(Design: Lily Yeh | Image Credit: Just_Super/ Getty Image/ Nvidia/chatGPT/Microsoft)
What if an AI could manage your calendar, or even your small business, without your constant direction? That's the promise of Big Tech's latest obsession, agentic AI. Though these "AI agents" are still in the early stages, the general idea is to move from a transactional relationship with an AI (user prompts, computer answers) to the AI being more autonomous. It can perform actions on your behalf with minimal or no input.
Self-driving cars are a great example of agentic AI, and one of the most advanced. These vehicles scan their surroundings using cameras and sensors, using the information to make driving decisions independently. Now, companies are looking to apply those same principles to the everyday tasks we do on our phones and computers.
Generative AI vs. Agentic AI: More Than a Chatbot
Whereas a basic chatbot conversation consists of a written back-and-forth, agentic AI works differently. The user gives the system a written or verbal prompt, telling it what to do. Then, the AI finishes the task, or agrees to complete it regularly over a specified timeframe.
One good introductory example of agentic AI is OpenAI's Tasks, available for ChatGPT Plus, Pro, and Team subscribers. The chatbot can now perform simple, routine actions without human oversight, like checking the stock market or delivering a news summary every morning (better hope it doesn't hallucinate).
Seems straightforward enough, right? But it doesn't always work as expected, based on our testing. We asked ChatGPT Plus to check the stock market every hour and send an update on how the major indexes moved. It did, which was impressive, but it continued to send annoying texts and emails on weekends, when the market is closed. We asked it to only send updates when the market is open, and it replied, "Ok, I'll do that," but it did not adjust its behavior. Then, we asked it to stop the task altogether; it confirmed it would but kept going anyway until we manually toggled it off in the Settings menu.
Coming Soon: Agentic AI From Amazon, Apple
Amazon's vision for the new Alexa+ is for it to be a well-rounded personal assistant. It's rolling out now with just a few features to start, such as calling an Uber or putting a meeting on your calendar invite. In the future, it could take that a step further by checking to see if a meeting conflicts with your schedule and perhaps suggesting another time when it knows you'll likely be on your computer already.
Apple has similar visions for the revamped version of Siri with "on-screen awareness," meaning the ability to take actions for you in your apps. It's already proving difficult to accomplish, leading to yearslong delays and a major executive shakeup. (Alexa+ has also experienced delays.)
These systems need a ton of data to succeed. In the case of Alexa, that's gathered from what you tell it, what it observes in the device's cameras, files you upload, and much more. The more data, the easier it will be to achieve the goal of making independent decisions along the way, IBM says. That's why many AI companies, such as Anthropic, OpenAI, and DeepSeek, are launching "reasoning models." They can process the data more like an open-minded human brain than a fixed algorithm.
Can AI Do Your Job? Businesses Love Agentic
In addition to these consumer tech examples, businesses are particularly interested in AI agents, which can boost productivity and reduce their payroll if they work properly.
Nvidia debuted its AI agents for businesses early this year, calling them "knowledge robots" that can "reason, plan, and take action." They are essentially white collar robots. In March, Nvidia introduced new AI models that can perform complex coding and logic-based tasks independently. This tech could automate a portion of a software developer's job or perhaps replace entry-level roles as we define them today.
(Credit: Nvidia)Microsoft is also leaning heavily into agentic AI products for businesses. This year, the company launched a suite of 10 AI agents, each "fine-tuned," or customized, for specific job functions. "You can build a very rich agentic world defined by this tapestry of AI agents, that can act on our behalf across our working life, across teams, business processes as well as organizations," CEO Satya Nadella told Ignite attendees.
Later, when debuting Microsoft 365 Copilot Chat, Nadella said the tech "gives each of us more agency and more leverage while freeing us from the drudgery of repetitive tasks."
Amazon released its Nova AI agent creation platform in March, pitching it as a way for "developers and tech enthusiasts" to build their own bots and explore the technology. "Today, such agents are still in an early stage," Amazon said.
The healthcare industry could be among the first to experience rapid change thanks to AI agents. Apple, for instance, is working on an "AI doctor" app, which will analyze data from Apple Watches and fitness devices to provide health recommendations. In January, Nvidia released its agentic AI healthcare models, which aim to "boost the efficiency of clinical trials and optimize planning for the launch of therapies and medical devices—ultimately improving patient outcomes." That's far more complex than early AI applications we've seen in healthcare, like transcribing patient notes and reading X-rays (sometimes with poor outcomes).
The Risks of White Collar Robots
Are we comfortable giving AIs this much control? We have reasons to be both hopeful and skeptical. If the system works well, it could save time and be immensely helpful. But if it suffers from hallucinations or performs an action incorrectly, the fallout could undermine any benefits.
Like humans, AIs can suffer from behavioral problems. One study found that large language models built by OpenAI, Anthropic, and Meta were more prone to escalating violence in a war video game simulation than a human player. (These are real risks as China and the US race to master AI warfare.)
While you may be more likely to witness AI wreaking havoc on your computer than the battlefield, the more we rely on these systems, the more critical it becomes to ensure your new coworker performs as expected.