Major banks are investing in AI startups not because the underlying AI models are irreplaceable, but because the software layer that connects those models to a bank's internal data, permissions and workflows is nearly impossible to replace once installed. This week, UBS backed Finster AI and HSBC invested in Model ML, joining JPMorgan Chase and Wells Fargo in a trend that signals how financial institutions view AI as a long-term operating layer rather than a temporary technology trend.

The distinction matters enormously. Banks can swap out the underlying large language model (LLM), an AI system trained on vast amounts of text data, from one provider to another. But the software sitting between the model and the bank, the software that knows how to access internal data, apply permission rules, and format outputs into presentations, financial models, or research notes, becomes deeply embedded in a bank's operations. Replacing that system would require massive recoding efforts and operational disruption.

What Are These AI Workflow Companies Actually Doing?

Companies like Finster AI, Model ML, and Rogo operate in what's called the "AI workflow layer." They sit between the raw AI models and the bankers who use them. FT Partners, which led Model ML's $75 million Series A funding round last year, is now rolling out the platform firmwide across its organization, demonstrating confidence in the model's value.

These platforms handle several critical functions that make AI useful in a banking context:

• Data Integration: They connect to internal bank databases and systems, pulling in proprietary information that public AI models have never seen.
• Permission Management: They enforce access controls so that junior analysts cannot see senior traders' strategies or confidential client information.
• Output Formatting: They transform raw AI responses into polished deliverables like client presentations, financial models, or research notes that bankers can actually use.
• Model Flexibility: They allow banks to swap the underlying AI model without rebuilding the entire system, providing some protection against vendor lock-in.

The practical implication is significant. Banks continue to build their own AI systems and sign deals with major technology providers like Google, OpenAI, and Anthropic. But the workflow layer companies are becoming the glue that makes those relationships valuable.

How Are Banks Approaching AI Infrastructure Decisions?

A separate research effort by Coalition Greenwich, which interviewed 15 technology professionals at major banks across North America, the United Kingdom, and Europe, reveals that banks are taking a cautious, hybrid approach to AI and infrastructure modernization. The findings show that despite public enthusiasm for cloud computing and AI, the mainframe, the large centralized computer that has powered banking for decades, remains deeply embedded in financial institutions.

More than 50% of applications at major banks still run on mainframes, with some institutions reporting concentrations above 75%. Consumer banking applications like deposits, loans, and payments are heavily deployed on mainframes, and banks contend they will remain there despite modernization initiatives. Senior executives at the largest banks often prefer to migrate only 5% of applications off the mainframe each year, a pace that means most chief information officers will retire before any mainframe plans are abandoned.

When it comes to AI specifically, the picture is nuanced. IBM, the largest mainframe provider, has developed AI capabilities for mainframe systems, including fraud prevention applications. Credit card transaction data often resides on mainframes, making them a logical place to run fraud detection AI. However, not all banking executives are aware of or focused on these new AI capabilities and need education about their potential.

The research found that more than half of respondents believe their mainframes can accommodate increasing AI demands, although the latest AI capabilities are not yet driving technology decisions. Banks continue to value the mainframe more for resilience and dependability than for speed of innovation.

What Does This Mean for the Future of Banking Technology?

The likely end state is not a choice between mainframe and cloud, but a hybrid environment in which each supports the workloads it handles best. Banks will continue modernizing core applications on mainframes while using cloud infrastructure where it provides greater flexibility, speed, or access to new capabilities. Among respondents, 27% plan to modernize legacy applications while remaining on the mainframe, and 20% plan to maintain their current applications without expanding them. Only 13% plan to migrate all software and applications off the mainframe.

Meanwhile, the investments by UBS, HSBC, JPMorgan, and Wells Fargo in AI workflow startups suggest that banks are thinking strategically about where AI adds value. The workflow layer is where competitive advantage lies, not in the underlying models themselves. As banks continue to adopt AI across consumer and commercial banking applications, the question is less about whether technology will replace mainframes and more about where AI workloads should run and how they integrate with existing infrastructure.

The broader implication is that AI in banking is not about ripping and replacing legacy systems. It is about layering new capabilities on top of existing infrastructure in ways that preserve what works while adding new functionality. For technology leaders at banks, this means the next decade will be defined not by dramatic transformations but by careful integration of AI into hybrid environments where mainframes, cloud systems, and new workflow platforms coexist and serve different purposes.