Watch the video with chapters & transcript →
AI IN FINANCE
In short: Claude for Excel built a 5-year DCF model for Nvidia directly inside the workbook, using live formulas linked to the historicals tab and no hardcoded numbers.
Anthropic's Claude for Excel sits inside your own workbook, reads the data on every tab, and writes formulas for you. To test it on realistic work, Ani Björkström, a tech consultant in the Stockholm financial industry, rebuilt the kind of DCF model junior analysts produce at large financial institutions, using Nvidia's public financials as the only input.
Claude only needs one tab of historical financials; every other tab can start empty. Because Nvidia is publicly traded, its income statement, balance sheet, and cash flow statement are freely available online and can be copied into a historicals tab, with all values in US$ millions.
In the test, the DCF, assumptions, and earnings tabs were cleaned of all numbers first, so every formula in the finished model was demonstrably generated by Claude.
The first prompt primes the model with a role: "You are a senior equity research analyst working directly inside this Excel workbook," followed by instructions to build a 5-year DCF on the DCF tab, taper Nvidia's growth down each year, and use live formulas with no hardcoded numbers. Crucially, the prompt asks Claude to confirm its understanding and present a plan before touching any cells.
Claude responded with a plan referencing both tabs, noting that Nvidia's growth rate was 65% in the last year while projecting lower growth forward. After a simple "yes, go with it," Claude produced the full model: unlevered free cash flows, valuation, and an implied value per share.
Claude initially set the terminal growth rate at 3%, above the roughly 2.5% long-run GDP growth a careful analyst would use as a ceiling. A follow-up prompt challenged the assumptions tab as "too aggressive or unsupported," told Claude to change terminal growth to 2.5%, and asked for the new implied value per share, which Claude recalculated and explained.
| Model input | Claude's first pass | After the analyst's challenge |
|---|---|---|
| Terminal growth rate | Claude proposed 3% terminal growth | Corrected to 2.5%, aligned with long-run GDP growth |
| Revenue growth | 65% growth in Nvidia's last historical year | Growth tapered down in each projection year |
| Share price check | Implied value per share from fundamentals | Compared against Nvidia's market price of 194 |
Comparing implied value per share with Nvidia's actual market price of 194 turned the exercise into a real valuation question: is the stock overvalued? Claude added the current share price to the assumptions tab and flagged the relevant cells in red and yellow.
The answer is yes and no: Claude did about 90% of the job, but not the judgment. It built the DCF, generated assumptions, and wrote the difficult formulas, yet it needed an experienced analyst to spot that 3% terminal growth exceeds long-run GDP growth. Claude can do the math; the analyst supplies context and skepticism.
No, when prompted correctly it uses live formulas linked to the historicals tab, so a changed input flows through the entire DCF.
No. Claude proposed 3% terminal growth, above long-run GDP growth of about 2.5%, and only corrected it when challenged in a follow-up prompt.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?