1 Month of Claude in Finance

Finance-focused summary (AI in FP&A workflows)

The presenter (an FP&A manager) explains how they’ve used Claude (an LLM) for about one month to improve core FP&A tasks. They emphasize that while AI can speed up work substantially, it also increases stress and requires heavy human validation and process redesign.

Disclosures / disclaimers

• No explicit “not financial advice” disclaimer was included. (The content is about FP&A work/process rather than investing advice.)

Key takeaways

• AI use increases stress because analysts must:

  • Learn what AI can do
  • Keep improving prompts/processes
  • Validate outputs There’s also no known “ceiling” for capability, and mistakes can be costly.

• Human-driven templates still matter: rather than asking AI to “start from scratch,” the presenter builds/maintains their own analysis templates and instructs AI to populate them.

• Common payoff: AI often delivers the first ~80% of an analysis, but the analyst must review and correct/augment the remaining work.

Methodology / step-by-step frameworks mentioned (FP&A process design)

A) Automate data extraction

• Build a Power Query connection to the company’s financial system report.
• Use Claude to generate Advanced Editor / API code (e.g., report ID and parameters).
• After setup, update by pressing “refresh” instead of manual download/copy/paste.
• Apply the same approach to other data sources (e.g., customer data, headcount).

B) Variance analysis with AI-assisted commentary

• Create a variance analysis template that links Actuals and Budget sources.
• Add helper fields/columns to clarify which fields AI should analyze for variance drivers.
• Instruct AI to generate first-draft commentaries.
• Human workflow: review/validate commentaries against underlying financial results, then edit/add/remove as needed.

C) Revenue analysis (driver-based, multi-source)

• Expand the working file so AI can analyze revenue drivers using data beyond the financial system (e.g., data warehouse, customer info).
• Make comparisons explicit with dedicated fields/columns such as:
  • Current year
  • Budget
  • Prior year
• Improve control/accuracy using explicit flags (e.g., an “exclude” field for products to omit).
• Add qualitative inputs (e.g., churn reasons) into the worksheet so AI can connect quantitative changes to qualitative context.

D) Ad hoc large-dataset analysis

• Provide AI with operational context and specify which fields to reference.
• Ask AI to identify relationships/correlations and highlight which relationships to be wary of.
• Use it as a rapid starting point, then validate.

Explicit recommendations and cautions

• Use templates you trust: experienced analysts have preferred style/logic; AI-generated templates may not match your methodology.
• Don’t over-trust AI “formula error checking”:
  • The presenter tried having AI review worksheet formulas; AI often said everything was fine, but mistakes were still found later.
  • They advise treating AI “error checking” with a grain of salt, especially when calculation results are numerically wrong even if formulas appear consistent.

Key numbers / timelines mentioned

• Timeline: testing Claude for about a month.
• Time-savings expectation:
  • Tasks taking 2–3 hours may be reduced to about 30 minutes (goal/expectation).
  • Example benefit: letting AI investigate first so the analyst can work while it loads (no exact minutes stated).
• Performance estimate: AI typically gets ~80% of the work done, with human validation for the remaining ~20%.

Financial / FP&A instruments mentioned

• None. The discussion is focused on internal FP&A processes (no stocks, ETFs, bonds, FX, commodities, or indices mentioned).

Presenters / sources

• Presenter: Unnamed individual speaking from experience as an FP&A manager.
• AI tool/source: Claude (LLM by Anthropic, implied by name).
• Related tool referenced: Microsoft Power Query (and Excel built-in AI features).