Agentic AI in Banking Needs Explicit Levels of Authority

As enterprises move from generative AI towards agentic systems, the engineering question changes. 

A generative system primarily produces an output. An agentic system can retrieve information, call tools, make decisions within a workflow, and initiate subsequent actions. 

That expanded capability makes authority design an important part of deploying AI in banking. 

Consider three customer requests: finding an IFSC code, investigating a failed UPI transaction, and reporting suspected fraud. 

An AI system may be capable of understanding all three. Their operational consequences are very different. 

For this reason, automation decisions cannot depend only on whether a model is confident in its interpretation. The system also needs an explicit definition of what actions are permitted for each class of request. 

One practical approach is to separate cases by resolution authority. 

Informational cases can qualify for autonomous handling when the supporting information has been verified. Operational cases can allow an agent to investigate, retrieve transaction information, and request an action while keeping execution separately authorised. Sensitive categories such as fraud, KYC, deceased accounts, or legal matters can require human review before communication or action proceeds. 

The distinction should continue below the workflow level. 

LLMs are useful for interpreting unstructured customer language: identifying intent, detecting language, extracting sentiment, and recognising signals that may indicate urgency. Regulatory classifications and consequential decisions can then be bound to deterministic rules using inputs such as transaction value, fraud flags or regulatory status. 

Tool permissions create another boundary. 

An agent investigating a transaction may need read access to customer and transaction data. That does not imply write access to the underlying financial system. The agent can gather facts and prepare the next step while the actual movement of money remains outside its authority. 

This is the approach we used in Samadhan, our banking grievance-redressal accelerator. The agent can check transaction information, retrieve customer data and consult policy material, with its investigation capped and its tools constrained. Deterministic rules establish resolution classes, sensitive cases are routed to human reviewers, and the AI has no authority to move money. 

This becomes increasingly relevant as banks deploy agents deeper into operational workflows. Model capability determines what an AI system can understand. A well-designed authority model determines what the institution is prepared to let it do.