AI in banking: Use cases, benchmarks, & benefits

What is AI in banking?

Artificial intelligence in banking refers to the application of machine learning, generative AI, natural language processing, and automation to financial tasks such as fraud detection, credit underwriting, compliance monitoring, and customer service.

This software learns patterns from data and generates language and analysis, rather than following a fixed set of rules.

The adoption of AI in banking is no longer in the pilot stage. IBM's 2025 annual banking outlook found that 78% of banks already take a tactical approach to using generative AI, deploying it in specific workflows, but only 8% have moved to a fully systematic, enterprise-wide deployment.

That gap between tactical and systematic illustrates where AI in banking stands today. Most institutions have found use cases that work, but only a few have scaled them significantly.

How AI is used in banking

Machine learning and AI can handle several banking use cases well.

AI in retail and commercial banking

Retail and commercial banking is often where AI operates at scale because the work is high volume and pattern-driven. Core applications include:

• Fraud detection: HSBC's AI system reduced false positives by roughly 60% while detecting substantially more genuine financial crime.
• Credit decisions and underwriting: AI facilitates faster approval processes by using a wider range of borrower data.
• Customer service: Chatbots and call summarization tools can reduce resolution times.
• Personalization: AI-generated product and pricing recommendations can be tailored to individual customer behavior.
• Debt collection and risk management: AI can help prioritize outreach and detect default signals earlier in the process.
• Regulatory compliance: Anti-money-laundering monitoring and automated reporting are common AI use cases in banking.

AI in investment banking

AI in investment banking supports the analyst's toolkit, which includes market research, deal sourcing, due diligence, financial modeling, and pitch preparation.

This is deal-oriented, judgment-intensive work, making it much harder for AI compared to high-volume retail tasks. A single valuation can involve numerous assumptions, source documents, and firm-specific conventions, so an error early in the reasoning process can compound.

This complexity is exactly what makes benchmark evidence so important for anyone evaluating AI in investment banking.

How Mercor trains AI for banking

Mercor partners with finance professionals, including investment bankers, analysts, and specialists from firms such as Goldman Sachs, Morgan Stanley, and JPMorgan, to build the datasets and evaluation frameworks that teach AI models real banking work.

These experts write realistic tasks, define what a correct answer looks like, and grade model outputs against professional standards. That expertise provides the foundation for both training frontier models and measuring their capabilities objectively through benchmarks.

What are the benefits of using AI in banking?

The main advantages of using AI in banking are primarily speed and cost. McKinsey & Company estimates generative AI could add between $200 billion and $340 billion annually to the global banking industry (roughly 2.8% to 4.7% of industry revenue), mostly through productivity gains rather than new revenue streams.

The practical benefits of AI in banking include:

• Better product personalization based on actual customer behavior.
• Faster customer service, often cutting response times from minutes to seconds.
• Lower fraud losses as a result of real-time transaction monitoring.
• Faster decision-making in underwriting and compliance reviews.
• More consistent regulatory monitoring across large transaction volumes.
• Increased analyst time for judgment-intensive work instead of document preparation.

Ultimately, AI doesn't decide who gets a loan or which deal to pursue. Instead, it means the people making those calls can concentrate their time on work that requires human judgment.

What Mercor’s benchmarks reveal about AI's real banking capability

One of the biggest questions that banking enterprises are facing is how much of the work today's AI can actually perform. While, ultimately, the most meaningful evaluation will be the one based on your own banking workflows rather than generic benchmark scores or vendor claims, here’s what Mercor's benchmarks are suggesting:

Mercor's APEX AI productivity index for investment banking as a representative banking workflow, shows that leading models perform well on discrete analytical work, demonstrating their ability to accelerate financial research, valuation, and financial modeling.

Mercor's APEX Agents benchmark for investment banking that measures end-to-end analyst workflows rather than individual tasks show a sharp drop in performance, indicating that while today's models are strong assistants, they remain far less reliable at independently completing complex, multi-step banking workflows.

Every bank's workflows are different. See how Mercor can help benchmark AI models and agents against your team's actual workflows so you can make investment decisions .

How to implement AI in banking

The implementation of AI works best as a narrow, step-by-step process rather than a broad platform rollout. It's best to start with a single measurable workflow, and then expand only after accuracy and controls are proven.

Step 1: Identify the banking problems AI should solve

Start with a workflow that's high-volume, repetitive, and has a clearly measurable outcome, such as transaction monitoring or first-pass document review. Avoid vague goals, as they tend to produce vague results.

Step 2: Prepare high-quality financial data

Models are only as good as the data underpinning them. Issues such as inconsistent transaction labels or incomplete customer records will often remain undetected at first but will later show up as errors downstream.

Step 3: Choose the right AI patterns and internal expertise

Match the AI approach to the task. Use retrieval-based systems for research, generative models for drafting memos, and classification models for fraud detection. Make sure to staff the project with people who understand the underlying banking work, not just the technology.

This is also where integration decisions matter. A model that can't connect cleanly to the core banking systems will never be successful at scale, no matter how accurate it is in testing.

Step 4: Protect sensitive financial data

Customer and transaction data are subject to regulatory obligations even before AI is applied. Encryption, access controls, and clear data-handling boundaries need to exist before a model starts using production data.

Step 5: Test, measure, and scale

Benchmark performance against a defined standard before a wider rollout, then monitor continuously. Scaling an unproven workflow too early on is the most common way these projects fail.

AI in banking: The risks and challenges

Poor AI models rarely fail dramatically. Instead, they tend to fail subtly, introducing errors in a memo nobody double-checks or a lending pattern nobody audits until a regulator investigates.

The risks of using AI in banking include:

• Hallucinations: A model may generate confident but incorrect information.
• Bias in lending decisions: Models trained on historical data can replicate discriminatory patterns.
• Data privacy exposure: Customer financial data may be compromised when it passes through third-party models.
• Regulatory mismatch: A model's decision logic might not be explainable to a supervisor or auditor.
• Overreliance: Staff may stop double-checking outputs that look polished but are still in need of verification.

The European Banking Authority's 2024 supervisory reporting notes that EU and EEA banks are increasingly using general-purpose AI. As a result, regulators are monitoring model explainability and fair-lending compliance more closely.

Agentic workflows (where a model takes multiple actions in sequence with limited human input between steps) raise additional risks that are harder to detect. If an early step in an agentic chain is wrong, the model won't necessarily flag it before compounding the error in subsequent actions.

That's why banks running agentic pilots tend to implement a human checkpoint after every few steps rather than at the end only.

What governance controls apply to artificial intelligence in banking?

AI in banking sits inside an existing, increasingly rigorous governance framework. In the U.S., the Federal Reserve and the Office of the Comptroller of the Currency's SR 11-7 model risk management guidance require independent validation, ongoing monitoring, and thorough documentation. A May 2025 GAO review confirmed these expectations apply to AI systems.

In the EU, the AI Act classifies credit scoring as a high-risk application, adding additional requirements around transparency and human oversight. Generative and agentic AI are still only partly addressed by these regulations, meaning more rules are likely to come.

What comes next? Future trends for AI in banking

The near-term direction for AI in banking points toward deeper integration alongside tighter oversight. Likely trends include:

• Agentic AI systems that execute multistep workflows rather than responding to individual prompts, although benchmarks show these systems still perform poorly on long-horizon professional tasks.
• Converging regulatory frameworks, such as SR 11-7, the EU AI Act, and the National Institute of Standards and Technology's AI Risk Management Framework, that will increasingly demand the same controls.
• Greater demand for explainable AI so that model decisions are auditable for fair-lending and examination purposes.
• Human-in-the-loop systems that embed expert review directly into high-stakes workflows rather than adding it afterward.

Will AI replace banking professionals?

AI in banking is strong at drafting, summarizing, and pattern detection at scale. However, it's weak at multistep reasoning and any task that requires expert professional judgment.

That judgment is what's making finance professionals more valuable, not less. Platforms such as Mercor work with experienced finance professionals to help train and evaluate AI models, one of several new roles emerging around AI for people with established banking backgrounds.

AI will not take over banking jobs anytime soon, but it is causing roles to change.