AI-driven economies are only as strong as their trust layer

As financial systems become more autonomous, trust has to be engineered into every decision a machine is allowed to make. 

At 2:13 a.m., a bank’s AI agent blocks a $2 million payment. 

It flags something unusual, checks the transaction against the bank’s policies, and uses the authority the bank has given it to stop the payment. 

By 8 a.m., the customer wants an explanation. By 10 a.m., the bank may also have to file a Suspicious Activity Report (SAR) with the regulator. 

The bank now needs answers to four questions. 

• Who made the decision? 

• What was the agent allowed to do? 

• Why did it make that decision? 

• Who is responsible for the result? 

These questions become harder as financial systems move from helping people make decisions to making and executing decisions themselves. The technology can act in seconds. Trust still has to withstand scrutiny long after the action is complete. 

Why AI changes the trust problem in banking 

Finance has always run on trust. A customer trusts a bank to protect their money, a bank trusts another bank to settle a payment, and a regulator trusts every institution to follow the rules. 

For most of modern banking, people and established processes carried much of that trust. 

AI changes the balance because software can now evaluate millions of transactions, make decisions in seconds, and act across several systems without waiting for a person to review every case. 

An agent can move from recommending an action to taking it. 

That shift creates a different requirement for financial institutions. When a person makes a decision, the institution can usually identify who made it, what authority they had, what information they considered, and who is accountable for the outcome. 

An autonomous system needs those same foundations built into the technology around it. 

The Bank for International Settlements said in its 2024 Annual Economic Report that the growing use of AI in finance raises new challenges around models, data, governance, and financial stability. 

For banks, the practical test is simple. When something goes wrong, can the institution prove what happened, why it happened, and whether the system acted within the authority it was given? 

That is where trust stops being something a bank can assume and becomes something it has to engineer. 

Cross-border payments show where trust gets harder 

Cross-border payments make this problem easier to see because the same transaction can move through multiple institutions, jurisdictions, settlement systems, and control frameworks. 

The World Bank found that the average cost of sending $200 across borders was 6.36 percent in the third quarter of 2025, more than double the United Nations target of 3 percent. 

That friction reflects correspondent banking networks, different settlement rules, multiple intermediaries, and processes that require institutions to coordinate and check each other’s work. 

AI can remove some of that friction by making decisions faster and coordinating actions across systems. But faster execution creates a second requirement. Every institution involved still needs to understand the context in which a decision was made. 

That context can include payment rules, sanctions requirements, approval limits, exception handling, settlement dependencies, internal policies, and operational knowledge built up over years. 

A model can understand how payments work in general. It cannot reliably know the specific rules and exceptions that govern a particular institution. That distinction becomes critical when the software is allowed to act autonomously. 

How context makes agentic AI explainable 

For an agent to make a decision that a bank can trust, three things have to work together. 

• Context gives the agent the institution’s rules, policies, processes, exceptions, and operational knowledge needed to understand the situation.  

• Ontology structures that context by defining the relationships between transactions, entities, rules, risks, and actions.  

• Explainability connects the agent’s decision back to the context and rules that shaped it, so the institution can understand and prove why the action happened. 

Consider an agent deciding whether to stop a payment. 

The model may recognize patterns associated with suspicious activity. The agent still needs to determine whether those patterns meet the bank’s own policies , whether an exception applies, whether the transaction falls within its authority, and what action the bank’s procedures require. 

That requires institutional context. In a bank, context includes policies, business rules, regulatory requirements, transaction definitions, approval limits, exception handling, and operational knowledge that may never have been captured in a model’s training data. 

The challenge is making that context usable by the agent. 

This is where an ontology becomes important. A domain-specific ontology provides a structured representation of the institution’s concepts, relationships, rules, and exceptions. 

For example, a payments ontology can connect a payment type to its parties, purpose, risk signals, applicable rules, approval limits, exceptions, and required actions. That gives the agent a structured understanding of the transaction and the institutional context in which it must act. 

The agent can then use that context to apply the model’s reasoning to the specific situation in front of it. 

With explainability, the bank not only has a record showing that the payment was stopped. It also can trace back the decision to the context, rule, exception, and authority that led the agent to stop it. That is the difference between recording an action and explaining a decision. 

It also explains why agent identity, permissions, and compliance matter as much as the intelligence of the underlying model. 

The BIS Innovation Hub has emphasized the importance of governance, explainability, and human oversight as financial institutions increase their use of AI. 

The more authority an agent receives, the more important it becomes for the institution to understand the context behind its decisions and prove that the agent stayed within its boundaries. 

How context, ontologies, and explainability become trust infrastructure 

When an agent has the right context to understand a financial decision, an ontology to structure that context, and explainability to show how the context shaped its action, trust becomes something the institution can demonstrate. 

That means trust has to be designed into the systems surrounding the agent from the start. 

Every autonomous financial action needs four foundations. 

• Identity. The institution must know exactly who or what is acting. 

• Authority. It must know what that actor is allowed to do. 

• Evidence. It must be able to reconstruct the context and reasoning behind the action. 

• Domain expertise. The system must reflect the financial processes, rules, dependencies, and exceptions within which it operates. 

These foundations create the environment in which autonomous decisions can be trusted. Identity establishes the actor. Authority defines its boundaries. Domain expertise provides the institutional knowledge the agent needs to operate within those boundaries. Evidence makes the resulting action traceable. 

This is where context, ontology, and explainability become part of the trust infrastructure. 

Context gives the agent the institutional knowledge required to understand a situation. An ontology structures that knowledge by defining the relationships between transactions, entities, rules, risks, and actions. Explainability then connects the agent’s decision to that context, allowing the institution to understand why a particular action was taken. 

That matters because financial decisions rarely exist in isolation. 

A payment decision can depend on a fraud rule. A settlement decision can depend on a reconciliation process. A credit decision can depend on an approval limit or regulatory requirement. 

An agent operating in these environments needs to understand those relationships before it acts. The intelligence of the underlying model does not remove that requirement. It makes the quality of the institutional knowledge surrounding the model even more important. 

Building that knowledge into financial AI requires engineering teams that understand the systems they are changing. 

An AI agent making a payment decision has to understand the rules governing that payment. A settlement system has to account for reconciliation and settlement dependencies. A fraud engine has to understand how financial crime controls work in practice. 

This is where domain expertise becomes critical. It brings deep banking and payments knowledge into the engineering process, so the context and ontologies used by AI systems reflect how the institution actually operates. 

The engineering challenge is therefore larger than connecting a model to an existing financial system. It is about building the relationships between the model, the institution’s context, its business rules, and the systems that allow an agent to act. 

The result is an AI system grounded in the institution’s actual operating reality and the financial processes it is designed to support. 

The model provides reasoning capability. The institution provides the context. The ontology structures that context. The agent uses it to act. Explainability shows why it acted. Domain expertise brings those elements together in the financial systems where the decision actually happens. 

That is how trust becomes something an institution can build into autonomous finance rather than something it has to prove after the fact. 

In AI-driven finance, trust will depend on whether institutions can prove that their machines acted within the context, authority, and rules they were given.