Note: this site and the author does NOT have any financial relationship with the service provider being reviewed.
THREE POINT OVERVIEW:
We asked Ramp Sheets to solve a basic analyst modeling test, a common part of recruiting processes for junior buyside roles.
The results from Ramp were disappointing, as the models produced were a mess.
We gave the same test to Shortcut AI and the results were better, but the models still had errors and formatting issues. Shortcut AI is definitely ahead of Ramp Sheets as it relates to financial modeling right now, but its performance is still below what you would expect from a competent analyst.
INTRODUCTION:
After publishing two reviews on the Shortcut AI spreadsheet tool this week (Links: Review 1, Review 2), I wanted to review Ramp Sheets, another spreadsheet AI tool that has been gaining popularity online recently.
Pictured Above: Typical hype posts on X
Unfortunately, because Ramp Sheets does not yet have the ability to process PDFs or pull financial info from SEC filings online (the tool is only able to work within Excel files at the moment), we could not test its ability to pull cap structure information and historical financials like we did with Shortcut.
Given this limitation, and to have a little fun, I used a basic analyst/associate-level LBO modeling test I’ve come across in the past. In a recruiting process, this type of test is meant to evaluate a candidate’s ability to quickly model pro forma cash flows and returns to various parts of a new capital structure (see Appendix for full model test instructions).
We tried two variations of the model test. First, I only provided the model test instructions and the income statement in excel (which includes one year of historical figures on top of which forecasts should be built) - this was meant to test Ramp’s ability to independently build the model from scratch. Second, I uploaded a sample excel model solution for a slightly different modeling test as a separate tab and asked Ramp to use the same model structure and formatting as in the template to solve the test - this was to test the model’s ability to use a template and apply it to another modeling task.
REVIEW OF RESULTS:
Test Results Without a Template:
The output for the open-ended variation of the test was not usable. Formatting was bad (even though I asked for formatting to be consistent with provided income statement in the excel file), some assumptions used did not match instructions, certain cells were completely mis-linked, and some numbers were stored as text and therefore not taken into account in the calculations.
In Figure 1 below, you can see that the sources and uses table totals are linked completely incorrectly and formatting is, to put it lightly, bad (also note that the required sponsor equity contribution is off by ~$70M but we can’t trace how it was calculated since the value is hardcoded):
Figure 1: The sources and uses table…could use a bit of work…
In Figure 2 below, you can really get a feel for the output - nonsensical formula links and nothing really working properly.
Figure 2: …the above is self explanatory…
But despite all of these errors, Ramp Sheets says the model is all good in the chat (Figure 3)….
Figure 3: Not sure I agree… perhaps more concerning than the errors is Ramp’s inability to recognize that there are errors…
Test Results With a Template:
To see if I could get an improvement in results, I included a new tab that included the solution for a slightly different modeling test. This was to provide a template and directly guide Ramp towards the model structure and output I needed - Figure 4 below gives you a sense of the general formatting and structure of the template (clean and simple). Note that this was essentially an answer key with a full cash flow model, including all formulas and formatting - Ramp just needed to apply the existing structure, formulas, and formatting to another variation of basically the same test.
Figure 4: Part of the template is shown above - this was effectively an answer key. In our prompt, we asked Ramp to “use this template as “an example of the formatting, model structure, and output structure that you should put together”.
But the results were still poor. In Figure 5 below, the sources and uses table shows a negative required sponsor equity contribution due to the “repay existing debt" figure on the uses side being stored as text and not being taken into account in the sum formula…
Figure 5: Sources and uses table showing negative ($11M) in required sponsor equity contribution. Note how the format is also nothing like the provided template.
In Figure 6 below, you can see more absolutely brutal formatting and a forecast that for some reason has SG&A declining by 11% per year when the provided assumptions called for 3% annual inflation in SG&A.
Figure 6: Unacceptable income statement output
As you can probably guess, the calculated returns output from all of this was also not great - see below (Figure 7) for the equity returns calculations…
Figure 7: Projected equity returns output (I’ve seen better)
Circling back to Shortcut AI. I thought I owed it to Shortcut to try the same exercise given my reviews earlier this week. So I gave it the same modeling test instructions along with the income statement and the model template. The results were generally better - Shortcut’s model was generally usable and I could at least see where mistakes were coming from.
But the formatting was still quite bad and did not at all reflect the reference template. And the biggest fundamental reasoning error came from Shortcut not recognizing that the existing debt being repaid as part of the LBO was also part of the Enterprise Value of the transaction - as a result Shortcut double-counted the existing net debt on the “uses” side of the table (Figure 8). This resulted in the required sponsor contribution being overstated by ~$300M.
Figure 8: Shortcut double counted existing net debt on the “uses” side, leading to an overstated required sponsor equity contribution. The cap table EV also did not match the net debt and sponsor equity figures shown. Also note that this does not really reflect the same clean formatting seen in the template (Figure 4) even though the instructions called for the same formatting to be used.
However, as Figure 9 and Figure 10 show, the model produced by Shortcut was at least complete and auditable (unlike Ramp’s). Other errors made by Shortcut included: income statement tax expense being calculated off of EBIT instead of EBT and an incorrect call premium being used in the 2L IRR calculation (used 101 instead of 104). But at least the mechanics of the income statement, cash flow projections, and debt schedule were generally right. Shortcut was also able to build in an interest toggle switch to turn on circularity for interest expense calculation which is nice.
Figure 9: At least the model did not #REF out like Ramp’s
The returns calculation for the equity tranche was very off due to the overstated initial equity contribution - but again, relative to Ramp at least the model was working.
Figure 10: Returns for the equity are very wrong, and returns for the 2L are only slightly wrong (wrong call premium used).
KEY TAKEAWAYS:
Ramp Sheets clearly needs a lot more work. Not only is it missing the capability to read and pull from PDF uploads or other external sources, its performance within a controlled and simple Excel analyst-level exercise was poor. It’s inability to recognize clear errors such as reffed out cells was also disappointing.
At the moment, Shortcut AI is clearly ahead of Ramp Sheets for finance/investing modeling purposes. But Shortcut also clearly needs improvement. Even when executing these simulated modeling exercises, which don’t have the complexity and messy data that can be present in real world situations, Shortcut did not produce a result that would be acceptable at an analyst/associate level.
But of course, it is important to remember that there is no such thing as a final verdict with AI given this technology is improving so quickly. Don’t worry, I will be testing Ramp and Shortcut again soon to monitor progress.
APPENDIX:
a.) Model Test Instructions
Appendix 1: Model Test Instructions
b.) Provided Income Statement:
Appendix 2: Income Statement Provided with Model Test