Can AI build your stock model in Excel? A banking trainer graded four tools against its analysts
Wall Street Prep, which trains investment-banking analysts, asked four AI tools to build Apple's full three-statement model and graded them on the same scale it uses for its trainees. The fastest tools were also the ones that invented historical numbers.
Retail investors who build their own valuation models now have AI agents that work directly inside Excel. Do they produce a model you could rely on? The most detailed public test we found comes from Wall Street Prep, a financial-modeling training firm, published February 12, 2026.
The setup
Each tool got the same prompt: build a fully integrated three-statement model for Apple with three years of history and four years of forecasts, using the latest 10-K, the Q4 press release and consensus estimates, in investment-banking format. Wall Street Prep says a strong analyst needs 2 to 3 hours to do this well from scratch. The versions tested were Shortcut v7.4 (an Excel add-in from Fundamental Research Labs), Claude Opus 4.6, Microsoft Copilot Agent Mode on GPT-5, and ChatGPT 5.2. Outputs were graded on the scale the firm uses for analyst trainees.
| Category | Shortcut | Claude | Copilot | ChatGPT | Low-tier analyst |
|---|---|---|---|---|---|
| Speed and understanding intent | 10.0 | 10.0 | 7.0 | 0.0 | 1.0 |
| Data extraction, formatting, best practices | 5.7 | 5.0 | 3.0 | 4.0 | 6.7 |
| Forecasting and model structure | 5.0 | 4.8 | 4.8 | 2.0 | 7.5 |
| Overall | 5.9 | 5.5 | 4.4 | 2.5 | 6.4 |
For reference, the firm's mid-tier analyst range averages 7.9 and its top analysts 9.4. Speed is the only category where any tool beat a top analyst.
What went right
- Speed. Shortcut and Claude finished setup in roughly 15 minutes and Copilot in about 25; ChatGPT took close to an hour. A decent analyst would need 1 to 2 hours.
- Clarifying questions. Claude and Shortcut asked about consensus versus management guidance, revenue segments, buybacks and layout before starting. Copilot and ChatGPT asked none.
- Explanations. Claude gave the best explanations of where data came from, and was the only tool to backsolve EBITDA correctly.
What went wrong, and why it matters
The most important finding is in historical data. Wall Street Prep reports that Shortcut and Claude, the two front-runners, 'hallucinated significant portions of historical data', with errors 'subtle enough to be dangerous': slightly wrong line items that still added up to correct subtotals. The firm reran both and saw improvement, but its conclusion is that fixing this would take longer than typing the numbers in yourself. Copilot and ChatGPT were more accurate on historicals (8 and 7 out of 10 versus 5 and 3).
Structure had its own gaps:
- No circularity. No tool forecast interest from cash and debt balances, so all four scored 0 on that subtask.
- Debt and shares. None handled debt correctly, and no agent modeled shares outstanding correctly.
- Plugs and hardcodes. Most relied on plugs instead of full integration. Claude and ChatGPT often hardcoded values that should have been formulas.
Wall Street Prep's practical advice follows from this: do not let the agent fetch the data. Upload the filings as PDFs or spreadsheets and have the tool work from those.
A second test points the same way
Shortcut also publicizes a May 2026 benchmark by Ankura's Office of the CFO across three financial models and five Excel AI tools, in which Shortcut ranked first. Even in that vendor-published summary, the stated conclusion is that no tool produced a production-ready model without meaningful human review. We read Shortcut's page, not Ankura's underlying report.
What it costs (read October 14, 2026)
Shortcut's pricing page lists a free tier with 20 weekly credits and Pro at $100 a month billed annually with 1,000 monthly credits; it says a message typically uses 2 to 15 credits. Claude Pro, which includes Claude for Microsoft 365, lists at $20 a month billed monthly or $17 a month annually (claude.com/pricing). We did not check Microsoft Copilot or ChatGPT pricing for this piece. For newer model versions on public finance benchmarks, see our AI Model Index.
How to use these tools without getting burned
- Feed the 10-K yourself. The test's biggest failures came from the agents sourcing historical data on their own.
- Tie out every historical line to the filing, not just the subtotals. The dangerous errors summed correctly.
- Check that interest, debt and share count are actually modeled, not plugged.
- Treat the output as a 0-to-60% draft, which is how Wall Street Prep describes the good use case Context: n: not applicable, qualitative guidance · Feb 12, 2026 · Wall Street Prep · margin: not applicable.
What it shows / What it doesn't
It shows
- How four AI tools performed on one realistic, demanding modeling task, graded on a human analyst scale.
- That speed and clarifying questions were strengths, while sourcing historical data was the weakest link for the top two tools.
- Concrete structural gaps shared by all four: no circularity, wrong debt, wrong share count.
It doesn't show
- Statistical weight. It is one prompt, one company and one grading team, with a rerun for only two tools.
- Current versions. Shortcut v7.4, Claude Opus 4.6, Copilot on GPT-5 and ChatGPT 5.2 have all been updated since February.
- How the tools perform when given clean source files, which the testers recommend but did not score.
- Whether a model built this way leads to better investment results.
Why it matters
An AI agent can get a valuation model started in minutes, which is real homework saved. The test shows the errors hide in the historical numbers you would least expect to check, so the tie-out against the filing, and the conclusions you draw, stay with you.
Primary sources
- VerifiedRanking the Best AI Tools for Financial Modeling (2026) (Wall Street Prep, February 12, 2026)Scorecard, subtask scores, versions and findings read October 14, 2026. Wall Street Prep sells AI training and advisory services.
- LinkedShortcut Ranked Top Excel AI Agent Ahead of Claude for Excel (Ankura Office of the CFO) (Shortcut)Vendor-published summary seen in search results; Ankura's underlying report not read.
- VerifiedShortcut pricing (Shortcut)Free and Pro tiers, credits per message, read October 14, 2026.
- VerifiedClaude plans and pricing (Anthropic)Pro price and Microsoft 365 inclusion read October 14, 2026.
In our directories
How we used AI: AI tools helped research and draft this article. Every figure was then checked against the primary sources listed above before publication.
Not financial advice. For information and education only. AI model outputs and research results can be wrong and do not predict future returns. Nothing here is a recommendation to buy or sell any asset. See our disclaimer and methodology.