Agentic AI moves beyond hype—but true autonomy is still a ways off

I’m on my way back from Fortune Brainstorm Tech, where I spent the week in the mountains of Park City, Utah, with about 300 AI-focused executives—and I’m still thinking about AI agents. (I probably should be thinking about the wind in the evergreens and the first hints of fall color, but that’s life on the AI beat.)

One of the most popular panels I moderated was on AI agents: autonomous software systems that can perceive their environment, make decisions, and take actions toward goals with little or no human intervention. Full disclosure: The session was sponsored by Salesforce, which has a major stake in the space with its Agentforce platform.

But the conversation went far beyond one company’s pitch. Leaders from Zillow, Experian, and Okta joined to discuss how agentic AI is beginning to take shape inside large enterprises. Their consensus: Agents aren’t just souped-up chatbots. While it’s still early days, they represent a shift from tools that simply follow instructions to systems that can act on context—ideally within the guardrails companies set.

Shibani Ahuja, senior vice president of Enterprise IT Strategy at Salesforce, shared what I consider to be a really useful “maturity model” that helped me understand the evolution of AI agents within enterprise companies:

• Level 0: The copilots we all know—tools like ChatGPT, Claude, or Gemini that retrieve information and respond to prompts.
  
• Level 1: Agents that not only surface information but also recommend next actions (say, suggesting how to resolve a customer support ticket).
  
• Level 2: Agents that begin autonomously handling routine tasks in siloed data environments—scheduling meetings, sending follow-up emails, or pulling from internal calendars and email systems.
  
• Level 3: The real turning point—agents that can orchestrate workflows across multiple domains. Think an agent managing a sales pipeline by combining CRM data, support tickets, and financial reports to deliver a full customer picture.
  
• Level 4: Still largely aspirational—a world of “any-to-any” agent interoperability, where systems talk to one another across organizations. Imagine agents from different platforms collaborating to process orders, manage inventory, and route customer feedback in real time.

Nicholas Stevens, vice president of product for AI and home loans at Zillow, offered a concrete example of where the company is on this maturity scale—which, to my ear, sounded like Level 1 edging into Level 2. Zillow owns Follow Up Boss, one of the largest CRMs for real estate professionals, where AI is being used to lighten the load of human real estate agents juggling calls, notes, follow-ups, and property tours.

At first, Stevens explained, the tools are basic—summarizing calls or drafting a message that a human agent can edit. But over time, they begin taking on more autonomous tasks, like sending a pre-approval letter or booking a tour. The progression isn’t just about the technology, he said, but about user trust. Many real estate agents prefer to review and customize outputs early on. Yet as the system learns their tone, quirks, and even emoji preferences, they become more comfortable handing off responsibility.

Kathleen Peters, chief innovation officer at Experian, said the company already has conversational agents that recommend actions and now carry them out—for example, walking a customer through the steps to boost a credit score. On the B2B side, Experian’s agents are helping financial institutions assess risk in loan and credit card underwriting, shifting decision-making power from data science teams to product managers and even retail bankers. Looking ahead, Peters said the next step is “agents talking to agents, especially with a number of our partners like Zillow.” 

One of the things I’m particularly concerned about, though, is agent security—especially as companies start implement protocols like MCP (Model Context Protocol), a framework that standardizes how LLMs connect to and interact with external data sources, tools, and services. When I asked how many in the room had already implemented MCP in their organizations, many raised their hand. Bhawna Singh, chief technology officer of cybersecurity platform Okta, put my concern into perspective: She pointed out that before MCP, there wasn’t even a common understanding to start from regarding AI agents. While it’s early days, she said, and she acknowledged that MCP opens up a new attack surface that boosts the risk of malicious or spoofed agents, having a standard is a big step forward. Clearly, though, many enterprises won’t be able to get to Level 3 or Level 4 without more confidence in the safety and security of protocols like MCP.

Salesforce’s Ahuja agreed that the future was not about building agents that are “wild and free.” Instead, it would be about setting deterministic parameters for the things that must be consistent and reliable, while allowing adaptive reasoning where richer, more context-dependent answers are valuable.

I came away a bit surprised at how seriously these executives are taking the journey towards implementing AI agents across industries. They seemed realistic yet optimistic, even though there is plenty of hype out there–just six months ago, my former colleague Sage Lazarro wrote in Eye on AI that some AI agent customers said the reality did not match the hype. Maybe it’s just like some of the young trees that were swaying outside my hotel window in Park City: growing fast, not yet fully mature, but still well worth watching.

With that, here’s the rest of the AI news.

Sharon Goldman
sharon.goldman@fortune.com
@sharongoldman

FORTUNE ON AI

Cohere looks to shed its underdog status with a star AI hire, new CFO, and $7 billion valuation—chasing ‘ROI over AGI’ —by Sharon Goldman 

For OpenAI’s chief revenue officer, research chops rather than ‘big marketing efforts’ is the key to winning customers —by Sharon Goldman

AI for chemistry startup CuspAI raises $100 million in new venture capital funding —by Jeremy Kahn

AI IN THE NEWS

Microsoft to buy AI from Anthropic in partial shift from OpenAI. Microsoft is breaking from its deep reliance on OpenAI by bringing Anthropic’s Claude models into Office 365, according to The Information, marking its biggest move yet to diversify AI providers. In the coming weeks, the company will announce that Office Copilot features in Word, Excel, PowerPoint, and Outlook will be powered by a mix of Anthropic and OpenAI technology. Executives say Anthropic’s models, hosted on AWS, outperformed OpenAI’s in key areas like generating financial functions in spreadsheets and producing more polished PowerPoint decks. While the shift could be seen as a bargaining tactic amid tense negotiations with OpenAI over its restructuring and IPO ambitions, insiders emphasize that Microsoft is simply going with the best-performing models. 

Oracle shares skyrocket on announcement of massive AI deals. The Wall Street Journal reported Oracle shares soared nearly 40% after the company disclosed a string of multibillion-dollar AI-related contracts, signaling far greater traction in the cloud infrastructure race than expected. The software giant reported $455 billion in outstanding contract revenue for the quarter ending Aug. 31—more than four times last year’s level—and said additional mega-deals could soon push that figure above $500 billion. CEO Safra Catz called it an “astonishing quarter,” citing demand for Oracle Cloud Infrastructure from OpenAI, Meta, and Elon Musk’s xAI, among others. Analysts noted the backlog, which includes the $500 billion Stargate supercomputer venture with SoftBank and OpenAI, makes Oracle’s long-term revenue targets highly achievable. Once seen as a cloud laggard, Oracle is now positioning itself as a central player in supplying infrastructure for the AI boom, aided by Larry Ellison’s close ties with tech leaders like Elon Musk and Nvidia’s Jensen Huang.

AI is coming for YouTube creators. According to reporting from The Atlantic, YouTube creators like Jon Peters, who built an audience of over a million subscribers by sharing woodworking tutorials, now face the threat of obsolescence as their videos are being quietly scraped to train AI. Researchers have identified more than 15 million YouTube videos in AI training datasets—often without creators’ knowledge or permission—including instructional, news, and educational content. Tech companies such as Microsoft, Meta, Amazon, and Nvidia have used these datasets to develop generative video tools that can churn out custom clips, raising unresolved legal questions about copyright and fair use. While companies argue the practice is lawful, critics warn it could undermine creators’ livelihoods as AI-generated videos increasingly compete with and even overshadow authentic human-made work.

AI CALENDAR

Oct. 6-10: World AI Week, Amsterdam

Oct. 21-22: TedAI San Francisco. Apply to attend here.

Dec. 2-7: NeurIPS, San Diego

Dec. 8-9: Fortune Brainstorm AI San Francisco. Apply to attend here.

EYE ON AI NUMBERS

116%

That's how much consumer spending on AI tools has surged year-over-year in the first half of 2025, according to a new report from intelligence firm Consumer Edge.

OpenAI leads the pack, grabbing 40% of the market as subscriptions to ChatGPT Pro jumped more than 1,000% over the past year. But some upstarts are gaining ground: Perplexity’s AI search tool is growing fast, while Elon Musk’s xAI has already reached two-thirds of its scale.

The biggest shift is in video, where tools like Runway and Invideo are making it easy for anyone to create explainer clips or even cinematic scenes, fueling more than 50% growth in the category. AI is also finding a home in the workplace: Apps like Otter.ai and Fireflies.ai are becoming everyday meeting companions, while health care and engineering tools are expanding at breakneck pace.

The report says the takeaway is clear—consumers aren’t just experimenting anymore. AI is becoming a staple in how people work, create, and spend.