Tomasz Krakowczyk, Financial Services General Manager at Software Mind
In many cases, improving the model is only part of the challenge. Scaling AI also depends on whether core systems can provide timely data and support the workflows around it. We call this constraint the core ceiling.
What the core ceiling looks like
Core banking and policy administration systems were built to keep records accurate, and many still do that job well after decades in service. Problems arise when an AI use case needs something the core was never designed to provide.
In our work with financial institutions, the ceiling usually appears in one of three ways:
•Batch data: Fraud detection or next-best-offer models need events within seconds. A core that reconciles overnight provides yesterday's picture.
•Closed interfaces: Without APIs or event streams, every new use case needs its own integration project, and each one joins a long queue.
•Tight change windows: When the core is released quarterly, an AI feature that needs to write back – update a claim, flag an account – waits for the next slot.
A better large language model (LLM) won't fix any of these. They sit in the architecture, so that's where they need to be addressed. Consider an insurer using AI to help triage claims. The model may be ready, but if policy information is updated overnight and claims handlers must manually transfer its recommendations into the claims system, the process remains slow and the benefits are limited. Making relevant policy data available sooner and integrating recommendations into the claims workflow can help turn the pilot into an operational capability.
Why replacing the core isn't the starting point
The instinctive response is a full core replacement. For many banks and insurers, a full replacement is a multi-year program with significant delivery risk. If AI initiatives are made dependent on its completion, they may wait years to deliver value.
The complexity is real. McKinsey describes insurance cores as holding "several decades' worth of sparsely documented embedded business rules" – logic that has to be rediscovered before it can be moved. Even if the work is worth doing, it shouldn't hold back every AI initiative in the meantime.
The pressure to act is growing. KPMG found that 71% of banking leaders now agree they need to invest in modernizing platforms to bring new products and services to market, up from 46% in 2025. The question is where to begin.
Modernize what AI touches first
A more practical route is to map an AI roadmap to the parts of the core it depends on and modernize those first. Three areas come up in almost every project:
1A data access layer. Replicate core data into a governed, near-real-time store, for example through change data capture (CDC). AI can read from this layer, instead of querying the system of record directly, which improves data availability and freshness where the source system supports it.
2APIs and events around the core. Expose key domains – customer, account, policy, claim – through well-defined APIs and publish business events. New use cases connect to a stable interface instead of adding another point-to-point integration.
3Decisioning outside the core. Where appropriate, separate selected rules and decisioning capabilities from the core. This can allow teams to test and update them more independently, while maintaining clear controls over how decisions are approved and recorded. AI can support these decisions while the core remains the system of record.
Each step can support use cases already on the roadmap and create reusable building blocks for a future core migration.
Sequence the work by AI use case
To decide where to start, work from an AI roadmap, not an architecture diagram:
1List the AI use cases planned for the next 12–18 months.
2For each, note what it needs from the core: which data, how fresh and whether it has to write back.
3Look for overlaps. Capabilities that serve several use cases – customer data, transaction events, claim status – go first.
44. Set a measurable target for each capability, such as data latency or the time needed to integrate a new use case.
This ties modernization to outcomes your business already wants. Instead of defending a core program with a distant payoff, teams fund the capabilities that unblock AI projects already approved.
Key takeaways
•In banking and insurance, legacy core systems can be a major barrier to scaling AI, particularly when use cases depend on timely data or changes to core workflows.
•The core ceiling usually shows up as batch data, closed interfaces and slow change windows.
•Full core replacement isn't a prerequisite for scaling AI, and making it one can stall the roadmap for years.
•Start with what AI touches: a data access layer, APIs and events, and decisioning outside the core.
•Sequence the work by overlapping use case needs, and track progress with measures like data latency and integration time.
To learn how you can accelerate your AI roadmap, get in touch with our team.
About the author
Tomasz Krakowczyk is Financial Services General Manager at Software Mind. With over 15 years of experience in IT management, Agile coaching and transformation leadership, he helps financial organizations align software delivery with business strategy.
About Software Mind
Software Mind, a global technology partner, provides consultancy and software development powered by AI accelerators. Delivering deep AI, cloud and data expertise, Software Mind turns technology into measurable business value so that companies across industries can thrive in competitive markets.
www.softwaremind.com
Recent Stories