Despite the hype around AI in recent years, the technology’s disruptive impact has been fairly modest. Experts say that’s likely to change next year as AI agents force their way into all aspects of our lives.
Since the surprise success of ChatGPT in late 2022, billions of dollars have poured into the AI industry as startups and big tech firms try to capitalize on the unquestionable promise of the technology.
But while hundreds of millions of people around the world are now regularly using AI chatbots, putting them to productive use is proving harder. Recent research from Boston Consulting Group found that just 26 percent of companies who have experimented with AI have moved past proof of concept to get real value out of the technology.
That could be because current iterations of the technology are, at best, a kind of copilot. They can help users accomplish some tasks more efficiently, but only with close supervision and the ever-present risk of mistakes. The situation could be about to change though, according to leading voices in the AI industry, who say that autonomous AI agents are poised to have a breakout year in 2025.
“For the first time, technology isn’t just offering tools for humans to do work,” Salesforce CEO Marc Benioff recently wrote in Time, a publication he owns. “It’s providing intelligent, scalable digital labor that performs tasks autonomously. Instead of waiting for human input, agents can analyze information, make decisions, and take action independently, adapting and learning as they go.”
At the core of all AI agents is the same kind of large language model (LLM) that powers services like ChatGPT. This makes it possible for humans to interact with agents via language, but the algorithm is also a “reasoning engine” that comes up with a step-by-step plan to tackle tasks.
Agents also typically have access to external data sources relevant to their application—for instance customer databases or financial records—and software tools they can use to achieve goals.
At present, the reasoning capabilities of LLMs are limited, which restricts where agents can be deployed. But with the advent of models like OpenAI’s o1 and DeepSeek’s R1, which are specialist reasoning models, there’s hope that agents could soon become much more capable.
Major players are investing heavily in that promise.
In October, Microsoft unveiled Copilot Studio, which allows companies to build customized agents capable of tasks like handling client queries and identifying sales leads. The same month, Salesforce rolled out its Agentforce platform, which also allows customers to create their own bots. And last month, Benioff told TechCrunch his goal is to have one billion agents deployed within a year.
Leading AI research labs are also increasingly focused on agents. Anthropic recently previewed a version of its Claude 3.5 Sonnet model that could take control of a user’s computer, and Google’s recently announced Gemini 2 has been trained to perform similar tasks. OpenAI also has plans to unveil an agent codenamed “Operator” early in the new year.
Startups are looking to get in on the action too. According to Pitchbook, the number of funding deals for agent-focused ventures was up more than 80 percent by September compared to the previous year. The median deal value was also up nearly 50 percent.
But there is some skepticism around how quickly agents are likely to burst onto the scene. As The Verge notes, AI companies have been ploughing billions into research and development with little revenue to show for it and are still searching for a killer app that justifies their sky-high valuations. Practical considerations could mean progress is slower than they hope.
For a start, these models are still prone to “hallucinations” where they generate incorrect or misleading responses to queries. This is problematic enough in a chatbot but much more concerning when it’s an agent capable of independent action.
Quartz notes this risk can create considerable overhead as companies have to implement many layers of security designed to catch mistakes. This could become incredibly complex as the number of agents increases and require investment in new platforms and even “guardian agents” to monitor their activities.
Agents can also be expensive because “reasoning” through problems requires they make multiple calls to the underlying LLM. This quickly adds up, either in terms of dollars spent with an LLM-provider or energy burned for companies that host their own models.
Nonetheless, many in the industry expect 2025 will be a turning point in deployment.
“I think 2025 is going to be the year that agentic systems finally hit the mainstream,” OpenAI’s new chief product officer, Kevin Weil, said at a press event ahead of the company’s annual Dev Day, according to The Verge.
Deloitte’s Global 2025 Predictions Report forecasts that of the companies already using generative AI, a quarter will launch pilots or proofs of concept with AI agents, growing to half by 2027. And the second half of the year could see full adoption of agents in some workflows.
Others are more bullish. Konstantine Buhler of Sequoia Capital told Bloomberg that 2025 will see the emergence of networks or “swarms” of AI agents working together within businesses. Kari Briski, vice president of generative AI software at Nvidia, agrees and thinks this will necessitate the emergence of AI orchestrators—essentially AI managers that oversee and coordinate numerous agents.
No matter who’s right, it seems certain that agents will be the major preoccupation of the AI industry in 2025. If it pays off, the world of work could look very different by the end of the year.
Image Credit: Gabriella Clare Marino on Unsplash
