Banks are among the most sophisticated data organizations on the planet. They have decades of experience managing risk, building models, and making consequential decisions at scale. However, the capabilities that make financial institutions world-class risk managers, such as rigorous governance, strict change control, and deep regulatory accountability, are the same ones that can slow AI adoption to a crawl.

Despite these considerations, research shows banks are bullish on generative AI (GenAI). Accounting firm EY’s March 2025 GenAI in Retail and Commercial Banking survey found that 77% of banks have launched GenAI applications, with 61% of respondents reporting substantial results from their deployments. The survey found 89% of respondents expect GenAI to deliver major transformative benefits within the next two years. At the same time, interest in agentic AI is growing.

The Challenge: This Tech Moment Is Different

Banks have experience navigating technology disruptions. They adapted their operations to the internet, mobile banking, and the first wave of machine learning. But the current moment has a different character, because three forces are converging at the same time:

  1. Competitive pressure: Fintechs and neobanks are unfettered by decades of legacy infrastructure, and they are not subject to the same regulatory overhead as traditional institutions. They can ship AI features in weeks. It can take traditional banks years to move a comparable capability from pilot to production. That asymmetry affects customer acquisition, product quality, and margin.
  2. Regulatory frameworks are changing fast: The Federal Reserve’s guidance on model risk management (SR 11-7) was written for traditional statistical models. In revised guidance (SR 26-2), regulators are increasingly applying the original logic to machine learning and GenAI systems and creating compliance requirements that most banks are still working out. The frameworks are evolving faster than many institutions can adapt.
  3. Consumer technology is increasing customer expectations: People experience personalized, real-time, frictionless service from the apps they use every day. They expect their bank to feel similarly responsive. According to research by Kantar, a market research firm, 77% of consumers use mobile apps for banking, and the top reasons consumers changed their banks in the previous 12 months were for convenience, ease of use, and speed.

The result is a situation where banks cannot afford to move slowly, but they also cannot afford to move carelessly. These risks pull in opposite directions.

Where AI Is Creating Real Value in Banking

Amid all the pilot programs and strategic roadmaps, certain AI use cases have moved from experiment to operational infrastructure:

Fraud detection and prevention is the most mature category. Banks are more digital and more interconnected than ever, which means a larger attack surface and faster-moving threats. AI-driven fraud detection now processes thousands of transaction variables in real time, incorporating behavioral patterns, geolocation data, and device fingerprinting.

According to a Claims Journal article citing the 2025 Mastercard and Financial Times Longitude report, 42% of card issuers and 26% of acquirers (who process transactions on the part of merchants) have each avoided more than $5 million in fraud losses over two years, and they attribute those savings to AI.

Clean, reproducible model pipelines that run reliably under production load are what make real-time detection possible at scale. That infrastructure requirement matters more than most teams realize until it becomes a problem.

Credit risk and underwriting with AI have moved past experimentation. Manual underwriting is too slow and too inconsistent for a competitive lending market. AI models can incorporate alternative data sources, improve predictive accuracy, and return decisions in minutes rather than days.

EnterCard, one of the Nordic region’s leading credit market companies, reduced overall credit risk model development time by 25% after moving to a platform to enable secure access to open source packages. The compliance documentation process for heavily regulated credit models dropped from several weeks to just a few days. That kind of operational improvement is what makes scale possible.

Regulatory compliance and anti-money laundering (AML) represent the third pillar. The regulatory burden in financial services is a permanent feature of the landscape, and it is intensifying. AI improves the accuracy and throughput of AML and know-your-customer (KYC) processes while reducing false positive rates that drain analyst time.

However, AI also introduces new model risk questions, specifically around transparency, audit trails, and reproducibility, and regulators are actively wrestling with these realities. Institutions that treat compliance as an infrastructure requirement rather than a box-checking exercise will be better positioned as the regulatory environment continues to evolve.

When and Why AI Breaks Between the Lab and Production

The most important structural problem in AI for banking is the gap between the environment in which a model is built and the environment where it needs to run. In regulated institutions, bridging that gap can take months.

Here is what that failure typically looks like: a data scientist builds a working model on their local machine or in a cloud notebook. But the model depends on specific package versions that are not approved for production. Security teams flag the open-source dependencies because their provenance cannot be verified. The model lineage documentation that regulators would require for sign-off does not exist. The ML pipeline is built on tooling that the production infrastructure team does not support. Six months later, the pilot is still a pilot.

These are common patterns in large financial institutions and occur for understandable reasons. Security teams are doing their jobs. Regulatory reviewers are doing their jobs. The problem is organizational and structural. It is also compounded by mergers and acquisitions, which can introduce incompatible technology and workflows overnight. Add to this that data science and IT teams often operate under fundamentally different assumptions about what “ready to deploy” means.

As long as the tools, environments, and governance frameworks used in development are disconnected from what production actually requires, the AI-to-production gap will persist.

3 Infrastructure Gaps that Keep AI for Banking in Pilot Mode

So what does success look like? Across independent research, it’s easier to identify what is common among financial services AI implementations that failed. Three gaps account for most of the distance between AI ambition and production reality:

  1. Lack of standardized tooling across teams: Every group can build its own stack. A May 2025 GAO report on AI use in financial services found that federal financial regulators are actively expanding AI oversight precisely because fragmented implementation across institutional functions creates supervisory blind spots. Only about one-third of organizations across industries report genuine enterprise-wide AI scaling, with disconnected infrastructure cited as a primary barrier. To close the gap: Provide secure access to curated tools across the organization.
  2. Reproducibility is an afterthought: In regulated environments, the question “can you recreate exactly what this model was doing on this date with this data?” will be asked. A Bank for International Settlements analysis of AI explainability in financial services found that the limited reproducibility of complex AI models poses significant challenges for both financial institutions and regulators, particularly when those models are used in critical business applications such as credit decisioning and risk management. To close the gap: Treat reproducibility as a compliance requirement.
  3. Governance is added at the end of development. McKinsey’s research on GenAI governance in financial services identifies this as a core gap: most financial institutions’ data governance controls do not sufficiently address AI development workflows, which creates downstream compliance exposure and slows the model approval process. To close the gap: Build governance into the development process.

Banks Use Anaconda as their One-Stop Solution

Standard tooling, reproducibility, governance, and shared infrastructure are precisely the capabilities Anaconda is designed to provide. The Anaconda platform helps organizations go from prototype to production faster and more safely, closing the execution gap between what data scientists build and what IT can trust in production. It addresses the full stack of requirements: secure package management and distribution, governed AI development environments, AI model development and deployment, and the reproducibility and audit trail infrastructure that regulated industries require.

Danske Bank, Denmark’s largest financial institution, described Anaconda as their “one-stop solution for the majority of data-science-related work,” supporting 90 production models across upselling, cross-selling, and marketing decisioning while meeting stringent regulatory requirements. Organizations using the platform have documented a 119% ROI within 8 months, a return driven by both cost savings and the compounding value of getting AI into production faster and keeping it there.

For banks trying to close the gap between AI ambition and production reality, the infrastructure question is often the right place to start. To learn more about AI in banking, see how financial institutions are scaling intelligent systems. To discover how we are building AI-native capabilities into our platform, see our Outerbounds acquisition announcement.