There’s plenty of buzz around AI agents, and many believe that they could fundamentally alter many workflows, but it could be a while before these agents are actually deployed in the real world.
Aravind Srinivas, CEO of Perplexity AI, recently offered a contrarian perspective amidst the hype surrounding AI agents. He outlined the challenges in building truly functional AI agents, particularly highlighting the limitations posed by operating system ecosystems like iOS and suggesting a more measured timeline for their widespread adoption. His comments also touched upon the potential market for such agents, comparing their value proposition to that of human executive assistants.
“We’re going to be building agents that let you book stuff, buy stuff, make travel plans, book hotels,” Srinivas explained, talking about Perplexity’s plans. “Kind of personalized to you – which hotels you like staying in, what kind of rooms you prefer, you like a certain view – and essentially build like a personal assistant or an executive assistant for you.”
He acknowledged that human EAs (Executive Assistants) will still have their place. “Now you still have your human EA, definitely, right?” Srinivas continued. “But there are some people who cannot afford one and this will be a great 80/20 for them. That’s that’s the idea. People pay their EAs personally like $100-200k dollars an hour.” He then positioned AI agents as a more affordable alternative, stating, “and so if that’s mainly available for 80% of the value at 10 to 100x lower price that’s a pretty good business to build.”
Srinivas linked the success of these agents to the broader vision of Perplexity AI. “We hope doing that will let us make our core answer engine product, this question answering machine that you have, as widely accessible and freely available as possible. So it’s a long-term vision,” he stated. Crucially, he cautioned against overly optimistic predictions: “It’s not going to work this year. It’s still… you know, anyone who’s saying agents will work in 2025 should be skeptical of it.”
He concluded by outlining the immediate focus: “But we are trying to figure out at least what are the simple use cases that we can make work, like buying stuff without having to, you know, go and enter your credit card details everywhere, your shipping information everywhere. So simple workflows like these,” he said.
It’s not every day that someone in the AI space says that it could take longer than expected to reach a breakthrough. In particular, timelines for AGI have become more aggressive over time, with OpenAI saying AGI could arrive before 2027, and Softbank’s Masayoshi Son saying it could arrive much sooner than that. But Aravind Srinivas has hinted that it could be a while before AI agents become a reality. AI agents do have some thorny issues to navigate. There could be issues around security with agents having access to users’ credit card and payment details, and there could be issues in how they navigate the web — it’s possible that some websites will not allow AI agents to operate in the same way as human users. And with an AI leader themselves saying that AI agents won’t necessarily be able to take on their final form this year, there might be some merit in tempering the hype that currently exists over agentic workflows.