Every major financial institution has published a statement about its AI strategy in the last two years. The statements are remarkably similar. Transformation. Efficiency. Enhanced decision-making. Competitive advantage. Most of them describe ambition rather than deployment. The gap between what gets announced and what is actually running in production is large enough to matter.
Here is the honest inventory.
What is genuinely deployed and doing real work: natural language processing on news and regulatory feeds, running in real time at most significant trading operations. This is not new technology. Banks and hedge funds have been extracting sentiment signals from text for over a decade. The models have improved substantially. The edge, where it exists, is in the quality of the training data and the speed of the pipeline rather than the underlying technique.
Research summarisation using large language models is deployed at scale across the bulge bracket. Analysts have tools that process earnings transcripts, competitor filings and macro reports faster than any human reading programme. The output is a starting point, not a conclusion. The models hallucinate with confidence, which in a financial context is not an amusing quirk: it is a material risk. A model that invents a revenue figure from a filing it did not read, and presents it with the same tone as a figure it did read, is not a research tool. It is a liability unless the human reviewing the output is genuinely checking rather than skimming.
Algorithmic execution and quantitative strategy have used machine learning for longer than the current conversation implies. Systematic funds have been running machine-learning-driven strategies for thirty years. What is new is the accessibility: the tools that required a team of PhDs and proprietary infrastructure in 2005 are now available to a two-person fund via cloud compute. The edge has compressed. The application is real.
Compliance and transaction monitoring is the unsexy deployment that generates the most genuine operational value. AI-driven AML pattern detection has improved substantially over rules-based systems. False positive rates have come down. The humans reviewing flagged transactions are spending more time on genuine cases and less time on noise. This is not the transformation narrative the press releases describe. It is a meaningful, compounding improvement that will not be in any keynote.
If a model could reliably predict directional market movements it would be the most valuable intellectual property in human history. It would not be a monthly SaaS subscription.
What is theatre: most vendor product announcements, most AI-powered market prediction claims, most co-pilot interfaces that wrap a general-purpose language model in a financial skin and call it a trading tool. If a model could reliably predict directional market movements it would be the most valuable intellectual property in human history. It would not be a monthly SaaS subscription at a fintech conference.
What changes next is more interesting than what currently exists. The direction of travel is toward agentic systems: AI that does not just answer questions but executes workflows. Portfolio rebalancing, compliance checking, report generation. The firms building proprietary models trained on their own data, rather than fine-tuning on public datasets, are building a compounding moat. The regulatory response is coming and will arrive as explainability requirements: if your model makes a decision that affects a client, you will need to show your working.
The talent that matters right now is not the pure quant or the pure technologist. It is the person who understands both the financial problem and the model's actual limitations with equal clarity. That intersection is small and the market for it is extraordinary.
Most of the room is still applauding the announcement. The real work is quieter than that.