Every bank I have worked with has a system nobody wants to touch.
Sometimes it is a mainframe core banking platform written in COBOL. Sometimes it is a card switch that has been processing ISO 8583 messages reliably for twenty years. Sometimes it is a batch file job that runs at 2 a.m. and that exactly one engineer understands. These systems are old, but they are not broken. They move real money every second.
And yet the business wants mobile apps, open banking APIs, real-time payments and cloud analytics. The question is never "should we modernize?" The question is "how do we modernize without taking the ATM network offline on a Friday payday?"
After roughly sixteen years building ATM, POS and digital banking middleware for banks across the Middle East, Asia and Europe, here is what I have seen actually work.
The tempting plan is to replace the old core in one program. It rarely goes well. Large core replacements have a long public history of delays, outages and customers locked out of their accounts. The reasons repeat:
The alternative is incremental modernization. Less exciting, far more survivable.
Before writing new code, build an inventory of every point where systems talk to each other: protocols (ISO 8583, SWIFT MT/MX, ISO 20022, flat files, SOAP, proprietary TCP), message volumes, peak times, and who owns each interface.
In most banks this map reveals a "spaghetti" of point-to-point links. The ATM switch talks directly to the core. The mobile app calls a SOAP service that calls another SOAP service that writes to a shared database. Every new channel added another direct connection.
You cannot modernize what you have not mapped. You also cannot secure it.
The single most valuable move is introducing a governed integration layer between channels and back-end systems. This can be an API gateway plus a payment middleware or message broker, depending on the flow.
The layer does four jobs:
With the layer in place, use the strangler fig pattern: move one business flow at a time from legacy to new, while the rest keeps running untouched.
A typical order that has worked for me:
For each flow, run the new path in shadow mode first: send traffic to both systems, compare results, and only switch over when mismatches are near zero. Keep the routing switch reversible for weeks, not hours.
Integration layers are where security quietly degrades. Patterns I see repeatedly:
This matters well beyond one bank. In the United States, CISA identifies financial services as critical infrastructure, and regulators such as the FFIEC and OCC have repeatedly flagged third-party and legacy integration risk. As U.S. banks connect older cores to real-time rails like the Federal Reserve's FedNow Service and to ISO 20022 messaging, the integration layer becomes the front line of payment security.
During migration, two systems may each believe they own the truth. Automate reconciliation between legacy and new ledgers daily at minimum, and in near real time for payment flows. Every mismatch is either a migration bug or a hidden business rule you did not know about. Both are worth finding before customers do.
Teams celebrate when the new system goes live and forget to decommission the old path. Two systems running forever means double the attack surface and double the cost. For each migrated flow, define:
Legacy banking systems are not the enemy. They are proof that something worked for decades. Modernization succeeds when it respects that, puts a disciplined integration layer between old and new, moves one flow at a time, and treats security and reconciliation as part of the design.
The banks that get this right do not make headlines. Their customers simply never notice the migration happened. That is the goal.