Earlier this month, our co-founder and CEO Adam Wenchel sat down with Noah Kravitz of the NVIDIA AI Podcast for an insightful discussion about the adoption of generative AI in the enterprise, guardrails on generative AI systems, bias and observability in AI systems, Arthur’s work in this space, and more. We’ve pulled out the highlights and main themes for this blog post, but if you’re interested, you can listen to the full podcast below.

‍

‍

The Current State of Generative AI in the Enterprise

While not every organization is at the tip of the spear when it comes to generative AI, Adam noted that even a lot of traditional organizations have spent the last four or five years trying to build up their AI capabilities in general—adding talent, putting infrastructure in place, getting data ready, and more. 

“It’s been a slow rise up the maturity curve for the last four or five years,” said Adam, “but generative AI has definitely caused an inflection in terms of the adoption.” This is partially due to the last few years of investment from these companies, and partially due to the omnipresence and broad accessibility of generative AI. “Everyone sees the possibilities with generative AI, so there’s a real sense of urgency—and [generative AI strategy] has become a board-level issue,” Adam added. 

“Everyone sees the possibilities with generative AI, so there’s a real sense of urgency—and [generative AI strategy] has become a board-level issue.”

This sense of urgency is motivated largely by efficiency, but surprisingly, has little to do with cost-cutting. With generative AI allowing organizations to increase efficiency at more tedious tasks, companies are simply reapplying those costs and resources to more strategic work and top-line growth.

‍

Enterprise Use Cases of Generative AI

Adam shared that many companies are starting with internal use cases of generative AI as a way to ensure they understand the technology and know how to deploy it responsibly before they put it in front of their customers and partners. 

One of these internal use cases is in the realm of human resources—for example, answering questions about benefits. Using generative AI for a task like this not only boosts productivity, but it also allows for better answers and increases team satisfaction as well. Legal and investment information are a few other common use cases that allow companies to reallocate resources to more strategic work.

For all of these different enterprise applications of generative AI, Adam says, “the price of an incorrect answer is quite high. If you give someone bad information about their benefits or if you make an investment based on bad advice … even if it’s right 95% of the time, 5% is not great.” 

“The price of an incorrect answer [from an LLM] is quite high. If you give someone bad information about their benefits or if you make an investment based on bad advice … even if it’s right 95% of the time, 5% is not great.”

‍

Concerns & Challenges with Generative AI Applications

As mentioned above, hallucinations (i.e. incorrect or misleading results) from AI applications are a universal concern for organizations deploying these systems. “Wrong answers are bad for everyone,” said Adam. “If you have an online cocktail recipe generator, you can probably live with a bad drink. But especially if you’re in any sort of business application, whether it’s HR information or investment advice or legal advice, you can’t live with wrong answers.”

It doesn’t end there, though. “When you go to deploy AI systems based on elements in the real world,” noted Adam, “you quickly discover there are some risks and some challenges that need to be managed.” Things like prompt injections, toxicity, and data leakage are other major concerns when it comes to inappropriate uses and potential failures of these systems. 

“When you go to deploy AI systems based on elements in the real world, you quickly discover there are some risks and some challenges that need to be managed.”

‍

Arthur’s flagship LLM product, Arthur Shield, was built to address exactly these concerns. It acts as a firewall for AI and LLMs, allowing organizations to apply policies around the usage of large language models that protect both companies and end users from harm. In today’s ever-changing technology landscape, the organizations that are able to deploy generative AI quickly and safely will ultimately be the ones that will succeed.

Adam and Noah also covered a number of other topics in the podcast, such as bias in AI, observability, and the future of work. To listen to the podcast in its entirety, click here.