2025 was the year of agentic AI. How did we do?

With nearly 16 million new vehicles sold annually in the U.S., this use case was right in the sweet spot of where Capital One focuses its agentic AI efforts. “We want to start off at the low end of the risk spectrum, but also find use cases with impact and enough complexity that we can learn from it,” says Prem Natarajan, head of enterprise AI at Capital One. He says that Chat Concierge has been embraced by dealers because it has dramatically increased customer engagement and is 55% more successful in converting leads into buyers.

But like most deployments of agentic AI, Chat Concierge did require a bit of massaging after it launched. Capital One kept a close eye on latency, which is the time delay between an AI system receiving an input and then generating a response. Since launch, Natarajan says, Capital One has reduced latency fivefold, an improvement he attributes to increased engagement and the company’s decision to build its own proprietary multi-agentic workflow. “We can really tune these things for latency, because we build our own stack,” adds Natarajan.

Capital One is among the many enterprises that have embraced agentic AI in 2025, the year when Sam Altman, CEO of AI startup OpenAI, said these systems would finally “join the workforce.” AI agents differ from chatbots like OpenAI’s ChatGPT and Google’s Gemini in that they are designed to tackle more complex tasks, have greater autonomy, and frequently require deeper integration with enterprise data to do so.

“The sense of an agent comes from the word ‘agency,’” says Natarajan. “These agents have a bit of agency.”

By 2028, 15% of day-to-day work decisions could be performed by AI agents, research firm Gartner has predicted, and a third of all enterprise software applications are expected to include agentic AI. But third-party studies published throughout 2025 have consistently shown that many organizations remain stuck in pilot mode. While 30% of organizations surveyed by consultancy Deloitte reported they are exploring agentic AI, only 11% say they actively use those systems in production.

Building the agent is only step one

Lari Hämäläinen, a senior partner at McKinsey, says 88% of the companies that the consultancy studies are using AI in at least one business function. But many of these AI tools are siloed, with only 20% working cross-functionally. “The major reason is just that it’s very complex,” says Hämäläinen.

Agentic AI requires not only a lot of effort to actually build the agents, but also that organizations establish new governance policies and make org structure decisions about who will lead implementation and oversee these agentic systems during the pilot and through the full production phases. “We need a bit of patience before we start to see enterprise adoption at scale,” says Hämäläinen.

Some larger enterprises say they are showing meaningful progress. Athina Kanioura, chief strategy and transformation officer at food and beverage giant PepsiCo, says she’s focused agentic deployments on three areas: the technology ecosystem, which includes data and software engineering; customer service; and ways to improve the overall employee experience. 

What Kanioura has been measuring is how much productivity PepsiCo can extract from agentic AI, at what cost, and also how customers and employees interact with these systems. “We found very quickly that both internal and external, they were keen to interact with the agents,” says Kanioura.

AI agents have been deployed to assist with software testing, and they have not only sped up the validation cycle, but have also identified some technical gaps that wouldn’t have been caught had humans worked alone. “It was better than expected in the pipeline,” says Kanioura. “So that gives us a lot of confidence.”

One especially bullish proponent of AI agents has been cloud-based software giant Salesforce, which debuted the Agentforce platform last year to help business clients including Reddit, Pfizer, and OpenTable build and launch AI agents across customer service, marketing, sales, and other functions. Salesforce has closed 18,000 Agentforce deals since the product launched in October 2024, and higher sales for the product helped the company raise its full-year guidance earlier this month, though investors have put pressure on Salesforce’s stock as they had hoped adoption would be faster.

“We believe customers are going to take the step from dabbling and experimenting to really start to push these agents to scale,” says Madhav Thattai, chief operating officer of Agentforce. He lauded the October introduction of Agentforce 360 this fall at the Salesforce Dreamforce conference, where the company promoted new features it says will make it easier to build, control, and deploy agentic AI. 

‘We had a lot to learn’

Salesforce has been on its own iterative journey when deploying the new tech, launching an agentic experience for the company’s support website, but quickly discovering that “we had a lot we needed to learn,” says Thattai. Customers were asking questions the agent couldn’t answer, and some responses were taking too long. Salesforce has made improvements to speed up the responsiveness of this agent, which has now engaged in more than 2 million interactions with customers.

“Once you start to get to scale, you want to monitor the agent’s performance, analyze what the agent is doing, and you want to continuously improve the agent,” says Thattai.

Commercial real estate and investment management firm JLL has 34 agents in discovery and development, including an add-on AI feature to the Prism property management platform that enables some autonomous tasks. Property managers can rely on the agentic tool to address tenant issues, including automatically adjusting the temperature of a building after a tenant complaint.

Yao Morin, chief technology officer at JLL, says AI agents are also supporting software development, resulting in more work for humans to manage these systems rather than write code themselves.

Her role as CTO, Morin says, is to think through, “How do we make sure that there are control checks and balances between humans and agents, and how do you manage those agents in an effective way so that they don’t go rogue?”