The 10,000-Hour Paradox: Why AI Alone Fails in Investment Analysis

There’s no question that AI, particularly models like Claude, Codex or Cursor, have fundamentally changed how software is written and deployed. The productivity gains are real. The quality improvements are real. Entire workflows have been redefined and some pretty smart people are predicting that all software will soon be produced by AI.

But one of the main reasons this has worked so well is rarely called out: the people building these systems are coding domain experts and the results are objectively measurable. They don’t just understand recursive self-improvement loops and AI evals; they understand what good software code looks like, how it breaks, how it should be structured, and most importantly, how to evaluate whether an AI-produced output is actually correct.

That last part is the difference between output that sounds right and something you can rely on for doing real work that matters.

Investment Analysis Is a Harder Problem Than Coding

In coding, correctness is often binary: the software works or it doesn’t. The same is broadly true for many aspects of data retrieval: the capital of France is Paris. Period.

In investing, that’s not the case.

You’re dealing with incomplete information, conflicting data, changing assumptions, and forward-looking judgments. There isn’t a single objective right answer. There are better and worse ways to frame the problem, structure the analysis, and reach a conclusion.

That’s why generic AI tools, even very good ones, tend to fall short in this domain. AI can retrieve information, generate outputs and summarize. They can even sound convincing.

But AI often doesn’t know what really matters.

And AI doesn’t may not know when something is wrong.

More pragmatically, iteration, reaching nuanced conclusions, evaluating risk and return potential,  and sometimes recognizing that the future is likely to diverge from what a synthesis of past evidence suggests, are important elements of building investment conviction. 

That’s the workflow Bullet Point Network was built to address. And it all comes down to the power of a purpose-built AI + 1 Platform. 

Where AI Actually Works (and Where It Doesn’t)

One simple way to think about it is: Trunk, Branches, Leaves.

“The trunk is the architecture of your thinking: where you decide what to explore, consider what actually matters, evaluate the risk/return and iterate. If you get that wrong, everything else breaks.”

• The Leaves (Execution): Data gathering, drafting, formatting, and calculations. AI is extremely effective here for immediate efficiency gains.

• The Branches (Intermediate Work): Analyzing markets, testing assumptions, and synthesizing information. Properly structured AI adds real value across many branches.

• The Trunk (The Foundation): Your investment thesis, model architecture, and view on what drives the business. This requires real domain expertise.

The 10,000-Hour Paradox

The reason this distinction matters comes down to system design. You need to prioritize sources that are relevant and accurate. You need to drive calculations from a trusted engine, and you need fluid, transparent ways for a trustworthy and competent person with real domain expertise to direct, supervise, edit, and own the conclusions. 

As Malcolm Gladwell framed it, real expertise is built over time, and his now-famous 10,000 hours of doing, repeating, revising, and learning what works and what doesn’t can be scaled with properly designed AI workflows, but it can not safely replace the trunk of any analysis.

In investment analysis, that shows up in very practical ways:

• Generating a thesis and recognizing when it doesn’t hold together;

• Seeing what’s missing in an analysis and searching for evidence to validate it;

• Knowing which data points actually matter most

• Quantifying the upside and downside

• Identifying when something looks right but really isn’t

Domain expertise allows you to design systems that work, and to evaluate their outputs correctly. The workflows, the prompt structures, the way sources are paired, the way models are selected, and how the team iterates as new information emerges all need to be purpose bult for  how real investment work actually gets done.

Human-in-the-Loop 

With fundamental, long-term investing, full autonomy should not be the goal. Fragmented piecemeal solutions are not scalable either. Fluid human-in-the-loop controls and transparency need to be intentional:

• Upfront: You define the thesis, structure the output, and decide what matters.

• In Between: AI does a significant amount of heavy lifting faster and more efficiently than traditional workflows.

• At the End: You evaluate outputs, identify inconsistencies, add nuance, and own the results.

The Bottom Line: What Investment Winners Actually Need

Prudent investment decisions require trust and confidence in the output. It’s best to have repeatable, scalable process that is easy to customize for each company, iterate as information emerges and views change, and transform raw information into a high-conviction investment conclusion faster, without compromising rigor.

That’s a very different problem from writing software code, retrieving data or generating (smart-sounding) text rapidly for you to copy & paste into other places. 

To win at investing, you don’t need to produce outputs faster. You need to spend more of your time on the things that matter, armed with the reliable information needed for smart decisions. You need:

• Workflows that mirror how investment analysts actually think

• Model selection based on the task, not a one-size-fits-all approach

• Multi-level prompt structures that interact with each other

• Source-to-prompt pairing that prioritizes relevant information

• Integration with spreadsheets as the source of truth for calculations

• AI+1 human-in-the-loop controls and fluid interation over time

At BPN, we’ve built exactly that: a connected platform to do one thing really well: deliver fundamental investment analysisyou can trust