AI agents are a moment of truth for tech

The successful deployment of agents—artificial intelligence that can take actions and make decisions on behalf of people or even themselves—is vital to the future of model developers such as OpenAI, enterprise software companies like ServiceNow and Salesforce, consumer giants such as Apple, plus AI infrastructure companies including Nvidia, cloud service providers and data center operators. That’s not even to mention all the public and private investment capital riding on those companies’ results.

AI agents are so important because they are expected to power a new generation of products and capabilities that will drive revenue, through grabby consumer applications such as Apple Intelligence and the transformation of business functions like coding, customer service and supply-chain management.

All of that activity should in theory spur demand for ever-more computing power and storage, as well as faster networks, especially as agents become imbued with greater reasoning power, Nvidia CEO Jensen Huang said this month at the company’s annual GTC developers’ conference. AI’s next-generation reasoning capabilities will require 100 times as much computing power than seemed necessary for AI a year ago, he said.

Investors are growing trepidatious, however, as the IPO of startup CoreWeave shows. The company, which rents out access to Nvidia AI infrastructure, priced its initial public offering below expectations late Thursday at $40, and opened even lower on Friday before closing at $40.

But AI agent adoption is limited so far, according to Tom Coshow, a senior director analyst at researcher Gartner. Just 6% of 3,400 people in a recent Gartner webinar on the subject said their companies had deployed AI agents, according to Coshow.

The survey isn’t a formal market analysis, Coshow cautioned. While the group potentially was predisposed to have an interest in AI agents, its responses may have actually been tempered by Gartner’s guidance that many so-called AI agents are mere assistants. During the webinar, Gartner shared its definition of agents as AI that makes a decision and takes an action.

The leading edge

I spoke to a handful of early adopters that the tech companies have highlighted in recent months to understand for myself how AI agents are being deployed, and to what effect. Human resources giant Adecco Group, based in Zurich, has deployed Salesforce’s Agentforce service agents to a regional recruitment hub in the U.K. to help fill jobs for its employer clients.

The system recently helped Adecco fill 100% of open positions for one client, up from a more typical 70% with the same employer previously, according to Greg Shewmaker, Adecco group senior vice president for operations and AI. Adecco Group announced Thursday the launch of a new company, backed by Adecco and Salesforce, that will partner with Salesforce to help organizations build an integrated workforce of humans and AI agents.

Shewmaker said he is generally optimistic about the prospects for AI agents—as long as companies approach agents less as a tech deployment, and more as the development of digital workers that need to be onboarded and trained.

Their training of these agents, which is often customer-specific, started two months ago and is ongoing, Shewmaker said. They are trained on a combination of structured data such as CVs and job descriptions and unstructured data including transcripts from recorded conversations and reverse interviews with recruiters that add previously unavailable context to traditional static data.

Payments giant Visa worked with ServiceNow to create an agent-based dispute resolution system for card-issuing banks, credit unions and other financial technology companies. The result, ServiceNow Disputes Management, mediates situations in which a cardholder contests a charge, such as in an instance of fraud.

The dispute management system helps banks and other institutions speed up dispute resolution for their customers, cut operational costs and boost employee productivity, according to Dorit Zilbershot, group vice president of GenAI and AI experiences at ServiceNow.

She said hundreds of Visa-issuing financial institutions are now using the system, but declined to share a specific number or identify any of the banks. It can take about 12 weeks to get a bank up and running with the ServiceNow platform.

The dispute resolution agent is deeply connected into the broader market for AI infrastructure, employing Nvidia’s NeMo framework for building and deploying gen AI models, Nvidia microservices that accelerate deployment, and Nvidia’s DGX cloud-based AI supercomputing service, according to ServiceNow.

Proof of concept

But most companies exploring AI agents are in the proof-of-concept stage. The 107-year-old National Hockey League is working with data infrastructure company VAST Data to digitize and tag its archive of film and photos, including half a million recorded games.

The idea is to train agents to autonomously spot potential stories and insights, such as whether a record is about to be broken, that human broadcasters can use during a game. An agent might let announcers and producers know, for example, that a certain player is on the cusp of becoming the first Canadian in years to score three times in one period, according to the NHL’s Dave Lehanski, executive vice president for business development and innovation.

The system wouldn’t require any prompting or questions: The AI would basically watch the game and come up with interesting ideas, Lehanski said. But there are plenty of challenges to overcome first, such as teaching models that have been trained on modern-day video to work effectively on vintage footage with lower-quality images and no metadata to help identify players, for example.

Perhaps the most surprising thing of all is that in Lehanski’s view, the NHL is pretty confident that those challenges can be overcome and the system can be ready for launch in one to two years. Says Lehanski: “There’s reason for optimism."

Investors who have collectively poured hundreds of billions of dollars into AI hope he’s proven right—and soon.

Write to Steven Rosenbush at steven.rosenbush@wsj.com