Banking adopted generative AI cautiously and for good reason, since regulated decisions demand explanation and auditability that generated output does not naturally provide. The applications reaching production share a pattern: they assist staff rather than decide, they operate on the bank’s own documents rather than open knowledge, and a person remains accountable for the output. The use cases below are grouped that way, with the governance requirements that determine whether any of them can actually deploy.
Document and Correspondence Processing
This is where banks deploy generative AI most successfully, because the work is high volume, the source material is the bank’s own, and errors are caught in review rather than reaching customers or regulators directly. Document handling consumes enormous staff time in banking and requires little judgement in most instances.
Extraction From Unstructured Documents
Pulling structured data from statements, contracts, and correspondence that arrive as documents rather than feeds. Reliable and immediately measurable against manual handling.
Summarising Long Documents for Review
Producing summaries of credit files, legal agreements, or case histories so staff read the relevant part rather than the whole.
Drafting Standard Correspondence
Generating first drafts of customer letters and internal notes from case data, reviewed and approved by staff before sending.
Classification and Routing
Categorising inbound documents and correspondence to the right team. High volume, tolerant of correction, and easy to measure.
Why This Category Works
Repetitive, grounded in the bank’s own material, and reviewed before consequence. Our generative AI development work builds these as retrieval-grounded rather than open generation.
Customer Service and Staff Support
The pattern that works here is assisting the person handling the query rather than replacing them. An agent-facing assistant that finds the right policy and drafts a response is valuable and low risk. A customer-facing system giving direct answers on regulated products carries considerably more exposure.
Agent-Facing Knowledge Assistance
Retrieving the relevant policy or procedure while an agent is on a call, grounded in the bank’s own documentation with the source shown.
Response Drafting for Agent Review
Producing a draft reply the agent edits and sends. Saves composition time while keeping accountability with a person.
Call and Case Summarisation
Generating case notes from call transcripts, which improves record quality while removing a task agents consistently dislike.
Complaint Handling Support
Assembling case history and relevant policy for the handler. Useful given the documentation requirements complaints carry in regulated markets.
Why Customer-Facing Needs More Caution
Statements about regulated products can constitute advice. Grounding, scope limits, and escalation thresholds are governance requirements rather than product refinements.
Internal Engineering and Operations
Banks run large technology estates with substantial legacy footprints, and generative AI helps with the documentation and comprehension work that surrounds them. This is a low-risk category because the output is reviewed by engineers before it affects anything.
Legacy Code Comprehension
Explaining what undocumented code does, which is valuable where original authors have left and documentation was never written.
Documentation Generation
Producing first-draft technical documentation from code and configuration, reviewed for accuracy rather than trusted directly.
Test Case Generation
Drafting test cases from requirements, expanding coverage faster than manual writing while keeping engineering review.
Migration Support
Assisting analysis during platform migration by mapping and explaining existing behaviour. Our legacy software modernisation work uses this as an analysis aid.
Operational Runbook Assistance
Surfacing the relevant procedure during an incident from internal documentation, which is faster than searching under pressure.
Compliance and Risk Support
Compliance is a strong candidate because the work involves reading large volumes of text against defined criteria, which suits the technology. The critical constraint is that the output supports a human decision rather than constituting one, since regulated determinations require accountable human judgement.
Regulatory Change Monitoring
Summarising regulatory publications and flagging what may affect which policies, for specialist review rather than as a determination.
Policy Gap Analysis Support
Comparing internal policy against regulatory text and surfacing potential gaps for a compliance officer to assess.
Transaction Monitoring Narrative Support
Assembling case narratives for suspicious activity review. The detection remains rule and model based, with generation assisting the write-up.
Audit Evidence Assembly
Gathering relevant documents and drafting responses to audit requests, with review before submission.
Never Automating the Determination
Regulated decisions need an accountable person. Our ai-consulting-services engagements draw this boundary explicitly during scoping.
Governance Requirements That Gate Deployment
In banking, governance determines whether a use case deploys at all, and it is usually the constraint rather than technical capability. Establishing these before building is the difference between a project that reaches production and a demonstration that stalls in model risk review.
Model Risk Management Inclusion
Generative systems fall within model risk frameworks in most institutions. Engage that process early rather than discovering it late.
Complete Decision Logging
Inputs, outputs, model version, retrieved sources, and human actions recorded. You must be able to explain a specific output months later.
Data Residency and Provider Terms
Confirm where data is processed and what the provider may retain. This constrains provider selection more than capability comparisons do.
Human Accountability Documented
Named accountability for output used in any regulated context, with a documented override path that is genuinely exercised.
Measured Baselines and Ongoing Review
Capture current handling times before deployment and sample output quality on a schedule. Our data analytics work makes that measurement routine.
FAQs
What generative AI use cases work in banking?
Document extraction and summarisation, correspondence drafting for staff review, classification and routing, agent-facing knowledge assistance, call summarisation, legacy code comprehension, test generation, and compliance research support. All assist staff rather than deciding autonomously.
Can generative AI make lending decisions?
Not appropriately. Regulated credit decisions require explanation and auditability that generated output does not provide, and they need accountable human judgement. Generative AI can assemble and summarise the case file that a person then decides on.
What is the main constraint on AI adoption in banking?
Governance rather than technology. Generative systems fall within model risk management frameworks, need complete decision logging, and require confirmed data residency and provider terms. Engaging those processes early determines whether a project reaches production.
Should banks use customer-facing generative AI?
With considerably more caution than staff-facing systems. Statements about regulated products can constitute advice, so grounding in approved material, hard scope limits, and escalation thresholds are governance requirements rather than optional refinements.
How should banks log AI usage?
Record inputs, outputs, model version, retrieved source documents, and any human action or override. The standard is being able to explain a specific output months later, which is what audit and regulatory enquiry actually require.
Where should a bank start with generative AI?
With internal document processing or agent-facing assistance. Both are high volume, grounded in the bank’s own material, reviewed before consequence, and measurable against current handling time, which makes the governance case straightforward to make.
