What’s the greatest AI challenge facing financial services?

As Deloitte has observed, becoming an AI-powered financial institution is fundamentally a technology modernisation challenge.

The AI capability exists. What needs to evolve alongside it is the data infrastructure, integration architecture, and technical foundations that allow AI to operate at scale: reliably, explainably, and in real time. AI is not exposing a lack of ambition in financial services. It is exposing decades of deferred decisions.

The scale of the legacy challenge

According to McKinsey, 70% of Fortune 500 companies still operate software over two decades old. Legacy systems consume 80% of IT budgets across enterprises, leaving only 20% for the innovation that is supposed to drive competitive advantage. And 85% of enterprises say legacy systems are actively blocking AI adoption.

Financial services specifically carry a heavier legacy burden than most industries - a consequence of building complex, mission-critical systems over decades of growth, acquisition, and regulatory evolution. The situation has become acute. Legacy system integration is cited by 41% of financial services firms as their single biggest digital transformation challenge. Data quality and silos cause 35% of project failures. And despite billions invested in AI tooling and pilots, only 7% of insurers have successfully scaled AI enterprise-wide.

These are not technology problems in the narrow sense. They are strategic problems with technology at their root - and they will not be solved by deploying another AI model on top of financial services architecture that was never designed to support it.

Why legacy systems and AI are fundamentally incompatible

To understand why legacy technology blocks AI adoption in financial services, it helps to look at what AI actually requires to function at enterprise scale, and what legacy systems were designed to do.

AI requires real-time data. Legacy systems operate in batches.

Most core banking platforms, particularly those built on COBOL mainframes, were designed around batch processing: collecting transactions, running end-of-day reconciliations, and producing overnight reports. This architecture made perfect sense in 1985. It is deeply incompatible with the real-time data streams that AI models need to make accurate, timely decisions.

An AI-powered mortgage decisioning system cannot deliver instant approval if the customer data it relies on is 12 hours old. A fraud detection model cannot identify anomalies in real time if transaction data is processed in nightly batches. The mismatch is not a configuration issue. It is an architectural one.

AI requires clean, unified data. Legacy systems create fragmented silos.

Customer data at most large financial institutions is not held in a single place. It is scattered across the core banking platform, the CRM system, the loan origination platform, the mortgage system, the insurance policy engine, and dozens of other applications accumulated through decades of growth and acquisition. Each system has its own schema, its own definitions, its own version of what a "customer" or a "product" means.

When AI models draw on this fragmented data, the outputs reflect the fragmentation. Inconsistent customer profiles produce inconsistent recommendations. Duplicated business rules produce contradictory decisions. And every inconsistency erodes trust in the AI, making it harder to move from supervised pilots to autonomous deployment. Data availability and quality remain a leading pain point hindering AI adoption, cited by 66% of AI vendors working in financial services.

AI requires agility. Legacy systems are expensive to change.

One of the defining characteristics of legacy codebases is the cost and complexity of change. Many major banks still operate COBOL applications where the knowledge required to safely modify the code resides almost entirely with engineers who are approaching retirement. When that knowledge is not documented - and in many cases it is not - even modest changes carry significant risk of unintended consequences.

This creates a particularly dangerous dynamic for AI adoption in financial services. The AI is capable of learning and adapting rapidly. The systems it depends on are not. The result is a financial institution that can generate sophisticated AI outputs but cannot act on them at the speed the model was designed to enable.

AI requires transparency. Legacy systems are black boxes.

Regulatory frameworks in financial services - APRA in Australia, the FCA in the UK, and equivalents globally - require institutions to be able to explain their decisions. When an AI model denies a loan or flags a transaction, the institution must be able to demonstrate how that decision was reached.

This is extremely difficult when the data feeding the model comes from legacy systems that lack documentation, have undisclosed interdependencies, and operate according to business rules that no one has formally documented. The AI produces an explainable output. The data it relied on is not explainable. The regulatory risk is real and growing - and regulators are paying attention.

A practical example: mortgage decisioning

Consider a mortgage provider that wants to deploy an AI model to accelerate application decisioning: reducing approval times from weeks to minutes and improving the quality of credit decisions.

The AI model itself is not the hard part. The hard part is everything underneath it.

To function effectively, the model needs unified customer data across deposit accounts, existing lending facilities, and credit history - likely held across multiple systems that do not communicate in real time. It needs property valuation data from an external provider, integrated via an API that the legacy platform was not designed to support. It needs compliance logic that reflects current regulatory requirements, not the requirements of 10 years ago when the rules engine was last updated. And it needs to write its decision back to a system of record that was designed for human data entry, not automated output.

Each of these challenges is solvable. But they require the legacy issues to be addressed first. 

What modern financial institutions are doing differently

The financial services organisations moving from AI pilot to enterprise deployment are not the ones with the most sophisticated models. They are the ones that have built the right foundations.

The agenda for a CIO serious about AI-powered financial services looks like this:

Modernising core financial services platforms - not necessarily through wholesale replacement, but by systematically decomposing monolithic architectures into modular, API-first components that can evolve incrementally and support real-time integration.

Reducing technical debt - addressing the accumulated complexity that makes change expensive and integration fragile. Deloitte identifies this as fundamental to moving from AI experimentation to enterprise deployment.

Creating trusted enterprise data - establishing reliable, consistent data that AI models can draw on with confidence. This requires investment in both data infrastructure and the governance disciplines that keep data accurate and accessible.

Building AI governance frameworks - ensuring that as AI is deployed at scale, explainability, auditability, and oversight requirements are built in from the outset, not added later.

Sequencing AI deployment with modernisation - recognising that the transition from pilot to enterprise scale requires evolved infrastructure, and planning the journey accordingly.

AI is accelerating the modernisation itself

One of the most significant recent developments is that agentic AI and generative AI can now be used to accelerate the modernisation process, not just the applications that sit on top of it.

AI-powered tools can analyse legacy codebases to generate documentation that does not exist, map dependencies that are not formally recorded, surface business rules embedded in decades-old code, and produce modern equivalents that can be validated through automated parity testing. What once took years of painstaking manual work can now be approached in months.

GenAI and agentic AI are cutting modernisation timelines by 40–50% by automating code translation, dependency mapping, documentation, and quality assurance. For institutions that have been deferring modernisation because the effort seemed prohibitive, this changes the calculation materially.

The opportunity is significant. Firms that fully embrace digital transformation achieve 15–25% operational cost savings, up to 300–500% ROI on AI projects, and 25%+ revenue uplift from personalisation. Global financial services digital transformation spending reached $596 billion in 2025 and is projected to reach $685 billion in 2026, with AI and cloud accounting for half of that investment.

A strategic imperative

For financial services leaders, the framing of this challenge matters.

Legacy modernisation used to be a technology programme - something the IT department managed, reported on quarterly, and gradually progressed over the long term. Today it is a strategic imperative, because the institutions that do not build the foundations for AI-powered operations in the next three to five years will find themselves at a structural competitive disadvantage that is extremely difficult to close.

AI is available to everyone. The data infrastructure, the clean integrations, the real-time architectures, and the governance frameworks that allow AI to operate safely at enterprise scale are not. They are deliberately built by institutions that treat modernisation as a business priority rather than an IT one.

The conversation in financial services has shifted. AI is no longer the biggest challenge. Legacy technology is. And the institutions that recognise this - and act on it - are the ones that will define what an AI-powered financial institution looks like for the next decade.

How PhoenixDX can help

PhoenixDX works with financial services organisations across APAC to navigate the journey from legacy infrastructure to AI-powered operations. Our approach combines strategic advisory - helping institutions understand where to start, how to sequence the journey, and how to prioritise the investments that unlock the most value - with hands-on delivery capability using AI-powered modernisation tools and proven methodologies to accelerate the modernisation process itself.

We are problem solvers. We work with the problems our clients actually face. If your organisation is ready to move from AI experimentation to enterprise deployment, we can help you build the foundations that make it possible.

Let's talk