A Banking AI Response Needs Provenance, Not Just Accuracy

A customer asks why a failed transaction has not been reversed. An AI system produces a clear explanation, provides the expected timeline, and tells the customer what happens next. 

The response may read perfectly. 

For a bank, there is another question to answer: what information was that response based on? 

This becomes particularly important when generative AI enters regulated customer communication. 

A banking response can depend on transaction facts, internal procedures and applicable RBI guidance. A model-generated answer that sounds reasonable cannot provide the same assurance as one whose factual and regulatory basis can be reconstructed. 

Retrieval-augmented generation provides part of the solution. 

Regulatory circulars and internal SOPs can be broken into searchable passages and stored in a retrieval layer. When a complaint arrives, the system retrieves material relevant to that particular issue and supplies it to the drafting process. 

The response can then carry the specific paragraph or policy reference supporting it. 

Grounding also needs to extend beyond documents. 

If the answer depends on whether a transaction succeeded, failed or remains pending, that fact should come from the transaction system. Customer details should come from the appropriate customer record. Policy interpretation should be based on the approved knowledge source. 

This creates a useful division of responsibility: systems of record provide the facts, approved documents provide the rules, and the model assembles those inputs into a usable response. 

Traceability completes the picture. 

Tool calls, retrieved evidence and intermediate agent steps can be retained alongside the case. Confidence and safety conditions can then determine whether a prepared response qualifies for automatic dispatch. Cases that fail those conditions move to a reviewer with the draft and supporting context already available. 

This design has implications beyond complaint management. Any regulated AI workflow involving customer communication eventually needs to answer the same questions: Which facts did the model use? Which policy supported the conclusion? What actions did the agent take? Why was the response allowed to reach the customer? 

These questions shaped the RAG and agentic architecture we built into Samadhan. Responses are grounded in retrieved RBI material, transaction and customer facts come through scoped tools, agent activity is traced, and defined safety gates control autonomous dispatch. 

For regulated AI, producing a good answer is only part of the engineering requirement. The organisation also needs to be able to show how that answer was produced.