Shrink to Win Big! Small Models, Small Data Sets, Big Impact

Financial services firms are in a sprint to harness the power of artificial intelligence, with early successes promising to revolutionize productivity. Yet, behind the headlines driving building their own solution [source], a complex reality emerges; building secure, scalable AI systems that reliably deliver value is a challenge.

For example, AI assistants were initially seen as a quick win for businesses, promising to boost productivity and impress stakeholders. However, the reality has been more complex. Tools like Copilot and Gemini have shown potential but require significant investment in data preparation and management to deliver accurate and reliable results [source].

There are two clear trends emerging from our conversations with customers. First, cleaning and organizing data have emerged as a critical first step in realizing the full potential of AI assistants. Companies are investing heavily in data curation to ensure that AI systems can access accurate and up-to-date information.

This has led to a growing trend of building custom AI assistants tailored to specific business needs, rather than relying solely on pre-built solutions. It is also remarkably interesting that curating smaller datasets for a specific repetitive task gives us better performance [source].

And second, as organizations seek to scale AI initiatives, a growing number are turning to smaller, more focused models [source]. These models, while less glamorous, offer a compelling value proposition: they are often cheaper to fine-tune and deploy [source], give better performance from RAG pipelines, consume fewer computational resources, can deliver impressive results for specific tasks and have a smaller security footprint.

In essence, they represent a more pragmatic approach to AI, prioritizing efficiency and ROI over headline-grabbing capabilities.

It is understandable that the desire to leverage existing investments in processes is driving the build approach. By integrating AI models into current workflows, organizations can maximize the value of their existing systems.

This strategy is particularly appealing in sectors like financial services, insurance and retail, where established processes can be greatly enhanced with AI capabilities.

While off-the-shelf products offer convenience, many of our customers, particularly those in highly regulated industries like finance, are opting to develop their own AI systems to safeguard sensitive data.

This approach allows for greater customization and integration into existing workflows. We can further validate this with the news around Morgan Stanley's debrief solution [source], a tool that summarizes meetings and generates email drafts, is definitive proof point of this trend.

The Role of Security and Trust

It is not surprising that the rapid development of AI has raised concerns about security, privacy, and compliance. As AI systems become increasingly sophisticated, the risks associated with data breaches and misuse grow.

Building trust in AI requires a robust security framework that protects sensitive information while ensuring transparency and accountability. Simple AI systems with smaller security footprints are easier to secure and protect, is a design decision that should be considered [source].

The convergence of smaller, more focused AI models with smaller, domain specific data sets is reshaping the AI landscape. At Fortanix, we can help solve simple end-to-end secure AI workflow orchestration, so organizations can focus on precision and performance of their AI systems to accelerate time-to-market irrespective of their “small” vs. “big” design choice.

Our pedigree is in building simple data security and cryptography products to solve complex security problems. With the growing adoption of AI, we are helping our enterprise customers focus on efficiency, security, and compliance, while they confidently explore the potential of AI while safeguarding their most valuable assets- their data!