Banking Is Necessary. Core Banking Systems Are Not.

01The empty banking hall

For thirty years banks have bought one vendor suite, such as Temenos Transact, Oracle FLEXCUBE or Infosys Finacle, to hold balances, post entries, run product maths and enforce limits. That model is ending. Each of those jobs is being absorbed by horizontal infrastructure and by software agents, and by the mid-2030s "core banking system" will describe a function rather than a product anyone buys. The strategic question for a bank's board is no longer whether to leave the packaged core, but how fast it can do so safely.

Start with the banking hall. At Equity Bank in Kenya, 97% of transactions already happened outside branches in 2019, 77% on mobile alone.⁴ India's UPI processed 172 billion payments in 2024, up from 38 billion in 2021.⁵ Brazil's Pix went from 9.4 billion to 63.5 billion over the same years, and to 79.8 billion in 2025.⁶ Money moves on the device, initiated by the customer and settled in seconds.

The same Equity figures show what is at stake. Branches handled 3% of transactions but 49.5% of the value.⁴ The large, complex, high-liability movements are still the ones a bank must get exactly right. Leaving the core is therefore a decision about where that accuracy will live, and who will own it.

02What the core really does, and why most of it is now a commodity

Strip away the vendor language and a core banking system is an audited state machine: it changes numbers in a ledger and can prove every change. It does seven jobs (Exhibit 3). The first five are well-understood mathematics and data management. Double-entry bookkeeping was set down by Luca Pacioli in 1494, and an amortisation schedule is a formula. Event-sourced ledgers such as Thought Machine Vault, Formance and TigerBeetle, and ordinary cloud databases, already do this work at scale. Paying a suite licence for an atomic debit and credit is increasingly hard to justify.

The last two jobs carry the licence. A real posting must be idempotent when a phone retries, reversible when a rail fails, provisioned under IFRS 9 and reported to the central bank in its format. That work does not go away when the suite does. It moves into what this paper calls the accountability spine, which the bank must own whatever it buys.

03Why speed pays

The run-the-bank tax. Celent estimates that banks spend about 61% of their technology budgets running the bank and 39% changing it.²¹ Much of the run cost sits in maintaining and customising the core. Every year spent on the old suite is a year of that budget not spent on growth.

The market has already voted. Microsoft, whose dominance rests on horizontal software that every office needs, never built a core. It launched Microsoft Cloud for Financial Services in 2020 with partners such as Finastra, and Temenos's banking cloud runs on Azure.⁷ SAP moved its banking business into SAP Fioneer in 2021, with the investor Dediq taking 80%.⁸ Monzo, Nubank and Revolut built their own cores.⁹ When JPMorgan Chase replaced its US retail core in 2021, it chose Thought Machine, a cloud-native engine founded to "eradicate legacy technology from the industry".^10,11

The machine customer is arriving. In September 2025 Google published the Agent Payments Protocol, backed by more than 60 firms including Mastercard, American Express and PayPal. It lets an AI agent pay on a person's behalf using signed mandates.²² Visa launched a Trusted Agent Protocol the same month.²² When a customer's agent can compare every bank's rates in milliseconds and move a balance under mandate, deposits become far more mobile. A bank whose core updates overnight is invisible to that agent. The cost of moving slowly is no longer just inefficiency. It is lost customers.

Speed compounds. A bank on product engines and a real-time ledger launches products in weeks, prices them per customer and lets its own agents service them. Each product moved off the core lowers the cost of moving the next.

Exhibit 4 shows how these pieces come together. A customer gives a delegate a goal and a mandate: "keep three months of expenses liquid and put the rest wherever it earns most". The delegate approaches the bank through an agent gateway, a new control point that verifies who is acting, under which mandate and within which limits. Behind it, the bank's own autonomous agents handle servicing, credit decisions, collections, treasury and fraud within parameters set by staff. They call product engines for the maths and post to a horizontal ledger utility that keeps every table correct in real time.

None of this works on a batch core. Agents need real-time balances, idempotent APIs and event streams; mandates need a registry the regulator can inspect. That is why the core must go first, and fast. Beside the stack sits the accountability spine, the part no vendor can own: licence and capital, mandate records, reconciliation, regulatory returns, model risk controls and liability to the customer.

05What to adopt: a five-year agenda

The fastest safe route off the core is not one leap. It is a sequence of contained moves, each of which pays for the next (Exhibit 5). Banks should plan to adopt six capabilities:

• A cloud-native ledger. Real-time, double-entry and event-sourced; bought from a specialist or built on a hyperscaler, and run alongside the old core during migration.
• An event backbone. A streaming platform that publishes every balance change, so agents and engines react in real time instead of waiting for end of day.
• Product engines. Separate engines for deposits, lending and cards, configured by product teams rather than coded by the vendor.
• A decision and policy layer. Rules and models in one governed place, with every decision explainable to a customer and a supervisor.
• An agent platform and gateway. Identity for agents, mandate verification, spending limits, kill switches and a full audit trail, accepting emerging standards such as AP2.
• The accountability spine. Continuous reconciliation, automated regulatory reporting, a model risk office and tested recovery plans.

06Move fast, but never in one leap

Speed is not the risk; the big-bang cutover is. TSB's 2018 migration moved all its customers at once. Services were disrupted for months, a significant share of 5.2 million customers were affected, and regulators fined the bank £48.65m.¹⁴ Commonwealth Bank of Australia's core programme, budgeted at A$580m in 2008, ended near A$1.1bn.¹⁵ Both argue for product-by-product migration with parallel running, not for delay.

Nor will the incumbents disappear quietly. Analysts put the core software market at $13–17bn in 2024 and still growing,¹⁷ an estimated $3 trillion a day still flows through COBOL systems,¹⁶ Bank of America is still opening branches,¹⁹ and Starling has turned its own core into a product, Engine, sold to other banks.¹³ The core will be dismantled by banks that choose to, not by the passage of time.

Three further risks need owners at board level. Agents will make mistakes, so liability, limits and kill switches must be designed before scale. Regulation is tightening: the EU AI Act classifies AI used for consumer credit scoring as high-risk, and DORA brings critical technology suppliers under direct oversight.^18,23 And moving to one hyperscaler swaps vendor lock-in for concentration risk, so the ledger should be portable.