Insights

State of AI in U.S. commercial banking

Use cases, outcomes, and the emerging agentic stack

Adnan Masood, PhD, Chief AI Architect, UST

Figure 1. State of AI in U.S. Commercial Banking — evidence-based deployment and outcomes (UST AI Research).

Commercial banks have always invested heavily in technology, but the current AI cycle differs in one important way : it is moving from experimentation to operational muscle memory. Over the last 24 months, U.S. commercial banks have begun to industrialize AI across the operating model—front office, middle office, back office, and technology—using a mix of predictive machine learning, generative AI copilots, and early forms of bounded “agentic” orchestration. What’s changing isn’t the existence of AI; it’s where it shows up, how tightly it is governed, and whether leaders can point to credible outcomes rather than aspirations.

The fastest way to separate what’s real from what’s rhetorical is to follow the evidence trail banks are willing to disclose: interaction volumes in digital assistants, reductions in internal ticket volumes, throughput gains in regulated operations, and fraud-loss reduction. These are the metrics that survive both the audit committee and the earnings call.

DIVIDER

A practical taxonomy: Where AI actually lands

In banking, technology only matters if it moves one of four levers: revenue growth, risk control, operational efficiency, or speed of change. That’s why it helps to organize AI the way banks fund and govern it:

Front office: customer channels, contact centers, relationship management, origination journeys. Middle office: fraud, financial crime (AML/KYC), risk, compliance, model governance. Back office: operations, payments ops, reconciliation, document processing, finance operations. Technology: platforms, data foundations, cybersecurity, and software delivery.

Across those layers, “copilot first” is the dominant pattern. AI drafts, summarizes, classifies, and routes; humans decide. Where automation expands, it tends to be tightly bounded steps—virtual assistant containment, IVR routing, document extraction, or reconciliation matching. True agentic AI (multi-step execution across tools and systems) is emerging, but it remains the exception, not the default.

DIVIDER

Front office: Servicing is the beachhead, not a sideshow

Customer service is where public evidence is strongest because banks can count on it. They can report volumes without disclosing proprietary models. And they can improve experience while taking cost out of routine interactions.

Bank of America’s Erica is still the clearest scaled example. As of April 2024, the bank reported Erica had surpassed 2 billion interactions and responded to 800 million inquiries from more than 42 million clients—proof that AI-mediated servicing is not a pilot but a core operating capability at consumer scale. By April 2025, Bank of America said clients had interacted with Erica more than 2.5 billion times, with 20 million clients actively using the assistant. The same disclosure notes that Erica delivered personalized insights and guidance more than 1.2 billion times, highlighting how servicing and “next-best insight” personalization often converge inside a single digital experience. In that same release, the bank said more than 98% of clients get the answers they need from Erica within 44 seconds, on average.

Regional banks are publishing credible servicing metrics as well. Truist’s 2024 corporate responsibility and sustainability report states that its virtual assistant handled an average of 300,000 conversations per month in 2024, signaling steady production usage rather than experimentation. Wells Fargo positions “Fargo” inside its mobile app as a virtual assistant for everyday banking needs, including helping customers find transactions and routing to customer service when needed.

Citi’s disclosures reinforce a complementary move: GenAI doesn’t just “talk to customers”; it increasingly supports the human agent. In June 2025, Citi described launching GenAI pilots in U.S. Personal Banking Operations, including “Agent Assist” with real-time transcripts and after-call summaries—classic levers to reduce handle time and after-call work in the flow of service.

Figure 2. Front office servicing has reached industrial scale (selected public metrics; UST AI Research synthesis).

The strategic implication is that the first durable value from GenAI in servicing is not an open-ended conversation. It’s structured containment plus structured augmentation—flows that can be audited, escalated, and measured.

DIVIDER

Personalization and next-best-action: High value, low transparency

Personalization is routinely cited as a high-value AI use case, particularly as banks fight for primary relationships and deposits. Yet public evidence tends to describe outcomes (“insights,” “guidance”) rather than governance (“why this offer?” explainability, bias testing, control thresholds). That caution is rational: personalization is adjacent to fair lending, discrimination risk, and reputational exposure.

Still, the signal is clear that personalization is embedded in everyday experiences. Bank of America’s description of Erica’s personalized guidance at a massive scale implies ongoing investment in segmentation and behavioral targeting within the digital channel. Wells Fargo’s Fargo is described providing personal insights like balance forecasts, reflecting the predictive side of next-best-action even when the underlying models are not disclosed. Capital One describes using proprietary AI to customize user experiences across digital and mobile channels, aligning with a test-and-learn personalization culture.

What we rarely see publicly is the control-plane detail—how models are constrained, monitored for drift, and defended under regulatory scrutiny. For outside observers, the absence of detail should be read as disclosure conservatism, not necessarily weak governance.

DIVIDER

Middle office: Risk, fraud, and KYC are where AI proves ROI

Fraud is one of the few domains where banks publish numbers that look like unambiguous ROI. PNC’s 2024 annual report states that since the deployment of its new account fraud platform, reported identity theft fraud decreased by 44%. The same report describes a 78% reduction in fraudulent text reports compared to 2023 and blocking more than a million unsigned calls in 2024—concrete outcomes tied to scam and impersonation mitigation. These disclosures highlight a core truth: AI’s most defensible value often shows up first as loss avoidance, not revenue uplift.

Regulated operations deliver similarly “hard math” results. At JPMorgan Chase’s 2024 Investor Day, management described major productivity gains in KYC file processing, including processing 230,000 files with 20% less staff and citing productivity improvements in the 80%–90% range. Whether the underlying stack is document intelligence, NLP, or workflow automation, the business outcome is straightforward: throughput rises without proportionate headcount growth, and backlog risk declines.

Figure 3. Fraud reduction can deliver measurable ROI (illustrative outcomes cited in public disclosures; UST AI Research).

Citi’s contact-center pilots mirror this approach to safe deployment: use GenAI for real-time transcripts and after-call summaries, keep humans in control, and focus on speed and accuracy in regulated conversations. That’s the pattern banks prefer in the middle office—augment decisions, don’t automate them end-to-end unless the controls are mature.

DIVIDER

Back office: Document AI and payments ops become products

The back office is where AI can feel less glamorous but often creates the most durable benefits. Banks have automated document flows for years; GenAI expands what “automation” can handle in semi-structured documents and exception paths.

PNC’s annual report makes the direction explicit, stating the bank is focused on “document AI” to drive efficiencies. Truist provides a more productized example: in February 2026, it launched an AI-enabled receivables platform designed to match payments to invoices, unify remittance data, accelerate cash application, and minimize exceptions—essentially exporting back-office automation benefits to commercial clients. The significance is easy to miss: AI is not only reducing internal costs; it is also being packaged as a differentiated commercial-banking capability. Technology: Copilots, platforms, and the real cost of “AI at scale”

If you want to understand why CIOs are enthusiastic about GenAI, start with internal support. These are high-volume, repeatable workflows where even modest automation yields immediate returns. Bank of America reported that over 90% of employees use “Erica for Employees” and that it reduced calls into the IT service desk by more than 50%—a rare, operationally concrete internal assistant metric. Citi reports similarly broad distribution: its proprietary AI tools reach 80 countries and more than 175,000 colleagues, signaling enterprise standardization of internal AI capabilities. Wells Fargo has framed its push as workforce enablement at scale, noting that it has trained over 90,000 employees and deployed AI tools to over 180,000 desktops. Engineering workflows are also being reshaped. J.P. Morgan Payments describes internal tooling that leverages an internal LLM capability (“LLM Suite”) for SDLC tasks such as pull request review and unit tests—work that accelerates delivery while remaining constrained by existing software approval gates. U.S. Bank’s GenAI “Developer Assistant” targets a different bottleneck: partner integration. It recommends APIs, generates sample code, and is intended to help embedded banking clients integrate faster and go live sooner. In both cases, the theme is the same: AI is being wired into work systems, not bolted on as a separate chat window.

DIVIDER

Model choice: Why “which LLM?” is a governance question

Banks do not “choose an LLM” the way they choose a core system. They build abstraction layers—internal gateways that can route tasks to different models by use case, with security and governance controls enforced centrally.

Where banks do name models, three patterns appear in public evidence. First is enterprise-wrapped access to OpenAI-class models via Azure OpenAI. In April 2024, Ally’s leadership described its platform’s initial LLM tool as Microsoft’s Azure OpenAI Service, naming GPT‑3.5 Turbo and GPT‑4, and framed the approach as a controlled internal platform. Second is multi-model usage: Banking Dive reported in February 2026 that Citi’s internal “Stylus Workspaces” is underpinned by Google Gemini and Anthropic Claude. Third is open-weight models for control: Capital One stated it developed “Chat Concierge” leveraging Meta’s open-source Llama model as a base and customizing it with proprietary data to meet performance, risk, and governance thresholds.

Figure 4. Model and provider matrix (examples of banks publicly naming models/providers ; UST AI Research).

The control plane is the story behind these choices. Ally’s tech blog describes building a PII masking workflow using LangChain to scrub sensitive data before it is sent to “any” LLM and then rehydrate it afterward—plumbing that turns GenAI into a compliant enterprise service rather than an unsecured experiment.

Figure 5. The control plane: guardrails, PII protection, and auditability that enable safe scale (UST AI Research).

DIVIDER

Agentic AI: Early, bounded, and control-heavy

“Agents” is the most abused term in AI conversations. In banking, the definition should be strict: an agent is a system that can plan and execute multi-step actions across tools and workflows, with controls, approvals, and auditability—not merely draft text.

By that standard, public evidence of true agentic behavior remains limited. Capital One provides one of the clearest bank-authored examples: its tech blog describes Chat Concierge as a multi-agentic system that can take action (including scheduling appointments) and explicitly validates plans to mitigate hallucinations and errors. Citi’s September 2025 press release describes upgrading Stylus Workspaces with “Agentic AI,” integrating with select Citi systems and web search, and streamlining multi-stage workflows into a single automated process. The promise is evident, but the disclosure does not specify the underlying model provider or the approval/audit mechanics, so outsiders should be cautious about over-interpreting autonomy from marketing language alone.

Figure 6. The shift from retrieval to action: the bounded agent loop (illustrative example; UST AI Research).

The pragmatic takeaway: agentic AI is arriving in banking, but it’s arriving with guardrails first. Expect bounded agents in domains where actions can be pre-approved, reversible, and logged—scheduling, case routing, document assembly, controlled system updates—long before autonomous movement of money or autonomous credit decisions.

The spend reality: Big numbers, thin disclosure

Most banks don’t disclose “AI budgets” as a standalone line item; AI is embedded in broader technology and transformation spend. Still, the disclosed spending envelopes show what it takes to operate AI at scale. JPMorgan Chase cited technology spend of about $17 billion. Bank of America has described annual technology investment of $13 billion, with roughly $4 billion allocated to new initiatives in 2025. Citigroup’s earnings materials cited $11.8 billion in technology investments and $2.9 billion in transformation investments in 2024. These are not pure AI numbers, but they define the runway for AI platforms, data, security, and change capacity.

Figure 7. AI spend is largely embedded in broader technology and transformation budgets (illustrative public disclosures; UST AI Research).

DIVIDER

Where the next wave of wins will come from

Across the evidence, the strongest verifiable outcomes today cluster in three places: digital servicing at scale, internal productivity copilots, and fraud/identity protection. The next wave of competitive separation will not come from those who have a chatbot. It will come from those who have an operating system for AI—a control plane that governs model choice, protects data, tracks provenance, evaluates outputs, and integrates AI into workflows without creating unmanageable risk.

The state of AI in U.S. commercial banking, then, is a story of disciplined deployment. Banks are proving value where measurement is clean, and automation is bounded. As agentic capabilities mature, the winners will be the institutions that treat autonomy as a product feature to be earned—one workflow, one approval gate, and one audit log at a time.

Ready to move from AI pilots to operational scale? Let's talk about what disciplined deployment looks like for your institution. >


DIVIDER

References

Bank of America, “BofA’s Erica Surpasses 2 Billion Interactions, Helping 42 Million Clients Since Launch,” Bank of America Newsroom (Press Release), Apr. 8, 2024. [Online]. Available: https://newsroom.bankofamerica.com/content/newsroom/press-releases/2024/04/bofa-s-erica-surpasses-2-billion-interactions--helping-42-millio.html (accessed Feb. 22, 2026).

Truist Financial Corporation, “Truist 2024 Corporate Responsibility & Sustainability Report,” 2025. [Online]. Available: https://www.truist.com/content/dam/truist-bank/us/en/documents/corporate-responsibility/truist-2024-corporate-responsibility-and-sustainability-report.pdf (accessed Feb. 22, 2026).

Wells Fargo, “Fargo: Your Virtual Assistant for Everyday Banking Needs,” Wells Fargo product page. [Online]. Available: https://sites.wf.com/fargo/ (accessed Feb. 22, 2026).

Citigroup, “Strengthening Our Commitment to AI Innovation,” Citi Insights, Jun. 23, 2025. [Online]. Available: https://www.citigroup.com/global/insights/strengthening-our-commitment-to-ai-innovation (accessed Feb. 22, 2026).

Capital One, “AI at Capital One,” Capital One Tech, n.d. [Online]. Available: https://www.capitalone.com/tech/machine-learning/ai-at-capitalone/ (accessed Feb. 22, 2026).

The PNC Financial Services Group, “2024 Annual Report,” filed Mar. 3, 2025. U.S. SEC EDGAR. [Online]. Available: https://www.sec.gov/Archives/edgar/data/713676/000162828025009250/pnc-20241231.htm (accessed Feb. 22, 2026).

JPMorgan Chase & Co., “Investor Day 2024: Highlights Transcript,” May 20, 2024. [Online]. Available: https://www.jpmorganchase.com/content/dam/jpmc/jpmorgan-chase-and-co/investor-relations/documents/events/2024/investor-day-2024-highlights-transcript.pdf (accessed Feb. 22, 2026).

Truist Financial Corporation, “Truist launches AI-enabled receivables platform to accelerate cash application and minimize exceptions,” News Release, Feb. 3, 2026. [Online]. Available: https://media.truist.com/2026-02-03-Truist-launches-AI-enabled-receivables-platform-to-accelerate-cash-application-and-minimize-exceptions (accessed Feb. 22, 2026).

Bank of America, “AI Adoption by BofA’s Global Workforce Improves Productivity, Client Service,” Bank of America Newsroom (Press Release), Apr. 8, 2025. [Online]. Available: https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/04/ai-adoption-by-bofa-s-global-workforce-improves-productivity--cl.html (accessed Feb. 22, 2026).

Citigroup, “AI Enters 80 Markets and Counting,” Citi Insights, Sep. 4, 2025. [Online]. Available: https://www.citigroup.com/global/insights/ai-enters-80-markets-and-counting (accessed Feb. 22, 2026).

Wells Fargo, “Wells Fargo names Saul Van Beurden to Lead Artificial Intelligence and Digital Transformation,” Wells Fargo News Release, Nov. 20, 2025. [Online]. Available: https://newsroom.wf.com/news-releases/news-details/2025/Wells-Fargo-names-Saul-Van-Beurden-to-Lead-Artificial-Intelligence-and-Digital-Transformation/default.aspx (accessed Feb. 22, 2026).

J.P. Morgan Payments, “How J.P. Morgan Developers Leverage AI,” J.P. Morgan Payments Developer Portal, Apr. 22, 2025. [Online]. Available: https://developer.payments.jpmorgan.com/blog/guides/ai-software-development (accessed Feb. 22, 2026).

U.S. Bank, “GenAI Assistant Helps Companies Embed U.S. Bank Solutions into Their Platforms,” U.S. Bank News & Stories, Jan. 26, 2026. [Online]. Available: https://www.usbank.com/about-us-bank/news-and-stories/article-library/genai-assistant-helps-companies-embed-us-bank-solutions-into-their-platforms.html (accessed Feb. 22, 2026).

J. Moore, “Ally’s Generative AI Strategy Eyes Multiple LLMs, AI Agents,” TechTarget SearchCIO, Apr. 17, 2024. [Online]. Available: https://www.techtarget.com/searchcio/feature/Allys-generative-AI-strategy-eyes-multiple-LLMs-AI-agents (accessed Feb. 22, 2026).

Banking Dive, “Citi Stylus Workspaces underpinned by Gemini and Claude,” Feb. 4, 2026. [Online]. Available: https://www.bankingdive.com/news/citi-stylus-workspaces-agentic-ai-turbocharging-productivity/752890/ (accessed Feb. 22, 2026).

Capital One, “Driving the Future of Car Buying with Agentic AI,” Capital One Tech Blog, Mar. 5, 2025. [Online]. Available: https://www.capitalone.com/tech/ai/future-of-ai-car-dealerships-shopping/ (accessed Feb. 22, 2026).

Ally, “Ally Gives Back to LangChain AI Community with PII Masking Module,” Ally Tech Blog, Jan. 11, 2024. [Online]. Available: https://www.ally.com/tech/ally-gives-back-to-langchain-ai-community-with-pii-masking-module/ (accessed Feb. 22, 2026).

Citigroup, “Citi Unveils Citi Stylus Workspaces with Agentic AI, Turbocharging Productivity at Citi,” Press Release, Sep. 22, 2025. [Online]. Available: https://www.citigroup.com/global/news/press-release/2025/citi-unveils-citi-stylus-workspaces-agentic-ai-turbocharging-productivity (accessed Feb. 22, 2026).

Citigroup, “Fourth Quarter 2024 Earnings Transcript,” Jan. 15, 2025. [Online]. Available: https://www.citigroup.com/rcs/citigpa/storage/public/Earnings/Q42024/4Q24-Earnings-Transcript.pdf (accessed Feb. 22, 2026).