Why data architecture is the foundation for AI in investment operations

Artificial intelligence is shifting the financial services industry, and as firms move towards adoption, there are questions around where to start. Firms are increasingly testing AI integration as a way to further automate human intensive activities such as reconciliation, document processing, reporting and client servicing. Much of the commentary on this shift focuses on models and use cases. Less attention is paid to the data architecture that determines whether those use cases can actually work. 

That gap matters more in financial services than in most other industries. This is a heavily regulated environment, where outputs need to be accurate, explainable and auditable, not just fast. Before a firm can build AI-native tooling, it needs a data foundation that can support it. 

What AI-native tooling needs to work in a regulated environment 

The starting point for most financial services software, traditional or AI-enabled, is data. Firms need trusted and reliable data sources. Without this, the quality of any AI output is at risk, regardless of how sophisticated the underlying model is. 

This is not a new problem. Data quality has always mattered in investment operations. What has changed is the consequence of getting it wrong. 

When a human analyst works from an incomplete data set, the error is usually visible and containable. When an AI system works from the same incomplete data set, it can generate a confident-sounding output that hides the underlying gap. That output can also move faster through a process and touch more decisions before anyone notices. 

Trusted data is only part of the picture. Firms also need repeatable, auditable processes. Relying on answers or protocols that can change from one instance to the next introduces a level of unpredictability that is difficult to justify in a regulated setting. 

Investment operations teams are accustomed to demonstrating exactly how and why a number was produced. AI tooling cannot be an exception to this rule, and the fallout of mistakes is not diminished if they were caused by AI. 

This aligns with how regulators are approaching the topic. International securities regulators, coordinating through IOSCO, have identified data quality, model explainability and human oversight as recurring themes in their review of AI use across capital markets (IOSCO, Artificial Intelligence in Capital Markets, March 2025). In the UK, the Financial Conduct Authority has taken the position that AI will be supervised through existing rules on governance, accountability and consumer outcomes, rather than through a dedicated AI rulebook (FCA, AI and the FCA: our approach). 

In practice, that means firms are expected to be able to explain and evidence their AI-enabled processes in the same way they would any other operational process. There is no separate, lighter-touch standard for a process simply because AI sits inside it. The data feeding that process, and the trail showing how it was used, need to hold up to the same scrutiny as everything else the firm does. 

Moving from legacy systems to AI-suitable architecture 

Recognizing what good tooling requires is one step. Getting there from a legacy environment is another. Most firms are not starting from a blank slate. They are working with systems and processes that were built for a pre-AI world, often layered over years of acquisitions, integrations and workarounds. 

The instinct in this position could be to build everything internally, to suit the specific environment. That is rarely right. Firms should evaluate their existing processes and decide case by case whether to build or buy. External providers offer access to existing investment in data infrastructure, quality and auditability alongside AI capability. 

Any AI and data roadmap should be assessed with the same rigor as any other strategic decision. Considerations include data ownership and portability, whether governance and auditability match what the firm’s regulators expect, and whether the AI direction fits what is feasible for the firm. 

What a solid data foundation unlocks 

Once the data foundation is in place, the benefits extend beyond the immediate AI use case. Securing the foundation changes what investment operations teams are able to do day to day. 

The first and most direct benefit is that talented resources are freed up to focus on higher-value tasks. Skilled analysts currently spend a meaningful proportion of their time reconciling, chasing and validating data rather than interpreting it. A stronger foundation reduces that burden and allows people to spend more time on judgment-based or client relationship work. 

Second, the process of improving the foundation creates a natural opportunity to revisit legacy processes and change them to fit today’s needs, rather than needs from years ago. Many investment operations processes were designed around the constraints of older systems. Rebuilding the data layers for different parts of the business forces a firm to ask whether those processes still make sense, or whether they were simply inherited. 

Third, a stronger data foundation leads to fewer errors. This follows directly from the first principle: firms need trusted, reliable data sources, and when they have them, the output built on top of that data is more reliable too. Fewer errors also mean less time spent on remediation, which compounds the resourcing benefit already described. 

Fourth, and closely related, a strong foundation supports stronger auditability. In a regulated industry, being able to show how a number was derived, what data fed into it and what process produced it is not optional. A firm that has invested in its data architecture is better placed to meet that standard consistently, across a growing range of AI-enabled processes, rather than firefighting it case by case. 

The challenge is not insurmountable 

Data foundations have long been a challenge. What differs today is the number of routes available. A firm can improve the data behind one process at a time and build from there, rather than running a single firm-wide programme. AI can also be pointed at the data estate itself, where validation rules, monitoring, lineage and governance all respond to the same tooling. And alternative investments carry data sets that resisted structuring for years, which extraction tooling now reaches. 

Financial services runs on precision, trust and the ability to show work. AI raises the stakes on all three. The practical path is to identify the use cases where better data architecture has the greatest effect and work through them in increments. Firms that start early will be better placed to adopt AI safely.