How I use AI as a research partner for fundamental analysis without sacrificing analytical discipline. The adversarial refinement technique.

I run a small concentrated value fund. Partnership-era Buffett methodology. Owner earnings. Capital allocation analysis. Margin of safety. I have been using AI as a research partner for years and I want to share one specific technique that I think every fundamental analyst should be using.

I am not going to talk about asking AI basic questions. If you are typing "is Company X a good investment" into ChatGPT you are wasting the technology. What I want to discuss is adversarial self-refinement applied to investment analysis, because I think it solves a problem that every single investor faces and nobody has a good systematic answer for.

The problem is confirmation bias on your own positions.

Munger said it best. "The human mind is a lot like the human egg, and the human egg has a shut-off device. When one sperm gets in, it shuts down so the next one cannot get in. The human mind has a big tendency of the same sort."

Once you develop conviction on a position, your analytical process starts working for the thesis instead of against it. You notice confirming data. You explain away disconfirming data. You read the earnings transcript and find what you expected to find. Every experienced investor knows this is happening and very few have a systematic process for counteracting it.

Here is mine.

Pass One: Restate and verify the thesis

The first pass is not new analysis. It is a structured restatement and verification of the existing thesis. I prompt the AI to articulate the original investment case in three paragraphs: the structural advantages identified, the growth drivers projected, and the valuation framework applied. Then I feed in the most recent financial data and have it assess how the business has actually performed against each element of the original thesis. Not against Wall Street estimates. Against my specific assumptions.

This step alone is valuable because it forces an honest accounting of whether the thesis is playing out on its own terms. Most investors compare results to consensus expectations. That tells you what the market thinks. Comparing to your own original assumptions tells you whether your analytical framework was right.

Pass Two: The adversarial challenge

Here is where it gets powerful. I switch the persona entirely.

"You are a short seller who has been studying this company for six months with the explicit goal of finding reasons the stock will decline 50% or more over the next two years."

Then I give five specific mandates. Identify the most fragile assumption in the bull thesis and explain how it breaks. Find the metric that looks healthy on the surface but conceals deterioration underneath. Name the competitive threat the bull case is underweighting. Describe the scenario where management incentives diverge from shareholder interests. Articulate what the current price is assuming and what happens if those expectations are not met.

The critical instruction at the end: "Do not soften the bear case. Make it as strong as the evidence allows."

This is not playing devil's advocate. This is a genuine persona shift. The model operating as a dedicated short seller with a mandate to find reasons for catastrophic decline activates different analytical patterns, emphasizes different data, and surfaces different risks than the same model operating as a long-only value analyst doing a "risk check." I have run this enough times to be confident the difference is real, not cosmetic.

Pass Three: Independent synthesis

The third pass steps back from both positions. I prompt the AI as an independent observer who has read both the bull thesis and the bear case and must assess four things. Which bear arguments are strongest and cannot be dismissed. Which bear arguments are theoretically valid but lack supporting evidence right now. What is the single most important question that would determine which side is right. And given the current price, whether the risk/reward is skewed enough to justify adding to the position.

The synthesis does not average the two views. It identifies which arguments survive scrutiny and which do not. That is a fundamentally different operation.

Why this works

Three passes. Three different personas. Three different analytical frames applied to the same position. The bull analyst, the short seller, and the independent observer are not the same thinker looking at the problem three times. They are three different thinkers with different training, different incentive structures, and different blind spots.

Graham talked about Mr. Market as an emotional counterparty. Munger talked about inversion as a problem-solving discipline. Buffett talked about the importance of understanding the bear case for every position you own. This technique operationalizes all three of those ideas into a systematic repeatable process. It gives you a structured short seller working against your thesis every time you review a position. That is a luxury that individual investors have never had before.

Practical notes

The quality of this technique depends heavily on the quality of your prompt architecture. You need a strong persona layer (specific analytical identity, not just "you are a short seller" but the full mandate and methodology). You need to provide the actual data rather than letting the model use its training data. You need constraints against false precision ("frame outcomes as scenarios, do not claim exact probabilities"). And you need a structured output format that forces ranked arguments rather than an undifferentiated list.

I have built a complete five-layer framework for structuring these prompts (Persona, Context, Task, Constraints, Output Format) and I use it for every analysis I run, not just the adversarial technique. But the adversarial refinement is the single technique I would recommend starting with if you have never used AI for serious fundamental analysis. It addresses the biggest vulnerability in every investor's process, the inability to genuinely challenge your own best ideas, and it does it systematically.

I wrote a full guide on the complete framework if anyone is interested. But the three-pass adversarial technique above is self-contained and you can try it on your highest-conviction position today. I would be curious to hear what the short seller persona surfaces that you had not considered.

Happy to discuss any of this. This is my favorite topic in the world and I will talk about it with anyone.