a customer, an AI intermediary and one or more regulated financial institutions. Noponen et al. (2024), although concerned more broadly with algorithmic management, highlight the underlying tension between automation and autonomy that becomes particularly salient when technological systems assume increasingly consequential decision-making functions.

The central proposition of this paper is therefore deliberately more nuanced than the claim that artificial intelligence will "replace" banks. Agentic AI may instead transform the architecture of financial intermediation by introducing a new layer between financial institutions and the formation of customer decisions. The strategic and regulatory significance of this development lies in the possibility that control over the financial customer will increasingly depend not on who holds the customer's money, but on who mediates the customer's understanding of where that money should go.

The remainder of the paper examines this proposition through the literature on digital banking, platform power, algorithmic decision-making and AI-based financial advice. It considers the mechanisms through which AI agents could acquire gatekeeping power, the implications for consumer autonomy and trust, the potential consequences for banks and financial infrastructures, and the regulatory challenges created by delegated financial agency. In doing so, it treats Kamm's (2026) proposition as a forward-looking hypothesis rather than an established outcome, using the emerging academic literature to assess both its plausibility and its limitations.

2. From Financial Intermediation to Interface Intermediation

Financial intermediation has traditionally been understood in terms of the economic functions performed by financial institutions between savers, borrowers and other market participants. Banks accept deposits, extend credit, facilitate payments and transform risk and maturity, thereby reducing the transaction and information costs associated with financial exchange. Their intermediary role, however, has never been confined to the movement of funds. Banks also mediate information: they determine which products are presented to customers, how those products are explained and, to a significant extent, the institutional context within which financial decisions are made.

Digitalisation has begun to separate these functions. The technological transformation of banking has allowed financial products, distribution channels, customer interfaces and payment infrastructures to become increasingly modular. Vives (2019) argues that digital disruption is fundamentally reshaping the competitive structure of banking, bringing incumbent banks into competition not only with traditional financial institutions but also with fintech firms and large technology companies. The resulting environment is increasingly characterised by platform-based business models in which customer access, data and digital interfaces become strategically important alongside the underlying financial products.

This development suggests that financial intermediation should not be understood solely as the mediation of capital or risk. It can also be understood as the mediation of access, information and choice. The distinction becomes increasingly important with the emergence of agentic artificial intelligence. Whereas earlier forms of digital banking largely transferred the customer's interaction with the bank from a physical to a digital environment, an AI agent could potentially remove the requirement for the customer to interact directly with any particular financial institution at the initial stage of a financial decision.

Kamm's (2026) discussion of an AI "digital concierge" illustrates this possibility. Consider a customer seeking to finance a new home. Under the conventional digital banking model, the customer is likely to identify potential lenders, visit their websites or applications, provide information and compare the resulting offers. The bank therefore occupies the initial point of interaction and has an opportunity to frame the customer's understanding of the available financial choices. Under an agentic model, by contrast, the customer could simply communicate an objective to an AI system. The agent might interpret the customer's financial circumstances and preferences, identify potentially suitable mortgage providers, compare their products, explain the relevant trade-offs and, subject to appropriate authority, initiate an application.

The significance of this hypothetical interaction lies not in the automation of the mortgage application itself. Banks already automate substantial parts of lending, onboarding and servicing. Its significance lies in where the customer begins the decision process. Under the traditional model, the bank's digital interface is the environment in which the customer encounters the bank's products. Under the agentic model, the customer may encounter the bank only after an external system has already interpreted the customer's objective and determined that the bank is relevant to it.

This represents a potentially important shift in the location of intermediation. The conventional bank acts as an intermediary between financial capital and the customer, but an AI agent could increasingly act as an intermediary between the customer and the financial market itself. The distinction is subtle but consequential. The AI system does not necessarily replace the bank's financial function; instead, it potentially inserts itself upstream of the bank's relationship with the customer.

The resulting structure can be understood as a movement from financial intermediation towards interface intermediation. Financial intermediation concerns the institutional mechanisms through which funds, risk and financial services are exchanged. Interface intermediation concerns the mechanisms through which customers encounter, interpret and select those services. The two functions may increasingly be performed by different organisations.

This distinction is important because control of the interface can generate economic value independently of ownership of the underlying product. A bank may continue to provide a mortgage, hold a current account or manage an investment portfolio while an external platform increasingly determines whether the customer considers that bank in the first place. The institution providing the financial product therefore retains operational and regulatory significance, but its influence over the customer's decision environment may diminish.

Platform scholarship provides a useful theoretical foundation for understanding this possibility. Qi (2022) argues that the economic power of Big Tech platforms arises in part from their ability to structure interactions between different groups of market participants. Platform power is consequently not dependent upon direct ownership of all the assets involved in an economic transaction. Instead, it can arise from occupying a strategically important position through which users and providers must interact. The platform's value derives from its ability to organise those interactions, control access and benefit from the network effects and information advantages generated by its position.

This perspective is particularly relevant to banking because financial services already contain multiple layers of infrastructure and intermediation. Robinson, Dörry and Derudder (2024), in their analysis of SWIFT, demonstrate how technological transformation can produce partial platformisation without necessarily eliminating established financial institutions. Their concept of an "obligatory passage point" is especially relevant: an actor can acquire structural significance by becoming a point through which transactions or interactions must pass, even when it does not itself provide the underlying service.

Agentic AI introduces the possibility that a similar dynamic could emerge at the level of customer decision-making. An AI platform would not need to become a bank, acquire a banking licence or hold customer deposits in order to acquire strategic influence over banking. If customers increasingly use the platform to identify financial needs, compare providers and determine appropriate actions, the platform could become an obligatory passage point in the decision process preceding a financial transaction.

The distinction can be illustrated through a conventional mortgage journey. In a traditional model, the customer identifies a need, searches for potential providers, evaluates products and approaches a bank. The bank then assesses the customer, recommends or offers products and ultimately executes the transaction. In an agent-mediated model, the customer could instead communicate the underlying objective to an AI system. The AI would interpret the objective, identify potentially relevant providers, compare available alternatives and potentially initiate contact or an application. The bank would still perform the regulated functions of underwriting, contracting and lending, but its position in the customer's decision process would have changed.

The difference is therefore not simply technological. It concerns the allocation of informational and cognitive authority. In the traditional model, the customer performs much of the work of search and comparison, while each bank controls the information environment within its own distribution channel. In an agent-mediated model, these activities could become concentrated within the AI platform. The platform may acquire information about the customer's preferences across multiple financial relationships while simultaneously developing a market-wide view of the products available from competing providers.

This creates the possibility of a particularly powerful combination of information advantages. A bank traditionally possesses detailed information about its relationship with an individual customer but relatively limited information about the customer's interactions with competing institutions. An AI intermediary could, subject to consent and technical access, potentially observe or coordinate across several financial relationships. At the same time, it could maintain information about a broad range of financial products and providers. The resulting informational position could be substantially different from that of either a traditional bank or a conventional comparison website.

The distinction between a comparison platform and an agentic intermediary is therefore significant. A comparison website generally requires the consumer to formulate a query, review results and make a decision. An agentic system potentially performs some or all of these cognitive activities on the consumer's behalf. It moves from facilitating comparison to participating in choice formation. This is precisely where the concept of interface intermediation becomes analytically useful.

The resulting relationship between the customer and the bank may consequently become more distributed. A customer could maintain a mortgage with one bank, a current account with another institution and investments with a third, while interacting primarily with a single AI platform. The customer's financial relationships would remain institutionally separate, but the customer's experience of those relationships could become consolidated within the AI interface. The platform would effectively become a cognitive gateway to finance.

This possibility also changes the meaning of customer ownership. In traditional banking, customer ownership is closely associated with the account relationship. The bank that holds the current account, provides the mortgage or manages the investment has direct access to the customer and controls the primary service interface. In an agent-mediated environment, these relationships may become decoupled. The financial institution may continue to own the contractual and economic relationship associated with the product, while the AI platform increasingly owns the interaction through which the customer discovers and manages that relationship.

The distinction can be expressed as a movement from account ownership to decision-interface ownership. Account ownership concerns the institution that provides and administers the financial relationship. Decision-interface ownership concerns the institution that mediates how the customer understands available choices and determines which action to take. These forms of ownership need not reside with the same organisation.

This possibility has important implications for competition. If the AI intermediary controls product discovery, banks may increasingly compete not only for customers but for representation within the agent's decision environment. A bank's competitive advantage could therefore depend partly on whether its products are accessible to AI systems, how those products are represented, the quality and timeliness of the data available to agents, and the criteria through which the agent evaluates competing offerings. Competition could consequently shift from competing solely for customer attention to competing for algorithmic consideration.

Such a shift would extend the platform dynamics identified by Vives (2019) and Qi (2022). Digital disruption initially weakened the exclusive relationship between banks and their physical distribution channels; fintech subsequently introduced new digital interfaces and specialised services; agentic AI could go further by mediating the customer's interaction with the entire financial ecosystem. The progression is therefore not necessarily from banks to technology companies, but from institution-controlled interfaces towards platform-mediated interfaces.

The strategic significance of this development should not, however, be overstated. An AI platform may become an important interface without becoming the dominant intermediary. Consumers may continue to prefer direct relationships with banks, particularly for high-value or sensitive decisions. Regulatory requirements may restrict autonomous financial activity, while banks may develop their own agentic capabilities or establish direct relationships with third-party agents. Moreover, as Robinson, Dörry and Derudder (2024) demonstrate in the context of financial infrastructure, platformisation can coexist with persistent institutional power rather than simply replacing incumbents.

Nevertheless, the possibility of interface intermediation provides a useful framework for analysing the emerging relationship between banks and AI platforms. The critical question is no longer simply whether technology companies will enter banking or whether banks will adopt artificial intelligence. It is whether the organisation that controls the customer's decision interface will become strategically more important than the organisation that provides the underlying financial product.

If that occurs, the competitive structure of banking could change without any corresponding disappearance of banks. Banks could remain the institutions that hold deposits, provide credit, manage investments and execute regulated transactions, while AI platforms become the institutions through which customers decide which of those services they want and from whom they want them.

The consequence is a potentially fundamental redistribution of power within financial intermediation. The bank may continue to own the account, but the platform may increasingly own the decision. It is this possibility—the movement from control of the financial relationship to control of the financial decision interface—that provides the foundation for examining AI platforms as potential gatekeepers of banking.

3. The Gatekeeper Thesis

The central proposition emerging from Kamm's (2026) analysis is that a technology platform does not need to become a bank in order to exercise significant influence over banking. This challenges the conventional framing of Big Tech's relationship with financial services, which tends to ask whether technology companies will enter regulated financial markets and compete directly with incumbent institutions. A potentially more consequential question is whether the opposite process is occurring: whether financial services are increasingly being mediated through the customer interfaces of technology platforms.

This distinction matters because the economic significance of an intermediary does not necessarily derive from ownership of the underlying product. As the literature on platform power demonstrates, an organisation can acquire substantial influence by controlling access between different groups of market participants (Qi, 2022). In financial services, this suggests that a technology platform could become strategically important without accepting deposits, underwriting loans or managing investments itself. Its power could instead arise from determining how customers encounter financial products and how those products are evaluated before a financial institution ever becomes directly involved.

Agentic artificial intelligence potentially intensifies this dynamic because it can occupy a position between a customer's intention and the financial market. A conventional search engine provides information in response to a query, while a comparison platform presents alternatives for the customer to evaluate. An agentic system could potentially perform a more extensive sequence of activities: interpreting the customer's objective, identifying relevant providers, evaluating alternatives, making a recommendation and, where authorised, initiating the resulting action. The system therefore has the potential to influence not only what the customer sees, but what the customer considers and ultimately what the customer does.

This progression can be conceptualised as three interconnected forms of gatekeeping: discovery, recommendation and execution. They should not be regarded as entirely separate forms of power. Rather, they represent successive points at which an AI intermediary can intervene in the formation and implementation of a financial decision. The significance of the gatekeeper thesis increases as the system moves from determining which alternatives are visible, through influencing which alternative is preferred, towards having the authority to act on the customer's behalf.

3.1 Discovery Power

The first form of gatekeeping arises through control over discovery. Before a customer can choose between financial products, a set of potential alternatives must first be constructed. In a conventional financial market, this task is performed partly by the customer, who searches for providers, visits websites, consults advisers and compares products. Financial institutions also influence discovery through advertising, branch networks, search-engine optimisation, partnerships and the design of their digital channels.

An AI intermediary could substantially alter this process. A customer who asks, "What is the best mortgage for me?" is unlikely to receive an exhaustive representation of every mortgage available in the market. The AI system must first determine which products and providers are relevant enough to consider. That determination may depend upon the customer's stated objectives, the information available to the system, the accessibility of provider data and the criteria used to construct the candidate set.

The economic significance of this process is considerable. Competition traditionally assumes that consumers can compare products that are meaningfully available to them. Yet if an AI intermediary determines which products enter the comparison set, competition may occur one stage earlier. The relevant question is no longer simply which bank offers the most attractive mortgage, but which banks and mortgages are sufficiently visible to be considered by the system in the first place.

This represents a form of pre-market selection. A provider can be disadvantaged without its product being explicitly rejected by the customer because the product may never enter the customer's decision environment. In a conventional search process, exclusion can be visible: a consumer may choose one bank after considering several alternatives. In an AI-mediated process, exclusion may occur invisibly at the stage at which the system constructs the alternatives presented to the customer.

The concept of platform power developed by Qi (2022) is relevant here because the power of a platform can derive from its ability to structure interactions rather than simply from ownership of the goods or services exchanged. An AI system that determines which financial products are surfaced could therefore acquire a form of gatekeeping power even though the underlying products remain owned and provided by independent financial institutions.

The implications are particularly important because AI-mediated discovery could potentially become personalised. A conventional search engine generally responds to an explicit query, whereas an agent may possess a continuing understanding of the customer's financial circumstances, preferences and previous decisions. Subject to the permissions granted to it, the system could therefore determine relevance not merely at the level of the market but at the level of the individual.

This creates a shift from market discovery to personalised discovery. The customer does not necessarily ask which mortgage products exist; the customer asks which mortgage is appropriate for them. The distinction transfers part of the responsibility for constructing the relevant choice set from the consumer to the AI system.

3.2 Recommendation Power

Discovery determines which alternatives become visible. Recommendation determines how those alternatives are evaluated.

An AI agent that simply retrieves financial products would already influence the customer's information environment, but an agent capable of ranking alternatives would exercise a deeper form of gatekeeping power. The system would move from deciding what the customer can see towards influencing what the customer should consider preferable.

This is where the distinction between information provision and financial advice becomes increasingly difficult to maintain. A system that tells a customer that five mortgage products exist is performing a different function from a system that states that one of those mortgages is "the best option for you." The latter requires an interpretation of the customer's objectives and an assessment of competing alternatives against those objectives.

Research on algorithmic decision-making provides an important foundation for understanding this transition. Sargeant (2023) argues that algorithmic systems in financial services raise not only questions of efficiency and accuracy but also normative questions concerning autonomy, discrimination and the conditions under which individuals make decisions. Algorithms can influence behaviour not merely by making decisions themselves but by structuring the environment in which human decisions are made.

This is particularly relevant to AI agents because their recommendations can be expressed through conversational interaction. Rather than presenting an abstract score or ranking, the system may explain why one product appears preferable to another in terms that are personalised to the customer. The AI therefore has the potential to function as a choice architect.

The term is significant because choice architecture does not require coercion. An individual may remain formally free to reject a recommendation while nevertheless being strongly influenced by the way alternatives have been selected, ordered and explained. The system can shape the decision environment without making the final decision itself.

The apparently simple question of what constitutes "the best mortgage" illustrates the problem. Lowest interest rate, lowest total cost, greatest repayment flexibility, fastest approval, lowest risk or greatest compatibility with the customer's future plans could each produce different answers. The process of determining which criterion should dominate is not merely a technical optimisation problem. It involves an interpretation of the customer's values and priorities.

Consequently, the interpretation of the objective becomes part of the financial decision.

This observation has broader implications for AI-based financial advice. Brüggen et al. (2025) argue that AI financial advice raises ethical questions concerning autonomy and responsibility precisely because the system does not simply communicate information; it participates in the construction of recommendations. Baeckström and Matkovskyy (2026) similarly examine differences between human and AI financial advice behaviour, highlighting the emerging importance of understanding how financial advice changes when the adviser is an artificial rather than human actor.

The issue is therefore not whether AI recommendations are inherently better or worse than human recommendations. Chen et al. (2025) provide evidence that FinTech-based systems can reduce certain forms of human bias in lending, suggesting that algorithmic decision-making may sometimes improve consistency and decision quality. The more fundamental question is whose objectives the system is optimising and under what constraints.

This introduces the possibility of a conflict between the customer's interests and the platform's commercial incentives. If an AI platform receives compensation for directing customers towards particular financial providers, the system's ranking of alternatives may have consequences that are not apparent from the recommendation itself. The platform could consequently occupy a position analogous to an intermediary adviser while possessing commercial incentives that differ from those of the customer.

The gatekeeper problem therefore extends beyond algorithmic accuracy. It concerns the governance of the criteria through which recommendations are generated and the transparency of the relationship between those criteria and the interests of the customer.

3.3 Execution Power

The third and most consequential form of gatekeeping arises when an AI system moves from recommending an action to undertaking it.

The distinction between recommendation and execution is fundamental. A customer can remain the ultimate decision-maker when an AI system proposes a course of action, provided the customer meaningfully reviews and authorises that recommendation. Once the system is capable of initiating transactions or submitting applications without contemporaneous human intervention, however, the relationship changes from assisted decision-making to delegated agency.

Kamm (2026) identifies an important problem at this stage: the authority granted to an AI agent may need to operate across multiple financial institutions. A customer could, for example, instruct an agent to manage a specified amount of money within defined limits. The agent might then interact independently with several banks or financial service providers. Each institution may be able to authenticate the agent and verify that it is acting on behalf of the customer, but that does not necessarily mean that each institution can determine the customer's remaining authority across the wider financial ecosystem.

This exposes a fundamental distinction between identity and authority.

Identity answers the question of who is acting. Authority answers the question of what that actor is permitted to do. In an agentic financial environment, establishing that an AI system is genuinely acting on behalf of a particular customer would therefore be insufficient. The financial institution would also need to determine whether the particular instruction falls within the authority delegated by that customer.

This distinction becomes increasingly important as agents become capable of coordinating multiple actions. An instruction to "optimise my cash position" might result in transfers between several accounts, changes to savings products or applications for alternative financial products. The agent may understand the customer's overall objective, while each individual institution sees only a fragment of the resulting activity.

The problem is therefore one of distributed authority. The customer may establish a set of permissions at the level of the agent, while financial institutions encounter individual instructions at the level of specific transactions. Unless there is an interoperable mechanism for representing and verifying the customer's delegated authority, the institutions may lack visibility into the wider context in which an instruction has been generated.

This creates questions that conventional authentication systems are not designed to answer. A bank may be able to establish that an instruction originates from an authorised agent, but it may not know whether the customer has already committed part of the relevant financial authority elsewhere, whether the transaction remains consistent with the customer's current objectives, or whether the agent has exceeded the scope of its mandate.

The resulting problem is not simply technological. It concerns the allocation of legal and economic responsibility between the customer, the AI platform and the financial institution. If an agent acts incorrectly, responsibility cannot necessarily be assigned solely on the basis of which institution executed the transaction. The underlying decision may have been formulated by the AI, authorised by the customer and executed by the bank.

Agentic finance therefore requires a conception of authority that goes beyond authentication. It requires mechanisms through which delegated authority can be defined, constrained, verified, monitored and revoked. The technical capacity to establish that an AI agent represents a particular customer is only the first stage. The more difficult question is whether the financial ecosystem can establish what that agent is authorised to do at any particular moment.

This is also where the gatekeeper thesis reaches its strongest form. Discovery power influences what enters the customer's field of consideration. Recommendation power influences which option appears preferable. Execution power potentially converts that influence into action. Each stage therefore represents a deeper form of participation in the customer's financial decision process.

The three forms of power can consequently be understood as a continuum rather than as isolated functions. At the discovery stage, the agent controls visibility. At the recommendation stage, it influences preference. At the execution stage, it may exercise delegated agency. The progression is significant because each stage increases the potential economic and regulatory consequences of the platform's behaviour.

The distinction also helps clarify why the emergence of AI gatekeepers could matter even if banks remain the ultimate providers of financial products. A bank can continue to control the balance sheet, undertake credit assessment, hold deposits and execute payments while simultaneously losing influence over the earlier stages through which customers arrive at those services. The financial institution retains control over the product, but another institution increasingly influences the path by which the customer reaches it.

This produces a distinctive form of platform power. The AI platform does not necessarily control the financial asset; it controls the route to the financial asset. It does not necessarily make the final decision; it shapes the conditions under which the decision is made. And it does not necessarily execute every transaction; it may determine which transactions the customer ultimately instructs the bank to execute.

The gatekeeper thesis should therefore not be interpreted as a prediction that AI platforms will simply replace banks. Its more significant implication is that the competitive boundary of banking could move upstream, towards the processes of discovery, recommendation and delegation that precede the traditional banking relationship.

In this sense, the central competition may no longer be solely between banks for the customer's account. It may increasingly be between financial institutions and technology platforms for influence over the customer's decision process. If AI agents become the primary systems through which individuals formulate financial questions, evaluate alternatives and delegate actions, the organisation controlling that interface could acquire a strategically important position within the financial system without ever becoming a bank itself.

The gatekeeper of banking, in other words, may not be the institution that controls the customer's money. It may be the system that controls which financial possibilities the customer sees, which one appears preferable, and when the decision becomes an action.

4. Platform Power and the Changing Economics of Banking

The gatekeeper hypothesis can be situated within the broader literature on platformisation, which highlights how economic power increasingly depends not only on ownership of assets or provision of services, but also on control over the infrastructures through which users access markets. In this context, the significance of agentic AI lies not necessarily in its ability to replace banks, but in its potential to occupy a strategically important position between customers and financial institutions.

Qi (2022) argues that platform power derives in part from the ability of platforms to organise relationships between users and other economic actors, creating forms of dependency around the platform as an intermediary. This perspective is particularly relevant to banking because financial services are already characterised by multiple layers of intermediation. Banks provide regulated products and balance-sheet capacity, while payments networks, technology providers, fintech applications and digital distribution channels increasingly shape how those services are accessed and consumed. Platformisation therefore does not necessarily eliminate existing intermediaries; instead, it can reorganise the relationships between them.

Research on financial infrastructure provides an important parallel. Robinson, Dörry and Derudder (2024), for example, describe SWIFT as an “obligatory passage point” within global payments, illustrating how control over infrastructure can generate strategic power even when the organisation exercising that control does not directly provide the underlying financial service. Their analysis also demonstrates that platformisation is not necessarily a straightforward process of displacement. New technologies can challenge incumbent structures while simultaneously becoming incorporated into, or dependent upon, existing infrastructures.

This suggests that the future relationship between banks and technology platforms should not be conceptualised simply as a contest between “banks” and “Big Tech”. A more plausible outcome is a form of layered intermediation, in which different actors control different stages of the financial process. The AI platform may occupy the customer-facing layer, interpreting intentions, coordinating decisions and mediating access to financial products. The bank may remain responsible for regulated financial provision, balance-sheet intermediation and the contractual relationship underlying the product. Payment networks and other financial infrastructures may continue to execute transactions, while regulators and supervisory institutions provide the institutional framework within which these activities remain permissible and accountable.

The significance of this structure is that economic importance may become increasingly detached from direct ownership of the financial product. Banks could retain control over deposits, loans, investments and payment accounts while losing some degree of control over how customers discover, evaluate and select those products. In other words, the bank could remain the provider of the financial asset while another platform becomes the primary mechanism through which demand for that asset is organised.

This distinction is important because customer relationships have traditionally been a major source of competitive advantage in banking. The value of a banking relationship does not arise solely from the underlying product. It also derives from the institution’s ability to understand customer needs, present relevant products, provide advice and retain the customer across multiple financial interactions. Digital platformisation may progressively separate these functions. If an AI agent becomes the primary interface through which customers formulate and execute financial decisions, some of the informational and relational value historically associated with the bank could migrate to the platform.

This does not imply that banks necessarily become less important. Indeed, control over financial infrastructure and regulated balance-sheet capacity may remain exceptionally valuable. Platformisation can even reinforce the strategic importance of infrastructure providers, particularly where access to regulated systems, payment networks, customer accounts or financial data remains difficult to replicate. The issue is therefore not whether banks disappear, but where economic value and bargaining power accumulate within the emerging financial architecture.

For banks, this creates a potentially significant shift in the economics of customer acquisition and retention. If an AI platform controls the initial interaction with the customer, determines which products enter the consideration set and influences which option is ultimately selected, then the platform may capture part of the economic value traditionally associated with the customer relationship. This could occur through subscription models, referral or distribution economics, transaction-related revenues, data advantages, or preferential positioning within AI-mediated recommendations. The precise commercial model remains uncertain, but the underlying strategic issue is clearer: control over the customer interface can create economic value even without ownership of the underlying financial product.

The implication is that competition may increasingly occur between different layers of the financial system rather than solely between banks offering competing products. Banks may compete not only for customers, but also for visibility and favourable representation within the decision systems that increasingly mediate those customers. This introduces a new dimension of platform dependence. A bank could offer a highly competitive mortgage, savings account or investment product while nevertheless becoming commercially disadvantaged if an AI intermediary rarely presents that product to relevant customers.

The changing economics of banking can therefore be understood as a potential separation between product control and relationship control. Banks may continue to control the financial product, the balance sheet and the regulated contractual relationship, while AI platforms increasingly control the interface through which customer demand is expressed. The institution that controls the financial product may consequently not be the institution that captures the greatest economic value from the customer relationship.

This represents a deeper transformation than a conventional shift in distribution channels. If the customer’s primary interaction is with an AI agent rather than with the bank, the platform may become the economic gateway through which financial demand reaches regulated providers. The strategic question for banks is therefore no longer simply how to develop better products, but how to remain economically relevant within an ecosystem in which another actor may increasingly mediate the customer’s path to those products.

Platformisation, in this sense, does not necessarily eliminate financial intermediation. It may instead reallocate the most valuable layer of intermediation. The bank can remain the institution that holds the account, extends the loan or executes the regulated financial service, while the AI platform becomes the institution through which the customer decides which account, loan or service to choose. The resulting shift is from competition over the provision of financial products towards competition over access to, and influence within, the customer decision process.

5. The Consumer Autonomy Problem

The most important counterweight to the gatekeeper thesis concerns consumer autonomy. The prospect of delegating financial decisions to an AI agent may appear inherently empowering. Consumers face increasingly complex financial products, limited time and attention, and substantial differences in financial knowledge. An intelligent agent capable of searching across providers, comparing alternatives and explaining financial trade-offs could reduce these burdens and make financial decision-making more accessible.

Recent research suggests that there may be substantial consumer demand for such capabilities. Jaffri, Jha and Butt (2026), using nationally representative evidence from the United States, find that interest in AI-based financial advice is associated with factors including digital readiness, risk tolerance, mobile banking use, time constraints and perceived financial constraints. Their findings suggest that consumers may view AI not simply as a substitute for human advisers, but as a form of cognitive support that can help them manage the complexity and demands of financial decision-making.

This potential benefit is important. The traditional financial system imposes significant cognitive costs on consumers. Understanding different products, comparing fees and conditions, assessing risk and determining whether a financial decision is appropriate can require considerable expertise. An AI agent that can perform these tasks continuously and at relatively low marginal cost could reduce search costs and improve access to financial information. In this respect, agentic AI may expand consumer capability rather than merely automate existing advisory functions.

However, convenience and autonomy are not synonymous.

The central issue is delegated cognition. When a consumer asks an AI agent to “optimise my finances”, the consumer is not simply delegating a mechanical task. The consumer is delegating part of the reasoning process through which the task itself is defined and evaluated. The agent must determine what “optimisation” means, which objectives should take priority, which constraints are relevant and which trade-offs are acceptable.

These questions are rarely purely technical. Optimising a financial position could mean maximising expected returns, reducing risk, increasing liquidity, minimising fees, paying down debt, increasing long-term wealth or preserving financial flexibility. Different objectives can produce entirely different recommendations. The problem therefore begins before the agent selects an action: it arises when the agent interprets what the consumer is trying to achieve.

This creates an important distinction between formal authority and practical control. A consumer may formally retain authority over their financial accounts and may even be required to approve significant transactions. Yet practical control can become weaker if the consumer relies heavily on an AI system to determine which options are presented, how they are evaluated and which course of action appears most appropriate. The consumer remains the nominal decision-maker, but the cognitive process underlying the decision has partially migrated to the agent.

Brüggen et al. (2025) identify a closely related tension in their ethical analysis of AI-based financial advice. Their discussion distinguishes individual autonomy, understood as the capacity to make independent and informed decisions, from questions surrounding the autonomy and operation of algorithmic systems themselves. This distinction is particularly important in agentic finance because the more sophisticated the system becomes, the less plausible it may be to treat the consumer as independently generating the reasoning that leads to a decision.

The concern is therefore not simply whether an AI recommendation is accurate. A technically correct recommendation can still create an autonomy problem if the consumer does not understand the assumptions underlying it, cannot meaningfully challenge its reasoning, or is unaware of the alternatives that the system has excluded. Autonomy requires more than the existence of a final approval button. It requires that consumers retain meaningful capacity to understand, contest and alter the basis on which important decisions are made.

This leads to what can be described as the delegation paradox. The more capable an AI agent becomes at making financial decisions on behalf of a consumer, the greater the potential benefit of delegation. At the same time, however, the more reasoning the agent performs, the less meaningful it may become for the consumer to retain only nominal responsibility for the resulting decision. Increasing agent capability can therefore simultaneously increase consumer convenience and reduce the practical exercise of consumer autonomy.

The paradox becomes particularly significant when delegation is continuous rather than episodic. A consumer who asks an AI agent to compare two mortgage products remains relatively close to the underlying decision. A consumer who authorises an agent to monitor accounts, identify opportunities, rebalance investments, refinance debt and optimise cash holdings has delegated a much broader set of judgements. In the latter case, the consumer is no longer simply outsourcing information retrieval. They are outsourcing an ongoing process of financial reasoning.

This does not imply that consumers should avoid delegation. Delegation is fundamental to modern economic and social life. Consumers routinely delegate decisions to professionals, institutions, software and automated systems because complete personal control over every decision would be impractical. The relevant question is therefore not whether delegation occurs, but what conditions make delegation compatible with meaningful autonomy.

For agentic financial systems, three conditions are particularly important. Delegation should be informed, meaning that consumers understand the nature and scope of the authority they are granting. It should be bounded, meaning that the agent operates within clearly defined objectives, permissions and constraints rather than possessing unlimited discretion. Finally, it should be reversible, meaning that consumers can modify or withdraw delegated authority without disproportionate difficulty.

These conditions also change the way consumer protection should be understood. Traditional financial regulation often focuses on disclosure, suitability and consent at particular points in a transaction. Agentic AI introduces a more continuous relationship in which the system may repeatedly interpret preferences and initiate actions. Consumer protection may therefore need to address not only whether a particular transaction was authorised, but also how the authority underlying that transaction was established, interpreted and exercised over time.

The consumer autonomy problem consequently provides an important qualification to the gatekeeper thesis. The power of an AI platform does not arise solely because it can control what consumers see or recommend. It may also arise because consumers voluntarily transfer part of their decision-making capacity to the system. Platform power and consumer delegation can therefore become mutually reinforcing: the more capable and convenient the agent becomes, the more consumers may rely upon it; and the greater that reliance becomes, the greater the platform's influence over financial decisions.

The central challenge is thus not to prevent delegation, but to ensure that delegation does not become a substitute for meaningful consumer agency. An effective financial AI system should make delegation easier without making control merely symbolic. The objective should be a relationship in which consumers can benefit from machine reasoning while retaining meaningful authority over the objectives, constraints and consequences of the decisions made on their behalf.

6. Automation Bias and the Problem of Trust

The gatekeeper problem is compounded by the possibility of automation bias: the tendency for individuals to place excessive reliance on recommendations or outputs generated by automated systems. This concern is particularly significant when agentic AI moves beyond providing information and begins to influence or execute consequential financial decisions.

Skitka, Mosier and Burdick (1999) demonstrated that users can over-rely on automated decision aids, including in circumstances where automated recommendations conflict with otherwise available evidence. Their research identified both omission errors, in which users fail to act because an automated system does not recommend action, and commission errors, in which users follow an inappropriate automated recommendation. The significance of these findings extends beyond the particular systems studied. They demonstrate that the presence of automation can alter the way individuals process information and allocate responsibility for decisions.

This dynamic may become particularly important in financial services because contemporary AI systems do not merely produce numerical outputs. They can communicate through natural language, explain their recommendations, respond to questions and maintain an apparently continuous understanding of the consumer's circumstances. These characteristics can create a perception of competence and understanding that is qualitatively different from the experience of interacting with a conventional algorithmic score.

Consider the difference between an opaque output such as “algorithmic score: 0.82” and an explanation such as: “I recommend Bank A because it provides the best overall fit with your liquidity requirements and risk profile.” The latter does more than communicate a result. It constructs a narrative of reasoning. The system appears to have understood the consumer's circumstances, considered relevant trade-offs and reached a conclusion on the consumer's behalf.

This apparent understanding may increase the persuasive force of the recommendation. Natural-language interaction can make algorithmic outputs easier to interpret, but it can also make them easier to trust. The danger is therefore not simply that consumers misunderstand an AI recommendation. They may understand it sufficiently well to accept it while failing to question whether the system's underlying assumptions, objectives or information are appropriate.

This distinction becomes particularly important when considering the relationship between trust and autonomy. Trust can reduce the cognitive burden associated with decision-making. A consumer who trusts an AI adviser does not need to independently verify every calculation or investigate every available alternative. In this sense, trust is an important condition for useful delegation. Yet excessive trust can also undermine the very autonomy that delegation is intended to support.

Recent research reinforces the importance of trust in AI-based financial advice. Sungkarungsri and Kiattisin (2026), for example, examine credibility-based and benevolence-based dimensions of trust in AI financial advisory services, highlighting that perceptions of whether an AI system is competent and whether it acts in the user's interests can influence adoption and advocacy. These dimensions are particularly relevant to agentic finance because the system is not simply answering questions. It may be entrusted with increasingly consequential decisions and actions.

However, perceived trustworthiness is not equivalent to institutional accountability.

A consumer may perceive an AI agent as credible, competent and benevolent while having little understanding of which organisation is responsible for the recommendation. This creates a potential accountability gap. The AI agent may be operated by a technology platform, obtain financial products from banks, rely upon third-party models or data providers, and ultimately initiate transactions through regulated financial institutions. Responsibility is consequently distributed across several actors even though the consumer experiences the process as a single interaction.

This creates an important asymmetry. From the consumer's perspective, there may be one trusted agent. From an institutional perspective, there may be several organisations with different responsibilities, incentives and legal relationships. The consumer therefore experiences integrated agency, while the underlying system may consist of fragmented accountability.

The problem becomes more acute when an AI agent makes a recommendation that produces a harmful outcome. The bank may argue that it merely provided the product requested by the customer. The platform may argue that it only facilitated access to the bank. The model provider may argue that it supplied a general-purpose technology rather than financial advice. Yet from the consumer's perspective, the decision may have been presented as a coherent recommendation by a single intelligent system.

This creates a fundamental governance question: who is accountable for a decision made through an AI-mediated financial relationship?

The answer cannot depend solely on whether the consumer formally authorised the transaction. As discussed in the previous chapter, authorisation does not necessarily demonstrate meaningful control over the reasoning that produced the decision. Nor should perceived trust in the AI system be treated as evidence that the consumer understood the allocation of responsibility between the different institutions involved.

The problem can therefore be understood as a three-way distinction between trust, authority and accountability. Trust concerns whether the consumer believes the system is competent and acts appropriately. Authority concerns what the consumer has permitted the system to do. Accountability concerns which institution is responsible when the system's behaviour produces an undesirable outcome. These three concepts may align, but agentic financial systems create circumstances in which they can increasingly diverge.

This divergence has important implications for the design of AI-mediated financial services. A trustworthy system should not simply appear reliable. It should make the boundaries of its authority intelligible and provide consumers with meaningful information about the institutions responsible for different parts of the decision process. Likewise, an AI system that is capable of acting autonomously should not be allowed to create the impression that responsibility disappears simply because the consumer delegated the decision.

The automation-bias problem therefore reinforces the central argument of this paper. The power of an AI financial platform may arise not only from controlling what consumers can see, but from influencing how readily they accept what they see. If natural-language systems become highly persuasive, the gatekeeper function may extend from controlling the consideration set to shaping the consumer's willingness to challenge the resulting recommendation.

The governance challenge is consequently not to eliminate trust. Trust is necessary for useful delegation and widespread adoption. The objective should instead be to prevent trust from becoming a substitute for accountability. Consumers should be able to rely on AI systems without being required to assume that the system is infallible, benevolent or ultimately responsible for every consequence of its actions.

In this sense, the most important question is not whether consumers trust AI financial agents. It is whether trust can coexist with meaningful oversight, contestability and institutional responsibility. As agentic AI becomes more capable of making and executing financial decisions, the quality of the financial relationship may increasingly depend on maintaining that distinction.

7. Efficiency Does Not Necessarily Mean Consumer Welfare

The case for agentic finance remains substantial. AI systems could reduce the search costs associated with increasingly complex financial markets, improve access to financial information, assist consumers with financial planning and automate administrative tasks that currently require substantial time and effort. By continuously comparing products and monitoring changing circumstances, agentic systems could also make financial services more responsive to individual needs.

The potential benefits extend beyond convenience. AI may also reduce certain forms of human bias within financial decision-making. Evidence from lending provides an important illustration. Chen et al. (2025) find that algorithmic decision-making can reduce certain cognitive biases among loan officers, particularly when algorithms influence decision defaults rather than merely provide optional recommendations. This suggests that automation does not necessarily undermine fairness or consumer welfare. Under appropriate conditions, it may correct some of the inconsistencies associated with human judgement.

Agentic AI should therefore not be characterised simply as a threat to consumer autonomy. The more important question is whether the objectives pursued by the system remain aligned with the interests of the consumer.

This can be expressed through a central question: who controls the optimisation function?

When a consumer instructs an AI agent to “optimise my finances”, the instruction appears straightforward but is conceptually ambiguous. Optimisation always requires an objective. The system must determine what should be maximised or minimised, over what time horizon and subject to which constraints. A system designed to minimise financial costs might prioritise lower fees and interest rates. A system designed to maximise expected wealth might accept greater risk. A system designed to preserve liquidity might recommend an entirely different portfolio.

These differences matter because the optimisation function determines the outcomes that the AI treats as preferable. The technical capability to optimise is therefore not sufficient to establish that the resulting decisions are beneficial to the consumer. Consumer welfare depends partly on who defines the objective that the system is optimising.

Consider an AI agent instructed to minimise a customer's financial costs. In principle, such a system could generate substantial consumer benefits by identifying cheaper insurance, lower mortgage rates, more competitive savings products or lower-cost investment options. If the agent can search across providers and compare products more efficiently than an individual consumer, it may substantially improve the consumer's position.

The situation becomes more complicated when the platform operating the agent has its own commercial interests. Suppose the platform receives commissions from particular financial providers, earns advertising revenue, operates preferred-provider arrangements or derives economic value from directing customers towards particular products. The system may still describe its recommendations as an optimisation of the customer's financial position, while the underlying commercial environment creates incentives that point in a different direction.

The problem is not necessarily that the platform will deliberately act against the consumer. Rather, the concern is that the consumer's objective function and the platform's commercial objective may become misaligned. A recommendation can therefore be technically rational according to the system's internal criteria while simultaneously failing to maximise the outcome that the consumer would have chosen if fully informed about those criteria.

This creates a distinctive form of principal-agent problem. In a conventional principal-agent relationship, one party delegates authority to another whose actions may be difficult to observe and whose incentives may not perfectly correspond with those of the principal. Agentic finance introduces a more complex architecture in which the consumer delegates decision-making capacity to an AI platform that simultaneously interacts with multiple financial providers.

The resulting relationship can be represented conceptually as:

Customer → AI platform → Financial providers

The AI platform occupies an intermediate position between demand and supply. It may possess detailed information about the customer's preferences, financial circumstances and behavioural patterns while also having access to information about the products, pricing and commercial incentives of multiple providers. This intermediary position can generate substantial informational advantages.

The platform may consequently know more about the consumer than any individual financial provider, while simultaneously knowing more about the available financial market than the consumer. This creates a potentially powerful informational position. The platform can observe what customers request, which recommendations they accept, which products they reject and how their preferences evolve over time. Such information may itself become an important source of competitive advantage.

The economic significance of this information asymmetry is that the platform may be able to influence both sides of the market. On the consumer side, it controls how financial options are presented and compared. On the provider side, it may control access to a significant source of customer demand. This creates the possibility of a feedback loop in which greater customer adoption generates more data, improved personalisation increases platform usefulness, and increased platform usefulness generates further customer dependence.

Platform power can therefore emerge even without explicit coercion. If consumers find the AI agent useful and providers increasingly depend upon it for customer acquisition, the platform may acquire bargaining power over both sides of the relationship. This is consistent with the broader literature on platform power, in which control over market access and information can become economically significant even where the platform does not directly provide the underlying service (Qi, 2022).

The key issue is therefore not whether agentic AI produces efficiency. It very plausibly can. The issue is whose interests determine how that efficiency is deployed. A system can reduce search costs while simultaneously narrowing the range of products consumers encounter. It can provide personalised recommendations while embedding commercial preferences into those recommendations. It can reduce human bias while introducing new forms of platform-level influence.

This distinction is particularly important because efficiency can make commercial influence harder to detect. If an AI agent consistently produces convenient and apparently high-quality outcomes, consumers may have little incentive to investigate how recommendations are generated. The very usefulness of the system can therefore increase reliance upon it. Efficiency and trust may reinforce one another, even where the platform's incentives are not perfectly aligned with those of the consumer.

Consumer welfare consequently depends not only on the accuracy of the AI's recommendations but also on the governance of the optimisation process. Consumers need to know, at an appropriate level of transparency, what objectives the system is pursuing, what constraints it operates under and whether commercial relationships influence the options it presents. Without such safeguards, the optimisation function risks becoming a hidden mechanism through which platform interests are incorporated into consumer financial decisions.

The central economic question is therefore not whether agentic AI will make financial markets more efficient. It is whether the efficiency gains generated by AI will accrue primarily to consumers, financial providers, AI platforms, or some combination of the three. The answer will depend substantially on who controls the interface, who defines the optimisation function and who bears responsibility when commercial incentives conflict with consumer interests.

Agentic finance may consequently create a paradox of its own. The more effectively an AI system reduces the friction of financial decision-making, the more influence the system may acquire over those decisions. Efficiency can therefore become a source of platform power. The challenge for consumer welfare is to ensure that this power remains aligned with the interests and objectives of the consumers whose decisions the system is designed to optimise.

8. Implications for Banks

If the gatekeeper thesis is correct, banks should not interpret agentic AI primarily as another iteration of digital customer service. The strategic significance of agentic AI lies deeper than the automation of existing interactions. It potentially changes the location at which the customer relationship is formed, managed and influenced.

For much of the digital banking era, banks have attempted to strengthen their relationship with customers by bringing more services into their own digital interfaces. Mobile applications, online banking platforms and integrated financial tools have allowed banks to control an increasing proportion of the customer journey. Agentic AI could challenge this model by placing another interface between the customer and the bank. The customer may continue to hold an account with the bank while interacting primarily with an external AI system that determines when the bank becomes relevant to a particular financial decision.

The strategic challenge is therefore not simply to build better chatbots. It is to remain economically and institutionally relevant when the interface through which customers make financial decisions may no longer belong to the bank.

8.1 Banks should own a financial confidence layer

Banks should seek to remain the trusted and authoritative institution through which customers can understand the financial consequences of their decisions, even if an external AI agent becomes the primary conversational interface.

This does not necessarily require banks to own the customer's interface. Attempting to compete directly with every general-purpose AI platform may be neither realistic nor strategically necessary. Instead, banks could focus on controlling the quality, reliability and verifiability of the financial information that their products and services expose to AI systems.

This would require banks to make relevant financial information accessible in machine-readable forms while preserving the contextual information needed for accurate interpretation. Product characteristics, fees, eligibility criteria, risks, contractual conditions and relevant explanations would need to be presented in ways that authorised AI systems can reliably access and interpret.

The objective would be to create what might be described as a financial confidence layer: an institutional layer through which the bank remains the authoritative source of information about its products and obligations, even when another platform controls the customer's immediate interface.

This distinction is strategically important. If banks allow external AI systems to become the primary interface without ensuring that those systems can reliably access authoritative information, the bank risks losing influence over how its products are represented. Conversely, if banks can provide verifiable information, transparent product logic and reliable execution capabilities, they may remain valuable participants within an AI-mediated financial ecosystem.

8.2 Banks should become agent-ready

Banks will also need to develop the technical and institutional infrastructure required to interact safely with authorised AI agents. Becoming “agent-ready” means more than allowing an AI system to log into an existing digital banking interface. It requires banks to recognise that an AI agent may act as an authorised intermediary with a defined but potentially complex scope of authority.

This creates a need for more granular forms of permission and authentication. A bank should be able to distinguish between an agent that is permitted to retrieve account information, an agent that may prepare an application, and an agent that is authorised to initiate a transaction. Authority should therefore be attached to specific actions and contexts rather than treated as a binary question of whether an agent is generally authorised to act.

Such an architecture would require machine-readable product information, standardised mechanisms for delegated permissions and transaction-level authorisation. It would also require real-time authority checking, comprehensive audit trails and reliable mechanisms through which customers can modify or revoke permissions. Limits on transaction values, product categories, frequency of actions or permissible counterparties could provide additional boundaries around delegated agency.

The underlying principle is that identity should not be treated as equivalent to authority. Knowing that a particular AI agent is acting on behalf of a particular customer does not, by itself, establish that the agent is authorised to perform a particular action. Agent-ready banking therefore requires the bank to verify not only who is acting, but what that actor is permitted to do in the specific circumstances.

Human confirmation will remain important for particularly consequential actions. However, requiring human approval for every interaction would undermine much of the value of agentic systems. The more appropriate approach is likely to involve graduated forms of authority in which low-risk, routine actions can be delegated while significant or irreversible decisions trigger additional confirmation requirements.

8.3 Banks should distinguish advice from execution

The emergence of external AI agents also makes it increasingly important to distinguish between recommendation, authorisation and execution.

An external AI system may recommend a particular mortgage, investment product or savings account, but the bank should remain responsible for determining whether the resulting application satisfies its regulatory, credit and operational requirements. The AI recommendation should not substitute for the institutional processes through which the bank establishes eligibility, suitability, compliance and contractual validity.

At the same time, banks should not assume that the authentication of an instruction automatically makes that instruction appropriate to execute. An authenticated request may still exceed the authority that the customer intended to delegate. This is particularly important where an AI agent is capable of interpreting broad instructions and converting them into specific actions.

Kamm's distinction between authority and execution is therefore significant. Establishing the identity of an agent does not establish that a particular action falls within the customer's authorised mandate. An agent may be genuine and technically authenticated while nevertheless acting outside the scope of the authority granted to it.

Banks will consequently need mechanisms for interpreting and enforcing delegated authority at the point of execution. This could become a fundamental component of agentic banking infrastructure. Rather than treating delegated authority as a one-time consent event, banks may need to treat it as an ongoing, verifiable set of permissions that can be checked against individual actions.

Such an approach would also provide greater clarity about institutional responsibility. The AI platform may assist with discovery and recommendation, while the bank retains responsibility for the regulated execution of the financial product. Clearly separating these functions could allow banks to participate in agentic ecosystems without surrendering their institutional accountability.

8.4 Banks should compete on trust, not only convenience

If AI agents increasingly commoditise financial product discovery, banks may find it more difficult to differentiate themselves through digital interfaces alone. A customer may no longer choose a bank because its application is easier to navigate or because its website provides a more convenient comparison experience. Instead, the customer may encounter the bank only through an AI-mediated recommendation.

This creates a strategic question that banks will increasingly need to answer: why should an AI agent recommend us?

The answer cannot depend solely on brand recognition. If the AI system evaluates competing providers according to product characteristics, cost, reliability and customer outcomes, established brand advantages may become less decisive. Banks may therefore need to compete on attributes that can be demonstrated and verified by both customers and AI systems.

These attributes could include product quality, transparent pricing, reliability, data governance, security, regulatory compliance and demonstrable customer outcomes. The ability to provide accurate and machine-readable information may itself become a competitive capability. A bank whose products can be reliably understood, compared and verified by authorised AI systems may be better positioned than one whose value proposition depends heavily on brand familiarity or human sales processes.

Trust may consequently become an even more important strategic asset. In an environment where AI agents mediate customer decisions, trust operates at two levels. Customers need confidence that their bank will protect their interests and execute authorised instructions appropriately, while AI systems need reliable evidence on which to base recommendations. Banks therefore need to become trustworthy not only to their customers, but also to the technological systems through which customers increasingly access financial services.

This does not mean that banks should attempt to maximise their visibility within every AI system at any cost. Such an approach could simply reproduce the problems of advertising and platform dependence in a new technological form. Instead, banks should seek to establish a reputation for reliable, verifiable and appropriately governed financial products.

The broader strategic implication is that banks may need to shift from thinking primarily about owning the customer interface to thinking about remaining indispensable within the customer's decision architecture. The bank may not control the conversation, but it can still control the quality of the financial product, the integrity of its information, the security of execution and the institutional accountability that ultimately stands behind the transaction.

If agentic AI becomes the dominant interface to financial services, the winning banks may therefore not be those with the most sophisticated chatbot. They may be those whose products, information and infrastructure can be trusted by both consumers and the AI systems acting on their behalf.

9. Implications for Regulation

The emergence of AI financial agents creates a fundamental challenge for financial regulation because responsibility may become distributed across several actors while the consumer experiences the process as a single interaction. Traditional regulatory frameworks are often organised around identifiable institutions: a bank provides a financial product, an adviser provides advice, a payment institution executes a transaction, and a technology provider supplies infrastructure. Agentic AI can blur these boundaries by connecting several functions within a single decision process.

The resulting regulatory problem is therefore not simply whether an AI system is accurate or compliant. It is whether the entire chain through which a consumer's intention is transformed into a financial action remains appropriately governed.

Consider a hypothetical sequence in which a consumer asks an AI agent to optimise their savings. The agent interprets what “optimise” means, searches across multiple banks, excludes some providers because their systems are incompatible, ranks the remaining alternatives and recommends one. The consumer accepts the recommendation, after which the agent initiates a transfer and the receiving bank executes the transaction. The consumer subsequently experiences a financial loss.

The conventional regulatory instinct might be to ask whether the bank acted improperly. Yet the bank may have executed an apparently valid and authorised transaction. The AI platform may argue that it merely followed the consumer's instructions. The underlying model provider may argue that it supplied general-purpose technology rather than financial advice. The consumer, meanwhile, may regard the entire process as a single AI-mediated decision.

The question therefore becomes: who is responsible when a financial decision is distributed across an AI-mediated transaction chain?

The answer cannot simply be “the bank”. Nor should responsibility automatically be assigned to the AI provider. A more appropriate regulatory framework would distinguish between the different functions through which consumer intention is transformed into financial action.

The first is interpretation responsibility: who determined what the consumer actually meant? If a customer asks an AI agent to “optimise my savings”, someone—or something—must translate that broad instruction into specific objectives concerning returns, risk, liquidity, time horizon and other constraints. Errors at this stage can influence every subsequent stage of the process.

The second is recommendation responsibility: who determined which alternatives should be considered and how they should be ranked? This includes not only the final recommendation, but also the composition of the consideration set. An AI system that excludes particular providers before comparison can influence the outcome even if the eventual recommendation is factually accurate. This connects directly to the gatekeeper function identified earlier in the paper.

The third is execution responsibility: who had the authority to turn a recommendation into an actual financial transaction? This is particularly important where an AI agent operates across multiple institutions. Authentication of the agent establishes who is acting, but it does not necessarily establish whether the specific action falls within the authority delegated by the consumer.

The fourth is outcome responsibility: who bears responsibility when the overall process produces consumer harm? This question is more difficult because harmful outcomes may arise from several interacting decisions rather than from a single identifiable error. A recommendation may have been reasonable given the information available, the bank may have executed the transaction correctly, and yet the consumer may still experience a significant loss.

These distinctions suggest that regulation should not focus exclusively on the AI model itself. Model-level governance remains important, particularly in relation to reliability, transparency, testing and risk management. However, the model is only one component of an agentic financial system. Regulating the model without regulating the way it interacts with consumers, platforms, financial institutions and payment infrastructure may leave significant gaps in accountability.

A more appropriate approach may therefore be described as transaction-chain governance. The relevant regulatory object is not simply the AI system, but the chain connecting consumer intention to financial action.

This approach would require regulators to examine where authority enters the system, how consumer objectives are interpreted, how alternatives are selected, how recommendations are generated, how delegated authority is verified and how transactions are ultimately executed. It would also require clear allocation of responsibility between the different institutions involved.

Such an approach is consistent with the broader logic of financial regulation, which frequently assigns responsibility according to function rather than technology alone. The fact that a financial activity is mediated by AI should not automatically remove it from existing regulatory categories. Instead, the emergence of agentic systems may require those categories to be reconsidered where the functions performed by different actors no longer map neatly onto traditional institutional boundaries.

The concept of transaction-chain governance also provides a basis for addressing the accountability problem identified in earlier chapters. If an AI platform interprets consumer objectives and ranks financial products, it should not necessarily assume the same responsibilities as the bank that executes the resulting transaction. Conversely, the bank should not be able to avoid responsibility for its own regulated activities merely because the customer arrived through an external AI system.

The allocation of responsibility should therefore follow control, capability and function. An actor that determines how a consumer's objective is interpreted should bear responsibility for that interpretive process. An actor that controls which products are presented or recommended should bear responsibility for the relevant recommendation mechanisms. An institution that executes a transaction should remain responsible for the integrity and legality of that execution. Where several actors materially contribute to an outcome, regulation may also need mechanisms for determining shared or sequential responsibility rather than forcing harm into a single institutional category.

This becomes particularly important as AI systems acquire greater autonomy. A relatively simple chatbot that answers questions creates a different regulatory problem from an agent that continuously monitors a consumer's finances, selects products, negotiates with providers and initiates transactions. The greater the system's delegated authority and capacity for autonomous action, the greater the need for clearly defined responsibility at each stage of the process.

Regulation may therefore need to place greater emphasis on traceability. An agentic financial transaction should ideally leave an auditable record of the consumer's instruction, the authority granted to the agent, the information and constraints considered, the alternatives evaluated, the recommendation produced and the final action taken. Such records would not eliminate disputes, but they would make it possible to reconstruct how a financial decision was produced and which actors exercised meaningful control at each stage.

This is also important for consumer protection. Consumers cannot meaningfully contest an AI-mediated decision if they cannot determine what the agent was authorised to do, why a particular recommendation was produced or which institution was responsible for executing it. Contestability therefore requires more than access to a human customer-service representative. It requires sufficient institutional traceability for the consumer, regulator or court to reconstruct the decision process.

The regulatory challenge is consequently not to prevent AI agents from participating in financial services. Agentic systems may produce significant benefits through lower search costs, improved accessibility, reduced administrative burdens and potentially more consistent decision-making. The challenge is to ensure that these benefits do not come at the cost of regulatory ambiguity.

The central principle should therefore be that delegating a financial decision to an AI system must not mean delegating away accountability. As decision-making becomes distributed across consumers, AI platforms, financial institutions and infrastructure providers, responsibility must remain identifiable even when the decision process itself becomes technologically complex.

Agentic finance may therefore require a shift from model-centric governance to transaction-chain governance. The critical regulatory question is no longer simply whether an AI model is safe or accurate in isolation. It is whether the entire institutional and technological chain connecting consumer intention to financial action is transparent, bounded, auditable and accountable. This represents a significant extension of existing AI governance, but it is also a necessary response to the changing structure of financial intermediation.

10. Towards a Theory of Delegated Financial Agency

The preceding analysis suggests that agentic AI may create a distinct form of economic and institutional relationship that can be described as delegated financial agency. The concept captures a situation in which consumers do not simply obtain information from an AI system, but delegate part of the process through which financial objectives are interpreted, alternatives are selected and actions are undertaken.

This distinction is important because not all forms of AI involvement in finance represent the same degree of delegation. An AI system that retrieves information for a consumer performs a fundamentally different function from an agent that interprets a broad financial objective, selects among competing providers and executes transactions on the consumer's behalf. The economic and regulatory significance of the technology increases as decision-making authority moves progressively from the consumer towards the agent.

Delegated financial agency can therefore be understood as a progression through four related stages.

The first is objective delegation. The consumer specifies a goal rather than a precise action. Instead of instructing an agent to purchase a particular financial product, the consumer may ask it to improve savings, reduce borrowing costs or optimise their financial position. The consumer establishes the desired outcome while leaving significant discretion concerning how that outcome should be achieved.

The second is interpretive delegation. The agent determines what actions could plausibly satisfy the consumer's objective. This requires the system to translate an ambiguous human goal into operational criteria. It may need to determine relevant time horizons, risk preferences, liquidity requirements, costs and other constraints. At this stage, the agent is no longer simply processing an instruction; it is participating in the construction of the decision itself.

The third is provider selection. The agent determines which financial institutions or products should be considered and how they should be ranked. This is where the gatekeeper function becomes particularly significant. The agent can influence not only which option is ultimately selected, but which options enter the consumer's consideration set in the first place.

The fourth is execution delegation. The agent is authorised to initiate or complete financial actions. The distinction between recommendation and action becomes less meaningful because the system can translate its own reasoning into transactions without requiring the consumer to manually implement every decision.

These stages should not be understood as rigid categories. Rather, they represent increasing degrees of delegated agency. A consumer may delegate some aspects of a decision while retaining control over others. The significance of the framework lies in identifying the point at which AI assistance becomes AI-mediated decision-making and, ultimately, delegated action.

This progression also provides a useful way of understanding the changing economic significance of AI in finance. Information retrieval represents relatively limited delegation because the consumer remains responsible for interpreting the information and deciding what to do. Financial explanation involves greater involvement because the system influences how the consumer understands the available options. Personalised recommendation increases the agent's influence further because it begins to determine which option appears preferable. Provider selection gives the system greater gatekeeping power by influencing which institutions compete for the customer's business. Transaction initiation represents a further transition because the agent acquires the capacity to convert its recommendation into an actual financial action.

The most significant stage, however, may be autonomous multi-provider optimisation. At this point, the agent is no longer making isolated recommendations. It may continuously coordinate financial decisions across several institutions and product categories. It could, for example, monitor a consumer's current account, savings, mortgage, insurance and investments simultaneously and make adjustments according to a common set of objectives.

This represents a qualitative change in the nature of delegation. The agent is no longer simply assisting with individual financial decisions. It becomes an ongoing coordinator of the consumer's financial position.

The distinction matters because financial products are interconnected. A decision to hold more cash affects investment choices; a change in mortgage debt affects liquidity; an insurance decision affects household risk exposure; and changes in interest rates can alter the relative attractiveness of multiple products simultaneously. An agent capable of optimising across these relationships could therefore perform a role closer to financial coordination than conventional product recommendation.

The economic significance of delegated financial agency consequently increases as the agent moves from information provision to decision formation and from decision formation to execution. This provides a useful conceptual continuum for distinguishing relatively low-impact AI assistance from forms of AI activity that could have substantial implications for financial markets.

At the individual level, the benefits could be considerable. A capable agent may reduce search costs, improve financial planning and enable consumers with limited time or financial expertise to manage increasingly complex financial relationships. Delegation could therefore expand practical access to sophisticated financial decision-making.

At the same time, however, increasing delegation creates a corresponding concentration of decision-making power. If consumers rely on a small number of AI systems to determine which financial products they encounter, which institutions they use and how their assets are allocated, those systems could become important coordination points within the financial system.

This is where the concept of scale becomes critical. An AI agent making decisions for one consumer has limited systemic significance. An AI platform making or influencing millions of similar decisions may have substantially different economic effects. If large numbers of consumers use the same or a small number of AI agents to allocate deposits, select mortgages, switch insurers or purchase investments, the optimisation logic embedded within those systems could influence the distribution of financial demand across institutions.

Such influence could arise without the platform directly owning financial assets or becoming a regulated bank. The platform's power would instead derive from its position as an intermediary through which large volumes of financial decisions are coordinated.

This possibility connects the concept of delegated financial agency to the literature on platform power. Qi (2022) argues that platformisation can generate forms of economic power that extend beyond the direct provision of goods or services, particularly where platforms occupy strategically important positions between users and other economic actors. Applied to finance, the implication is that control over the decision interface could become economically significant even where control over the underlying financial assets remains with conventional institutions.

The systemic dimension should nevertheless be treated carefully. It would be premature to assume that AI platforms will inevitably acquire control over financial allocation at macroeconomic scale. Consumers may continue to use bank-owned applications, human advisers and multiple competing AI systems. Regulation may constrain the degree of delegated authority that agents can exercise, while banks may develop their own agentic capabilities.

The more immediate concern is therefore not that an AI system will independently control the financial system. It is that a relatively small number of AI interfaces could become important coordination points for millions of individual financial decisions.

This creates the possibility of correlated behaviour. If many AI agents rely on similar objectives, data sources, optimisation techniques or provider rankings, they could direct consumer demand towards or away from particular institutions in ways that are difficult for individual consumers to observe. A change in an AI platform's recommendation logic could consequently have effects extending beyond the individual customer whose decision it directly influences.

The concept of delegated financial agency therefore connects three levels of analysis. At the micro level, it concerns the relationship between consumers and AI systems and the extent to which financial decision-making is delegated. At the meso level, it concerns the changing competitive relationship between AI platforms and financial institutions. At the macro level, it raises questions about concentration, coordination and the possibility that AI-mediated decisions could influence the distribution of financial flows across the wider system.

This suggests that delegated financial agency should not be understood merely as a technological feature. It represents a potential reconfiguration of financial intermediation in which decision-making itself becomes an object of delegation. The critical economic resource is no longer only capital, customer deposits or financial infrastructure. It may increasingly be the capacity to interpret, coordinate and execute financial choices on behalf of large numbers of consumers.

The theoretical contribution of this perspective is therefore to shift attention from the question of whether AI will replace banks towards a more precise question: who will exercise delegated agency over financial decisions?

Banks may continue to own accounts, provide credit and hold regulated balance sheets. AI platforms may nevertheless acquire significant economic power if they become the systems through which consumers formulate objectives, evaluate alternatives and authorise actions. The future structure of financial intermediation may consequently depend less on which institution owns the financial product than on which actor controls the decision process that determines where financial demand is directed.

Delegated financial agency thus provides a conceptual bridge between consumer-level AI assistance and system-level platform power. At low levels of delegation, AI may simply make financial services easier to understand. At higher levels, it may begin to shape financial choices. At sufficient scale, it could become a mechanism through which financial demand is coordinated across markets.

The central issue is therefore not whether consumers will delegate financial decisions to AI. Some degree of delegation is already economically plausible and potentially beneficial. The more consequential question is how much authority will be delegated, to whom, under what constraints, and at what scale. Those questions determine whether agentic AI remains primarily a tool for consumer empowerment or develops into a new form of financial gatekeeping with broader economic significance.

11. The Gatekeeper Thesis Should Not Become a Gatekeeper Fallacy

The gatekeeper thesis provides a useful framework for understanding how agentic AI could alter the structure of financial intermediation. However, its explanatory power depends on recognising that technological capability does not automatically translate into economic or institutional dominance. The possibility that AI platforms could become important financial gatekeepers should therefore not be treated as an inevitable outcome.

There are several reasons why the concentration of financial decision-making within a small number of AI platforms may fail to materialise. The first concerns consumer preferences. Financial decisions are often characterised by high levels of uncertainty, emotional significance and perceived consequence. Consumers may therefore continue to prefer direct relationships with established financial institutions, particularly for decisions involving mortgages, retirement savings, investments or other areas in which trust and accountability are especially important.

The existence of a capable AI agent does not necessarily eliminate the value of institutional relationships. Indeed, the more consequential the decision, the greater the potential value of having a clearly identifiable institution responsible for the resulting financial relationship. Consumers may consequently use AI for information gathering and comparison while continuing to rely on banks or human advisers for final decisions.

The second constraint concerns the delegation of execution. The ability of an AI system to recommend a financial action does not imply that it will be permitted to execute that action autonomously. Banks and regulators may impose significant restrictions on the authority that external agents can exercise, particularly where transactions involve substantial financial commitments, irreversible consequences or sensitive personal information.

This distinction could produce a form of bounded agentic finance in which AI systems become powerful recommendation and coordination tools without acquiring unrestricted transactional authority. Such a model would still create significant platform influence, but it would limit the extent to which AI systems could independently reshape financial relationships.

The third constraint is interoperability. The gatekeeper thesis becomes strongest when a small number of AI platforms control access to financial providers. If, however, financial institutions can interact with multiple competing AI systems through common standards and interoperable interfaces, no single platform may acquire decisive control over financial discovery and execution.

Interoperability could therefore prevent the emergence of a single obligatory passage point. Instead of consumers having to access financial institutions through one dominant AI platform, banks could make their products and services available across a competitive ecosystem of agents. In such a structure, AI platforms would compete with one another for customer trust and usage, while banks would retain multiple routes to market.

The fourth constraint concerns the economics of financial intermediation itself. Financial institutions operate under extensive regulatory, capital, operational and liability requirements. Becoming deeply involved in financial intermediation may therefore impose costs and responsibilities that technology platforms have strong incentives to avoid.

An AI platform may benefit economically from controlling customer discovery and recommendation without wanting to assume the balance-sheet, regulatory and liability obligations associated with becoming a bank. This creates an important distinction between controlling financial decisions and becoming a financial institution. The former may be commercially attractive precisely because it allows a platform to capture value from financial activity while avoiding some of the institutional constraints associated with directly providing financial services.

This also reinforces the argument that platformisation should not be understood as a simple process of technological displacement. Existing financial institutions may adapt to platform structures rather than disappear because of them. Robinson, Dörry and Derudder (2024), in their analysis of SWIFT, demonstrate how incumbent financial infrastructures can retain strategic importance while adapting to technological and organisational change. Their analysis illustrates that platformisation can simultaneously challenge existing structures and provide mechanisms through which incumbents preserve or reorganise their positions.

The same dynamic may apply to banking. Banks could develop their own agentic capabilities, establish interoperable interfaces for external agents and use their regulatory and institutional advantages to remain central to financial relationships. Rather than surrendering the customer entirely to technology platforms, banks may seek to occupy multiple layers of the emerging architecture: as financial providers, trusted institutions, data custodians and execution infrastructures.

The resulting competitive environment may therefore be considerably more complex than a simple contest between banks and technology companies. AI platforms may control aspects of customer discovery and decision formation, banks may control regulated financial provision and execution, and infrastructure providers may control important elements of transaction processing. Each layer may possess distinct sources of power while remaining dependent upon the others.

This possibility suggests that the most plausible future is not necessarily:

AI platforms replace banks.

A more plausible scenario is:

AI platforms, banks and financial infrastructures become mutually dependent.

Mutual dependence, however, should not be confused with equal bargaining power. Different actors may control different critical resources, and the distribution of economic value may depend on which layer becomes most difficult to bypass. An AI platform with access to millions of consumers could possess substantial bargaining power over banks. A bank with scarce regulated infrastructure or highly valuable customer relationships could retain significant leverage over platforms. Regulators could also influence the balance by determining which forms of delegation, interoperability and data access are permissible.

The strategic contest is therefore likely to concern which layer captures the greatest share of value and decision-making power, rather than which institution simply replaces another. This returns the analysis to the central distinction developed throughout this paper between ownership of financial products and control of the decision interface.

The gatekeeper thesis should consequently be understood as a structural possibility rather than a prediction of technological inevitability. Agentic AI creates the technical conditions under which a new form of financial intermediation could emerge, but institutional incentives, consumer preferences, regulation, interoperability and incumbent adaptation will determine how far that possibility develops.

This qualification strengthens rather than weakens the thesis. The important question is not whether AI platforms will inevitably become the gatekeepers of banking. It is whether the emergence of systems capable of interpreting financial objectives, selecting providers and executing delegated actions creates a new location of potential gatekeeping power. Even if that power remains contested, distributed or constrained, its emergence would represent a significant change in the economics and governance of financial intermediation.

12. Conclusion

The central argument of this paper is that agentic AI may change financial intermediation not by eliminating banks, but by altering where intermediation takes place. Traditional banking combines the provision of financial products with an institutional relationship through which customers discover, evaluate and access those products. Agentic AI introduces the possibility that these functions become separated. Banks may continue to provide the regulated financial product, hold the balance sheet and execute transactions, while an AI platform increasingly mediates the customer's path towards those products.

This possibility represents a shift from conventional financial intermediation towards what this paper has termed interface intermediation. The critical resource is no longer solely control over capital, risk or financial infrastructure. Increasingly, it may be control over the interface through which financial choices are formulated. An AI agent capable of interpreting a customer's objective, constructing a consideration set, ranking providers and initiating an authorised transaction can influence financial outcomes without itself becoming a bank.

This is the basis of the gatekeeper thesis developed in the paper. Gatekeeping can occur at several stages. At the discovery stage, the AI system determines which providers become visible to the consumer. At the recommendation stage, it influences which alternatives appear preferable. At the execution stage, it may convert delegated authority into financial action. The significance of this progression is that influence over financial markets can arise before an AI platform owns any financial asset or assumes the balance-sheet functions traditionally associated with banking.

The analysis also demonstrates why the gatekeeper thesis cannot be understood purely as a story of technological efficiency. Agentic AI could produce substantial consumer benefits. It could reduce search costs, improve access to financial information, automate administrative tasks and potentially mitigate some forms of human cognitive bias. For consumers facing increasingly complex financial markets, delegation may therefore be a valuable form of cognitive support.

Yet the same capabilities that make delegation attractive create new questions concerning autonomy. When a consumer asks an AI system to “optimise” their finances, they delegate more than a discrete task. They delegate part of the reasoning through which the objective itself is interpreted. The resulting concept of delegated financial agency captures the progression from objective delegation and interpretive delegation to provider selection and execution delegation. As this delegation deepens, the consumer may retain formal authority while exercising less practical control over the reasoning that determines the outcome.

This creates what the paper has described as the delegation paradox: the more capable an AI agent becomes at making financial decisions on behalf of a consumer, the greater the potential benefits of delegation, but also the greater the possibility that consumer responsibility becomes detached from meaningful consumer control.

Automation bias and trust intensify this problem. Consumers may reasonably rely on systems that appear knowledgeable, personalised and responsive, but perceived competence should not be confused with institutional accountability. A conversational AI system can create an impression of understanding that makes its recommendations more persuasive than conventional algorithmic outputs. If responsibility for the resulting decision is distributed across an AI platform, model provider, financial institution and payment infrastructure, the consumer may encounter a single trusted interface while the underlying accountability remains fragmented.

The paper therefore argues that efficiency does not necessarily imply consumer welfare. The critical question is not simply whether AI can optimise financial decisions, but who controls the optimisation function. A platform may be capable of finding cheaper products or better financial outcomes while simultaneously possessing commercial incentives that influence which products are considered or recommended. The consumer's objective and the platform's economic objective may consequently diverge.

This has significant implications for banks. If AI platforms become important customer interfaces, banks cannot rely solely on conventional digital distribution or assume that ownership of the financial product guarantees ownership of the customer relationship. Banks will need to remain authoritative sources of financial information, develop infrastructure capable of interacting safely with authorised agents, establish granular and auditable mechanisms for delegated authority, and distinguish clearly between recommendation and regulated execution. Their competitive advantage may increasingly depend on trust, transparency, verifiable product quality and the reliability of the infrastructure they provide to both customers and AI systems.

The regulatory implications are equally significant. Traditional financial regulation often assigns responsibility according to institutional roles, but agentic AI can distribute a single financial decision across several actors. The paper therefore proposes transaction-chain governance as a complementary regulatory perspective. Rather than focusing exclusively on whether an AI model is safe or accurate, regulation should consider the complete chain connecting consumer intention to financial action: who interprets the objective, who selects and ranks alternatives, who authorises the transaction, who executes it and who remains accountable when harm occurs. Delegation should not become a mechanism for delegating away responsibility.

At the same time, the paper does not argue that AI platforms will inevitably become dominant financial gatekeepers. Such a conclusion would overstate the evidence and underestimate the capacity of existing institutions to adapt. Consumers may continue to prefer direct relationships with banks for high-stakes decisions. Regulators may restrict autonomous execution. Interoperability may prevent individual platforms from becoming obligatory passage points. Banks may develop their own agentic capabilities and use emerging architectures to reinforce rather than surrender their strategic positions.

The more plausible outcome is therefore not the replacement of banks by AI platforms, but a reconfiguration of interdependence. AI platforms may increasingly mediate customer discovery and decision formation; banks may retain control over regulated financial provision, balance-sheet capacity and execution; and payment and regulatory infrastructures may continue to constrain and enable both. The strategic question will be which layer captures the greatest economic value and exercises the greatest influence over financial decisions.

The broader contribution of this paper is consequently to shift the analytical focus from the question of whether AI will replace banks towards a more fundamental question: who will control delegated financial decision-making? If consumers increasingly rely on AI agents to formulate objectives, evaluate financial alternatives and execute authorised actions, then the competitive boundary of banking may move towards the decision interface.

The bank may continue to own the account. It may continue to provide the loan, investment or insurance product. It may continue to hold the regulated balance sheet. Yet another institution may increasingly determine which financial possibilities the customer encounters and which of those possibilities appears preferable.

The future of financial intermediation may therefore be shaped not by the disappearance of the bank, but by the emergence of a new intermediary between the customer and the bank: the agentic decision interface.

References

Baeckström, Y. and Matkovskyy, R. (2026) 'Financial advice behaviour: humans versus AI', Journal of Corporate Finance, 101, 103044.

Brüggen, E.C., Gianni, R., de Haan, F., Hogreve, J., Meacham, D., Post, T. and van der Werf, M. (2025) 'AI-Based Financial Advice: An Ethical Discourse on AI-Based Financial Advice and Ethical Reflection Framework', Journal of Interactive Marketing.

Chen, Y., Dong, Y., Hu, J. and Huang, Y. (2025) 'Does FinTech reduce human biases? Evidence from advisory vs. automated FinTechs in lending', Journal of Banking & Finance, 181, 107093.

Jaffri, A., Jha, A. and Butt, H.A. (2026) 'Household interest in artificial intelligence for financial advice: Evidence from the United States', Journal of Behavioral and Experimental Finance, 51, 101212.

Kamm, A. (2026) 'OpenAI's "Dots" Could become Banking's Gatekeeper', digital age, 5 October.

Noponen, N., Feshchenko, P., Auvinen, T., Luoma-aho, V. and Abrahamsson, P. (2024) 'Taylorism on steroids or enabling autonomy? A systematic review of algorithmic management', Management Review Quarterly, 74(3), pp. 1695–1721.

Qi, H. (2022) 'Platform power: monopolisation and financialisation in the era of big tech', Cambridge Journal of Economics, 46(6), pp. 1289–1314.

Robinson, G., Dörry, S. and Derudder, B. (2024) 'Preserving the obligatory passage point: SWIFT and the partial platformisation of global payments', Geoforum, 156, 104007.

Sargeant, H. (2023) 'Algorithmic decision-making in financial services: economic and normative outcomes in consumer credit', AI and Ethics, 3, pp. 1295–1311. doi:10.1007/s43681-022-00236-7.

Skitka, L.J., Mosier, K.L. and Burdick, M. (1999) 'Does automation bias decision-making?', International Journal of Human-Computer Studies, 51(5), pp. 991–1006. doi:10.1006/ijhc.1999.0252.

Sungkarungsri, P. and Kiattisin, S. (2026) 'How trustworthy AI fosters adoption and user advocacy in financial advisory services', Discover Artificial Intelligence.

Vives, X. (2019) 'Digital disruption in banking', Annual Review of Financial Economics, 11, pp. 243–272. doi:10.1146/annurev-financial-100719-120854.

Contact

Reach out via email for inquiries.

Email

Subscribe to newsletter

info@grcadvisory.ch

© 2025. All rights reserved.