Plausible vs. Optimal: The Danger in using LLMs to write Investment Memos

When I talk to finance industry leaders about what they want AI agents to do, at the top of their list is for the AI to write investment memos. It’s just text, the thinking goes, and ChatGPT seems to do such a good job at it. So how hard can it be?

It’s hard. 

If you think it’s easy, it’s likely because you haven’t actually evaluated the goodness of the memos being created by the tools that you use. Memos created by many AI tools sound correct and confident, but they are not aligned to what your firm would do or how it would make decisions. Large Language Models (LLMs) are exceptionally good at sounding right, even when they are analytically wrong.

When integrating AI into high-stakes financial environments, you must distinguish between two distinct properties of generated output: plausibility and optimality. Prompting a foundation model will consistently yield a plausible memo, but in venture capital or institutional investing, relying on "plausible" is a fast track to mediocre returns.

Here is why deploying individual AI productivity tools across your firm might be degrading your firm’s decision quality, and how to architect agentic systems that actually preserve your firm's alpha.

The Multiverse of Plausible Outcomes

To understand the gap between plausible and optimal, let’s separate an investment memo into two distinct aspects:

1. Style: The syntactic formatting, tone, and structural fluency of the document.

2. Methodology: The logical rigor and the correctness of the decision advocated within the text.

When you naively prompt an LLM to evaluate a deal and write a memo, you are essentially asking it to sample from a multiverse of possible documents. The model operates probabilistically, predicting the next statistically likely word based on its vast, generalized training data. It selects one "universe" from a near-infinite set of mathematical possibilities.

Suppose some firms lead with an executive summary and others end with a conclusion. The LLM-generated memos will be a probabilistic mix of the two styles. Some of your memos will have an executive summary at the top, and others will have the summary at the end.

The same thing is true of the methodology, although it is harder to conceptualize. As a firm, what you will get is a distribution of all plausible analyses and decisions. But a lexical average methodology is not what drives returns. Rather than getting the optimal answer, you are getting a highly articulate guess.

The Fluency Trap and the Loss of Human Signal

Because the style of a memo is so visible, it receives the bulk of tool providers’ attention. Individual productivity tools—custom GPTs, agent skills, or custom AI applications—have been designed to easily customize style. You can deploy these tools across your associates, instruct them to upload a successful historical memo as a template, and the AI will flawlessly mimic that formatting.

This creates a dangerous illusion where fluency is confused with correctness. You can easily generate a beautifully structured document, but evaluating a real-world deal involves subtle, complex variables. Relying on basic few-shot examples or Chain-of-Thought (CoT) prompting to assess market dynamics or founder psychology makes analytical correctness a complete crapshoot.

Furthermore, this dynamic destroys a critical operational signal: talent development. Historically, partners learned which associates possessed superior analytical judgment by reviewing the nuances in their work. Today, because associates are largely just "vibing with prompts," that signal is lost. The AI artificially smooths over everyone's work, making it look uniformly polished and completely masking the variance in actual analytical rigor.

‍The irony of the current AI wave is that by arming your associates with individualized productivity tools, you have vastly sped up memo and deck creation, but you have systematically reduced the overall quality of analysis within your firm.

Moving to the Optimal: Building Firm-Wide Intelligence

To fix this, you must start constraining the LLM to reasoning within a strict, verifiable workflow. Your agentic memo-writing application cannot start with a basic system prompt. It must be anchored by a Point of View (POV). This POV is programmatic intelligence, derived through machine learning or deep data analysis of your firm’s past deals, successful exits, and historical failure modes. Then, you need to ensure that the context includes the current state, such as the fund’s holdings, exposure, and mark-to-market trends. You also need to verify that the resulting memo is benchmarked appropriately against past memos at your firm, for example, that the language is not stronger than the language for a past deal whose economics were comparatively better. You need to do the refinement iteratively in a loop to ensure that the memo is iteratively improved to fit within your firm’s parameters. 

Here is how you incorporate your point of view, firm-wide context, deterministic verification, and iterative refinement into an agentic process:

• Separate Synthesis from Adjudication: Use the LLM for what it excels at—parsing unstructured data rooms, extracting key metrics, and formatting the final text. Do not rely on it to make the final mathematical or strategic judgment because an LLM's internal probability weights are a black box and cannot be audited.

• Embed Intelligence in Tools to Ensure Verifiability: When the agentic workflow needs to evaluate unit economics or valuation multiples, it should not guess. Instead, the agent must call specific, deterministic tools, machine learning models, and backend APIs. This guarantees verifiability. If an investment committee questions a metric, they can trace the logic back to the exact code execution and data source, rather than hitting a dead end in an AI's probabilistic hallucination.

• Verification-Driven Rewriting: Once the deterministic tools produce verifiable outputs, the system uses those results to actively rewrite the memo to ensure strict alignment with your firm’s parameters. The agent systematically evaluates the output against your internal benchmarking, risk tolerances, and specific "credit box" criteria. If a deal's unit economics fall outside your approved credit box, the system forces a rewrite of the investment thesis to explicitly address—or reject the deal based on—that verifiable gap, ensuring the narrative matches the hard data rather than spinning a plausible but false positive.

• Implement Agentic Debate: Use a multi-agent architecture where one agent drafts the initial thesis, while a separate "red-team" agent—armed with your firm’s specific historical loss data—challenges the methodology. This creates a verifiable trail of logic and dissent, showing exactly how risks were identified, debated, and mitigated before the final memo was ever approved.

In engineering terms, you will hear these referred to as prompt engineering, context engineering, harness engineering, and loop engineering or iterative refinement. 

Build, Don't Buy Your Alpha

Firm-wide intelligence cannot be bought off the shelf. If your firm relies on the same generic AI productivity tools as your competitors, your decision-making process is already being commoditized.

To preserve and scale your alpha, you must build custom agentic architectures. By embedding your proprietary data and methodological POV into the application itself, you transition your AI from a tool that generates plausible prose into an engine that consistently drives optimal decisions.

At Obin, we build bespoke Agentic workflows for financial institutions. To go beyond individual productivity tools and start employing firm-wide intelligence, contact us.