Generative AI in Banking: Useful Workflows and Review Limits

Generative AI in banking creates drafts, summaries, explanations, and other content from instructions and permitted source material. It is useful for preparing knowledge work, but a generated answer does not establish that a bank policy, financial conclusion, or customer action is correct. fenbs can retain approved process instructions and public source references between assistant sessions, without acting as a banking system.

A sensible first use is an internal draft that someone can check completely: a training outline, a public-source research summary, or a procedure review agenda. Start with the review boundary, then decide which tool and information are suitable. Broad access to customer records is unnecessary for testing whether the basic workflow helps.

What does generative AI add to banking knowledge work?

Its distinctive contribution is producing and transforming content. Given approved reference text, an assistant can propose an outline, simplify wording, or organize questions. It is different from a calculation engine applying a defined formula or a predictive model estimating an outcome. Those systems can coexist, but they need evaluation for their actual jobs.

Broader definitions, model distinctions, and banking use categories are covered in what generative AI is, generative AI applications, generative AI versus LLMs, AI in the banking industry, and uses of AI in banking. The difference between creating content and taking a sequence of actions belongs in agentic AI versus generative AI. Here, the practical scope is supervised preparation for bank employees.

How should a bank employee build a reviewable draft?

Specify the question, audience, approved sources, and output format. “Draft a staff discussion outline using these three public publications; cite each material statement and list unresolved questions” gives the reviewer something concrete to inspect. “Explain banking compliance” creates a broad answer whose missing boundaries are hard to see.

Ask for factual claims, interpretations, and open questions to be separated. Open each source and check the relevant passage. Verify publication status, jurisdiction, covered institutions, and whether the document describes a rule, guidance, a proposal, or an individual speaker’s view. A regulator’s name alone does not tell you the document’s authority.

A process owner should compare the resulting draft with the original instructions before considering whether the writing is good. Did it add an unauthorized approval step? Did it remove an exception? Did it convert “may” into “must”? Those changes can matter more than whether the summary reads smoothly.

Keep the approved assignment and source references in fenbs project context when your institution permits that use. The next assistant can then prepare a revision against the same boundaries instead of inferring the process from a partial chat summary.

Why are fluent answers and citations insufficient?

An assistant can present a wrong answer with confident language and a plausible explanation. The NIST generative AI risk profile↗ (opens in a new tab) discusses confabulation: content presented as reliable even when it is false or unsupported. A citation must be checked for both existence and relevance.

Say an assistant summarizes a public document for a training session. It might accurately describe one paragraph while omitting an exception that changes the conclusion. The reviewer needs to inspect the scope and supporting passages, not merely click the first link and confirm that the page loads.

Use a test set containing ordinary questions, ambiguous requests, missing sources, and questions the tool should decline to answer. Check whether it acknowledges insufficient evidence, preserves important qualifications, and directs unresolved issues to the appropriate employee. Record the failed examples so a later prompt revision can be tested against the same problems.

For higher-impact work, specify a reviewer who has enough expertise and time to challenge the result. A checkbox saying “human reviewed” is weak if the reviewer cannot inspect the evidence. The human review guide explains how to make that responsibility part of the workflow.

Which banking decisions need a firmer boundary?

Keep a content experiment separate from authority to approve credit, move funds, change account records, or release regulated communications. Those activities require their own institutional approvals and controls. A tool that writes an explanation should not gain the ability to perform the underlying action simply because both appear in the same conversation.

For example, Regulation B’s notification provisions↗ (opens in a new tab) address notices and specific reasons for adverse credit actions. An assistant must not invent a reason to fill a letter template. The bank’s approved process must establish the actual action, required content, and applicable timing before any communication is released.

Avoid assuming that one familiar model-risk document governs every generative AI use. The OCC’s revised model risk guidance↗ (opens in a new tab) expressly places generative and agentic AI outside its scope. That is not permission to use them without controls; your risk, legal, and compliance teams must identify the requirements and policies applicable to the intended activity.

Approve information handling separately. Customer records, account details, credentials, internal security material, and restricted reports should remain outside an unapproved workflow. Review who can access inputs and outputs, how information is retained, and whether an external provider may use it. Public and fictional materials make a useful starting test without introducing those records.

What context should survive between assistants?

Save the permitted task, approved source list, terminology, required reviewer, and unresolved issues. Record people’s decisions about what the assistant may do and when it must stop. Keep sensitive source content and formal banking records in their approved systems; a process context record does not need to duplicate them.

Distinguish prepared from checked and approved. “Training outline drafted; specialist review pending” gives the next session an accurate starting point. “Training complete” could cause another assistant to reuse an unchecked explanation. Keep the next reviewer’s action explicit so a fresh session can continue without guessing whether the material is ready.

How to do this with fenbs?

1. Create a free account and, after institutional approval, add a project for a public-source training exercise.
2. Add tasks for drafting the outline, verifying the sources, and obtaining the required professional review. Write the evidence expected for each check.
3. Record the approved instructions in Decisions and rules. Save reusable drafting instructions in Skills and prompts, keeping restricted data outside the project.
4. Connect a supported assistant using its guide on the Integrations page. MCP is a standard connection that lets an AI app use an outside tool; have it read the context before drafting.
5. Save progress with unresolved issues and the next reviewer’s action. Test a fresh session to confirm that it recognizes what is still awaiting a person.

fenbs is not the right fit for processing banking transactions or storing restricted customer information without institutional approval. It does not replace a core banking platform or a compliance records system. Its role is the approved work context that helps people and assistants prepare and review the same task consistently.

Judge the exercise by whether an employee can trace the draft to its sources and see the remaining review work. Start with one task whose results you can inspect, then revise the instructions based on observed errors. Find comparable supervised examples in the AI by role hub.