A general-purpose model can produce fluent language without knowing which bank policy, product rule or procedure is currently authoritative. Grounding narrows that problem by supplying relevant approved information when the answer is generated, but the retrieval and generation steps still need independent controls.
The use case defines which sources may be authoritative
The bank first identifies the decisions or questions the system may support and the documents permitted for that purpose. Product terms, approved procedures, regulatory interpretations, service scripts and public disclosures can have different owners, audiences and effective dates, so they should not be mixed into one undifferentiated collection.
Each source has ownership, approval status, version, effective date, access classification and retirement rules. Drafts, superseded documents and unapproved employee notes remain outside production retrieval unless a deliberately bounded use case requires them.
Retrieval selects context before the model writes
When a user asks a question, the system searches the approved knowledge base for passages judged relevant and sends selected context with the instruction to the model. Retrieval may use keywords, semantic representations, metadata filters or a combination of methods.
A retrieval result is not automatically correct. Poor chunking, missing metadata, vague queries or a document that uses similar language for a different product can supply the wrong context even when every source in the collection is legitimate.
The answer should remain traceable to the supplied evidence
The system can instruct the model to answer only from retrieved material, identify its supporting passages and state when the evidence is insufficient. Citations help a user inspect the basis for an answer, while the interface distinguishes source text from generated interpretation.
A citation is not proof by itself: a model can attach the wrong source, overstate what a passage means or combine facts that do not support the conclusion. Material uses therefore require tests that compare each claim with the underlying document rather than checking only whether a link is displayed.
Access controls follow the source, not the model's convenience
Retrieval enforces the requesting user's permissions before content reaches the prompt. A model should not reveal restricted policy, customer, employee or supervisory information merely because that material sits in the same technical platform as public content.
Logs preserve the query, source versions, retrieved passages, model and configuration used for the response, subject to privacy and retention requirements. When a document changes or access is removed, indexes and caches are updated so stale or unauthorized material does not remain available through the AI layer.
Evaluation tests retrieval and generation separately
Test sets cover common, ambiguous, outdated, adversarial and out-of-scope questions. Reviewers measure whether the right evidence was retrieved, whether the response stayed within it, whether citations were accurate and whether the system declined when no reliable answer was available.
Production monitoring tracks unsupported claims, missing or stale sources, user corrections, escalation and changes in retrieval or model performance. Grounding can reduce unsupported answers, but it does not eliminate hallucination, data-quality, security or model-risk responsibilities.
Read the primary material
Banking Explained prioritizes regulators, official publications and first-party announcements.