AI in Banking: Deployments, Costs and Measured Results | October 2026

Welcome to AI Press Review. I am David Perron. This edition examines how banks are putting artificial intelligence to work, using disclosures published between late April and 18 October 2026. The most useful evidence concerns software engineering, mortgage preparation, financial crime investigations and support for advisors. We will look at the models and operating arrangements banks actually identify, the results they report, and the distinction between adoption, targeted savings and financial benefits. The latest earnings disclosures are especially revealing: a growing deployment count can coexist with benefits that remain difficult to identify in quarterly earnings.

Citigroup provides a concrete account of how AI can change the economics of technology delivery. On its 13 October earnings call, the bank said 14,000 engineers were using Devin, its principal agentic coding tool. Management linked the decision to bring more work in-house with a 30% year-on-year reduction in technology vendor spending. That identifies both a tool and an expenditure category. The mechanism matters: engineers use an agent to undertake more development internally, and some work previously purchased from vendors moves back into the bank. The reported reduction therefore concerns technology suppliers. Applying that percentage to the whole cost base would produce a different claim. For a technology leader, the useful comparison is the cost of delivering the same work internally, including the agent, review and infrastructure, against the supplier expenditure displaced.

JPMorgan Chase described a different part of the equation in its 13 October earnings discussion. Chief financial officer Jeremy Barnum said the bank combines proprietary models with open-weight and open-source models, including models run in its own data centre. The strategy is designed to preserve flexibility and manage costs. Open weights make model parameters available under the relevant licence, while a proprietary service normally gives access through the provider's interface. Those arrangements create different hosting and operating responsibilities. Barnum also said the benefits were not observable in current quarterly earnings, and that token costs would offset some gains. The bank is therefore discussing deployment and economics together. Its disclosure supports a strategy of using several kinds of model. It supplies neither a complete model inventory nor a measured net return for every deployment.

Bank of America's July results make the deployment stages unusually explicit. Its presentation distinguishes more than 300 approved AI and machine-learning use cases, 114 live generative-AI use cases and 34 fully deployed cases. These populations answer different questions. Approval says that a case has passed an internal decision point. A live capability can already serve users while rollout continues. Full deployment describes a later stage. Machine learning also covers techniques beyond generative AI, so the approved inventory has a wider technology scope. The practical lesson is to keep the denominator attached to each number. A programme report that combines approved initiatives, live pilots and completed rollouts into a single total can obscure delivery progress. This disclosure allows those stages to be examined separately, while leaving project-level financial returns to a different assessment.

CIBC's 31 July announcement illustrates the same distinction at platform level. The bank describes its new agentic workspace, CIBC AI 2.0, as being tested. Employees can delegate research, preparation and document-comparison tasks to agents through a framework built internally. The release also says that the preceding CIBC AI assistant had been rolled out to more than 50,000 employees, with an average of 20,000 employees using it daily. Those figures establish substantial use of the earlier assistant. They do not measure adoption of the new workspace. For implementation teams, the distinction concerns the step from answering a question to coordinating work across data and tools. A widely used conversational assistant can provide a foundation for that development, while the permissions, review arrangements and operating results of the agentic version still require their own evidence.

Lloyds Banking Group's June recruitment announcement puts the staffing question in operational terms. It said more than 700 colleagues were already shaping AI use cases and more than 1,000 roles were planned during 2026 to support the build-out. Almost 300 agentic-AI-related positions were to be filled through internal and external appointments, including engineering, responsible-AI and product roles. This is evidence of a planned change in capabilities and the composition of work. It cannot be read as a count of completed external hires. Internal appointments also mean that gross recruitment activity and net employment growth can differ. The concrete implication for a bank's workforce plan is to track the movement of people into these functions alongside roles removed or changed elsewhere. The disclosed programme identifies demand for specialist skills, rather than establishing the net employment effect of AI across the group.

HSBC announced a Global AI Centre of Excellence in Singapore on 27 July. The centre was expected to employ more than 100 AI specialists and initially focus on wealth conversations, agentic treasury solutions and AI-enabled digital payments. A separate programme concerned 100 wealth relationship managers. Keeping those programmes separate matters when assessing the AI recruitment total. The announcement also describes Singapore-built capabilities intended to scale across HSBC's global network, with human judgement and accountability retained. The operating choice is to concentrate specialist development in a centre that works with business teams. That can support shared capabilities across markets, while the deployment of each service remains a separate task. At the date of the announcement, these were planned positions and intended applications. The document does not establish that every role had been filled or that every planned solution was already operating.

TD Bank Group describes a much more specific production workflow. Its July account says all new personal mortgages pass through an agentic preparation process. The system scans client documents, extracts information, calculates income, checks selected policy requirements and consent, searches for discrepancies, and prepares a memo for an underwriter. TD reports that work previously averaging 15 hours per client could take about 3 minutes. The task boundary is crucial: the comparison concerns preparation of application information for adjudication. Credit approval, funding and the wider customer journey include additional activities. The human underwriter still receives the findings. For an operations executive, the significant feature is the combination of document handling with a defined output that another person reviews. This is a more concrete deployment description than an assistant's general ability to answer mortgage questions, and it offers a specific workflow against which performance can be evaluated.

Customers Bancorp's second-quarter release provides a separate lending example. The bank said selected commercial loans in an AI-supported closing pilot had closed in 7 days, compared with a typical 30 to 60 days. Management described the pilot as completed ahead of its previous schedule. A commercial-loan closing involves coordination and documentation after several other lending decisions, so its elapsed time answers a different question from the time required to prepare an underwriting memo. The word selected also limits the population. These were pilot transactions, with broader deployment described as a subsequent objective. For a lending executive, the result gives a reason to examine which coordination steps were removed or accelerated and whether the pilot population resembles ordinary production. The release supports the bank's account of a faster pilot. It does not establish that every commercial loan can now close within the same period.

NatWest's May announcement identifies a use case close to the daily work of trade-finance operations. The bank said it would implement Cleareye.ai's ClearTrade platform to extract and classify information from complex trade documents in digital and paper form. The announced scope also includes examination against International Chamber of Commerce rules and checks for trade-based money laundering. These activities connect document interpretation with the controls that determine whether a transaction can proceed. The operational implication is that extraction accuracy alone is an incomplete measure: the bank also needs reliable identification of discrepancies and control findings. NatWest describes an implementation programme and its intended scope. The announcement supplies no measured processing-time reduction. For trade teams assessing a comparable project, it identifies the workflow to examine and the provider selected, while leaving realised performance to subsequent evidence.

BMO's second-quarter presentation reports that frontline chatbots, including Lumi, help staff navigate more than 8,000 policies and have reduced internal helpdesk call volumes by 60%. That is a defined service metric. The same presentation says AI use has scaled to more than 2,000 software developers, with a targeted productivity improvement of 20% to 30%. The presentation therefore places an observed operational result alongside an ambition for another function. They should keep their respective labels. Fewer helpdesk calls can indicate that employees obtain information through a different channel. Assessing the full benefit also requires knowing whether answers are correct and whether unresolved questions reach support. The coding percentage, meanwhile, remains a target in this disclosure. For management reporting, this example shows why an operating indicator and a planned productivity uplift belong in different columns even when both sit within the same AI programme.

Wells Fargo introduced AI Teammate in July as part of Advisor Gateway. The initial capability lets financial advisors and support teams ask questions in ordinary language, search the firm's resource centre and summarise product, process and workflow information. It also provides guided next steps. The implementation detail is its location inside the platform advisors already use. Employees can obtain information at the point of work instead of moving between several resource pages. Wells Fargo describes later capabilities for preparing client conversations and surfacing opportunities as future development. The release gives a clear account of the initial service, but no measured reduction in hours or increase in revenue. For wealth-management teams, it illustrates a deployment centred on advisor support and information access, with a stated boundary between current functionality and the roadmap.

Starling's August Smart tools release takes the interface closer to the customer. The bank described an initial set of tools for personalised savings plans, scam guidance, automated VAT saving and support for Making Tax Digital. It also announced an intention to release new tools each week. The activities are specific enough for a customer to recognise the task being performed: allocating money for a tax obligation differs from asking a general assistant for financial information. Starling separates these initial capabilities from a later ability for customers to create their own tools. For a retail or small-business banking team, this is evidence of product packaging around recurring financial tasks. The announcement establishes the capabilities offered and the planned release cadence. It leaves customer uptake and realised financial outcomes for later measurement.

Citizens Financial Group's July Reimagine the Bank update describes a portfolio with different levels of maturity. It reports AI tools in credit-card contact centres and employee human resources, alongside pilots for commercial prospecting and complaint handling. Its engineering-productivity ambition comes from an early pilot. This sits inside a broader transformation programme, whose annualised benefit targets include work beyond AI. Start-up costs are also expected to offset benefits within the year. For a finance team, the accounting perimeter is as significant as the use-case list. A transformation programme's projected annualised benefit is not automatically an AI-only saving, nor is it the amount already recognised in the current period. The disclosure is useful because it lets the reader separate deployed tools, pilots, programme targets and implementation costs before drawing a conclusion about financial performance.

U.S. Bank's expanded collaboration with Amazon Web Services identifies Amazon Bedrock and Amazon Nova Sonic among the technologies supporting its modernisation and customer-experience plans. Naming the service and model provides a firmer architectural reference than an announcement that mentions generative AI alone. Bedrock provides access to model capabilities through AWS, while the release identifies Nova Sonic for conversational customer experiences. The document uses planning language for the customer-experience enhancement. It consequently supports an account of the technology programme and selected components, rather than a completed result for every customer interaction. For a bank choosing between operating models itself and consuming a cloud service, this is an example of the latter route. The commercial and control assessment still needs the bank's own workload, integration requirements and service arrangements.

Amalgamated Bank's 15 October release describes its work as a design partner for the FIS AML Investigations Agent. The agent gathers relevant data, analyses customer activity and prepares investigation summaries supported by an audit record. The bank projects a reduction of more than 90% in initial alert-review time and approximately 98% in deeper case-investigation time, based on its current workflows and the agent's capabilities. Those percentages are projections in the bank's announcement. Investigators retain authority over final decisions. This matters for a financial-crime programme because preparing evidence and deciding whether activity is suspicious carry different responsibilities. The proposed gain concerns the workload around investigation. The disclosure does not show that an investigator's decision has been automated, and it does not present the projected percentages as a completed production study.

The MVB Bank arrangement described by Bretton AI changes the purchasing unit as well as the technology. Under the managed service, analysts use the AI platform to review alerts and complete enhanced due diligence. Bretton says a trained analyst reviews every AI-assisted output before delivery to MVB. Pricing is tied to completed work rather than analyst hours. The supplier's account therefore links efficiency to the bank's cost of the service instead of the number of hours the provider can bill. That is a concrete commercial distinction for procurement teams assessing automation. It also preserves a human review stage in the service design. The account is a supplier declaration about its arrangement with a named bank. It establishes neither a measured reduction in MVB's total employment nor a general conclusion that internal compliance teams can be removed.

Foreign-exchange forecasting provides a useful example of the distance between a model benchmark and a bank outcome. Ant International's release names Barclays' BARX NetFX and Standard Chartered's SCALE FX among applications of its forecasting technology. HaiPay's analysis of the associated leaderboard reports Falcon first among 35 entries labelled pretrained, but tenth among 122 entries with sufficient data for scoring on its normalised error measure. The technical paper also identifies a limitation concerning conditioning on additional variables. The ranking depends on which population is being compared. For a treasury or markets team, that makes the evaluation design material: selecting the pretrained subset answers a different question from comparing all scored approaches. A supplier's named banking applications and a benchmark result also remain separate from measured savings or trading performance inside either institution.

The Bank of England's July Financial Policy Committee record brings the operational-risk perspective into the review. It says rapid advances in frontier AI capabilities have increased financial-stability risks associated with cyber and operational resilience. The assessment also identifies a complication in responding quickly: rapid patching of vulnerabilities can itself lead to errors and outages. That connects the pace of threat discovery with the institution's ability to test changes, deploy them safely and recover services. For bank management, the issue is the capacity of the full response process. A faster identification of vulnerabilities is useful only if remediation can keep pace while preserving service continuity. This is a dated supervisory assessment of risk. It should be read as such, rather than converted into a count of actual AI-caused banking incidents.

The Financial Stability Board published its final report on responsible adoption of AI on 15 October. It identifies 12 sound practices intended to help financial institutions capture benefits while managing risks in support of financial-system resilience. The report says application should reflect each institution's context and risk profile. This provides a dated reference for reviewing governance and control arrangements around a bank's deployments. It also makes the perimeter clear: the publication is a set of practices, not evidence that a particular institution has implemented them. For a bank already expanding its AI inventory, the relevant exercise is to compare those practices with the actual allocation of responsibilities, assessment and controls across its systems. The report adds a current supervisory reference to that work without establishing the effectiveness of any individual bank's programme.

The final operational signal comes from separate accounts of agent monitoring and financial-services implementation. Salesforce's engineering team describes how conventional green status indicators can conceal a request that never reaches an agent or an outage that leaves a user without an answer. In a separate interview, Raymond James describes comparing agent results with trusted internal sources through a knowledge graph and evaluation process. The connection is specific: checking that a service is available and checking its answer require different evidence. An agent can respond promptly and produce the wrong information. It can also have a healthy component dashboard while the complete request fails. These accounts support that bounded distinction. They do not establish a Salesforce deployment at Raymond James or a sector-wide adoption rate. For production teams, the useful implication is to test the complete user request and the content returned.

The disclosures in this edition show where the evidence is becoming concrete: a named coding agent linked to supplier spending, a defined mortgage-preparation workflow, an internal service metric, and financial-crime tools with an explicit review stage. They also show why the unit of measurement matters. Deployment counts, task duration, vendor costs and projected programme benefits describe different outcomes. Keeping those distinctions visible makes comparisons between banks more useful. Thank you for listening. You can find more at www.aequitus.net. Subscribe to the channel for the next edition of AI Press Review.