Financial teams have plenty of work that looks small on paper and takes hours in practice.
A compliance analyst may need to review a customer file, collect information from several documents, check transactions, prepare a case summary, and then pass everything to another person for review. The same pattern appears in financial research, fraud investigations and reporting.
This is where AI agents are becoming more interesting than a basic chatbot.
Anthropic has introduced financial-services agents that can work across multi-step tasks, use external tools and data, and keep a human involved before an important action is taken. The company has also announced integrations with financial data providers and examples involving KYC, fraud, financial research, modelling and AML investigations.
Here are five workflows where that approach can reduce manual work.
1. KYC and customer due diligence
KYC is a natural candidate for workflow automation because the work often starts with documents and ends with a reviewable case.
Anthropic lists KYC screening among its financial-services agent templates. The workflow can involve building an entity file, reviewing source documents and preparing an escalation package for a compliance professional.
The interesting part is what happens between those steps.
An agent can work through the available material and organize the relevant information before a compliance analyst makes the final call. That leaves the person with a prepared case instead of a collection of documents that still needs to be sorted manually.
There is a concrete example from Parcha. According to Anthropic, the company used Claude for customer due diligence and reduced one workflow from three months to five minutes.
That figure comes from a specific customer case, so it should not be treated as a general benchmark for KYC teams. Still, it shows why document-heavy compliance work is being considered for agentic systems.
2. Fraud investigation
Fraud detection is another area where the investigation itself can involve a lot of repetitive work.
A suspicious case may require looking at transaction information, documents and other risk signals before an analyst can understand what happened. An AI agent can help assemble that context and produce a case summary for review.
Inscribe is one example cited by Anthropic. The company uses Claude for fraud detection, document verification, risk analysis and KYC/KYB checks. Anthropic reports that the time required for a fraud review went from 30 minutes to 90 seconds in the cited case.
In another customer example, Inscribe reported a 70-fold increase in processing output.
Again, these are company-specific results rather than an industry-wide measurement. They are useful mainly because they show the type of work being delegated to an AI system: collecting and assessing information around a fraud case, rather than simply generating a piece of text.
That distinction matters.
A fraud analyst still needs to understand the evidence and decide what should happen to the case. The agent can take care of part of the preparation.
3. Financial research
Financial research presents a slightly different problem.
The difficult part is often not writing the final report. It is getting the right information first.
Anthropic has connected Claude with financial research and data platforms including FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, LSEG and Daloopa. Other connectors include Dun & Bradstreet, Fiscal AI and Financial Modeling Prep.
A research workflow can therefore look something like this:
Research question → financial data → analysis → draft report → analyst review
The model does not need to rely entirely on information already inside the model. It can work with external sources provided through the available tools.
This also connects with research from the Bank for International Settlements. BIS has discussed the use of generative AI in economic analysis, financial supervision and payment oversight, and has pointed to retrieval-augmented generation (RAG) as one way to improve the reliability of outputs by connecting models with verified and specialized information.
Cabe destacar one practical point here: source access does not remove the need to check the result. It changes where some of the manual work happens.
Instead of searching through several systems and then starting the analysis from scratch, an analyst can review the material assembled by the workflow.
4. Financial models and reporting
Spreadsheets are another place where AI assistance can fit into an existing financial workflow.
Anthropic says Claude can work with Excel, PowerPoint and Word through Microsoft 365 add-ins. Context can also move between applications. For example, work started in a financial model can continue when preparing a presentation.
Citadel is one of the examples Anthropic gives. Its investment professionals use Claude for Excel to create and update coverage models, separate signal from noise and pressure-test their work.
The useful part here is not simply asking an AI to “analyze an Excel file.”
A more practical workflow is closer to:
Financial model → analysis → scenario work → checks → summary → presentation
Some of those steps are repetitive. Others require financial judgment.
The division between them is important. A model can help with the preparation and analysis while the professional remains responsible for the financial interpretation.
5. AML and compliance investigations
AML investigations combine several of the problems found in the other workflows: large amounts of information, repeated checks and a need to document what happened.
Anthropic says FIS is working with the company on an agent intended to reduce AML investigations from days to minutes. The companies are also working on agents for areas including credit decisioning, fraud prevention and deposit retention.
The proposed workflow is fairly straightforward:
Alert → gather relevant information → review transactions and context → organize evidence → prepare the case → compliance review
The last step should not be overlooked.
Financial institutions operate in an environment where decisions can have regulatory and customer consequences. An agent preparing an investigation is a different proposition from an agent independently deciding what happens to an account.
That is one reason human review remains part of the financial-services agent model described by Anthropic.
Where should the human stay in the workflow?
This is probably the most important part of the discussion.
The Financial Stability Board has identified several risks associated with AI in financial services, including model risk, data quality, cyber risk, third-party dependencies and concentration around service providers.
The FSB has also noted that generative AI in financial institutions has so far been used heavily for operational efficiency.
That suggests a practical way to think about agentic AI in finance.
For a lower-risk task, an agent might prepare a draft that someone checks.
For an investigation, it might gather evidence and organize the case before a specialist makes the decision.
For a consequential financial or compliance decision, the level of human oversight needs to be much higher.
The Office of the Comptroller of the Currency, together with the Federal Reserve and FDIC, issued revised model risk management guidance in 2026. The guidance emphasizes areas such as validation, monitoring, governance, access controls, data management and auditability. It also notes that generative and agentic AI are developing quickly and are not directly within the scope of that particular guidance.
So the regulatory picture is still developing.
What makes a good fintech AI workflow?
Looking at these examples, a pattern appears.
The strongest candidates for agentic workflows tend to involve:
repetitive manual steps;
large amounts of financial or documentary information;
several tools or data sources;
outputs that can be reviewed;
a clear point where a human takes responsibility.
That does not mean every financial process should become autonomous.
Quite the opposite. For regulated work, knowing where the agent stops can be just as important as knowing what it can do.
The practical takeaway
Claude's financial-services use cases are mostly interesting when they are treated as workflows rather than isolated AI features.
KYC can involve document review and case preparation. Fraud teams can use AI to speed up investigations. Researchers can connect Claude with financial data providers. Analysts can work with financial models and reporting tools. AML teams can use agents to prepare investigation material.
The common thread is fairly simple: the agent handles part of the work around a decision, while the professional remains responsible for the decision itself.
For fintech teams, that may be a more useful starting point than asking whether AI can replace an entire role.
Sometimes saving time means removing the five steps nobody wanted to do manually in the first place.
Top comments (0)