During the filming of Jaws, Steven Spielberg and his film crew had constant problems with the movie's most important prop - the mechanical shark - which kept breaking down as a result of salt water corrosion. To continue filming, Spielberg decided the shark would have to be hidden under water for most of the scenes. Only later did the film crew realize that the shark was made more frightening by the very fact that it was not visible - the abstract idea of the shark within the human imagination was scarier than the shark itself.

To a certain extent, that's what is happening with the back and forth around the impact of AI (and more specifically Anthropic) on the future of white collar work. My sense is that there are a lot of people following the doomsday predictions that haven't had time to even try Claude Cowork yet. But they have the idea that Cowork is some sort of magic solution that can already take their job tomorrow and vaporize entire industries, a shark quickly approaching under water (instead of teeth it has Skills).

Here are my initial thoughts and sample outputs after trying out Claude Cowork. Maybe lifting the big fish out of the water will make it a bit less scary.

I asked Cowork to build a preliminary screening model and memo to evaluate a leveraged loan investment using only the information contained in a 74 page lender presentation. I pointed Cowork to a folder that included detailed instructions on what to include in the institutional style model/memo, a model template, and a memo template (along with a finished example deck for another investment). The final output that I requested was an 11 page deck including: exec summary, transaction summary, EBITDA adjustment summary, business description, credit risks/highlights, and a financial forecast that incorporated a macro slowdown.

Here are some of the issues that came up for me. If this is a skill issue let me know.

1.) Hallucinations are rare but not extinct when Claude is confused.

The instructions and presentation template included a section where LTM EBITDA adjustments should be detailed. Instead of pulling exact information from the QofE slide in the Lender Presentation showing the adjustments included in LTM Adjusted EBITDA , Claude hallucinated adjustments between two already-adjusted EBITDA figures for different periods three months apart in FY 2020 (likely due to confusion about fiscal year end date). (Note: a very stale lender presentation and not a live deal was used for this example).

2.) Issues with formatting and transferring exhibits from Excel and PDF materials to Powerpoint.

Cowork seems to be missing the ability to paste Excel model outputs as well as exhibits from PDF files into Powerpoint. As you can see above, formatting is definitely lacking relative to the output example provided to Claude. The formatting issues and lack of proper model outputs and exhibits was an issue across all 11 slides.

3.) Missing component of credit thesis and lack of depth behind forecast assumptions.

Claude omitted the significant real estate value embedded within the Company's footprint (market value of real estate equal to ~70% of the 1L TL at face value) - this info was included on a dedicated slide as well as the executive summary of the Lender Presentation.

The revenue forecast was also built by geography rather than end market (which would be a better way of forecasting given 70%/30% between steady maintenance and cyclical new build end markets). These are both omissions that I would have been scolded for as an analyst.

So that's my initial experience with Claude Cowork, which at least for me helped put some of the hype into perspective.

I won't spend time on grand predictions (what do I know). But across finance use cases specifically (in contrast to coding), I do think people underestimate (i) the amount of time it takes to check/correct output and (ii) the willingness of senior folks to actually do the checking. And given the black box nature of LLM reasoning, the checking needs may not scale down as fast as AI capabilities scale up.

Who knows how much better Cowork will get this year. But testing it out was good for my sanity amid all the doomsday predictions and rebuttals.