Rafael Henrique | Lightrocket | Getty Images
Earlier this year, financial services company Klarna said its artificial intelligence agent, powered by OpenAI, had taken over two-thirds of customer chats and was doing work equivalent to that of 700 full-time agents. That was after just one month of use.
Alexander Kvamme, CEO of customer engagement startup Echo AI, told CNBC that Klarna’s announcement in February may have been the first sign of AI agents “having their ChatGPT moment.”
OpenAI released the ChatGPT chatbot to the public in late 2022, giving the public a taste of how new generative AI chatbots could provide much more thorough, creative and conversational answers to web queries compared with traditional search, which is how consumers sought online information for the prior 25 years. Google, Microsoft and others followed with rival products.
The industry quickly moved past text responses and into AI-generated photos and videos. Now comes the rise of AI agents.
Rather than just providing answers — the realm of chatbots and image generators — agents are built for productivity and to complete tasks. They’re AI tools that are able to make decisions, for better or worse, “without a human in the loop,” Kvamme said.
Grace Isford, a partner at venture firm Lux Capital, said there’s been a “dramatic increase” in interest among tech investors when it comes to startups focused on building AI agents. They’ve collectively raised hundreds of millions of dollars and seen their valuations climb alongside the broader generative AI market.
Generative AI exploded in 2023, with $29.1 billion invested across nearly 700 deals, a more than 260% increase in deal value from a year earlier, according to PitchBook. Meanwhile, the non-AI investing landscape has been in an extended lull for well over two years following record financings during the Covid pandemic.
If 2023 was the year of peak AI hype, 2024 is the year of early deployments.
“It has really been a torrent of innovation that has hit the market since the introduction of ChatGPT,” Jared Spataro, Microsoft’s corporate vice president of AI at Work, told CNBC. Microsoft is the biggest backer of OpenAI and has invested billions of dollars on its own generative AI models and products, in addition to the billions it’s poured into the ChatGPT developer.
The term AI agents isn’t neatly defined across the tech sector. Industry experts who spoke to CNBC about the emerging trend generally viewed agents as a step beyond chatbots, in that they’re typically designed for specific business functions and can be customized on the big AI models. Think of J.A.R.V.I.S., Tony Stark’s multifaceted AI assistant from the Marvel Universe.
AI agents are often described as advanced generative AI tools that can do multistep, complex tasks on a user’s behalf and generate their own to-do lists, so that users don’t have to walk them through the process step-by-step.
“An assistant is not just giving you the answer, but automating a series of steps,” said Francois Ajenstat, chief product officer at digital analytics company Amplitude.
How Microsoft and Google are playing
Microsoft CEO Satya Nadella said on an earnings call earlier this year that he wants to offer an AI agent that can complete more and more tasks on a user’s behalf, though there is “a lot of execution ahead.” Executives from Meta and Google have also touted their work in pushing AI assistants to become increasingly productive.
At Google I/O in May, Google announced Project Astra, the company’s latest advancement toward its AI assistant that’s being built by Google’s DeepMind AI unit.
In Google’s demo video, the assistant, using video and audio, was able to help the user remember where they left their glasses, review code and answer questions about an object that it was shown. It’s just a prototype for now, but Alphabet CEO Sundar Pichai said he hopes to roll it out to users later this year.
The demo came a day after OpenAI showcased a similar audio back-and-forth conversation with ChatGPT, positioning it more as an AI assistant that can function as a conversationalist, language translator, math tutor and co-writer of code.
Microsoft followed at its Build developer conference by announcing a partnership with Cognition AI, which will bring Cognition’s own AI agent, called Devin, to customers. Cognition bills Devin as the “first AI software engineer.”
Devin quickly caused a stir on social media for its ability to handle multistep processes. Instead of just generating simple lines of code, Devin creates a problem-solving process, writes the code, tests it and then ships it.
Martin Kon, operating chief of enterprise AI startup Cohere, said AI agents could start doing work such as booking a plane ticket and expensing it, offering a suggested interest rate on a loan, or emailing a customer about arrival time and updating Salesforce accordingly.
To date, the tools have largely been limited to tasks such as helping write code. At Microsoft’s GitHub, for example, roughly 46% of all code “across all programming languages” was AI-generated, CEO Thomas Dohmke wrote in a blog post in early 2023.
While the line between an AI coding tool and a true AI agent is blurry, most experts who spoke with CNBC said the defining characteristic of an agent is that it goes well beyond a single use case and starts to approach an all-capable personal assistant.
Anthropic and other startups are already working toward that goal. The first step is giving their chatbots the ability to interact with external tools and services on behalf of the customer.
Microsoft’s Spataro said the process of developing his company’s Copilot coding agent has “kind of been like being strapped to a rocketship.” A big part of what Microsoft is doing, he said, is moving from one- or two-step tasks to multistep tasks. That could involve looking at a user’s calendar and giving a 30-second outlook on what to prioritize for the day.
Fred Havemeyer, head of U.S. AI and software research at Macquarie, wrote in a recent note to investors that the firm is looking forward to seeing more AI agents.
“We think agentic AI, which can self-direct towards achieving tasks, will be the tools that unlock the value of GenAI for everyday users,” Havemeyer wrote.
Romain Huet, OpenAI’s head of developer experience, told CNBC that the concept of AI agents came into focus last year, but people quickly realized there was work to be done to make the tools more autonomous.
“We have the models that become more and more powerful, so we can now capture user intent much better than before, but we’re also still pretty early on that journey at building agents,” Huet said.
The big advancement, he said, will be when an AI agent can know your preferences and “take action on your behalf” without you asking.
Startups raise big money
AI agent startups are reeling in hefty piles of cash from investors. They’re not the billion-dollar-plus financings that have been going into the AI model companies, but valuations are still far ahead of business fundamentals.
Adept, which is led by alumni of OpenAI and Google, received a valuation of over $1 billion last year. The company says on its website that its technology “navigates the complexity of software tools so you don’t have to.”
H, a French AI agent startup, raised a $220 million seed round in May from investors including Amazon, Samsung, UiPath and Google ex-CEO Eric Schmidt. Artisan AI, a Y Combinator-backed startup working on AI agents that it bills as “AI employees for enterprise,” recently completed a $7.3 million seed round and says it’s onboarded more than 100 companies so far.
Artisan AI founder and CEO Jaspar Carmichael-Jack said it wasn’t possible to begin working on true AI agents until 2022 because that’s when chatbots such as ChatGPT first made it possible for the average consumer to interact with such tools.
“People talk about how the VC market is down in general,” Carmichael-Jack said. “But for us it’s like 2021 in AI startups.”
Braden Hancock worked at Facebook Research and Stanford’s Artificial Intelligence Lab before co-founding Snorkel AI in 2019. He said the market is in a “similar hype cycle” to that of self-driving cars. And broader AI agents will similarly take a long time to hit the mainstream, he said.
Hancock said agents must be “many times” better before people are “willing to accept putting something on autopilot.” He added that, when it comes to having technology sign your name and make money transfers on your behalf, “there’s a really high bar.”
Kanjun Qiu’s three-year-old startup, Imbue, has been valued at more than $1 billion, with backing from Amazon’s Alexa Fund and Eric Schmidt. Based on the company’s own user research, Qiu said the current characterization of AI agents — as generally intelligent personal assistants that handle delegated tasks — is not what users actually want, since, by design, they’re “not fully trustworthy.”
“Even as CEO, it’s hard for me to delegate things to my executive assistant,” Qiu said. “I’ve had her for two years, and she’s amazing.” For new things, Qiu said, “It’s still hard for me to fully know, ‘Okay, is this going to come back the way I expected?'”
Imbue is developing ways for people to make their own AI software agents — without coding — to run in the background for their personalized needs, whether it’s creating a way to track the news or building a bot to book travel. These types of AI models wouldn’t need to train on user data, since each use case would be personalized.
Instead of delegating tasks to an agent built by the likes of OpenAI or Google, which would be centralized and controlled by those companies, Imbue imagines agents putting control in the hands of users.
“There’s a way of thinking about agents as enabling every person to make software,” Qiu said. The user is “asking the agent to write code on the computer, to make the computer do what I want to do.”
