Asset managers are waking up to an AI limitation that could reshape how they compete, according to a recent industry survey.
Data quality and compliance are top concerns
When asked what keeps leaders up at night, three issues dominate: the “black box” nature of models, data quality, and regulatory constraints. The survey shows that 69% of firms point to data quality and access as the main barrier to adopting AI, while 59% highlight compliance worries.
Both challenges are gradually being addressed as governance frameworks evolve and model transparency improves. Yet the deeper problem lies in the infrastructure that underpins most AI deployments in asset management.
Commoditisation and the hidden systemic risk
Nearly all managers—91% according to the same study—plan to broaden AI use within a year, making the claim “we use AI” a weak differentiator. Most rely on generic foundation large language models (LLMs) that are trained on public data and offered for a subscription fee. As a recent industry report notes, baseline analytical capabilities are becoming commoditised, shifting any real edge toward judgment and proprietary data.
Related: European ETFs eye new markets
When many firms run similar models on similar data, a new risk emerges: herding. The study found that 24% of managers see system‑level risks, such as crowd‑following trades, as the biggest regulatory blind spot. If a single firm’s model fails, the damage is limited; if the market’s underlying model aligns, the risk spreads across the system.
Specialised models, trained on unique data sets, avoid this pitfall because they cannot be replicated through a subscription. However, building such models demands resources that many firms lack.
Invisible biases in shared models
Commercial LLMs are increasingly used to process alternative data—earnings transcripts, news flow, broker research. Each model carries biases absorbed during training, many of which are hard to detect from the outside. For example, because U.S. equities have dominated markets for a decade, models trained on recent text may systematically favor American companies, mistaking historical prominence for lasting advantage.
Biases can be subtler. A model might associate certain spelling conventions in earnings calls with better performance, rewarding form over substance. These hidden errors sit within billions of parameters and only surface under detailed, domain‑specific testing. At portfolio scale, even small systematic mistakes can compound.
Related: Grants for innovation with R&D tax relief and loans: avoiding double funding
Foundation models that stop learning
The final limitation is more philosophical. Most firms build AI capabilities on commercial foundation models—large systems like Claude, Llama, and the GPT series—that are trained once and then frozen. Markets, by contrast, shift daily, leaving these models lagging behind the reality they aim to analyse.
Retrieval techniques that feed fresh information to a model offer a partial fix, but they do not update the underlying knowledge the model has learned. In investment management, where regimes change and signals decay, a model that cannot adapt between retraining cycles is structurally outpaced by the market it seeks to interpret.
AI adoption is accelerating.
Firms that invest in AI architectures designed to evolve with market conditions—training on proprietary data, rigorously testing for hidden biases, and enabling continuous learning—stand to gain a competitive edge. Those that rely solely on off‑the‑shelf solutions may find themselves stuck with outdated insights, a situation that could keep executives awake long after the markets close.
Leave a Reply