“We’re saying, ‘Look, we’re not going to show up and ask you to explain your process to me. You can do that. We have faith that you know what your process is, but we’re going to show up and use AI to actually discover your process,’” says Naveen Sharma, Cognizant’s global head of AI and analytics.
Global services provider Cognizant is laying the groundwork for its bid to be a leader in the burgeoning agentic AI market with the release of its new Cognizant Agent Foundry.
The introduction of the Cognizant Agent Foundry comes as the IT industry moves past the hype around ChatGPT and the use of AI to increase the efficiency and optimization of software coding, said Naveen Sharma (pictured), global head of AI and analytics for the Teaneck, N.J.-based company
The next phase is agentic AI, which comes from employing AI agents as “digital workers,” Sharma told CRN.
[Related: Cognizant Ties AI Agents From Multiple Apps For Scale, Autonomous Operations]
“The best phrasing I’ve heard used to describe it is a ‘digital worker,’ someone that is not replacing a human worker but is supplementing a human worker in terms of productivity,” he said. “And what’s different about this generation of digital workers is they can communicate with us using natural language. They have memory, they have context, they have the ability to integrate with other systems and other agents. When you abstract it far enough, it starts to look very much like how humans cooperate and collaborate and engage in teams and produce meaningful things together.”
Cognizant said it believes that simple work processes are increasingly being codified with digital workers being engaged, but they’re rarely uniformly understood by businesses, Sharma said. That, he said, led to the development of the Cognizant Agent Foundry.
“The notion of the Cognizant Agent Foundry is simple,” he said. “We’ll show up much like any consulting engagement. You come in, you assess, you make a recommendation, you either build, or you take something off the shelf, you configure it, and then you operate it. That part of the business isn’t changing.”
What is changing, Sharma said, is how it’s executed. For instance, in the assess phase, people in nearly any business will say they have documented processes, but the minute you dig into them the nuances of the processes start to really become interesting and intricate.
“We’re saying, ‘Look, we’re not going to show up and ask you to explain your process to me. You can do that. We have faith that you know what your process is, but we’re going to show up and use AI to actually discover your process,’” he said.
Cognizant is doing that through a couple of different methods, including working with partner companies doing some cutting-edge work in observing how work really gets done, Sharma said.
“This is not ‘big brother,’” he said. “This is just making sure there is transparency on tracking what the human does when they interact with a system or systems. We use AI to capture the process.”
The other way of doing this for businesses not comfortable with that approach is to have AI observe and record workflow processes, Sharma said.
“So opposed to having someone sitting over your shoulder taking notes, which is creepy, we’ve picked a slightly less creepy method which is we’ll just record video of what a worker does,” he said. “This is not hidden. There’s a little camera watching you ‘swivel chair’ between systems, computers and monitors, talking to people, all of that. Then we use AI to extract what is happening, to capture the processes that are to be supplemented by AI.”
After the assessment stage, Cognizant Agent Foundry then looks at what workflow processes can be automated, Sharma said.
“Maybe you’re an insurance company paying claims for a car crash,” he said. “Let’s say there are 15 steps in that process. Not every step is a candidate for agentification. Some will remain as is. Some absolutely require humans. Some are candidates for process automation, maybe robotic process automation. And some are candidates for agentification. So we create that ‘to be’ state as opposed to the ‘as is’ state.”
That “to be” state doesn’t mean going away to build the agents and coming back six months later with the results, Sharma said.
“We will have a very large menu card of options, essentially a library of prebuilt digital labor or agents that we will then configure and contextualize for whatever the client needs are,” he said. “We’re not just saying, Agent 1, Agent 2, Agent 3. We’re building a layer that sits in the middle that orchestrates what Agent 1 does and what Agent 2 does and when do you go to Agent 3 and when do you raise a digital flag to say some human needs to intervene or the process is completed.”
The last piece is to look at the results, Sharma said.
“Now that you’ve built this, how do you run this?” he said. “How do you govern this? How do you keep an eye on the costs? How do you make sure that these are all aligned to expected behavior in terms of not just AI and agents, but also the enterprise itself? And how do you make sure you improve this on a going-forward basis?”
Sharma said Cognizant already has multiple agents up and running, some of which were built with the company’s own tools and some with support from partners including agent builder platform Writer, as well as with the top three hyperscalers.
He also said the company’s agent library currently has a dozen or so agents up and running, with more on the way. Some will be horizontal, which work across every industry, and others will be targeted at specific verticals.
“I don’t know that we’ll ever be at a state where we have agents for every scenario under the sun because that will just be ridiculous,” he said.
The agent library has been up and running for about three months, and Cognizant already has two clients working with it, Sharma said
“It’s not just a science project in a conference room,” he said. “It’s something that we’re seeing working well with clients.”
Cognizant is also working on deploying the Cognizant Agent Foundry internally and has prioritized its own legal, finance and human resources functions for agentification, Sharma said. He said the company has also tested agent-to-agent collaboration on its own intranet.
“We built a series of agents, and we tested out a seemingly simple scenario where an employee is going to get married, looking at what needs to change and whether the system can you make those changes in all the various Cognizant systems,” he said. “We’ve run tests for up to 10,000 agents being orchestrated and managed using our own IP, and it seems to scale and run well.”
As Cognizant develops new agents for clients, many will go into the Cognizant Agent Foundry library, Sharma said.
“It’s transaction- or client-specific,” he said. “Not every client wants us to do it that way. The traditional model for consulting firms has always been a ‘talent for hire’ sort of setting. You pay us, we build something, then this is your IP. Other clients are more willing to invest, to give us the business knowledge while we bring the technology knowledge. We build it. We don’t charge dollars for it. But the understanding is, we walk away with the IP. That model is less about trying to make money for the month or the quarter than it is to build out our overall end-to-end flow and quickly build up libraries of agents.”
