A client asked a simple question this week.

How much liquidity does each portfolio company have?

Simple to ask. Not simple to answer.

The portfolio has more than 50 companies. Many carry multiple tranches of debt. Senior secured. Second lien. Mezz. Unsecured. Some information is in credit agreements buried 80 pages deep; some of it is in an Excel file someone built two years ago and hasn't touched since.

To answer the question properly, you have to read hundreds of pages of PDFs, pull numbers from a dozen Excel files, parse legalese that was written to be precise rather than readable, and then apply judgment about what actually counts as liquidity for each company.

This is the kind of work that normally gets pushed down to a junior analyst on a Friday with a "can you pull this together by Monday" energy. Three days later they have a spreadsheet that's 80% right and nobody fully trusts.

The client and I are building it as an artifact instead. They are doing the building. I am sitting next to them, guiding the iterations.

A few things we are learning as we go.

1. The PDFs are the easy part now. If you'd asked me 18 months ago whether AI could read a 200-page credit agreement and pull the right covenant thresholds, I would have hedged. Today it's table stakes.
2. The Excel files are harder than the PDFs. Every analyst builds their model differently. Tabs are named whatever they felt like that day. The same company shows up as "ABC Holdings" in one file and "ABC Holdings, LLC" in another and "ABC" in a third. Fuzzy matching matters more than people think.
3. The capital structure logic is where the real work lives. You can't just sum up the debt. You have to rank it. Senior secured first, then second lien, then unsecured, then anything subordinated to that. A company with 200 million of cash and 500 million of senior secured debt is a different situation than a company with 200 million of cash and 500 million of unsecured holdco notes. The artifact needs to know the difference, and teaching it that difference is most of the work.

Getting Claude to replace some of these tasks, or even just augment them well, takes a real investment of time. You have to iterate on the prompts. You have to test the outputs against documents you already understand. You have to feed it the intelligence about your business, your conventions, your edge cases. It is not a one-shot exercise. The first version is always wrong in ways you didn't expect.

So we are working on it. We are several iterations in. The client is doing the building, asking the questions, finding the gaps. I am there to nudge the architecture and catch the moments where the logic breaks.

I'll keep you posted on how it's going and whether the ROI is worth the time we are putting in. Right now my bet is yes. But the only way to know is to keep building and see what comes out the other side.

The bigger lesson, even at this stage, is this one:

Find the work nobody wants to do. The work that's high stakes but low joy. The work where the answer matters but the process of getting there is brutal. That's where AI earns its keep, if you're willing to put in the iteration time to get it there.

More soon.

Alex