“Alpha comes from idiosyncratic risk”.

It’s the opposite of buying an index fund, where you own a bit of everything and ride the big trend. And that principle holds no matter if you’re investing in public markets, credit, venture capital or private equity.

Which leads to this version for Private Equity:

Private equity runs on single deals, not deals in the aggregate

And single deals really come down to a handful of specific decisions and assumptions. Not a portfolio of averages.

That single perspective has enormous implications for how and where to use genAI in the investment process. Because by their very design, LLM produce a statistically average opinion based on the training data.

With that mental model, it’s easy to see how relying on a machine that generates averages while you are looking for situational specifics has some inherent contradictions.

What PE actually underwrites Link to heading

The old Barbarians at the Gate days where value creation was mainly financial engineering and LBOs have faded. What PE really underwrites these days is usually an operational transformation thesis, financed by an aggressive capital structure. Importantly in that, the value doesn’t come from the financial structure alone; but rather from the operational improvement baked into the thesis.

Being able to build a good LBO model and deeply understanding cashflow are still crucial. But understanding the art of the possible relies more and more on industry depth, pattern matching previous operations, understanding that each company has its unique struggles and that operational reality is often threading the needle between contradicting facts.

An example, a business I worked with simultaneously had equally good reasons to keep a loss making division open AND to close that division down. Knowing which one plays out better isn’t a spreadsheet decision, but something that needs lived experience.

And it’s exactly that unique reality where an average opinion will do more harm than good.

LLM are incredible aggregators. They will aggregate priors, language patterns, scenarios, plausible situations. However, they also lack operator perspective to separate a reddit thread, a tourist opinion or lived reality; nor do they have a sense of how people will react to change or what a realistic timeline looks like. Instead they’ll default to common slop observations for the most part, including on business strategy.

Investment returns on the other hand come specifically from being able to understand the art of the possible in a single business situation, knowing how an operator might go about achieving those changes, and understanding asymmetry in the model.

That difference isn’t going to be reconciled by adding “be unique” in your prompt unfortunately. The whole question is anti-thetical to the system’s design.

None of the above means LLM are of no value. In fact, they are excellent upstream of some of those specific decisions.

If what you’re looking for is an overview, something “broadly right, precisely wrong”.

Need to quickly understand an industry or supply chain? Sure.

Want to expand with a quick cross-read through earnings from public companies in the sector to get a beat on industry themes? Excellent choice!

Note that I didn’t just say “broadly right”, I also said “precisely wrong”. Because often that’s exactly where LLM end up. They look shiny and insightful at first pass and pretty shallow once you’re over that. That’s not a bug, it’s a feature of the machine learning engine underneath built on sources you probably wouldn’t trust if explicitly credited in a memo. And while that is incredibly powerful and regularly useful, don’t confuse that appearance of fluency with correctness or added conviction to your idea.

A system that produces eloquent bs is still producing bs. Don’t mistake fluency for signal

Why judgment is irreducibly human Link to heading

So, great you can get more information summarized faster. And you can get a broader surface covered, with some caveats that you shouldn’t take it all at face value. Call that the side that is well-documented.

Success however often lives in interpreting complex unstructured and incomplete information. Exactly the kind of material that is poorly documented, if at all.

Case in point: you can feed any company policy into an LLM and have it rate, scrutinize and summarize it. And you can observe what actually happens with interviews and spot checks. Rarely are those 2 realities the same.

The edge from judgment lives in understanding and interpreting the gap between what’s documented and what’s true. And an LLM can’t really access that gap. Even if you fed it a transcript of a management meeting, or of all the employee interviews. They’d still struggle to pick up those particularly human clues in a conversation. And by the way, that doesn’t live in the training data either.

The practical perspective Link to heading

My bigger point is that the constraint in the investment process is judgment. And, as a good systems thinker, that means the best path is organizing the process around that constraint.

Applying AI is not a task level exercise. Instead, think about rewiring a process for the new constraint

An investment process isn’t linear, but conceptually it helps if you imagine the situation as things that are input to judgement (i.e. upstream) and things that are outcomes of judgment (i.e. downstream). Feel free to apply that at various levels of abstraction.

The most obvious upstream benefit is the ability to get a broader perspective. Quickly sorting through historical precedents, sector patterns, universe screening. Again, they’re about presenting you with a broader aggregate to judge on, not presenting you with a solution. That is the crucial nuance and where upstream meets judgment. You decide what is and isn’t relevant, or which assumptions carry more weight than others.

Another key point of judgment is the investment committee. One easy way is to have the investment memo vetted by an in-house built agent. Instruct the agent to pressure test the memo, ask the obvious questions, strengthen it for the fund thesis, … . Frame the whole exercise to get the analyst to better explain and solidify, don’t rely on the LLM explaining.

Anecdotally, one investment team I spoke to mentioned this probably saved one round-trip to the IC on average for them, because memos skipped the obvious first round questions.

And while most people in this space are very comfortable with excel, and increasingly with basic python; there is obvious usefulness in having an AI assistant inside excel/python. The perspective however isn’t “chatgpt, make me an LBO model”, but rather it can help with complex formulae, scenario analysis, data synthesis. Mechanical work that preps data as input to analysis. Not replacement of analysis.

Those are from lived experience. The key is that all of them improve the conditions under which judgment is made; but none are replacing the judgment.

In summary Link to heading

Whatever your process is. The key in getting the most out of new tools is in identifying the constraint and then optimize around that, rather than blindly chasing tool adoption.

Ironically perhaps, PE operators know exactly how to fix this. Identify the constraint, and underwrite the returns from operational transformation. It’s time for PE to get the PE treatment.