Banks are spending heavily on artificial intelligence. The measurable returns are beginning to emerge - but they are concentrated in specific workflows rather than evenly across the institution.
For the past three years, banking executives have spoken about generative AI as if a productivity revolution were inevitable. The investment has certainly arrived: banks are buying models, cloud capacity, data infrastructure, coding assistants and enterprise copilots at scale. What has been harder to see is the corresponding movement in the cost base.
That is beginning to change. The first credible disclosures suggest that AI is generating real economic value, but not in the simplistic form of an immediate, bank-wide reduction in expenses. The savings are appearing first in narrow, high-volume workflows where work can be measured: software development, transaction screening, know-your-customer checks, employee search, call summarisation and customer-service handling.
The pattern matters. It implies that the near-term AI productivity story is less about replacing entire departments and more about removing minutes, reviews, hand-offs and queues from millions of repeated tasks. When those small gains are multiplied across a large bank, they can become material. But the path from task-level efficiency to a structurally lower cost-income ratio is neither automatic nor immediate.
The adoption boom is real; the productivity boom is uneven
The most comprehensive regulatory evidence still shows a gap between adoption and realised economics. In the UK, the Bank of England and Financial Conduct Authority found that 75% of surveyed financial firms were already using AI, with another 10% planning adoption within three years. Firms expected some of the largest future benefits to emerge in operational efficiency, productivity and the cost base, even though current benefits were more evident in areas such as data analytics, fraud prevention and cybersecurity.
That caution remains relevant in 2026. The Bank of England's July 2026 business intelligence reported that AI was increasing productivity in finance, administration, customer service and software development, but that realised gains varied significantly by firm and depended on implementation quality, staff capability and human oversight. It also noted that savings were being partly offset by higher spending on software, cloud services and AI licences.
That is the central tension in banking. AI can make individual tasks cheaper while simultaneously raising the institution's technology bill. The productivity question therefore cannot be answered by counting licences or users. It has to be answered by locating where fewer hours, fewer manual checks, fewer errors or lower incremental staffing needs are actually being recorded.
Software engineering is producing some of the clearest measurable gains
Coding is one of the most mature enterprise use cases because output is digital, repetitive and measurable. JPMorganChase reported in its 2024 annual report that software engineers were seeing productivity gains of roughly 10% to 20% in the development lifecycle from AI coding assistants. By 2025, the bank said more than 90% of its engineers were using AI code assistants.
At DBS, generative AI was used to compress work such as test-case generation and documentation of user stories from months to weeks. Its CodeBuddy tool produced time savings of up to 20% on certain coding tasks. DBS also reported that code deployment time had been cut by 25% and model deployment cycles reduced to seven to ten weeks as part of its wider technology transformation.
These gains matter for more than IT budgets. Banks operate enormous software estates. Faster coding, testing, documentation and deployment can reduce project backlogs and shorten time to market, while allowing a given engineering workforce to maintain more systems or deliver more change. The immediate benefit may therefore appear as additional capacity rather than a visible decline in headcount.
The back office is where AI can remove the most repetitive work
Transaction screening provides one of the sharpest examples. In its 2025 annual report, JPMorganChase's Commercial & Investment Bank said AI enabled it to review more than twice the transaction volume while cutting the number of manual operator checks by half. That is a direct productivity gain: more activity processed with less human intervention, combined with faster turnaround for clients.
DBS reported an even more explicit workflow saving in customer due diligence. Its 2025 Consumer Banking and Wealth Management report said generative AI improved name-screening efficiency in know-your-customer processes by about 70%, freeing employees to focus on higher-value customer interaction and risk work.
These are precisely the processes where banks have historically carried large operational teams. Screening, reconciliation, documentation and exception handling involve high volumes, clear rules and expensive human review. AI does not need to replace the control; it only needs to reduce the proportion of cases that require a person. Even modest reductions in manual review rates can translate into significant capacity when applied to millions of transactions or customer records.
Customer service savings are emerging as avoided workload
Another area where the economics are becoming visible is customer service. Bank of America has expanded AI across both client-facing and employee tools. Erica, its consumer virtual assistant, has handled billions of interactions, while Erica for Employees supports internal technology and service queries. The key productivity effect is not simply that the bot answers questions; it is that routine requests can be resolved without entering a human queue.
The same logic is visible at NatWest, which said generative AI, simplification and cloud migration contributed to £100 million of investment capacity in 2025. Its 2025 results materials also pointed to AI call summarisation saving tens of thousands of staff hours. The important word is capacity: banks can redirect the time released by automation into customer-facing work or growth initiatives instead of immediately converting it into redundancies.
Lloyds Banking Group has gone further in attaching a monetary value to the new generation of tools. Lloyds said generative AI delivered around £50 million of value in 2025 and that it expected more than £100 million of additional value from next-generation AI in 2026. Those numbers remain small relative to the operating expenses of a major bank, but they are meaningful because they represent a shift from pilot activity to quantified business value.
The most important saving may be headcount that never needs to be added
AI productivity does not always appear as a line-item cost reduction. In a growing bank, the benefit may instead be visible in slower hiring, natural attrition, redeployment or the ability to absorb more business without increasing staff at the same rate.
DBS made this explicit in its 2025 financial reporting. Its CFO statement said automation and generative AI supported productivity and contained resourcing needs, even as the bank continued to invest and grow. The Bank of England has also reported that firms across the wider economy are increasingly using automation and AI to reduce the need for future headcount growth.
This is economically important because banking costs are heavily labour-based. If AI allows revenue, transaction volumes or compliance requirements to grow faster than staffing, the cost benefit compounds over time. The effect may initially be difficult to distinguish from ordinary cost discipline, but over several years it can reshape the operating leverage of a bank.
Economic value is broader than cost savings
The most detailed bank-level disclosure comes from DBS's 2025 annual report, which said more than 2,000 AI models across more than 430 use cases generated approximately SGD 1 billion in economic value during the year. The figure is deliberately broader than cost savings: it includes value generated through analytics, AI and machine learning across revenue, risk, customer engagement and productivity.
That distinction matters. An AI system that improves fraud detection can reduce losses. A personalisation engine can increase sales. A pricing model can improve margins. A coding assistant can save employee time. All four create economic value, but only one may show up as a traditional operating-cost saving.
JPMorganChase similarly said in its 2025 consumer banking shareholder letter that the value generated from AI and machine learning increased by nearly 60% year on year, while its operations organisation is targeting more than 40% gross productivity efficiency by 2030. The wording is significant: the bank is measuring value and gross efficiency before assuming those gains will translate one-for-one into lower reported expenses.
Why the savings are not yet transforming every cost-income ratio
There are at least four reasons the productivity boom can be real without yet appearing as a dramatic fall in total costs.
• Banks are reinvesting the gains. Capacity released by automation is often redirected into new products, controls, data remediation and growth rather than removed from the organisation.
• AI itself is expensive. Model access, cloud consumption, specialised chips, cyber controls, data engineering and governance create new costs. The Bank of England has observed that software, cloud and AI licence spending can offset part of the savings.
• Legacy technology limits automation. AI performs best when data is accessible and workflows are digitised. Fragmented systems force banks to spend heavily on integration before the model can change the process.
• Regulated decisions still require controls. Human review, explainability, model validation and audit trails can reduce the percentage of theoretical time savings that becomes cash savings.
The BIS Financial Stability Institute has highlighted the data-quality, privacy, security and third-party dependency challenges that accompany deeper AI adoption in finance. These are not side issues: every additional control can reduce the speed of deployment, while poor data can eliminate the productivity advantage entirely.
The next phase is workflow redesign, not more copilots
The first phase of generative AI in banking largely put assistants on top of existing work. Employees wrote emails faster, summarised documents, searched policies and generated code. That creates useful but bounded gains because the underlying process remains unchanged.
The larger opportunity comes when banks redesign the process itself. The European Central Bank has noted that intensive AI users tend to generate more value when AI is applied to core production processes rather than confined to peripheral tasks. In banking, that means connecting models to workflow systems, decision engines and controlled actions - not merely giving staff a chatbot.
That transition is already visible in agentic AI. A service agent can collect information, check policy, update a case and escalate only the exception. A software agent can generate code, test it, document it and route it for approval. A financial-crime agent can gather evidence across multiple systems before a human investigator sees the file. Each step removes hand-offs, and hand-offs are where much of banking's operational cost resides.
The productivity metric banks should watch
The industry may ultimately need better measures than "number of AI use cases" or "employees with access to a copilot." Those figures describe adoption, not economics.
The more revealing indicators are operational: cases handled per employee, manual-review rates, software release frequency, average handling time, false-positive rates, turnaround time, error rates, cost per transaction and the amount of business that can be absorbed without additional headcount. These measures make it possible to separate a genuinely transformed workflow from an expensive demonstration project.
The latest central-bank evidence supports that more discriminating view. The BIS has cautioned that the productivity payoff from AI could be large but remains uncertain and uneven across sectors and countries. Banking is likely to be similar: a few institutions with mature data, modern architecture and redesigned processes may capture disproportionate benefits while others spend heavily simply to keep pace.
Conclusion: the savings are real, but they are hiding inside workflows
The banking AI productivity boom is becoming measurable, but it does not yet look like a sudden collapse in industry employment or expenses. It looks like a 20% coding-time saving here, a 50% reduction in manual screening there, a 70% improvement in one KYC process, tens of thousands of call-centre hours removed, and millions of routine requests handled without a human queue.
Those numbers can sound modest beside the enormous projections attached to artificial intelligence. Yet banking is a scale business. When small improvements are applied to millions of transactions, thousands of developers and vast control functions, they can become strategically significant.
The bigger question is whether banks can turn released capacity into structural operating leverage. That requires more than giving employees AI tools. It requires redesigning workflows, cleaning data, modernising infrastructure, changing control models and deciding which savings should be reinvested and which should fall to the bottom line.
The first wave of AI made bankers faster. The next wave will determine whether the bank itself becomes cheaper to run.
References
1. Bank of England & FCA - Artificial intelligence in UK financial services (2024)
2. Bank of England - Agents' summary of business conditions, July 2026
4. DBS - CEO Reflections, Annual Report 2025
5. DBS - CIO Statement, Annual Report 2025
6. DBS - Consumer Banking / Wealth Management, Annual Report 2025
7. DBS - CFO Statement, Annual Report 2025
8. JPMorganChase - Annual Report 2024, Jennifer Piepszak letter
9. JPMorganChase - Annual Report 2025, Commercial & Investment Bank letter
10. JPMorganChase - Annual Report 2025, Consumer & Community Banking letter
11. Bank of America - AI Adoption by Global Workforce Improves Productivity and Client Service (2025)
12. NatWest Group - Banking through technology, data and AI (2026)
13. Lloyds Banking Group - Group expects over £100m in value from next-gen AI in 2026
15. ECB - What separates firms that use AI intensively from firms that don't? (2026)
16. BIS - AI and the global economy: implications for central banks (2026)