It’s not uncommon for institutions to announce AI breakthroughs at conferences, only to return to their offices and struggle to get anything past the compliance committee. Some institutions spend millions on ambitious AI pilots that are ultimately abandoned, while their competitors deploy at scale.
The problem isn’t a lack of ambition or budget, it’s a failure to recognise that financial services do not operate like Silicon Valley. You can’t just “move fast and break things” when regulators are breathing down your neck, and your core systems were built when fax machines were cutting-edge technology.
Integrating AI into institutions’ customer and back-office platforms requires a secure and scalable environment for cloud-based infrastructure to interact with and embed AI into organisations’ technology architecture.
The integration and data challenge
Here’s where most institutions get stuck. They think AI adoption means ripping out their entire tech stack and starting fresh. That’s not just expensive, it’s dangerous, especially when their core banking or insurance systems are built on the cloud. They already process billions in transactions and handle regulatory reporting, a mechanism institutions must maintain and enhance.
The smart approach is to build on top of what you have, not replace it. Modern composable platforms can layer AI capabilities onto existing systems without significant changes, unlocking the personalisation and automated decision-making benefits of AI adoption without the pains of wholesale migration.
Financial institutions also think their data challenges are insurmountable. Handling and storing customer data causes requirements for GDPR compliance, robust security and audit trails, but these aren’t reasons to avoid AI; they’re requirements to build into your AI strategy from day one.
The key is selecting platforms that are built from the ground up for financial services. Generic AI tools struggle with regulatory requirements and necessitate substantial development expenditures to build workflows for personalisation and decision-making. Purpose-built solutions handle data governance and compliance reporting as core features, while enabling institutions to boost efficiency and personalisation with AI in a secure and scalable environment.
The expertise and explainability gap
The biggest barrier isn’t technical, it’s organisational. Most financial institutions don’t need more data scientists. They need people who understand how to identify high-impact use cases and implement AI solutions that scale.
This is where low-code approaches shine. When business users can build and modify AI-powered workflows without writing code, you democratise innovation across the organisation. Marketing can optimise customer journeys. Operations can automate risk assessments. Product teams can launch new offerings in weeks, rather than months.
Obsessions over AI explainability are common, but most institutions approach this issue incorrectly. Instead of trying to explain complex neural networks to regulators, focus on AI applications where transparency is built into the design:
-
Decision engines that combine AI insights with clear business rules.
-
Risk models that show their work.
-
Customer engagement tools that track the reasons behind recommendations.
Another consideration is hallucinations; LLMs have not yet solved their tendency to fabricate a small number of outputs. Air Canada learnt this the hard way after its customer service chatbot misled a customer into paying for full-priced tickets when the customer was eligible for the airline’s reduced bereavement rates. Employing additional AI tools to check the output of the original model reduces the likelihood of users interacting with hallucinations. However, for high-importance processes such as loan application approvals or large insurance claim processing, financial services firms should implement human-in-the-loop (HITL) capabilities, enabling trained employees to verify the output of AI-based tools.
The institutions getting it right
After working with over 50 financial institutions worldwide, from Groupe Société Générale to Admiral Group, I’ve seen what works versus what sounds good in boardroom presentations.
The pattern is clear. Start with a composable architecture that integrates seamlessly with existing core systems. Focus on business outcomes, not technology features. Enable business users to iterate quickly without depending on IT for every change. Think of products, not projects, to build reusable capabilities that can power multiple use cases.
Successful institutions layer AI capabilities on top of existing infrastructure without replacing core systems. They target specific business challenges, such as improving underwriting speed, personalising customer experiences, and automating compliance reporting. They measure success in terms of ROI and operational effectiveness, not technical achievements.
The race is on
The institutions getting this right aren’t just improving operationally; they’re expanding into new markets and business models that weren’t possible before. I’ve seen insurers launch entirely new product lines in months instead of years and banks create personalised financial products for individual customer segments, all without touching their core systems.
This isn’t about keeping up with fintech startups. It’s about leveraging the stability and trust of established institutions while gaining the agility customers now expect. AI adoption in financial services doesn’t have to be a massive, risky transformation. The institutions winning today are taking a composable approach by building AI capabilities that enhance rather than replace existing infrastructure.
The technology exists. The regulatory frameworks are becoming clearer. The competitive pressure is mounting. What’s stopping most institutions isn’t technical capability, it’s organisational inertia and fear of disrupting systems that currently work. But some institutions are already moving. The question isn’t whether to embrace AI-driven transformation, it’s how quickly you can do it without breaking what already works.
