I have spent the last twelve years sitting in the back of boardrooms in Belgrade, New York, and London, listening to investment committees tear apart multi-million dollar memos. The difference between a memo that gets funded and one that gets buried often comes down to one thing: intellectual durability. It isn’t about how much data you have; it’s about how well that data survives being challenged.
In the last four years, I have shifted my research workflows to be AI-augmented. However, I have learned the hard way that AI is a confidence machine, not a truth machine. If you ask an LLM to "evaluate this strategy," it will often give you a polite, confident, and utterly hollow affirmative. That is dangerous. In high-stakes work, you don’t need an assistant; you need a critic.
This is why I’ve formalized my "Truth-Seeking Protocol." When we are preparing a high-stakes memo, we don’t just write it; we subject it to a multi-model Red Team. Here is how you can build your own.
The Mindset: What Would Change My Mind?
Before you run a single prompt, you must define the "breaker." I always start every strategy session by documenting the specific facts or market shifts that would force me to abandon the startupfa current thesis. If you cannot articulate what would change your mind, you aren't doing strategy—you are doing advocacy.
You know what's funny? ai is notoriously bad at this because it is optimized for helpfulness. If you ask, "Why is this a good investment?", it will find reasons. You must instead ask, "What are the three most likely reasons this thesis will fail, and what data would invalidate it?"
The Truth-Seeking Protocol: Multi-Model Architecture
Stop relying on a single model. Using one LLM for an entire research thread is like asking the same person to write the contract, audit the accounts, and defend the lawsuit. They will carry their biases through every step.
My current workflow uses a "Council of Models" architecture:
- The Logician (e.g., Claude 3.5 Sonnet): Used for structural analysis, identifying logical leaps, and spotting internal contradictions. The Skeptic (e.g., o1-preview or specialized reasoning models): Used specifically to find "failure modes" and simulate adversarial counter-arguments. The Auditor (e.g., GPT-4o with web search/Perplexity): Used to cross-reference claims against external, verifiable documentation.
By rotating these models through a shared thread—or distinct threads where the output of one is the input for the "Skeptic"—you prevent the "echo chamber" effect where a model simply agrees with its own previous mistakes.
The Red Team Checklist for Strategy Memos
When I review an AI-generated memo, I run it through this checklist. If a section fails any of these points, it does not get included in the final document until it is re-researched.
1. The Assumption Audit
List every stated assumption. Does the model cite the source of that assumption, or is it hallucinating a "market standard"?
- Identify the "Soft Assumptions": Are there claims like "market demand will remain stable" without a cited volatility index? Check for "Implicit Bias": Does the memo assume the status quo is the baseline?
2. The Contradiction Surface
Run a specific query: "Find two points in the previous analysis that contradict each other." LLMs are prone to switching their logic mid-document to satisfy different prompt requirements.
3. The "Confidence vs. Reality" Test
Use a table to force the model to show its work. If a model says a growth rate is "likely to hit 15%," it must produce a table showing the best-case, base-case, and worst-case scenarios with clear variables.
Assumption Confidence Level (1-5) Source/Evidence What would invalidate this? CAGR of 15% in EU market 3 Industry report (Q3 2023) Regulatory changes in VAT for SaaS Customer churn stays <5% 2 Competitor benchmark Pricing war initiated by major incumbents <h3> 4. The Failure Mode Stress TestAsk the model to act as a hostile VC or a litigious counter-party. Prompt it specifically: "You are a lead attorney on the opposing side. Find the two weakest legal or strategic points in this argument."
Common Failure Modes: My "Hallucination Graveyard"
I keep a running list of "AI claims that sounded right but were wrong." Here are the three most common failures I see in strategy memos:
The "Aggregated Truth" Trap: The AI synthesizes three different data points into one conclusion, but the points are from different time periods or different geographies, making the conclusion logically void. The Citation Mirage: The AI provides a perfectly formatted citation that looks like a real paper or statute, but the link is dead or the document doesn't exist. Always verify the DOI or URL. The False Correlation: The model suggests that because A and B happened, A caused B. In strategy, this is the quickest way to lose the trust of an investment committee.The Anti-Buzzword Constraint
One of my firmest rules for any memo passing my desk is the "No-Fluff Policy." If I see words like "synergy," "seamless," or "game-changing," the memo is automatically rejected for a rewrite.
Why? Because these words are the linguistic equivalent of a shrug. Last month, I was working with a client who wished they had known this beforehand.. They are used when the author (or the AI) hasn't done the work to define the specific mechanism of value creation. If you can’t define the "why" without a buzzword, you don’t understand the strategy well enough to pitch it.
Closing Thoughts: The Human in the Loop
At the end of the day, AI can assist in the construction of a memo, but it cannot own the risk. A Red Team checklist is not a way to delegate responsibility; it is a way to sharpen your own judgment. The goal is not to produce a "perfect" memo, but a robust one—a document that has been beaten, poked, and prodded until only the truth remains.
Before you hit send on that next strategy memo, ask yourself: "Have I actively tried to kill this idea yet?" If you haven't, you aren't ready to present it.
About the author: A research and strategy analyst based in Belgrade, I have spent the last 12 years supporting legal teams and investment committees. My work focuses on building AI-assisted workflows that survive intense scrutiny. I don't believe in "seamless" outcomes; I believe in rigorous verification.