Ask an asset management leader what worries them about AI, and three answers come up regularly: the “black box” problem of explainability, data quality and regulatory compliance.
According to Mercer’s 2026 survey, over two thirds (69%) of firms cite data quality and access as their main barrier to AI adoption, while 59% flag regulatory or compliance concerns. Each are legitimate and each are becoming solvable as governance frameworks mature and model transparency improves.
The issues that deserve more attention lie one layer deeper and rarely appear on due diligence questionnaires.
The commoditisation paradox
With 91% of managers planning to expand AI use within 12 months, claiming to “use AI” has lost its signalling power in a pitch.
The reason lies in infrastructure. Most firms rely on generic foundation large language models (LLMs), trained on the same public data and available to every competitor for a subscription fee. As BCG’s 2026 Global Asset Management Report notes, baseline analytical capabilities are commoditising, pushing the real edge towards judgment and proprietary data. The faster the industry adopts AI, the less value adoption itself holds.
Generic models carry a second, systemic risk: herding. When firms across the industry rely on similar models trained on similar data, chasing similar signals, AI can amplify crowded trades and correlated positioning. Notably, 24% of managers in Mercer’s survey identified system-level risks, such as herding, as today’s largest regulatory blind spot. If a single firm’s model fails, the damage is contained. If the entire market’s underlying model agrees, the risk becomes systemic.
Specialised models are different. Trained for prediction on proprietary data, they can’t be replicated with a subscription.
The bias we cannot see
There is an invisible dimension to this convergence – what shared models have quietly learned. Firms increasingly use commercial LLMs to process alternative data like earnings transcripts, news flow, and broker research. Yet, each model carries biases absorbed during training, some of which are highly relevant to investors, but almost impossible to inspect from the outside.
Consider a simple example: US equities have dominated global markets for over a decade, and most of the text used to train today’s models was produced during that period. A model scoring sentiment may systematically lean positive towards American names, mistaking a historical regime for permanent truth.
Distortions can be subtler still. A model might have learned that earnings transcripts in American English correlate marginally with historical outperformance and quietly reward spelling conventions rather than the substance of management’s words. These biases hide within billions of parameters and surface only under careful, domain-specific testing. A firm relying on an off-the-shelf generic model inherits errors it cannot see. At portfolio scale, small systematic errors compound.
The model that stopped learning
The final limitation is more philosophical. Most firms build their AI capabilities on commercial foundation models: large systems trained on vast datasets (text, code, images) that learn general-purpose capabilities. Therein lies the problem: markets change every day. Foundation models, such as Claude, Llama, and GPT series, do not. They are trained, frozen and deployed, then refreshed months or years later.
A biological comparison helps illustrate the point: evolution moves over thousands of years, yet individual organisms must learn and adapt quickly, sometimes even in hours. Today’s foundation models have achieved the evolutionary part without lifetime learning. Retrieval techniques that feed models fresh information offer a partial remedy, but supplying new facts differs from genuinely updating what a model has learned.
Contrast this with what is emerging in other technology areas. Tools like Cursor, the AI coding assistant, fine-tunes its model almost daily on feedback from tens of thousands of developers, perhaps the first real-world example of a system that learns in near real time.
In investment management, where regimes shift and signals decay, a model that cannot adapt between retraining cycles is structurally behind the market it claims to analyse.
Looking beyond the vendor demo
Draw these threads together and the limitation that matters most comes into focus. Generic AI, adopted off-the-shelf, runs on the same data as everyone else’s, carries biases users cannot inspect, and stands still while markets move.
The implication for investment leaders is clear: AI has transitioned from a procurement exercise to a strategic decision. Just as species risk extinction when their ecosystem shifts, durable differentiation will belong to firms whose AI architecture can adapt in real-time. Leaders investing in systems designed for financial markets from the ground up – trained on the right data, tested for the biases that matter, and built to adapt as markets do – will ultimately enjoy the best night’s sleep.
