Claude for Finance: How Anthropic Moved From Research Lab to Enterprise Product

How Anthropic Built Claude for Financial Services

This is Anthropic's financial services team explaining how they built Claude for Finance - not a generic product pitch but a framework for understanding how code capabilities translate to specialized domains.

"0.5% of the world's population are software engineers. Code is just one sliver." Nick Lynn (former investment banker turned PM) makes the key insight: Claude's code excellence isn't just about programming - it's the universal shortcut. "Code is how a company is run. That means Claude is great at interacting with complex systems, exposing thinking and logic." The file creation feature that makes Excel and PowerPoint is just Claude accessing a virtual machine, running Python code at scale.

Three verbs govern everything: retrieve, analyze, create. Retrieval connects to core data sources (S&P, Factset, Pitchbook). Analysis happens through code or spreadsheets - financial models are "a way for analysts to inject their own judgment of what the future looks like." Creation is making outputs client-ready and boardroom-ready. The goal: end-to-end agentic autonomous system.

BCI example: from static spreadsheets to live dashboards. Analysts traditionally refresh comp analysis weekly or quarterly in static Excel. BCI connected Claude artifacts directly to S&P and Factset datasets - now it's a live dashboard that updates with one prompt. Managing directors interface with these directly. "Not just acceleration of work but transformation of work."

Memory systems are foundational. Claude needs to understand patterns and preferences across tools - cloud AI, Excel, browser, MCP connections. If an analyst prefers S&P for EBITDA calculations, "Claude will remember those preferences too, just like a good intern would."

Evals aren't mystical. Nick's encouragement to enterprise customers: "Evals sound like mystical concepts but they're really simple. Tasks you care about and an articulation of what good looks like." Bring those directly into training - that's how Claude gets better for your specific use case.

10 Insights on AI in Finance From Anthropic

• Code as universal shortcut - Excel/PowerPoint creation is Python in a VM; code capabilities transfer to every domain
• Retrieve → Analyze → Create - Three verbs governing finance product; end-to-end agentic system
• BCI live dashboards - Artifacts connected to S&P/Factset; comps analysis that updates with one prompt
• MCP adoption - S&P and Factset already published functional MCP servers in 6 months
• Subvertical differences - PE vs hedge funds vs insurance vs IB; different workflows require different components
• Memory across surfaces - Claude in browser, Excel, Chrome needs to maintain context
• Evals from customers - "Tasks you care about + what good looks like" fed into training pipeline
• NBIM example - Norwegian wealth fund, 9,000 portfolio companies, MCP integrations for daily insights
• Safety three components - Secure deployment, accurate answers, trust/verification/auditability
• Finance-specific training - Now investing in pre-training and post-training specifically for finance

What Code Capabilities Mean for Financial Analysis

Code is the universal shortcut to every domain. Claude's Excel and PowerPoint creation is just Python running in a VM. The retrieve-analyze-create framework for finance could apply to any knowledge work - the analyst's job isn't disappearing, it's transforming from weekly static spreadsheets to live dashboards.