AI Agents: Real or Hype?
“Even if the technology were ready tomorrow, trust is still lacking.”
The next AI wave is here—agents. Or, so companies say. But are these agents truly agents, or are they just advanced assistants? SVNB spoke to some human AI experts to understand AI enterprise agents' current state and future.
This Fall’s AI agent wave began at Dreamforce in San Francisco, where Salesforce CEO Marc Benioff introduced Agentforce. Soon after, Microsoft expanded its Copilot suite with autonomous agents, IBM launched its Granite 3.0 AI models, and DataRobot unveiled new tools for scalable enterprise AI agent development.
However, one question remains: Are today’s AI agents really agents?
“The agents we have seen so far answer questions and provide information but don’t act on their own. A true agent would complete tasks without needing constant human oversight. For instance, it might autonomously book travel or manage projects based on preset instructions,” says Eva Nahari, who recently left the venture capital firm DNX Ventures to become Chief Product Officer at AI startup Vectara in Palo Alto.
According to Nahari, most applications are still assistive, not autonomous. However, she sees some signs of assistants beginning to reach agent-level functionality
“DNX Ventures has invested in a few startups that deliver impactful agentic AI-based products. For instance, a company called Genesys Computing does agentic workflows on Snowflake. Since it operates in a limited domain and runs within customers’ own Snowflake environments, it’s a good approach to safe agentic AI,” she says.
Still, Nahari believes it will likely take a few more years before agents can reliably handle high-stakes tasks.
Yana Welinder, founder and CEO of AI SaaS startup Kraftful, says their tools qualify as an “agent solution”.
“Kraftful autonomously collects and analyzes customer feedback for companies, turning that feedback into actionable product requirements by writing PRDs and user stories,” she says.
A major challenge for the AI industry, according to Welinder, is defining appropriate use cases.
“Currently, some agent solutions on the market fall short because they attempt to deliver general-purpose capabilities. For these cases, the technology isn’t quite there yet. Foundational models will need to evolve to support general-purpose agents that can operate effectively across diverse enterprise settings,” Welinder says.
Technology Is Not Everything
Technology is one thing, but trust is another. For AI agents to be used autonomously in companies, they must be reliable every time. This requires deep knowledge of the company, the market, and its clients. While this is important for AI assistants, it is crucial for AI agents.
“Why I am so excited about Vectara is that we are building capabilities around Gen AI applications to serve enterprises’ needs for both types of applications.For example, the ability to avoid hallucinations, the ability to secure and protect data, and the ability to transparently understand where answers come from and enable auditing,”, Eva Nahari says and adds:
“These are all very hard things to implement correctly, and many startups will not be able to meet enterprise needs as they lack the right approach or experience with enterprise regulations,” she adds.
Trond Arne Undheim, a Stanford research scholar and self-described futurist focused on technology and societal change, will address this topic this week at the Post Industrial Summit Fall 2024 in Menlo Park, where he is scheduled to speak about the AI economy.
He says that the future of AI agents is not only a question of technology. For AI agents to work as enterprise tools and scale effectively, they need market support.
“The market for AI agents isn’t fully established yet,” he says, pointing out that most enterprise use cases are still experimental.
Trond Arne Undheim believes that large-scale deployment will require a new “market architecture” involving collaboration between industry and government to create safe and effective conditions for AI agents.
Undheim concludes that widespread adoption will also require public trust and regulatory clarity:
“Even if the technology were ready tomorrow, trust is still lacking.”