NEW YORK — Enthusiasm and hype around generative AI grew last year as companies and individuals began experimenting with new tools like ChatGPT and GitHub Copilot.
But in 2024, organizations are facing a difficult implementation reality as enterprise AI efforts move from small-scale pilots to real-world deployments.
At this week's AI in Finance Summit New York 2024, speakers and attendees expressed a mix of excitement and anxiety about the rapid pace of generative AI development.
“This is the kind of technology we've seen before,” John Chan, director of AI/ML technology at Raymond James, said in his presentation, comparing generative AI to the emergence of technologies like mobile and cloud. “It's a wave of.” “But I think the only difference this time is that it's actually much faster.”
Sarah Hoffman, vice president of AI and machine learning research at Fidelity, said in a post-presentation Q&A that while these technologies tend to have distinct early adopters, generative AI is more widely used across demographics. He said that he seemed to have a certain charm. This widespread transition from experimentation to practical application tempers the initial excitement about generative AI in finance.
Applications of generative AI in financial services
Practitioners cited a wide range of use cases for generative AI in finance and banking, including chatbots and virtual assistants, fraud detection and prevention, credit risk assessment, personalized marketing, investment management, and document analysis and processing. I am.
The most common of these applications today is text and code generation, a broad category that includes everything from email drafts to SQL queries to synthetic data, Chan said.
Text summarization and analysis are also very popular. Financial services companies often have large amounts of data, including customer transactions, financial reports, and regulatory filings. However, much of this information is unstructured and segregated across different departments, making it difficult for organizations to use it effectively.
“In large companies, it's very difficult to share knowledge across the company, and it's difficult to collaborate,” Hoffman says. In-house generative AI tools could help users by summarizing information about a company's structure, goals and projects in other business units, she said.
A technique called search augmentation generation (RAG), a popular topic across conferences and in the AI industry, can be particularly helpful here. RAGs link generative models with databases or document repositories, allowing the models to find documents relevant to a user's query and use that data to construct better-informed responses.
Prompt engineering is about fine-tuning work that improves how questions are asked and improves the domain knowledge of large language models through training, while RAG focuses on improving accuracy and access to knowledge. . It also requires less user effort compared to prompt engineering and can be a cheaper, model-independent fine-tuning alternative.
Challenges in moving from pilot to production
To date, financial services companies’ generative AI efforts have primarily taken the form of internal proofs of concept or pilot projects.
Sahil Agarwal, CEO and co-founder of Enkrypt, a generative AI security platform vendor, said in an interview with TechTarget Editorial that of his startup's financial and banking customers, “more than 95% are still very inward-looking. I can say that.” “Only a handful have gone outward.”
This is partly because external projects are more risky. While internal failures can be frustrating in terms of wasted time and resources, external failures can be very public, especially for those working in highly regulated, high-stakes industries such as finance. It can be embarrassing and excessively expensive.
Several presenters and attendees referenced a February incident in which Air Canada was ordered to compensate customers who received misleading information from its chatbot. Agarwal also mentioned New York City's AI-powered chatbots, which a recent study found regularly advise businesses to break the law. “That's where you see it's not ready for production yet,” he said.
In addition to these reputational and compliance risks, building production-scale generative AI comes with practical challenges such as collecting and cleaning data and acquiring the necessary technical talent and computing infrastructure. Agarwal said there is often strong pressure from executives and leadership to implement generative AI, but the people actually implementing it struggle.
“Everyone wants to work on AI projects, but if you're an executive, how do you prioritize?” Brennan, co-founder and CEO of cybersecurity startup BLodgic・Mr. Lodge said this in his presentation. He said it is difficult to determine which projects are viable, profitable, and realistic given organizational resources.
Understand the limitations of generative AI
Experts also emphasized the importance of recognizing the limitations of generative AI. Generative AI has a notable tendency to fabricate false or misleading responses, known as hallucinations, which remains the technology's biggest problem, Chan said. And while technologies like RAG can help, they are not a panacea.
“What people think is that RAG-based systems don't create hallucinations,” Agarwal says. “They are trying to detain [the model] When it comes to vector databases and sets of documents…but this technology can still cause hallucinations. It is still possible to fabricate or mix up answers between documents. And that becomes a huge challenge for anyone trying to put these things into production. ”
Even with guardrails in place, generative AI tools can produce harmful output, including output that is biased against marginalized groups, or output that is dangerous or explicit. “Even if this technology doesn't cause hallucinations, you're getting information from the Internet, and it may not be your values,” Hoffman says.
Generative AI can also pose security and compliance issues.
Lodge noted that generative AI systems raise two intellectual property concerns. Using generative AI tools exposes your organization's IP to third parties. However, sometimes the system itself generates copyrighted content, such as when an AI coding tool like GitHub Copilot generates a snippet of its own code.
This could have costly consequences down the road as the regulatory environment around generative AI solidifies. “When I talk to security and compliance officers at these large financial institutions, their main concern — in fact, their only concern — is that auditors and regulators will fine me. ''Agarwal said. “They're taking a very cautious approach. You have to do that with any technology.”
Managing these risks is critical to the success of AI initiatives as organizations move toward real-world deployments. For enterprises, this process includes thorough preparation before embarking on an AI initiative, keeping humans up to date on the deployment of generative AI, and carefully selecting the right use cases. It will be.
“This technology can be used as a tool, but it is far from superhuman,” Chan said. “AI can’t solve every problem.”
Lev Craig covers AI and machine learning as the site editor for TechTarget Editorial's Enterprise AI site. Craig graduated from Harvard University with a BA in English and has written about enterprise IT, software development, and cybersecurity.
