Why 84% of banks now prioritize Open Source in their AI strategy
Moveo AI Team
February 1, 2026
in
🏆 Leadership Insights
The global banking sector is undergoing a quiet but significant shift. According to the State of AI in Financial Services 2026 report, produced by NVIDIA in partnership with Evident Insights, 84% of financial institutions now consider open source models and software important to their AI strategy. Even more revealing: 42% are already using or evaluating autonomous AI agents.
This data reflects a shift in mindset. Banks want flexibility to use different models, avoid dependency on a single vendor, and maintain control over their data. But there's a critical detail many overlook: open source models alone don't solve the compliance, security, and orchestration challenges the financial sector demands.
Alexandra Mousavizadeh, co-founder and co-CEO of Evident Insights, summarizes the issue: "Open source models can help banks close the gap with early movers, unlock cost efficiencies, and safeguard against vendor lock-in, but they're not without their limitations. Proprietary approaches can unlock superior performance for domain-specific tasks. Leading banks need to demonstrate proficiency in both approaches, applying the right kind of model to the right problem, in the right context."
Over the past two years, many financial institutions adopted generic AI solutions, hoping to solve multiple problems with a single tool. Some opted for open source models seeking flexibility. Others remained on closed proprietary platforms. Reality revealed limitations in both approaches when implemented in isolation.
The central problem isn't the model itself, but the architecture. According to O'Reilly Radar analysis, most of the market still relies on what experts call the "prompt-and-pray" approach: sending an elaborate prompt to a model and hoping it does the right thing. It works well in demos. It fails in production.
Generic AIs do not understand the specific context of the banking sector. Every conversation starts from zero. There is no memory of previous interactions, behavioral patterns, or relationship history. For collections operations, for example, this means losing valuable signals that could indicate the best approach for each customer.
Compliance also suffers. According to American Banker, generic solutions treat regulation as an additional layer, not as part of the design. When something goes wrong, there's no fallback. The system can't self-correct, and there's no way to verify if the AI just broke a policy or promised something it shouldn't have.
The most subtle problem, however, is structural. Many organizations operate with three disconnected systems: one for segmentation, another for conversation, and another for determining next actions. Intelligence does not flow between them. This is why the choice between open source and proprietary is a false dichotomy. What matters is the orchestration layer that connects everything.
Lean more → The Champion Strategy for Debt Collection: AI Trifecta
What defines solutions that deliver results
The NVIDIA/Evident report presents a noteworthy finding: 89% of respondents state that AI is helping increase revenue and reduce costs. Among those reporting gains, 64% saw revenue increases exceeding 5%, and 29% above 10%.
What differentiates those who achieve results?
First, the ability to apply the right model to the right problem, in the right context. This means understanding that different tasks require different approaches. A model optimized for document analysis is not the same as one that works best for customer conversations. Leading banks build architectures that allow the use of multiple models in a coordinated manner.
Second, fine-tuning on proprietary data. Helen Yu, CEO of Tigon Advisory, highlights: "The real value capture happens when institutions fine-tune these models on their proprietary transaction data, customer interaction histories, and market intelligence, creating AI capabilities that competitors cannot replicate." The difference between using a generic AI and one trained on business-specific data is the difference between having a generalist assistant and a specialist who deeply understands the operation.
Third, agents that learn, remember, and connect. The report shows that 21% of banks have already deployed AI agents, with another 22% planning to do so within the next year. According to CIO Dive, these agents do not just execute tasks. They reason, plan, and learn from each interaction. More importantly, they connect information across systems, eliminating silos that fragment organizational intelligence.
But which capabilities truly matter? Our research found that only 6% of banks have the seven critical capabilities needed to scale agentic AI successfully. Discover what they are in our report, 7 Critical Capabilities for Agentic AI in Banking.
Architecture matters. As Backbase points out, solutions that treat AI as an additional layer (AI-bolted) work in demos but struggle in production. Solutions built from the ground up for AI (AI-native) integrate governance, compliance, and continuous learning into the design.
The bridge between flexibility and enterprise security
The 84% of banks prioritizing open source are not abandoning security or compliance. They are seeking platforms that offer the best of both worlds: the flexibility to choose and switch models as technology evolves, combined with the orchestration, governance, and security layer that the financial sector demands.
This means architectures that support "bring-your-own LLM", allowing the use of open source models like Llama or Mistral, or proprietary models fine-tuned for specific tasks. It means privately hosted LLMs, where banking data never leaves the institution's control. And it means multi-agent systems with compliance built into the design, not added as an afterthought.
In collections and revenue recovery, this specialization shows direct impact. According to Neurons Lab's analysis on agentic AI in financial services, AI agents that understand the sector context can identify signs of financial difficulty before accounts become delinquent. They can adjust approaches in real time based on behavior, channel preferences, and payment history.
The pattern repeats across different areas: solutions that combine model flexibility with enterprise-grade orchestration consistently outperform approaches that prioritize only one side of the equation.
The question leaders need to ask
The adoption of open source by 84% of banks is not a rejection of proprietary solutions. It is the recognition that flexibility and control are fundamental requirements. But flexibility without governance is a risk. Control without intelligence is a waste.
Before evaluating any AI solution, technology and operations leaders should consider some fundamental questions: Does the solution learn from each interaction, or does every conversation start from zero? Can I bring my own models, or am I locked into a single vendor? Is compliance built into the architecture, or is it an additional layer? Are the segmentation, conversation, and action systems connected, or do they operate in silos?
If you want to assess where your organization stands on this journey, we created a free tool that analyzes your readiness for agentic AI in less than 5 minutes. Take the AI Agent Readiness Assessment and discover the next steps for your strategy.
The future does not belong to those who use the most advanced model. It belongs to those who apply intelligence in a connected way, with the flexibility to choose the best models for each task and the governance to operate securely at scale. In a sector where relationships and trust define success, the ability to remember, learn, and adapt is not a differentiator. It is a requirement.