Every bank's strategy deck promises a modern core. Far fewer have one. The reason is not a lack of vendors or ambition; it is that the traditional route — replace the core in one programme and migrate everything at once — concentrates years of change into a single, high-stakes cut-over. When it goes wrong, it goes wrong in front of customers and supervisors.
What big bangs cost
The most cited cautionary tale is the April 2018 migration at TSB in the UK. The data moved, but the new platform failed on contact, disrupting branch, telephone, online and mobile banking. A significant proportion of the bank's 5.2 million customers were affected, and business as usual did not return until December 2018. In December 2022 the FCA and PRA fined the bank a combined £48.65 million, after it had paid £32.7 million in redress; regulators found that it had failed to organise and control the migration adequately and to manage the operational risks of its IT outsourcing1.
Even programmes that land successfully tend to cost more and take longer than planned. When a large Australian bank announced its core banking modernisation in April 2008, it forecast a cost of around A$580 million over four years, and chose a staged migration explicitly to mitigate risk2. It declared the programme complete in October 2012, after five years and more than A$1 billion of investment3. That was a strategic success — and still close to double the original estimate.
Core replacement: the plan and the outcome
Australian bank core banking modernisation, A$ million (A$m)
Note: Forecast from the bank's April 2008 release [2]; completion figure reported as 'over $1 billion' [3], so the bar is a lower bound.
Source: Australian bank investor release, “Core banking modernisation (media release)” (2008)
These cases are consistent with broader transformation data. One analysis of digital transformations found that 70% fell short of their objectives: 30% met or exceeded their targets, 44% created some value but missed them, and 26% created little or none4. One of the six success factors it identified was a modular, business-driven technology and data platform4.
Most large transformations miss their targets
Outcomes of digital transformations, % of cases (%)
Source: Published research, “Flipping the odds of digital transformation success” (2020)
The progressive alternative
Progressive modernisation accepts that the core will be replaced over years, and designs for that. Instead of one migration, the bank executes a sequence of smaller, reversible moves, each of which delivers value and retires some legacy. The pattern has four elements:
- Data layer first. Stream data from the legacy core into a governed, real-time data layer with a common model, lineage and quality controls. New products, analytics and AI read from this layer, not from the mainframe — which immediately reduces load on, and dependency upon, the core.
- Hollow out the core. Move capabilities with the highest change rate — pricing, product configuration, limits, customer servicing — into modern services outside the core, leaving it as a thinner system of record.
- Integrate through APIs and events. Put a stable API and event layer between channels and the core, so that front-end change stops requiring core change and the core behind the interface can be swapped segment by segment.
- Migrate by cohort. Move products or customer segments one at a time onto the new platform, running old and new in parallel with automated reconciliation between them, and keep a tested route back.
The data layer also changes the business case. Traditional core programmes deliver most of their benefits at the end, after the final migration. A data-layer-first approach starts paying back early: real-time analytics, regulatory reporting and AI use cases can all run on the new layer long before the last product leaves the legacy core. That early value is what sustains sponsorship through a multi-year journey.
Where AI changes the economics
The most expensive part of legacy modernisation has always been understanding what the old system actually does. Anthropic argued in February 2026 that the discovery and analysis phases — mapping dependencies, tracing execution paths, documenting business logic nobody remembers — can now be substantially automated, and that COBOL, which it estimates handles about 95% of US ATM transactions, can be modernised in quarters rather than years5. Markets took the claim seriously: IBM shares fell 13% on the day, their steepest one-day fall since 2000, although IBM countered that code translation captures little of the real complexity of mainframe estates6.
Early bank evidence is encouraging, if vendor-reported. A Swiss private bank's generative-AI-assisted modernisation with MongoDB reported migrating code 50 to 60 times faster than previous migrations and moving applications off legacy relational databases 20 times faster7.
The right framing is that AI accelerates comprehension, test generation and incremental refactoring — the work that makes progressive modernisation practical — rather than making big-bang replacement safe. A faster translation of the old core into a new language, cut over in one weekend, carries the same operational risk as before.
AI does not remove the need to migrate carefully. It removes the excuse for not knowing what you are migrating. (SCIKIQ view)
Governance for a multi-year journey
Progressive programmes fail differently from big bangs: not in a single outage but in drift, where the bank ends up running old and new cores indefinitely. The antidotes are a fixed decommissioning plan with dates, a business case measured in retired legacy as well as new capability, and operational-resilience testing of every migration step — the discipline whose absence the TSB findings highlighted1.