Financial intermediation is usually associated with institutions: banks, brokers, advisers, payment providers, exchanges and platforms.
AI introduces a different kind of intermediary. An intelligent agent can become economically important before it becomes legally autonomous or holds a financial product itself.
It can sit between the customer and the institution, interpret an objective, decide which information matters, filter providers, recommend an action and potentially prepare the next step.
That is enough to raise questions about market power, accountability and access—even if the final transaction still takes place inside a regulated bank.
The policy challenge is that the intermediary may be distributed across several actors. The customer may use one interface, the model may be supplied by another company, product data may come from several institutions, identity may be verified elsewhere and execution may remain with the bank. The user experiences one decision path; the system underneath may contain several responsibility boundaries.
Intermediation can happen before execution.
Expresses an objective.
Interprets, filters and recommends.
Provides the regulated product and controls.
Authenticates, executes and settles.
Economic intermediary, legal category
The strategic mistake would be to assume that an agent matters only once regulation recognises it as a distinct intermediary or once it is allowed to move money autonomously.
Influence can arrive earlier.
If the agent controls which providers are considered, how trade-offs are framed, which recommendation is presented and how the customer reaches the next action, it already occupies part of the decision architecture.
That does not answer how a specific legal regime should classify the agent. It does mean policymakers and institutions may need to observe economic intermediation separately from formal product provision.
This distinction is particularly useful for international institutions. Financial systems differ in licensing, data portability, payments infrastructure and consumer protection, but the economic question travels across jurisdictions: which layer interprets the customer's objective and shapes access to the market?
That layer can influence competition even when it does not take deposits, extend credit or hold customer assets.
An agent can reshape the financial relationship before it acquires the authority to complete the transaction.KEVIN MOSERI
Where does accountability sit?
Traditional financial systems make responsibility imperfectly but recognisably legible. The customer instructs. The adviser advises. The bank authenticates and executes. The infrastructure settles.
Agent-mediated systems can distribute those roles across several technical and institutional layers.
Four questions become inseparable.
The deeper the agent moves into the workflow, the more these questions need to be answered by architecture rather than disclosure alone.
A disclosure can tell the customer that AI is involved. It cannot, by itself, enforce a monetary limit, prevent an unauthorised tool call, preserve a transaction-specific approval boundary or reconstruct the chain of events after a disputed outcome. Those are system properties.
Access becomes a policy question
An intelligent intermediary is only as useful as the information and actions it can reach. That makes access to product data, customer-permissioned financial context, identity systems and execution interfaces strategically significant.
A closed system may preserve institutional control but limit interoperability. A highly open system may increase competition and innovation while creating new dependencies on identity, consent, security and accountability mechanisms.
The policy problem is therefore not simply “open versus closed”. It is how to create useful access without making authority ambiguous.
That matters for competition policy as well as consumer protection. If only the largest platforms can obtain sufficient data, identity access and execution connectivity to operate useful agents, the agent layer could become concentrated. If access is broad but poorly governed, the system may create different risks around security, provenance and responsibility.
The design objective is therefore not maximal access. It is contestable access with explicit authority boundaries.
Five questions worth separating.
- 01
Discovery: Can agents obtain reliable, comparable product information?
- 02
Context: What customer data can be accessed, under what permission and for how long?
- 03
Recommendation: How are conflicts, ranking logic and provenance made visible?
- 04
Execution: Which actions require fresh customer authorization?
- 05
Delegation: If standing mandates emerge, how are limits, revocation and auditability designed?
The new concentration risk may sit above the bank
If a small number of agent environments become the default interface through which customers express financial objectives, market power could accumulate at the decision layer even while many banks continue supplying products underneath it.
That would resemble neither traditional banking concentration nor a simple platform marketplace. The intermediary would influence choice because it understands context and controls the path from objective to recommendation.
For banks, that creates a strategic reason to think about interoperability and distribution now. For public institutions, it creates a reason to look beyond the regulated balance sheet when considering where influence may accumulate.
The relevant indicators may include who controls the customer interface, who can aggregate financial context, who sets ranking criteria, whose identity system is trusted and which intermediary can convert a recommendation into an authorised action. These are different measures from conventional banking scale, but they can still shape market outcomes.
The policy question becomes concrete when customer-facing banking AI is observed institution by institution.
This essay looks at the agent layer as a form of intermediation. Torinel® follows the German evidence: where AI sits, what it can do and what authority remains with the customer or bank.
German banking research, methodology and institutional evidence.
The institution may remain. The intermediary may change.
AI agents do not need to replace banks to alter financial architecture.
They can change which products are seen, how objectives are translated into options, where consent is collected and which institution receives the customer's next instruction.
That makes the agent layer a serious strategic and policy subject even before autonomous finance arrives.
The next financial intermediary may not hold the money. It may control the path to the decision.KEVIN MOSERI