A leading brokerage strategist, who has spent years deploying artificial intelligence in real-world investment settings, warns that the current approach to "distilling" expert knowledge into AI may be fundamentally flawed, potentially creating new risks for the financial industry.

Wang Kai, Chief Asset Allocation Strategist at Everbright Securities and co-chair of the FinAI Street Financial AGI Salon, recently shared his team's extensive hands-on experience with AI integration. Instead of repeating common industry narratives about AI empowerment, he detailed the practical boundary problems that have emerged after deployment, pushing the conversation forward.

The core pitfalls of "distilling" a master investor

Many institutions are attempting to feed an AI all of a veteran investor's research reports, speeches, and books to replicate their decision-making logic. Wang calls this approach misguided. The problem is not that the technology is inadequate, but that mixing an investor's thoughts from their 20s, 30s, and 40s creates internal contradictions. An investor's framework evolves with market cycles, and what worked 20 years ago may conflict with recent experience.

The solution lies in restructuring the knowledge base. Wang's team labels information along different timelines, splitting it into a "backbone" and "slices." The backbone consists of stable, long-term frameworks like PB-ROE or the Merrill Lynch Clock. The slices are the analyst's specific market judgments from each period. The AI is then tasked with learning from the slice most historically analogous to the current market environment, allowing for more precise sampling rather than broad, contradictory data retrieval.

The three layers of information and the source of alpha

Wang categorizes the information AI can process into three distinct layers. The first is "known knowns," market information already fully priced in, which AI can quickly organize. The second is "unknown knowns," fragmented public information scattered across filings and sources that AI can piece together to reveal clues beyond consensus. However, this still relies on public data.

The third layer, "unknown unknowns," is where AI is powerless. This information comes from offline research, on-the-ground observations, and tacit knowledge developed through long-term industry tracking. This is the critical source of excess returns. AI's lack of emotion is a double-edged sword. It prevents chasing rallies or panic selling during right-side trades. However, in left-side trades, where the market has yet to form a consensus, AI lacks the text corpus to signal a potential opportunity. This requires human judgment based on deep industry understanding. AI is excellent at scaling from 1 to 100, but it cannot make the leap from 0 to 1.

The hidden risk of "cognitive offloading"

Wang highlighted a rarely discussed risk: AI is creating a "cognitive offloading" that leads to a talent gap. Newcomers who rely on AI from the start without building their own analytical frameworks lose the ability to detect when the AI is wrong. The risk is not just that AI makes mistakes, but that people gradually lose the capacity to identify those mistakes. This creates a dangerous cycle where incorrect information forms a false foundation of knowledge.

To combat this, Wang suggests that organizations must continue to give junior analysts real training opportunities. While AI can handle basic tasks, allowing new talent to build their own experience is essential for long-term competitiveness. The best use of AI is for experts who can verify its output, not for novices who cannot.

Strategic advice for institutional leaders

Wang offers three key recommendations for decision-makers. First, avoid the extremes of either completely rejecting AI or fully trusting it. Institutions should establish clear boundaries, using AI for process-driven tasks while maintaining human confirmation and review at critical points, especially in trading where there is zero tolerance for error. Second, leverage the AI proficiency of younger generations, like Gen Z, by pairing them with experienced professionals.

Third, do not simply embed AI into existing workflows. Many institutions have created AI departments, purchased tools, and run training, but the traditional approval and research chains remain unchanged, making AI just another portal. Instead, Wang advocates for a "enclave" strategy: create a separate space where the new AI-powered workflow can fully operate and prove its value, and then use that success to restructure the broader organization.

The enduring moat for human analysts

When asked about the future role of senior analysts in an AI-dominated world, Wang identified three irreplaceable skills. The first is deep domain knowledge that resides in the mind and is never fully captured in text. The second is a unique understanding of a company's business logic, moving beyond short-term metrics to see the bigger picture through macro trends, demographics, and policy linkages. This ability to see through appearances is something AI lacks.

Third, and most importantly, is the value of frontline research. While AI can read thousands of meeting transcripts, it cannot feel the atmosphere of a factory floor. Wang recounts an anecdote where experienced analysts check the dust on manufacturing equipment. A dusty line is a hidden risk signal. These subtle, non-verbal cues are the true source of alpha and can only be obtained through human observation. As AI expands desk research capabilities, the best analysts will spend *more* time in the field, not less.

A final word on AI adoption

Wang advises financial professionals not to feel anxious about AI. Learning too quickly may mean dealing with buggy open-source projects and constant debugging. For most, it is fine to wait for more mature business models. However, he strongly encourages everyone to use AI themselves to understand its true capabilities. Finally, he stresses the critical importance of data security, especially for local private knowledge bases and trading data, which must be meticulously protected and isolated from cloud-based systems.