The first measurable gains are showing up in coding, customer service, transaction screening and knowledge work. The more difficult leap is turning time saved into durable operating leverage without allowing AI costs, controls and reinvestment to absorb the benefit.
Bank executives increasingly describe artificial intelligence as a productivity technology rather than an experiment. That claim is finally becoming testable. Large banks are disclosing minutes removed from service calls, hours recovered from coding and policy searches, fewer manual checks in transaction screening, and broader measures of economic value from AI programmes. The evidence is no longer limited to pilot counts or vendor demonstrations.
Yet the savings are not appearing in the way a conventional cost-cutting programme would suggest. At many banks, AI is creating capacity before it creates visible expense reduction. A developer writes more code. A customer-service representative handles a case faster. A compliance team reviews more transactions with fewer manual touches. A relationship manager spends less time preparing for meetings. The bank can then choose what to do with that capacity: reduce staffing, avoid future hiring, absorb higher volumes, improve service, strengthen controls, or reinvest the time in revenue-producing work.
That distinction matters because a productivity gain and a cost saving are not the same thing. The Bank for International Settlements said in July 2026 that the productivity payoff from AI could be large but remains uncertain and uneven across sectors and countries. The Bank of England similarly reported that firms are achieving real gains in software development, finance, administration and customer service, while also seeing part of the benefit offset by spending on software, cloud services and AI licences.
For banking, the emerging picture is therefore more interesting than a simple story about replacing employees. AI is beginning to alter the unit economics of specific workflows. The savings are appearing first in time, throughput and manual effort. Whether they become lower expense ratios, smaller workforces or higher returns depends on what management does next.
The measurable gains are concentrated in a few workflows
The strongest evidence so far comes from tasks that are high-volume, repetitive enough to benchmark and expensive enough that a small improvement matters at scale. Public disclosures from major banks point to the same clusters: software engineering, service operations, knowledge retrieval, transaction screening, client preparation and document-heavy controls.
| Workflow | Disclosed evidence | What it actually measures |
| Software engineering | BofA: ~19,000 developers using real-time coding assistance; productivity reported at >20%. | Output capacity and speed, not automatically payroll reduction. |
| Customer service | BofA: EricaAssist used by >18,000 service staff; average call time reduced by nearly one minute. | Shorter handling time and higher service capacity. |
| Transaction screening | JPMorganChase: more than double the screening volume with half the manual operator checks. | Fewer manual touches per unit of control work. |
| Knowledge work | UBS: in-house assistant Red reported to save users an average 80 minutes per week. | Time recovered from search, drafting and information retrieval. |
| Enterprise AI value | DBS: ~SGD 1 billion of 2025 economic value from data analytics and AI/ML across >430 use cases. | A broad value measure that can include revenue, risk and efficiency—not pure cost takeout. |
Coding is where AI productivity is easiest to see
Software engineering is becoming banking’s clearest productivity laboratory because output can be measured against relatively familiar baselines: time to generate code, complete tests, document changes, resolve defects or deliver releases. Bank of America’s second-quarter 2026 investor presentation says roughly 19,000 developers are using real-time coding assistance and that productivity has increased by more than 20%. The same presentation reports about 200,000 active users of general-purpose AI tools across the company, generating more than 400,000 prompts a day.
Citi offers another unusually concrete measure. In its 2025 annual report letter to shareholders, the bank says AI-assisted coding tools are creating approximately 100,000 hours of developer capacity each week. Citi frames that capacity as time that can be redirected toward higher-value innovation rather than as an announced reduction in developer headcount.
DBS reports a similar pattern at the process level. In its 2025 CEO reflections, the bank says technology teams used generative AI to reduce time spent on tasks such as test-case generation and user-story documentation from months to weeks. Across the bank, DBS says more than 2,000 models and 430-plus use cases generated approximately SGD 1 billion in economic value during 2025.
The coding use case is attractive because the work is digital, the feedback loop is fast and the developer remains available to review the output. But even here, gross productivity is not the same as net saving. Faster code generation can increase testing demand, security review, documentation requirements and the volume of software the bank chooses to build. A 20% improvement in a task does not necessarily mean a 20% reduction in engineering cost; it may instead mean the same team delivers more.
In customer service, seconds and minutes become capacity
Customer service is the second place where AI gains are becoming visible because banks already measure call duration, resolution rates, transfers, repeat contacts and queue times. Bank of America said in July 2026 that its EricaAssist tool is used by more than 18,000 service employees, provides contextual guidance in under three seconds and reduces average call time by nearly one minute per interaction.
That is a meaningful operational gain, but its economic value depends on volume and management choices. A one-minute reduction can lower the number of service hours required for a fixed call volume. It can also be used to absorb higher demand without adding staff, shorten queues, allow employees to spend more time on complex cases, or improve first-contact resolution. The productivity benefit is real even when the payroll line does not immediately fall.
HSBC is pursuing the same logic at a broader scale. In its June 2026 AI partnership announcement with Google Cloud, the bank said an AI-powered decision assistant was already cutting administration and client-meeting preparation from hours to minutes for thousands of users. HSBC expects the partnership to enable more than 200 additional use cases over two years and estimates benefit value above US$100 million through a mix of revenue gains and efficiency improvements. That figure is an expectation, not a realised cost saving, which is precisely why it should not be treated as one.
Controls may deliver the most defensible savings because they remove manual touches
Some of the most bank-specific productivity opportunities sit in controls rather than customer-facing work. Screening, onboarding, trade confirmation, credit review and KYC are costly because they combine high transaction volumes with regulatory obligations and exception handling. The economic prize is not eliminating control; it is reducing how much human labour is required to achieve an equal or stronger control outcome.
JPMorganChase provides a useful example. In the Commercial & Investment Bank section of its 2025 annual report, the bank says AI in transaction screening has enabled it to review more than double the volume while halving the number of manual operator checks. More than 90% of its engineers use AI code assistants, and more than 65,000 CIB employees actively use the bank’s LLM Suite.
Citi’s annual report disclosure points to the same category of work: automation of trade confirmations, AI-supported KYC and onboarding, and faster wholesale-lending underwriting and credit review. These are not glamorous use cases, but they are precisely where banks have accumulated large operations and control teams over decades of regulatory and process complexity.
The potential saving is therefore partly an avoided-cost story. If transaction volumes rise while manual reviews do not, unit cost falls even if absolute staffing remains flat. If AI reduces false positives, the bank can redirect investigators toward genuinely complex cases. If onboarding cycles shrink, revenue may arrive sooner. These benefits are economically real but can appear across several lines—cost, revenue, risk and client retention—rather than as one visible “AI saving” in the accounts.
Knowledge work is creating an invisible layer of time savings
A large share of white-collar banking work consists of finding information, reconciling internal rules, drafting routine documents and preparing for conversations. Those tasks rarely appear as a separate cost category, which makes AI savings easy to miss even when they are widespread.
UBS’s current AI and innovation disclosures illustrate the scale. Its in-house assistant Red has been rolled out to around 100,000 employees and has handled more than 25 million queries. Based on user feedback, UBS says Red saves employees an average of 80 minutes a week. Its MyPolicies assistant, which provides role- and location-specific answers from more than 1,100 policies and guidelines, has generated approximately 48,000 hours of reported time savings since launch.
Bank of America’s 2Q26 AI operating model shows similar diffusion beyond developers: more than 20,000 wealth professionals use AI for meeting-related activities, around 4,000 corporate and investment banking employees use it for research and presentation preparation, and thousands of payments and call-centre employees use role-specific tools. The bank’s disclosure is important because it separates general-purpose AI from function-specific tools and multi-process systems. The deeper the integration into a workflow, the more likely a time saving can be converted into a measurable operating result.
EVIDENCEMajor banks are now reporting measurable reductions in time, manual checks and cycle times, alongside broad economic-value estimates.
INFERENCEThe most likely near-term financial effect is not wholesale headcount elimination but slower hiring, higher volumes per employee, redeployment and better operating leverage. Public disclosures support that direction, but they do not yet prove a uniform industry-wide reduction in absolute costs.
Why bank expenses have not collapsed
If AI is creating so much capacity, why are bank cost bases not falling dramatically? The first answer is that the technology itself is expensive. Bank of America says it spends about US$14 billion a year on technology, with more than US$4 billion allocated to new initiatives including AI. JPMorganChase’s 2025 annual report puts its 2026 technology budget at approximately US$19.8 billion. Model access, cloud infrastructure, data engineering, cyber security, evaluation, governance and training all sit between a successful demo and a safe production system.
The second answer is reinvestment. JPMorganChase explicitly says it expects productivity gains to free capacity that can be reinvested in growth and that the objective is not simply fewer employees. In Consumer & Community Banking, the 2025 annual report says the value generated by AI and machine learning increased nearly 60% year on year and sets an operations objective of more than 40% gross productivity efficiency by 2030. “Gross” is the key word: management can capture some of that benefit as expense reduction and spend some on growth, control enhancement or service.
The third answer is that bank financial statements are driven by many forces at once. Bank of America’s 2Q26 results show a 59% efficiency ratio, improved from 63% a year earlier, while noninterest expense still rose 8% year on year and revenue rose 15%. That is a useful reminder of what operating leverage looks like: a bank can become more efficient without cutting absolute expense. It would be wrong to attribute the improvement to AI alone, but it is equally wrong to search only for falling costs as proof that AI is working.
The labour saving is likely to appear first as avoided hiring and a different workforce mix
For banks, the most consequential productivity question is labour. Personnel remains one of the industry’s largest cost pools, but regulated institutions cannot simply remove people from processes that still require judgement, escalation and accountability. The near-term mechanism is therefore more likely to be a lower labour requirement per unit of activity than a sudden collapse in headcount.
The Bank of England’s July 2026 business intelligence supports that interpretation across the broader economy. It found that firms deploying AI effectively can increase output without a corresponding increase in employment and in some cases reduce staffing requirements. The same report says demand is weakening for some entry-level and administrative tasks while shifting toward AI, data, judgement and oversight skills.
That creates a banking workforce effect that may be visible only gradually. A bank can let attrition reduce a team rather than announce layoffs. It can hold headcount flat while transaction volumes rise. It can redeploy analysts from document preparation to exception review, move service employees toward more complex interactions, or use AI to make senior specialists available to more clients. From an investor’s perspective, these are productivity gains even when the headline employee number changes little.
The counterargument: AI can create new work as quickly as it removes old work
The productivity case should not be romanticised. AI can reduce drafting time and simultaneously create new validation work. It can make customer service faster while generating more digital interactions that need monitoring. It can lower the cost of producing code while increasing the amount of code that security teams must review. It can automate financial-crime screening while raising new model-risk, explainability and third-party dependency questions.
The Bank of England notes that realised productivity gains vary materially with implementation quality, employee capabilities and the amount of human oversight required. It also reports that low-cost AI generation is creating “AI slop”, inauthentic applications and higher volumes of requests and complaints, while spending on software, cloud services and AI licences offsets part of the savings. In other words, the gross benefit can be large while the net benefit is smaller.
Regulators are also preparing for a more operationally AI-intensive financial system. The FCA’s July 2026 Mills Review identifies the transformation of firm operations as one of four major AI-driven shifts in retail financial services, alongside changes in consumer journeys, competition and fraud/cyber risk. The IMF has similarly warned that AI is becoming embedded in the operational core of finance and that shared dependencies on cloud, data and model providers can turn efficiency infrastructure into a source of systemic vulnerability.
What boards and investors should measure instead of “AI spend”
The quality of an AI programme is increasingly measurable, but the relevant metrics sit below the headline technology budget. The most useful evidence is unit economics: handling time per service request, manual reviews per thousand transactions, developer cycle time, onboarding days, false-positive rates, adviser preparation time, incidents per deployment and cost per completed workflow. These measures reveal whether AI changes the production function of the bank rather than simply adding another software layer.
The next level is conversion. Management should be able to explain what happens to the hours saved. Are they used to serve more clients, reduce contractors, slow hiring, shrink backlogs, improve controls or launch more products? Without that bridge, an impressive time-saving metric can coexist indefinitely with an unchanged cost structure.
The final level is net economics. Banks should compare realised benefits with inference, cloud and licence costs; data remediation; model evaluation; cyber security; governance; training; and the cost of running legacy and AI-enabled processes in parallel during transition. That discipline is particularly important because bank disclosures often combine revenue uplift, loss avoidance and efficiency into one “value” figure. DBS’s approximately SGD 1 billion of 2025 economic value is a credible indicator of scale, but the bank itself describes it as value from data analytics and AI/ML broadly—not as a billion dollars of expense removed.
Implications for banks, fintechs, regulators and investors
For banks, the strategic challenge is to move from individual copilots to end-to-end workflow redesign. General-purpose assistants create useful time savings, but the larger economics appear when AI is connected to trusted data, permissions, process rules and downstream systems. That is why the strongest disclosures increasingly concern multi-step processes such as screening, onboarding, service and development rather than generic chat.
For fintechs and technology providers, the value proposition is shifting from model quality to measurable integration. A bank will care less about whether a tool can draft a good answer in isolation than whether it can reduce a specific cycle time, preserve auditability, work with the bank’s data and controls, and demonstrate net benefit after infrastructure costs. Vendors that can prove operational outcomes may gain an advantage over those selling broad productivity claims.
For regulators, rising productivity can strengthen resilience if it removes error-prone manual work and gives skilled employees better information. It can also weaken resilience if control teams become too thin, staff stop understanding the processes they supervise, or multiple banks depend on the same models and cloud providers. Supervisory attention will therefore have to follow the operating model, not just the algorithm.
For investors, the most credible AI story is unlikely to be the bank with the largest number of use cases. It is the bank that can show sustained operating leverage while maintaining service quality and control performance. Over the next several reporting cycles, the signal to watch is not simply falling headcount. It is whether revenue and transaction volumes can grow faster than the cost base without a corresponding deterioration in risk.
Conclusion: the savings are real, but they are arriving as capacity before cash
Banking’s AI productivity boom is becoming visible in practical places. Developers are writing and testing faster. Service representatives are shaving time from interactions. Screening teams are processing more with fewer manual checks. Employees are spending less time searching for policies and preparing routine material. These are real operational changes, supported by increasingly specific bank disclosures.
What has not yet emerged is a simple industry-wide line showing AI cutting bank costs by a fixed percentage. The economics are more complicated. Banks are still investing heavily in technology, data and controls; many are reinvesting capacity in growth; and some of the gross savings are offset by new software, cloud, governance and verification costs.
That does not make the productivity boom illusory. It changes where to look for it. The first phase of AI value is measured in minutes, hours, manual touches and cycle times. The second phase will be measured in avoided hiring, better operating leverage and more scalable service. Only after those effects compound will the full saving become obvious in the income statement. The banks that understand that conversion problem—and manage it deliberately—are the ones most likely to turn AI enthusiasm into durable economics.
References
1. Bank of America — 2Q26 Financial Results / AI workforce enablement presentation (14 July 2026)
2. Bank of America — EricaAssist generative AI employee-service update (21 July 2026)
3. Bank of America — 2025 shareholder letter on technology and AI investment (March 2026)
4. Citigroup — 2025 Annual Report, Letter to Shareholders (2026 publication)
5. UBS — Innovation and AI at UBS: enterprise productivity and advisor tools
6. JPMorganChase — 2025 Annual Report, Commercial & Investment Bank letter
7. JPMorganChase — 2025 Annual Report, Consumer & Community Banking letter
8. JPMorganChase — 2025 Annual Report, Chief Operating Officer letter
9. DBS — 2025 Annual Report, CEO reflections on AI and operating model transformation
10. HSBC — HSBC and Google Cloud announce transformative AI banking partnership (17 June 2026)
12. Bank of England — Agents’ summary of business conditions, Box A on AI productivity (July 2026)