
Andrew Morgan
Capital markets firms know they need an artificial intelligence strategy, but many are still working out what that looks like in practice. They spend enormous amounts of time discussing, piloting and governing AI. So why is it still so difficult to turn that into practical business value?
The problem is not the technology. Excellent models are now widely available. The obstacles are architectural. The data foundations, governance frameworks and control structures that make AI defensible to regulators, clients, risk committees and chief information security officers (CISOs) inside a financial institution are not yet in place.
Full value at scale
We know that machine intelligence can be deployed successfully in this industry because systematic hedge funds have been building it into their workflows for years. They have also spent years investing in data infrastructure, software engineering and disciplined product governance. In many respects, the foundations came before the AI.
With the broad rollout and optimisation of large language models, even the most cautious institutions now have access to AI. Many use it to synthesise research, generate code and improve productivity.
But that is not realising the full value of AI at scale. What we want is to embed AI recommendations more broadly in the business. I do not mean unsupervised agentic trading desks. I mean being able to ask questions like “Show me my largest exposures across these portfolios” or “Which of my clients are most affected by today’s rate move?” and get back an answer that is accurate, relevant and grounded in business data.
For organisations without the software and data heritage of a systematic trader, AI can be daunting. Technology is moving quickly, and a steady stream of headlines about rogue systems, security failures and data leaks hardly inspires confidence among the people who have to sign it off.
These stakeholders have four main concerns about AI.
The first is repeatability. AI must produce output that can be explained, validated and reproduced. If a chatbot drafts two different versions of the same email, that is seldom an issue. But two different value-at-risk calculations generated from the same portfolio, inputs and model version are unacceptable. Systems producing poorly understood outputs that cannot be replicated are difficult to validate before they reach real-world workflows.
The second is data entitlements. Financial institutions operate complex permission structures designed to prevent sensitive information crossing desks, teams and client mandates. AI must respect those controls before data is retrieved. A useful answer is still a breach if it contains information the user was never entitled to access.
The third is semantics. Financial organisations do not lack data, but they do lack a consistent way of understanding it. Without shared definitions and business context, AI can produce answers that sound plausible but are fundamentally wrong.
The fourth is the data perimeter and alpha leakage. Trading data is intellectual property. Positions, orders and execution patterns reveal how firms generate returns. Sending that information outside the firm’s security perimeter would raise concerns about confidentiality, control and competitive advantage.
None of these four concerns is about model capability. Rather, they illustrate why so many AI initiatives struggle to move beyond the proof-of-concept stage: firms lack a way to connect AI capabilities to sensitive business data while satisfying critical governance requirements.
So how can AI be made usable inside a regulated capital markets organisation?
The architecture
Start with data. If you have data integrity, completeness, lineage and governed definitions, you have already done much of the heavy lifting. AI is only as useful as the information it is trusted to work with.
But data alone is not enough. AI also needs context. Data ontologies and semantic layers connect data across systems and functions, providing the business context that allows AI to reason consistently rather than simply retrieve information.
Once data and context are in place, AI can be incorporated into existing systems. Probabilistic AI and deterministic systems should remain separate, connected only through a controlled interface.
AI can interpret, reason and propose answers. Risk calculations, exposure calculations and benchmark methodologies remain inside deterministic software that produces the same result every time: creativity on the probabilistic side, controlled software on the deterministic side.
AI operates within existing controls, audit trails and governance processes. By default, anything consequential requires an explicit human decision, and the rate of delegation is determined workflow by workflow. Once an action is approved, it passes through the same pre-flight controls as before. Outcomes remain traceable and auditable.
Finally, sensitive data stays inside the firm’s governed boundary. The model is brought to the data rather than the data being sent to the model. Modern deployment architecture makes it possible to keep proprietary information inside the firm’s security perimeter and under the firm’s control, with no retention by model providers and no use of firm data for training.
Race to value
We can think about AI success in terms of “time to value”: how long it takes to move from “this looks interesting” to “I can use this in my business in a defensible way”. Yet many firms remain stuck between proof of concept and production.
Few organisations are limited by model quality. To pull ahead, what they need is governed definitions, deterministic calculation, lineage back to origin, entitlements enforced at the data layer, and all of it running inside the organisation’s own perimeter. The firms that move fastest will treat AI as an architecture problem, not a model-selection problem.