Generative AI Use Cases in Financial Services (2026)

The generative AI use cases in financial services that reach production are the ones where a model reads, drafts, or summarizes and a named person stays accountable for the result: service and advisor assistants, KYC document intake, credit memo and claims file preparation, AML investigation narratives, policy Q&A with citations, finance commentary, and code modernization. Credit, pricing, and fraud scoring stay with governed, explainable models; generative AI handles the language-heavy work around them.

This guide organizes the use cases by function — front office, operations, risk and compliance, finance, and engineering. For each one it covers what the system does, the data it needs, the controls a regulator or model risk team will expect, and how to tell whether it is paying off. It applies across banking, insurance, and wealth management; where a sector differs, we say so.

How Is Generative AI Being Used in Financial Services Today?

Adoption is broad but still shallow. In the Bank of England and FCA's 2024 survey of UK financial firms (opens in new tab), 75% were already using some form of AI and another 10% planned to within three years, yet foundation models accounted for only 17% of all AI use cases. Firms reported the biggest current benefits in data and analytical insight, anti-money laundering and fraud, and cybersecurity. Four of the five top risks they named were about data: privacy, quality, security, and bias.

The value case is large enough that few institutions are waiting. McKinsey estimated (opens in new tab) that generative AI could add $200 billion to $340 billion a year in value to banking, mostly through productivity. Regulators, for their part, have not written a separate rulebook. FINRA's Regulatory Notice 24-09 (opens in new tab) says its rules are technology-neutral and apply to generative AI "just as they apply when member firms use any other technology or tool." In practice, supervision, communications, recordkeeping, and model risk obligations attach to each use case on day one.

It also helps to separate generative AI from the machine learning banks have run for years. Traditional models score, classify, and predict: probability of default, likelihood of fraud, next best product. Generative models produce language and structure: a summary, a draft, an extracted field, an answer with citations. The strongest use cases pair the two, which is how Sphere approaches generative AI solutions for banking and financial services: the predictive model makes the call, and the generative layer prepares, explains, and documents the work around it.

Front Office: Service, Advice, and Onboarding

Customer service assistants

What it does: answers account and product questions, routes requests, and drafts replies for human agents. The public example everyone cites is Klarna, which reported in February 2024 (opens in new tab) that its assistant had handled 2.3 million conversations — two-thirds of its customer service chats — in its first month, doing work equivalent to 700 full-time agents. By May 2025 its CEO was telling reporters (opens in new tab) that "investing in the quality of human support is the way of the future for us," after concluding that the cheaper AI-led model had lowered service quality.

Data needed: product terms, fee schedules, and policies, plus authenticated API access to core systems for anything account-specific, scoped to the customer in the session. Controls: an always-available route to a person — the CFPB has warned (opens in new tab) that "deficient chatbots that prevent access to live, human support can lead to law violations" — no unapproved advice, and transcripts retained under your communications policy. Measure: resolution rate paired with repeat-contact rate, escalation rate, and complaint volume. Containment on its own rewards the wrong behavior.

Advisor and relationship manager copilots

What it does: prepares meeting briefs, takes notes, drafts follow-ups, and searches the research library. Morgan Stanley's AI @ Morgan Stanley Debrief (opens in new tab), launched in June 2024, generates meeting notes with client consent, drafts a follow-up email for the advisor to edit and send, and saves a note into Salesforce. Data needed: CRM records, meeting transcripts, and approved research and product material. Controls: consent for recording, advisor review before anything reaches a client, books-and-records retention, and suitability judgments that stay with the licensed person. Measure: advisor hours returned per week, time from meeting to follow-up, and CRM note completeness.

Onboarding and KYC intake

What it does: reads identity documents, bank statements, and corporate filings, extracts fields, flags gaps, and assembles the file for a KYC analyst. In insurance the same pattern applies to first notice of loss and broker submissions. Data needed: document images, onboarding rules, and the reference data you verify against. Controls: field-level confidence thresholds, human review of every exception, and sanctions screening that still runs through your established screening engine. Measure: time to onboard, touches per file, and the share of files sent back for missing information.

Operations: Lending, Claims, and Underwriting Files

Much of the cost in lending and insurance operations sits in reading. Credit analysts spread financial statements and write credit memos. Claims handlers summarize medical records, repair estimates, and correspondence. Underwriters rekey broker submissions into rating systems. Generative AI suits that preparation work: extracting figures, condensing a long file into a structured brief, and drafting the memo sections that describe the borrower or the loss.

The line to hold is the decision itself. Under the EU AI Act (opens in new tab), systems that evaluate the creditworthiness of individuals, and systems used for risk assessment and pricing in life and health insurance, are classified as high-risk. In the US, adverse action rules still require specific reasons when credit is denied. A language model that writes the memo is fine; a language model that is the decision engine of record is a model risk and fair-lending problem. Keep the score in a validated model and let the generative layer cite it. Our look at how AI is changing insurance claims handling covers the claims side in more depth.

Measure: cycle time from application or first notice to decision, analyst touch time per file, and how often a reviewer has to correct an extracted figure.

Risk and Compliance: AML, Surveillance, and Regulatory Change

AML and fraud teams are short of investigation capacity, not detection logic. The detection layer should remain machine learning and rules; in one Sphere engagement, a risk analytics platform built on Amazon SageMaker for a major bank improved fraud detection accuracy by 20%. Generative AI earns its place one step later: summarizing an alert's transaction history, pulling customer context together, and drafting the investigation narrative that an analyst edits and signs.

Two other compliance use cases are mature enough to start with. Policy and regulatory Q&A uses retrieval-augmented generation over your policy library, procedures, and regulator publications, returning answers that cite the exact clause; our guide to compliance-first RAG for banking and insurance explains the architecture. Regulatory change management uses the same retrieval to compare a new rule against existing policies and draft a gap list for the compliance owner. Communications surveillance teams also use language models to triage flagged messages by risk before a person reviews them.

Controls here are the strictest in the institution. Each use case belongs in the model inventory with an owner, a validation plan, and ongoing monitoring; most US banks apply SR 11-7 model risk management, adapted for LLMs. NIST's Generative AI Profile (AI 600-1) is a practical checklist for risks specific to these models, such as confabulation and information security. For the regulation-by-regulation view, see what FINRA, MiFID II, DORA, and the EU AI Act require of AI. Measure: alert handling time, narrative quality-assurance pass rate, and time to answer a policy question with a citation.

Finance and Treasury: Close, Reporting, and Analysis

Inside the finance function, generative AI drafts variance commentary for management reporting, explains reconciliation breaks, summarizes contracts for revenue recognition, and lets non-technical staff query financial data in plain English. One rule makes all of these safe: the numbers come from the ledger or the warehouse through deterministic queries, and the model writes the words around them. It never generates a figure. Our payment reconciliation agent walkthrough shows that split in a working AWS architecture.

Data needed: general ledger and subledger extracts, the chart of accounts, prior-period commentary, and access controls that mirror who may see which entities. Controls: reviewer sign-off, a trace from every sentence back to the query that produced its number, and segregation of duties the assistant cannot bypass. Measure: days to close, analyst hours spent on commentary, and the number of corrections made in review.

Engineering and IT: Modernization and Developer Productivity

Financial institutions run large estates of legacy code that few people still understand. Code assistants help document COBOL and older Java services, generate test cases before a migration, translate business rules into specifications, and speed up routine development. This is often the lowest-risk place to start, because engineers review the output through an existing change process — and it supports the core modernization and integration work that sits under most financial services technology programs.

The controls are software controls: mandatory code review, secret and dependency scanning, license checks, and clear rules on which repositories an assistant may read. EU firms in scope of DORA should expect AI coding services to fall within their ICT third-party risk management. Measure delivery outcomes — lead time for changes, change failure rate, escaped defects — rather than lines of code generated. The checks in due diligence on AI-generated code are a useful standard for your own repositories, not only for acquisitions.

Which Generative AI Use Case Should You Start With?

The table compares the use cases on the three factors that usually decide the order: how ready the data is, how close the output gets to a customer or a regulated decision, and how cleanly the value can be measured.

A sensible sequence starts with one internal, cited, human-reviewed use case to build the shared plumbing — permissioned retrieval, logging, evaluation — and then reuses it for the second and third. Customer-facing assistants come later, once you have evidence of how the system fails. Where the model runs matters as much as what it does; our guide to private LLM deployment for banks compares the hosting options, and our comparison of enterprise AI development platforms sets out how the major clouds differ on regional processing and governance. If you are planning the full build — data, integration, evaluation, and operations — see how Sphere structures enterprise AI development from use case to production.

How Do You Measure Whether a Use Case Is Paying Off?

Most pilots report time saved and stop there. The number is real but incomplete, and it is the one a CFO trusts least. Four habits make the business case hold up:

1. Baseline before you build. Measure current handling time, touches, error rates, and volumes for the exact workflow, from system logs rather than estimates.
2. Pair every efficiency metric with a quality metric. Faster claims summaries mean little if reviewer corrections rise; higher chat containment means little if repeat contacts rise.
3. Count the review cost. Human review is a control, not overhead to engineer away, but its time belongs in the return calculation.
4. Track unit cost in production. Model, retrieval, and infrastructure cost per handled case tells you whether the value survives at full volume.

Run the pilot long enough to see volume swings and edge cases, and agree with risk and finance on the success threshold before the first result comes in. A use case that clears a threshold set in advance is far easier to fund than one defended after the fact.