Migrating Core Banking Data Without Downtime: A Phased Approach for Bank Technology Leaders

Core banking modernization guide: run legacy and new cores in parallel, sync data continuously and migrate in reversible waves without downtime.

|
Nov 16, 2026

Summarize with AI:

Core banking modernization succeeds or fails on data movement, so the sync and reconciliation layer deserves the same scrutiny as the core vendor, because a cutover weekend is where configuration errors and code defects meet live customers with no route back.

Avoiding downtime takes six stages: assess every consumer of the legacy core, choose phased or componentized change unless one big-bang cutover can be rehearsed end to end, run both cores in parallel aligned by change data capture and daily reconciliation, migrate customers in waves starting with staff accounts, cut over each wave behind a rehearsed rollback, and decommission the legacy platform. Supervisors hear about the plan before vendor selection and get results after every wave.

Summary

  • Choose a phased or componentized approach unless the bank is small enough, or its product set simple enough, to rehearse a single cutover end to end.
  • Keep the legacy core as system of record during coexistence, and feed the new core and a separate reconciliation store from its transaction logs.
  • Move customers in cohorts that start with staff accounts and simple products, and close each wave only when record counts and balances tie out.
  • Give every wave a rehearsed path back: a freeze point, a reverse sync and one named person with authority to call it.
  • Brief supervisors before a vendor is chosen and keep test results, reconciliation reports and exit plans ready as evidence.

Learn more about Hybrid Deployment

What Core Banking Modernization Involves

Core banking modernization involves more engineering effort between coexistence and cutover than at the start or end of the program. Those middle phases carry most of the risk because customer records exist in two places at once, and both copies have to agree before the legacy system can be retired.

The core itself is the system that holds deposit and loan accounts, posts transactions, calculates interest and fees, and produces the balances every other system trusts. Core banking modernization changes that system of record, either by replacing it, by moving functions out of it one at a time, or by standing up a new core beside it. The approach chosen sets how long the bank runs two cores, and a longer dual-running period means more engineering to keep them consistent.

Assessment deserves more time than most programs give it. A bank that cannot list every downstream consumer of the legacy core, from statement printing to regulatory reporting, will discover the missing ones during a wave, when each discovery costs a rollback.

Why Legacy Cores Have to Change

Bank leaders weighing cost, skills and competition should scrutinize the batch-oriented core rather than the hardware under it. Separating the processing model from the hardware helps them decide what the bank has to replace.

Focus replacement decisions on batch processing and the skills needed to maintain the core. IBM mainframe use is growing again and exit initiatives are declining, while the mainframe skills shortage and ISV pricing remain open problems. COBOL is the sharper constraint: both Treasury systems that GAO selected for modernization run on COBOL and Assembly, languages with a shrinking pool of support staff.

Maintenance cost compounds the skills problem. Banks spend up to 70% of IT budgets maintaining technical debt alone, and banks still spend more than 80% of hosting budgets on legacy infrastructure rather than cloud services. Money spent keeping a monolith alive is money unavailable for the products customers notice.

Competitive pressure from fintechs and neobanks is now a direct driver of core modernization. Mature fintechs hold 17 percent of industry revenues, by one measure, and Revolut has more than 65 million customers worldwide while Nubank has more than 122 million across Brazil, Mexico and Colombia, with cloud-native technology and agile development models that position them to take more customers from banks. Leading neobanks are also moving from single products to broader financial platforms: Chime entered lending in 2025, Revolut received a full UK bank license, and Nubank grew credit in Latin America. Banks feel the pressure, as one in three retail banks treats product and proposition innovation as its single biggest IT spending focus, tied to customer retention and the fintech challenge, in a 2025 survey of 245 retail banking respondents. A bank that keeps funding a monolith cedes ground to rivals that already hold that share, and because how fast a new core can launch products now matters commercially, the migration approach is a business decision, not only a technical one.

Choosing an Approach: Big-Bang, Phased or Componentized

Phased migration is the right default for most banks because it turns one irreversible event into many reversible ones. Big-bang still makes sense for a small institution with a narrow product set, a single legacy core and a vendor that has converted many banks of that shape, where rehearsing the full cutover several times is realistic.

The cost of getting big-bang wrong is documented. The failed 2018 TSB migration cost the bank £412.16 million in 2024 prices, including £156.28 million in redress, rectification and remediation. Big-bang cutovers are exposed to three weaknesses: testing compressed to hit a public date, test environments that differ from production, and no practical way back to the old platform once customers are live.

Phased programs carry their own cost, mostly in dual running. BancoEstado moved 14 million customers to a cloud-native core in a phased migration over a few years, running both platforms in parallel. Those years of parallel operation are the price of keeping each step reversible.

Comparing Four Core Modernization Options

Each option trades program length against how concentrated the cutover risk is:

OptionRiskTimelineCost profileDowntime exposure
Big-bang replacementHighest; one cutover carries every defectShortest program, longest single outage windowLower dual-running cost, very high cost of failureConcentrated in one event; rollback often impractical after go-live
Wrap and extendLowest near term; legacy core staysMonths to first APIsLowest upfront; legacy maintenance continuesMinimal, because the core does not change
Strangler or hollow-outModerate, spread across wavesMulti-yearDual running plus integration workLimited to each wave and reversible while sync holds
Greenfield parallel bankModerate; new customers go firstFast launch, slow migration of the existing bookTwo banks operated in parallelLow for new customers; existing book still has to move

Wrap and extend buys time without retiring anything, so treat it as a bridge to one of the other three. Greenfield works when a bank wants a digital brand quickly, but it postpones the hard part: the existing book still has to migrate, and that migration needs everything the strangler approach needs.

Build, Buy or Partner for the New Core

Buying a vendor core is the better call for almost every bank, and building only pays when the core itself is how the bank competes. A digital bank whose product model no packaged core supports may justify a build; a regional bank with deposits, mortgages and cards does not.

The trade-off runs across three axes. Building gives full control of the data model and release timing but demands a large in-house engineering team for years. Buying a licensed core the bank hosts keeps control of hosting and data while leaving upgrades and integration to the bank. A SaaS core shifts operations to the vendor and shortens time to value, at the cost of release timing and dependence on the vendor’s roadmap.

Vendor dependence is the axis easiest to underweight. Contracts for packaged cores can limit the rights to modify or move the code, so the exit terms belong in the selection criteria from the first meeting, not in the renewal.

Running Legacy and New Cores Side by Side

Coexistence avoids customer-facing downtime: while both cores run, the legacy core stays the system of record for every customer not yet migrated, and ownership of a defined cohort transfers only after its data is validated in the new core. While sync holds, customers do not wait for a cutover window, because each account stays live on whichever core is authoritative for its cohort.

The engineering burden shifts to synchronization. Every posting on the legacy core has to reach the new core, the reconciliation store and downstream consumers in order, and every discrepancy has to surface before customers see it.

Syncing Records With Change Data Capture and Event Streaming

Log-based CDC minimizes added load on production tables compared with query-based extracts, which makes it the best fit for keeping up with a live core. Batch extracts query the tables directly, compete with posting jobs and leave hours of drift between systems. Log-based capture reads the database’s transaction log instead, so each committed change arrives with a log position that preserves commit order within that database.

Ordering holds within one database log, and it gets harder everywhere else. Clustered databases with several redo threads or data-sharing members have to merge their log streams before global order is reliable.

Partitioned event streams keep order per key, so keying the stream by account number keeps all events for one account in order, but nothing orders events across keys. Long-running transactions add another wrinkle, because a capture process has to buffer changes until it sees the commit or rollback. Key streams by account, design consumers around per-key ordering and idempotent writes, and do not assume a global sequence.

The initial load needs the same care as the stream. Interleaving chunked table reads with log events and bracketing each chunk with watermarks produces a consistent snapshot without table locks. The alternative, freezing tables for a full export, recreates the outage that coexistence is supposed to prevent.

Run sync one way per entity wherever possible. For each cohort, exactly one core accepts writes, and the other receives replicated changes. Bidirectional sync on the same records forces conflict detection, loop prevention through tagged writes and a rule for which side wins, and value-based conflict checks can miss real conflicts. If reverse sync is needed, and it will be for rollback, keep it dormant until a wave fails rather than running both directions continuously.

Reconciling Balances Every Day

Coexistence is working only if reconciliation ties out at three levels every day. A single nightly total proves very little; a total that ties at every level is the strongest daily evidence that the next wave can proceed.

Reconcile at these layers every business day:

  • Ledger totals: General ledger balances per product and currency match between the legacy core, the new core and the reconciliation store.
  • Account balances: Every migrated and shadowed account has the same closing balance, available balance and accrued interest in both cores.
  • Transaction counts and values: Postings per account per day match by count and by summed value, which catches a missing transaction that a balance check can hide.

Break thresholds should differ by field class. Balances, accrued interest and transaction values get zero tolerance, because a one-cent difference is a defect with a cause. Reference data such as addresses or marketing preferences can carry a small documented tolerance during early waves. Timing differences, where a posting lands on either side of a cutoff, get their own category so they do not hide real breaks.

Every break needs an owner by the next morning. Assign product operations to balance and posting breaks, the data platform team to missing or duplicated records, and the core vendor to calculation differences, with a single migration lead who decides when a break pattern pauses the next wave. Unowned breaks accumulate, and accumulated breaks are what make a wave impossible to roll back cleanly.

Where Airbyte Fits in Coexistence

Airbyte fits as the capture and replication layer that feeds the reconciliation store, not as the path that posts transactions into the new core. Airbyte replicates from relational databases with CDC that does not lock production tables, into destinations such as Snowflake, BigQuery, Databricks, Postgres or Iceberg on S3, where the reconciliation queries above run against an independent copy of legacy data. Schema change detection flags when the legacy schema moves, and checkpointing and retries let interrupted syncs resume from saved state.

Airbyte Flex matters here because bank data rarely gets to leave the bank, and its hybrid deployment model places the data plane inside the bank’s own environment. Airbyte operates the control plane while the data plane, along with your records, credentials, keys, and compute, runs in your environment, though some metadata such as cursor and primary-key values sits in the control plane. Flex runs on AWS, Google Cloud, Azure or on-premises infrastructure, and because the connectors are open source, the bank’s security team can inspect the code that reads its ledger instead of relying on a vendor’s description of it.

The limits are worth stating plainly. Replication runs as scheduled syncs and the schedule is plan-dependent, so freshness is as good as the last sync rather than that of a sub-second event bus. Checkpointing does not amount to an exactly-once guarantee, which is one more reason the reconciliation layer exists. Connector support varies by source, so check the documentation for the specific Db2, Oracle or SQL Server version in the catalog of 700+ connectors before designing around it. Mainframe file stores such as VSAM and mainframe databases such as IMS may need a mainframe-specific capture tool upstream. Ledger writes into the new core belong to the core vendor’s migration tooling, and conflict resolution between two writable cores belongs to whatever replication product the bank chooses for that job.

Migrating Customer and Account Data in Waves

Cohort-based migration is the main risk control in the whole program because it caps how many customers any single defect can reach. A wave that moves 2,000 staff accounts and breaks costs an awkward week; a wave that moves the entire retail book and breaks costs the bank its reputation.

Sequencing Waves by Product or Customer Cohort

Start with staff accounts, then simple retail deposits, then products with more moving parts. Staff accounts exercise real postings, cards and statements while the affected customers sit inside the bank and can report problems directly. Simple savings and current accounts follow because their interest and fee logic is easiest to verify. Mortgages, overdrafts, joint accounts and business customers with complex mandates come last, once the reconciliation tooling has proven itself.

Wave size should follow four constraints, whichever binds first. The reverse sync has to be able to restore the cohort to the legacy core inside the agreed fallback window. Contact center capacity has to absorb the calls a defect would generate. The reconciliation team has to clear every break for the cohort within one business day. And the cohort should contain one product family at a time, so a calculation defect points at one mapping. Early waves should be small enough that all four constraints hold comfortably; later waves grow as each constraint is shown to have headroom.

Validating Data Before and After Each Wave

Validation runs twice per wave, once in rehearsal and once on the live cohort, and the same checks apply both times. The professional internal-audit expectation is that converted data reconciles with the legacy system, that data owners verify dry-run results, and that no urgent or high issues remain open from the dry run.

Each wave’s validation pack should cover:

  • Field mapping: Every source field maps to a target field or is explicitly retired, with mapping rules documented and tested.
  • Record counts: Customers, accounts, mandates and open transactions match by count between source extract and target load.
  • Control totals: Summed balances, limits and accrued interest per product match to the cent.
  • Data quality rules: Mandatory fields are populated, dates fall in valid ranges and every account links to a customer.
  • Sign-off owners: A named business owner per product signs the pack, alongside the data platform lead and risk.

Business sign-off must not be treated as a formality. The person who signs for mortgages should be someone who would personally answer for a wrong repayment amount, because that sign-off is the last check before customers see the data.

Cutover and Rollback Planning

Every wave needs a tested way back, and “tested” means executed in rehearsal with production-sized data, not written in a runbook. The FFIEC IT Handbook’s development, acquisition and maintenance booklet includes a rollback or back-out plan among its IT change topics, and a core migration is the largest change most banks will ever make.

Dress Rehearsals and Go/No-Go Criteria

Run at least two full mock cutovers per wave on an environment configured identically to production, including both data centers if the new core runs active-active. A rehearsal that skips the configuration live customers will hit proves only that the rehearsal environment works.

Time every step. The mock cutover should show how long extraction, load, validation and switch-over take, and those measured times set the live window. If rehearsal timing overruns the window, shrink the cohort rather than compressing validation.

Go/no-go criteria should be numbers a steering committee cannot argue with: zero unresolved balance breaks in the final rehearsal, record counts matched for every entity, all priority-one and priority-two defects closed, reverse sync executed successfully within the fallback window, and performance tests passed at peak volume. A program that reaches the decision meeting with exceptions to its own criteria has already answered the question.

Fallback Steps If a Wave Fails

Fallback works only if the bank keeps the legacy core able to take the cohort back. That means a defined freeze point before cutover, after which no legacy changes for the cohort are accepted without capture, and a reverse sync that replays any postings made on the new core back to the legacy core during the stabilization window.

Decision authority has to be settled before the weekend. Name one executive who can order a rollback without convening a committee, give that person explicit triggers such as unresolved balance breaks or a posting backlog past a set threshold, and set a deadline after which rollback is no longer possible because the new core has accumulated too much activity to unwind.

Customer communication belongs in the fallback plan, written in advance. Customers in a failed wave need to know whether their balances are correct, which channel works and when the next attempt happens. Drafting that message during an incident is how banks end up issuing statements that later prove wrong.

Keeping Regulators Informed Throughout the Program

Early, documented engagement with supervisors costs a bank a few meetings and saves it from discovering requirements after a cutover date has been announced. The program’s evidence trail is also its regulatory file.

When to Notify Supervisors

Notification rules differ by jurisdiction, and they change, so confirm the current text for each regulator before the program starts. The pattern across regimes still gives a usable plan: tell supervisors at program start, again before the first live wave, and after each wave with results.

In the EU, DORA has applied since January 17, 2025, and requires firms to inform the competent authority in a timely manner about planned ICT contracts that support critical or important functions. For non-ICT third-party arrangements, EBA guidelines require timely notice of material changes or severe events affecting arrangements behind critical functions. An EU bank should therefore map these notice triggers before it signs any vendor contract.

The UK is the most explicit about migrations. Under the PRA’s updated outsourcing statement, a firm planning a major migration program may need to give notice before a provider is selected, with notifications submitted through FCA Connect; that update takes effect on 18 March 2027. The FCA’s parallel regime for material third party arrangements applies from the same date. A UK bank should therefore treat vendor selection as the first notification point rather than an afterthought.

US supervisors work through different triggers. FDIC-supervised banks must notify the agency of a service relationship within 30 days of the contract or first performance, and any computer-security incident that materially disrupts operations requires notice to the primary regulator within 36 hours. Neither rule is a pre-approval of a migration plan, so in practice the engagement runs through the examination relationship, and a failed wave starts the 36-hour clock only if it amounts to a computer-security incident that materially disrupts operations.

Evidence Supervisors Expect

Supervisors expect to see the same artifacts the program uses to make its own decisions, which is the best argument for producing them properly. Build the evidence file as the program runs rather than assembling it before an examination.

The core of that file looks like this:

  • Test results: DORA requires at least yearly testing of ICT systems that support critical or important functions, and the UK requires firms to keep testing and lessons-learned records for at least six years.
  • Reconciliation reports: Daily break reports, sign-off packs and break resolution records for every wave show that data moved accurately.
  • Exit plans: The bank documents and tests exit strategies for the new core vendor, covering both a stressed exit after vendor failure and an orderly one.
  • Third-party risk reviews: The bank completes a pre-contract risk assessment at entity level covering operational, legal and ICT risk, and monitors the vendor on an ongoing basis.
  • Vendor test participation: Management should request and review results and remediation plans from the core vendor’s own resilience testing.

US guidance is in flux. The agencies proposed replacing the 2023 interagency guidance with more tailored third-party guidance in September 2026, and that proposal is not final, so a US bank should recheck the current text before each wave rather than build its evidence file around a draft.

Data Residency and Cloud Targets

A cloud-hosted core turns data residency into an architecture decision that has to be made before vendor selection, because some jurisdictions restrict where banking data may sit at all. For EU personal data transferred outside the EEA, GDPR transfer rules require an applicable transfer mechanism or derogation, and the EU-US Data Privacy Framework survived an annulment action at the General Court in September 2025. Where the bank can place its core therefore depends on the jurisdictions whose data it holds.

Several jurisdictions impose harder limits. India requires payment system data to be stored only in India. Saudi banks need SAMA approval before using cloud services, and use of cloud outside Saudi Arabia needs explicit approval. For banks in these markets, the cloud region cannot be picked until the approval or location rule is settled.

Indonesia goes further by tying approval to specific systems. Indonesian banks need an OJK permit to place systems in a data center abroad or to hand transaction processing to a provider outside Indonesia. In the migration plan, both the cloud target and the processing provider depend on that permit.

Concentration is the other cloud question. The EU designated AWS, Google Cloud, Microsoft and Oracle entities among critical ICT third-party providers in November 2025, and the Basel resilience principles treat third-party dependency management as a core part of operational resilience. The migration data path inherits all of these constraints: a bank with EU and Indian books running Airbyte Flex would use one workspace per region, because each Flex workspace maps to one region and one data plane.

Target Architecture: APIs, Microservices and Cloud

A componentized core that posts in real time and exposes accounts through APIs is a useful target for meeting interface obligations, but those obligations do not prescribe the core’s architecture. Under the PSD2 technical standards, a bank’s dedicated interface must offer the same availability and performance as its own customer channels. The UK Open Banking Standard has reached version 4.0.1. These standards measure the interface the bank exposes, not how the core is built.

The US sets an explicit numeric bar. The CFPB’s Section 1033 rule sets a 99.5 percent monthly response rate for developer interfaces, though a federal court stayed the compliance dates on October 29, 2025 while the Bureau reconsiders the rule. US banks should plan the new core’s API layer to the published bar, but recheck the rule’s status before committing dates.

Beyond meeting interface obligations, the bank can also reuse its coexistence work after migration. Its data platform team can maintain the reconciliation store that Airbyte replicates into, with data as fresh as each sync, as the governed copy of customer and transaction data for analytics, risk models and AI applications, under the same access controls, audit logging and field-level hashing or encryption the bank already applied to migration data. For an AI application, the team must answer the same questions it addressed during migration: how fresh the data is, who can read which fields, and where the records physically sit.

Timelines, Costs and Why Core Projects Fail

The 18 to 30 month figure often quoted for core replacement appears in a 2012 Infosys implementation table reproduced by Celent, where that band applies only to banks above US$100 billion in assets; it is an old vendor figure for the largest banks, not a benchmark for phased programs.

Commonwealth Bank of Australia moved its core banking to cloud and cut restore time for critical incidents by 60%, so a cloud core can improve recovery, though this is one bank’s own result and not a benchmark for every program.

Budget for the new core, continued legacy maintenance and the integration work that keeps both consistent during dual running. Core conversions can be expensive and involve significant operational complexity, risks addressed in the Federal Reserve’s proposed guide for community banks, so the integration line deserves its own budget rather than a share of the core vendor’s estimate.

The arithmetic of a phased plan shows where the time goes. If a bank allows four months for assessment, six months for a coexistence pilot with live reconciliation, and five waves of about three months each including stabilization, the migration itself lands near 25 months, before the legacy core is decommissioned. Decommissioning comes after that, and archives, regulatory reporting and edge-case products can keep the old system running longer than planned.

Overruns are common enough to plan for. Among 17 payments modernization projects at 12 European banks, eight were financially successful, six delivered partial results and three failed, and about 20% of banks had abandoned at least one modernization project. Vendor performance, misfit between packaged products and bank requirements, and budget pressure all appear among the reasons.

Programs can run into recurring problems: scope expands as hidden legacy dependencies are discovered, and it is hard to find engineers who can read the COBOL that encodes decades of product rules. When leaders hold teams to fixed public dates, testing can be compressed first, so schedule pressure lands on the phase that protects customers.

By migrating in waves, teams can discover hidden dependencies early and limit how much legacy logic they must decode at once. Leaders can then use measured results to decide when to move a date.

Getting started with a zero-downtime core migration

Start by identifying every consumer of the legacy core and every product rule it encodes, because the migration team needs that inventory to size every later wave. Then make the approach decision on the evidence: phased or componentized unless the bank can rehearse a full big-bang cutover several times on a production-identical environment.

Before any customer moves, run a coexistence pilot. Stand up log-based CDC from the legacy databases into an independent reconciliation store, run the daily ledger, balance and transaction reconciliation for at least a full month-end cycle, and fix every break the team identifies during the pilot. Brief supervisors at the same point, while the plan can still change.

The first wave should be staff accounts with simple products, two full rehearsals, measured timing, numeric go/no-go criteria and a reverse sync that has been executed, not described. When that wave closes with zero unresolved balance breaks, the bank has proven the machinery every later wave depends on.

Get a demo to see how Airbyte keeps governed replication inside your infrastructure boundary.

Frequently Asked Questions

Duration, approach, synchronization, supervisory evidence and rollback all affect how much disruption a migration can contain. The practical decisions depend on the bank’s product mix and its ability to rehearse recovery.

How long does a core banking migration take?

Plan in waves rather than in years. Once the coexistence pilot is stable, each wave of rehearsal, cutover and stabilization takes a few months, so the total depends mostly on how many cohorts the product mix requires and how quickly early waves prove the tooling. Decommissioning the old core adds time on top of the migration itself.

Is big-bang or phased migration better?

Phased migration suits most banks because each step can be reversed. Big-bang is defensible for a small bank with few products, one legacy core and a vendor experienced with that conversion, provided the full cutover is rehearsed repeatedly on an environment identical to production.

How do you keep two cores in sync during migration?

Give each customer cohort exactly one writable core, replicate changes from the legacy core’s transaction log to the new core and a separate reconciliation copy, and compare the two every business day. Keep reverse replication ready but idle until a wave has to be undone.

What do regulators expect during a core migration?

Supervisors want to hear about the program early, before vendor contracts are signed, and they want proof as it proceeds. That proof is the same material the program uses internally: rehearsal outcomes, daily reconciliation records, vendor exit plans and third-party risk reviews.

Can a bank roll back a failed migration wave?

Yes, if the plan keeps the old core able to take the cohort back. That requires a freeze point before cutover, a rehearsed reverse replication of postings made after go-live, a deadline after which rollback stops being practical, and one executive authorized to make the call.

Why do core banking projects fail?

They fail through compressed testing, hidden dependencies discovered late, scarce legacy skills and a fixed public date that leaves no room to stop. Small, measured waves address each of those causes by limiting what any single step can break.

Integrate with 700+ apps using Airbyte

Move data from 700+ sources into warehouses, lakes, and beyond. Set up pipelines in minutes with pre-built connectors and the Connector Builder.