Industry Guide: Banking
How to Use Claude in Banking: A Practical Guide for Bank Teams
Faster casework, with a person still signing the file.
This guide is for people in retail, commercial, and private banks who want to put Anthropic's Claude to work on compliance, KYC and AML casework, fraud investigations, operations, customer communication, and legacy code. It sticks to what Anthropic has published for financial services and what named banks have said in Anthropic's own posts and customer stories.
Knowing how to use Claude in banking starts with the document-heavy work that slows every bank down: reviewing KYC files and flagging gaps, drafting AML investigation narratives from case data, summarizing regulatory changes into policy updates, writing fraud case notes, reconciling operational breaks, and modernizing legacy code. Claude reads, structures, and drafts quickly. Bank staff make the decisions, file the reports, and stay accountable to regulators, with every AI-assisted step inside the bank's model risk and data controls.
Published
Key Takeaways
- The strongest banking uses are KYC file review, AML case narratives, regulatory change summaries, operations reconciliations, and code modernization.
- Anthropic's May 2026 agent templates include a KYC screener, general ledger reconciler, and month-end closer.
- Anthropic reports named banking work at N26, bunq, BNY, Citi, and FIS, from process automation to developer platforms.
- AI assists the analyst. Suspicious activity decisions and regulatory filings remain human responsibilities.
- Every use sits inside the bank's model risk management, data protection, and recordkeeping rules.
Copy and Paste
Prompts You Can Use Today
KYC file gap review
You are a KYC analyst at a commercial bank. Review the onboarding documents summarized below for a corporate customer against our checklist: legal entity documents, ownership structure down to beneficial owners above our threshold, controllers, source of funds, expected activity, and sanctions and adverse media results. For each checklist item state present, missing, or inconsistent, and quote the document that supports your answer. List the follow-up requests for the relationship manager. Checklist and documents: [paste].
Paste your bank's actual checklist. The output should mirror your procedures, not a generic list.
AML investigation narrative
Draft the narrative section of an AML investigation case file from the alert details, transaction summary, and analyst notes below. Structure it as: subject and relationship, alert and scenario triggered, activity reviewed with dates and amounts exactly as given, analyst findings, and the open questions. Use neutral, factual language. Do not reach a conclusion about whether activity is suspicious; that decision is mine. Material: [paste].
Keeping the conclusion with the investigator is both good practice and what regulators expect.
Regulatory change summary
Summarize the regulatory text below for the bank's compliance committee. Give: what changes, who it applies to, effective dates as stated, which of our existing policies listed below are likely affected, and questions for legal. Keep it under four hundred words. Quote the text for every requirement and do not add any requirement that is not in it. Regulatory text and our policy list: [paste].
Always paste the actual text. Claude's training data may not include the latest version.
Fraud case note and customer letter
Here are the details of a card fraud case: the disputed transactions, the customer's statement, device and location data, and the investigator's findings. First, write an internal case note summarizing the evidence and the investigator's decision. Second, write a plain-language letter to the customer explaining the outcome and next steps, without revealing detection methods. Keep the letter under two hundred words. Details: [paste].
Splitting the internal note and the customer letter keeps detection details out of customer communication.
Legacy code explanation
Explain what the COBOL program below does, section by section, in plain language for a Java developer who will reimplement it. List every business rule it enforces, every file or table it reads and writes, and any behavior that looks unintentional. Then propose a set of test cases that would confirm a reimplementation behaves identically. Program: [paste].
Test cases derived from the old behavior are what make a modernization safe.
From Anthropic
What Anthropic offers for Banking
Claude for Financial Services
Anthropic's financial services offering, announced in July 2025, combines Claude models, Claude Code, and enterprise access with connectors to data platforms such as Snowflake, Databricks, and S&P Global, plus implementation partners. Anthropic says data is not used for training by default.
· Source
Agents for financial services
Ten agent templates released in May 2026, including a KYC screener, general ledger reconciler, month-end closer, and statement auditor, shipped as plugins in Claude Cowork and Claude Code and as cookbooks for Claude Managed Agents.
· Source
Anthropic and TCS partnership
Anthropic announced a partnership with TCS in June 2026 to build Claude-powered offerings for regulated industries, including lending advisory for banks and claims processing for insurers.
· Source
In Practice
How Banking Teams Use Claude
According to Anthropic's published customer story
N26
According to Anthropic's customer story, the digital bank N26 uses Claude through AWS Bedrock across more than a dozen internal use cases and has automated a large share of tasks in targeted processes while reducing manual processing.
According to Anthropic's published customer story
bunq
Anthropic's customer story on bunq reports that Claude Code helps new developers contribute meaningful code much faster than before, alongside fast account opening and a high rate of automated support resolution.
According to Anthropic's published customer story
FIS
On Anthropic's financial services page, FIS's chief executive describes building an agent with Anthropic for anti-money laundering investigations.
According to Anthropic's published customer story
BNY
In Anthropic's May 2026 financial services post, BNY's CIO describes AI digital employees that work a case from start to finish.
According to Anthropic's published customer story
Citi
Anthropic's October 2025 financial services update says Citi uses Claude in its AI-powered developer platform for planning and agentic coding.
Illustrative walkthrough
A KYC refresh review with a person signing off
- An analyst opens a periodic review case and loads the customer's documents into the approved workspace.
- Claude checks each document against the bank's checklist and lists items that are present, missing, or inconsistent.
- The analyst confirms the gaps, spotting one ownership change that needs a new beneficial owner declaration.
- Claude drafts the request to the relationship manager and a case note in the bank's format.
- The analyst reviews both, sends the request, and records the decision under their own name.
Banking and Claude: the Published Numbers
| Figure | Context | Source |
|---|---|---|
| automated up to 70% of tasks across targeted processes | N26, per Anthropic's customer story | Source |
| reduced manual processing by up to 50% | N26, across specific processes, per Anthropic's customer story | Source |
| 16 hours instead of 40 | Time for new bunq developers to contribute meaningful code with Claude Code, per Anthropic's customer story | Source |
Core Workflows
Claude Banking Use Cases Across the Bank
These are the places where AI for banking saves the most analyst and operations time.
AI for KYC and AML
KYC file reviews against your checklist, beneficial ownership summaries, and AML investigation narratives drafted from case data for the analyst to decide.
AI fraud detection in banking
Claude does not replace your transaction monitoring models, but it helps investigators summarize cases, write notes, and explain outcomes to customers.
AI for bank compliance
Regulatory change summaries, policy gap analyses, control testing documentation, and regulatory reporting commentary drafted from source text.
Operations
General ledger reconciliations, break investigations, month-end close commentary, and procedure documents, traced back to the system of record.
Code modernization
Explaining legacy core banking code, writing tests that capture current behavior, and helping developers reimplement it with Claude Code.
Customer communication
Complaint responses, letters, and support drafts in plain language, reviewed by staff before they reach a customer.
Compliance and Fraud
How Banks Use AI for Compliance and Fraud Detection
Ask how banks use AI for compliance and fraud detection and you will hear two different things. Detection itself usually runs on specialized transaction monitoring and fraud models. Generative AI like Claude helps with what happens after an alert fires: gathering the case facts, drafting the investigation narrative, checking the KYC file for gaps, and writing up the decision. That is where most investigator hours go.
Anthropic has built for this workflow directly. Its May 2026 agent templates for financial services include a KYC screener alongside reconciliation and close agents. On Anthropic's financial services page, FIS's chief executive describes building an agent with Anthropic that compresses AML investigations from days to minutes. BNY's CIO describes digital employees that work cases end to end.
Can AI reduce false positives in fraud detection? Specialized models can, and banks report results from them, but that is a different tool from a general assistant. What Claude can do is make each alert faster to clear by assembling the evidence and drafting the disposition for the analyst to confirm.
Agentic AI
Agentic AI in Banking and Generative AI in Banking Operations
Agentic AI in banking means multi-step workflows where Claude gathers documents, applies rules, and prepares a result without a person prompting each step. KYC screening, general ledger reconciliation, and month-end close are good candidates because the steps are defined and the output is checkable. Anthropic's agent templates include a general ledger reconciler and month-end closer for exactly this reason.
Generative AI in banking works best with a clear boundary: the agent prepares, a person approves. In practice that means every agent output lands in a queue for review, the agent's steps are logged, and the model sits inside your model risk management framework, such as SR 11-7 in the United States. Regulators will ask how the tool was validated and monitored, not just what it produced.
Code modernization is the quieter opportunity. Many banks run critical processes on decades-old code that few people still understand. Claude can explain that code, document its business rules, and help write tests before anything is changed. Anthropic's October 2025 post quotes Citi's CTO on using Claude in its developer platform for planning and agentic coding.
Choosing Tools
Claude vs ChatGPT for Banking
The Claude vs ChatGPT for banking question is best answered on your own files. Take a closed KYC case, a past investigation, and a regulatory summary your team already wrote. Run the same prompts in each tool and compare accuracy, how each quotes its sources, and how each behaves when information is missing.
In a bank the deciding factors are usually governance rather than raw capability: deployment options, data residency, retention terms, audit logging, and whether the tool fits your third-party risk and model risk processes. N26, for example, uses Claude through AWS Bedrock according to Anthropic's customer story. Your procurement and risk teams will decide the approved path.
Whichever tool you use, the prompts and review habits on this page carry over. The skill that matters is writing precise instructions and checking the output against the source.
Getting Started
Introducing Claude Into a Bank Team Safely
Start in the approved workspace
Use only the deployment your bank has approved for customer data. Test prompts on synthetic or closed cases first.
Pick one queue
Choose one repetitive process, such as KYC refresh reviews, and measure time and error rates before and after.
Write the procedure into the prompt
Paste your checklist and policy text into the prompt or project. Output then follows your procedure rather than a generic one.
Keep a reviewer on every output
Design the workflow so a named person approves each AI-drafted decision, note, or filing.
Limits
Where Banks Must Keep a Person in Charge
Claude can misread a document, miss a beneficial owner hidden in a footnote, or summarize a regulation that has since changed. It does not know your customer the way your relationship manager does. Every AI-assisted review needs a named reviewer, and every figure in a regulatory filing needs to be tied to the system of record.
Decisions with legal weight, such as filing a suspicious activity report, exiting a customer, or declining credit, stay with bank staff under your policies. Fair lending and consumer protection rules also apply to any AI-assisted customer decision, which is another reason to keep Claude on preparation and documentation.
Customer data is protected by law and by contract. Use only approved deployments, follow data minimization, and never paste customer information into a personal account.
Build the Skill Before the Rollout Reaches Your Team
Using AI well in a bank is a practical skill: precise instructions, checkable output, and good review habits. The free Nightschool AI curriculum is hands-on from the first lesson.
Frequently Asked Questions
How do banks use AI for compliance and fraud detection?
Detection usually runs on specialized monitoring and fraud models, while generative AI such as Claude helps investigators after an alert: gathering case facts, checking KYC files, drafting narratives, and writing up decisions. Anthropic's posts quote FIS on an agent that compresses AML investigations from days to minutes. The analyst still decides the outcome.
How do banks use AI?
Banks use AI for compliance and KYC reviews, fraud investigations, customer service, operations such as reconciliations and close, software development, and internal knowledge search. Anthropic's customer stories describe N26 automating tasks across targeted processes and bunq using Claude Code to get new developers productive faster.
What jobs will AI replace in banking?
AI is taking over tasks rather than whole roles so far: document review, first drafts of case narratives, reconciliation work, and routine customer queries. Roles built around judgment, relationships, and accountability to regulators remain. People who learn to direct and check AI output tend to take on more complex work.
Can AI reduce false positives in fraud detection?
Specialized fraud and monitoring models can, and that is where false positive reduction usually comes from. A general assistant like Claude helps differently: it makes each alert faster to review by assembling the evidence and drafting the disposition for an investigator to confirm.
Is Claude used by banks?
Yes. Anthropic's published posts and customer stories name banks including N26, bunq, BNY, Citi, and Commonwealth Bank of Australia, covering process automation, developer platforms, and agent workflows. Anthropic also publishes financial services connectors and agent templates such as a KYC screener.
What is agentic AI in banking?
Agentic AI means an AI system that carries out a multi-step process, such as gathering KYC documents, applying rules, and flagging gaps, without a person prompting each step. In banking it should end in a review queue where a person approves the result, with logging and model risk oversight.
Is it safe for banks to use AI with customer data?
It can be, through deployments the bank has approved after third-party risk, data protection, and model risk review. Anthropic states that its financial services offering does not use data for training by default. Staff should never use personal accounts for customer information.
How should banks validate AI-assisted work?
Treat the AI as part of a process under model risk management: define the intended use, test it on closed cases, measure error rates, require a named reviewer for every output, log the steps, and monitor quality over time. Regulators will ask how it was validated.
Keep going
Related Guides and Courses
Claude in Banking, by role
Spend Less Time Assembling Files
Hands-on lessons for people who work in regulated teams.