This post was originally published on this site.
At Dreamforce 2024, Enterprise Times had the opportunity to talk to Jayesh Govindarajan, the EVP of Salesforce AI. Jayesh is responsible for the development of AI at Salesforce. While this is a technical role, he is able to explain the complexities of AI in a simple way that business leaders can understand.
Dreamforce was all about Agentforce, the Agentic AI system that Salesforce has developed and has now launched. Jayesh explains his view of Agentic AI and describes it as being a spectrum. Salesforce is well along that spectrum and is now moving towards multi-modality, the ability to use audio, voice, video and image rather than just text.
Agentforce, like other Agentic AI solutions, is only capable of breaking down these modes into tokens that are associated with text. It does not yet “think” in terms of images. Will AI get to the point of thinking in terms broader than text?
He answered, “It may, but that’s a big architectural change, and I know that our researchers are looking at other mechanisms to encode than just tokens, but I think there’s still more gains to be had with this architectural approach, as can be seen from the new LLMs that are coming out so pretty much every day.”
Allaying the fears about Agentic AI and how to implement it
Along with bringing increased productivity to organisations, there are also concerns amongst employees that Agentic AI will threaten jobs. The technology is well short of General AI still. Humans are needed in the loop still.
Jayesh explained, “You need humans to be able to set the guardrails. Often, the LLMs need clear instructions on what not to do when faced with ambiguity. This is very much a human judgment domain.”
Jayesh goes on to explain how Salesforce ensures that humans are included when they should be. He explains the confidence that Salesforce uses. This ensures that if an AI is unsure of the answer, it will not just deliver its best response but will flag how accurate it believes the response is. Below a certain level, the decision is shared with a human in the loop.
Jayesh also explains what organisations must consider before implementing Agentic AI and why prompt engineering is still critical. He advocates that a phased approach is used.
To hear more of what Jayesh had to say, listen to the podcast here or on your favourite platform.
Where can I get it?
You can listen to the podcast by clicking on the player below. Alternatively, click on any of the podcast services below and go to the Enterprise Times podcast page.