AI in Finance: The Edge You Can't Download Sits in a Trader's

The core claim. The most valuable asset in financial investing isn't a better model or more data—it's the contextual judgment a human trader builds over decades. Annika Sokhal, founder of Moshi and former derivative trader, argues that the current AI race misses this point by trying to replace human decision-making with agents.

The data paradox. The world generated 175 zettabytes of data in 2024, but human attention is flat. For financial investors, more data doesn't equal more information—it widens the gap between what's available and what they can process. The real bottleneck is not compute but synthesis. Traders already have the most efficient reasoning engine available (a human brain running on 20 watts), yet most AI tooling tries to push that engine out of the loop rather than feed it.

The failure of replacement. Current AI tools fall into two categories: pull tools where the user provides context to a frozen model, and push tools that flood the user with information. Neither solves the actual problem. Sokhal highlights a specific failure: arguing with an LLM about whether FOMC is today or tomorrow wastes a trader's time. The bigger hidden cost is that one-person pods of 'builders with API calls' destroy the apprenticeship model where junior traders absorb context by sitting next to seniors.

The architectural split. Sokhal's solution is to separate the work into two distinct loops: the grind (data gathering, cleaning, cross-referencing) and judgment (the actual decision). Machines excel at the grind, humans at judgment. Moshi implements this with a deterministic team of six agents that monitor 2.4 billion data points across markets 24/7, measuring hallucination and traceability for every claim, while leaving the final decision with the human.

The operational reality. The gap nobody talks about is productionization. Building a weekend dashboard is easy; running it reliably for even one more person requires security, databases, storage, and API credit management. Pricing has shifted from tool-based to per-call, which changes the economics. Sokhal emphasizes that in finance, when something goes wrong, a human must stand up and say 'I made this decision for this reason'—an agent that self-corrects after leaking an API key is not acceptable.

The organizational cost. If every trader runs their own one-person AI team, the future generation loses the ability to absorb context through proximity and consistency. Sokhal calls this the real un-priced cost of current AI adoption patterns. Moshi's approach is to deliver institutional-grade analysis without requiring each person to run their own quant and data science team, claiming up to 35% efficiency gains through time saved on data work.