Banks enter the 2027 budget round in unusually good shape. Earnings are strong, capital is plentiful and, for the first time in years, most bank equity outside China trades above book value4. But the income statement hides an awkward fact. Over a decade of digital investment has barely moved the cost of running a bank relative to the size of its balance sheet. The improvement in returns has come mostly from income and from holding costs flat, not from changing how the work gets done.
A productivity puzzle in the numbers
BCG's 2026 analysis of 1,498 financial institutions describes profit growth driven by ‘positive jaws’, with income rising faster than costs, while the opex-to-assets ratio has not improved materially34. Financial industry headcount has grown by about 2% a year over the past three years4. McKinsey's 2026 review reports a modest improvement, with costs falling from 1.31% of assets in 2024 to 1.23% in 20257. Its 2025 review noted that banks spend more on technology than the next four sectors combined, yet productivity has lagged6.
The US picture is similar. In the second quarter of 2026 the efficiency ratio for all FDIC-insured institutions, meaning non-interest expense as a share of revenue, was 55.4%. For community banks it was 60.5%2. Scale helps, but it does not solve the problem.
Efficiency still varies widely by business model
Efficiency ratio (non-interest expense as % of revenue), FDIC-insured institutions, Q2 2026 (%)
Note: Lower is more efficient. Asset concentration groups as defined by the FDIC.
Source: Federal Deposit Insurance Corporation, “Quarterly Banking Profile: Second Quarter 2026” (2026)
Why the cost plans look optimistic
EU/EEA banks' funding plans assume real cost relief. Expenses stood at 22.5% of equity in 2025 and banks project 22.4% in 2027 and 21.8% in 2028, driven by lower staff and other administrative expenses1. The EBA is sceptical. ICT costs are the largest part of other administrative expenses, at 32.1% in 2025, up from 31.2%. 72% of banks name core banking and payment systems as a major source of ICT spending, and new technologies such as AI are likely to require more investment, not less1.
Banks plan for cost relief that ICT spending may not allow
EU/EEA banks' total expenses as % of equity, actual and funding-plan projections (%)
Note: EBA funding plan data. Staff expenses were 9.9% of equity and other administrative expenses 7.4% in 2025.
Source: European Banking Authority, “Risk Assessment Report – June 2026” (2026)
This is the core of the productivity puzzle. Technology spending has been used mainly to digitise existing processes and run legacy estates, not to remove work. Every new channel added cost, while the operations behind it remained largely unchanged.
Part of the problem is how technology budgets are managed. Most banks separate spending on running existing systems from spending on change, but few link change spending to costs actually retired. New digital channels have often been added alongside old ones rather than replacing them, so customer-facing modernisation has sat on top of largely unchanged back offices. The result is higher technology cost without a matching fall in operations cost, which is exactly what the EBA's figures on rising ICT spend show1.
Where AI actually moves cost
The evidence on AI points to a different kind of lever. McKinsey estimates that AI could reduce some cost categories by as much as 70% before offsetting technology costs, and the industry's aggregate cost base by 15–20% net6. It expects some of that benefit to be competed away to customers over time6. BCG reports agentic AI delivering more than 50% productivity gains in retail lending and more than 30% uplift in wealth management fee income3. Institutions plan to invest 2% of revenue in AI this year4.
In our experience, the cost moves where three conditions hold together:
- High-volume, rules-based work. Reconciliations, onboarding and KYC refresh, payments exceptions, credit memo assembly, regulatory returns and servicing requests.
- End-to-end redesign. Agents take on whole workflows under human supervision, rather than copilots speeding up individual steps. McKinsey describes one person supervising 20 to 30 agents running end-to-end workflows5.
- Decommissioning. Savings count only when a queue, a team, a vendor contract or an application is actually retired.
Workforce redesign, not just reduction
Some banks have started to put numbers on the workforce effect. A large US bank told investors in May 2025 that AI could reduce headcount in its operations and account services teams by 10%, and its consumer banking chief said she would ‘take the over’8. A large Asian bank said in February 2025 that it expected about 4,000 temporary and contract roles to fall away over three years as AI took on more work, while it added around 1,000 AI-related roles9.
The lesson from both is that the workforce changes shape as well as size. Fewer people do repetitive processing. More people supervise agents, handle exceptions and complex cases, own data and controls, and serve customers. Banks that plan this as a redesign, with reskilling, new roles and attrition-led transition, avoid the reversals seen when cuts outrun the evidence.
Measurement is what separates productivity from cost-cutting. Banks that track cost and cycle time per unit of work, such as cost per account opened, per payment exception cleared or per regulatory return filed, can see whether AI is changing the economics of a process or just shifting effort elsewhere. Those that track only headcount and total cost find out too late.
The window is favourable. Strong earnings give banks the capacity to invest and the credibility to ask investors for patience. The risk is using good years to fund more of the same, adding tools to unchanged processes. The banks that emerge with structurally lower cost-to-income ratios will be those that use this period to rebuild how work is done, one end-to-end process at a time.