Anthropic recently rolled out a finance-dedicated AI agent, targeting scenarios such as investment banking, asset management, insurance, credit analysis, and corporate finance. According to an Anthropic announcement, these agent templates can be used for time-consuming financial-industry tasks like making pitchbooks, conducting KYC reviews, and closing month-end accounts, and they can be integrated with Claude Cowork, Claude Code, and Claude Managed Agents.
But saying that this will replace financial researchers may be premature. On Facebook, the podcast “Hardcore Finance Basics” pointed out one specific pain point Anthropic is targeting in the financial research industry: a large amount of important information updates that are highly repetitive. In finance, however, data issues often aren’t explicit errors. Over time, junior analysts build “data sense.” A lot of information isn’t just about “being able to get it”—it also requires knowing what the company changed this time, which metrics can’t be compared directly with prior periods, and which figures are merely management’s framing.
Anthorpic can help with information update work in the financial research industry
With this release, Anthropic has introduced 10 financial-services agent tools that can perform tasks such as creating presentations, reviewing financial statements, and writing credit memos.
Paku, the host of “Hardcore Finance Basics,” who previously worked in a domestic large financial holding company’s trading desk, said the market’s reaction to tools like this is easy to swing to two extremes: on one side, “the end of finance,” “AI cracking the investment holy grail”; on the other, a lot of users showing off how they used only a few hours of vibe coding to produce an investment engine with backtest performance that looks astonishing. But he believes both narratives oversimplify the reality of financial research work.
Paku said the pain point Anthropic is directly tackling in the financial research industry is this: massive volumes of important but highly repetitive information update work. In fundamental research, whether buy-side or sell-side, financial statements, earnings calls, databases, presentations, models, and client reports are tightly interconnected. Before an analyst can build a model, the data must be in place; and because company-specific differences are extremely large, the research process almost inevitably requires supplementary help for organizing information across databases and documents.
Especially during earnings season: if sell-side analysts cover an entire sector, they must update huge amounts of financial statements, earnings calls, key metrics, financial models, and research reports at the same time. Even with junior analyst support, the whole workflow still feels like hell: each company cares about different metrics, models must be adjusted in different ways, and many of the clients are extremely time-expensive large funds—so analysts must extract genuinely valuable best ideas in a short period.
The biggest absurdity in financial research: 80% of the time spent on low-value work
Paku believes the truly ironic part of financial research is that outcomes often depend heavily on the initial direction of judgment—such as which key metrics to look at, which trends to focus on, how to handle missing data, and how to compare across companies. But in practice, analysts spend a lot of time on data collection, pulling Excel sheets, updating reports, and making presentations.
In other words, research output may be 80% determined by judgment, but 80% of the working time is eaten up by data preparation and format updates.
This is also where Anthropic’s finance agents fit in. It isn’t trying to directly find the investment holy grail for analysts; instead, it aims to produce a research workflow that’s about 60% complete: first help analysts capture data, connect databases, update models, and organize presentations and documents, and then humans use natural language to point out what’s wrong, what needs to be added, and which steps require new data.
Paku described it as “a junior who works very fast, but still needs one command and one action at a time.” The value isn’t in replacing senior analysts; it’s in lowering the amount of low-value hours so that the real research judgments return to humans.
The biggest risk: financial data errors are usually implicit
However, Paku also emphasized that the biggest challenge for finance agents isn’t whether they can write reports, but whether they can ensure the data is correct.
He noted that the most troublesome part of updating financial data is that errors are often not explicit errors. Numbers may “all appear to exist,” but they are actually misplaced, logically inconsistent, or defined incorrectly. What’s worse is that the further the error propagates downstream, the higher the tracking cost becomes—exponentially. When models, presentations, reports, and investment memos are all built on incorrect data, the cost of going back to find the mistake is far higher than if humans had initially judged the data source and definitions.
That’s exactly where junior analysts improve over time—that so-called data sense. Many key pieces of information aren’t in structured databases, but are hidden in management presentations, earnings call transcripts, financial statement notes, and company-defined metrics. This information isn’t just about “being able to get it”—it also requires knowing what the company changed in its wording this time, which metrics can’t be compared directly with earlier periods, and which numbers are just management packaging.
Similar problems also show up in financial AI benchmarks. Recently, research by BankerToolBench found that even the best frontier models still fail nearly half of the scoring items in end-to-end workflow tests at the investment-banking junior analyst level, and that bankers’ evaluations found 0% of their outputs reaching a client-ready standard. This shows that AI agents can handle some tasks, but there remains a clear gap from directly delivering high-risk financial outcomes.
AI can write SQL, but can’t freely play with LTV and churn rate
Paku also pointed out that if the task is just simple data retrieval, AI can indeed be very effective. Especially since modern ETL tools are already quite mature, if paired with good interfaces and systems with human intervention, finance agents do have a chance to improve research workflow efficiency.
But what’s truly dangerous is when users ask the AI to calculate more complex metrics on its own or metrics that heavily depend on definitions—such as segmented LTV, churn rate, or unit economics models. If humans don’t first inject clear formulas and benchmarks and instead let the AI run freely, the results could be extremely risky. The reason is that these metrics aren’t just math problems; they depend on business definitions, data conventions, and industry context. If the formula is even slightly wrong, the entire investment judgment could be skewed.
Anthropic’s finance agents are not “the AI investment holy grail,” and they’re not a toy that lets users spend two hours vibe coding an annualized 2000% backtesting engine. Instead, they are an industry tool trying to reorganize the financial research workflow.
The biggest change they’re most likely to bring is freeing analysts from massive amounts of data updates, Excel整理, report formatting, and presentation production—so people can put their time back into judgment: which metrics matter, which trends are worth following, which data can’t be trusted, and which comparison approaches will be misleading.
This article about Anthorpic promoting finance-dedicated AI agents, and insiders revealing the key point that Claude can’t replace analysts, first appeared in Lian News ABMedia.
