Cisco's Four Rules for Agentic AI: Redesign the Work First

Cisco runs a $15 billion services business with 20,000 employees, and Liz Centoni wants to redesign most of its workflows with agentic AI. Her first rule: don't just bolt AI onto the processes you already have.

Centoni is EVP and chief customer experience officer at Cisco Systems, a 26-year company veteran who led the $28 billion Splunk acquisition in 2024. Her organization handles professional services, customer success, technical support, and the renewals engine for software and services. She laid out four principles for making agentic AI actually pay off.

What happens when you automate the wrong thing?

When Centoni took her current post two years ago, she read customer feedback. The top complaint: the path to the right expert. Customers got handed from engineer to engineer, re-explaining the problem each time.

Her team's first move was an agent to speed up finding the right engineer. It backfired.

"It just upset our customers more because it showed that we actually did five handoffs behind the scenes and it surfaced it to them faster," Centoni said. "We just bolted AI onto the wrong thing and said, 'automation.'"

The right question, she argues, is why the handoffs exist at all. Workflow steps designed decades ago may solve problems an agent can simply cut through. You have to understand the desired end result first, then design the workflow around current capabilities. And you have to involve the people doing the work.

"Users are an essential part of the technology stack," Centoni said. "You have to bring in the subject matter experts and the users to outline the workflow, and you have to get close to the work. If you can't, you have no right designing."

Cisco's old model threw experts at every engagement — the more complex the problem, the more experts. Expert availability throttled speed. But the company has solved these problems tens of thousands of times over 40-plus years. It has now built that history into a knowledge base and playbooks that agents and humans share, customizing deployments against a customer's own KPIs. Human experts stay in the loop, with more bandwidth for iteration conversations.

How do you get from reactive to preventive?

Speed isn't the only goal. Centoni's benchmark used to be root-cause analysis time — cutting three hours to 30 minutes. Now she wants issues stopped before customers notice.

Agents monitor customer environments continuously, using decades of failure data to spot clusters and recommend next best steps before problems turn critical. "In 12 months, I'd like to be able to sit down with a customer and say you would've had 10 outages last year and we prevented eight of them," she said.

Why does model choice matter?

Boards increasingly demand ROI on AI, not adoption statistics. Centoni's discipline is simple: before starting an AI project, ask whether plain automation will do.

"If it's purely rules-based, I'm doing automation. If it requires bringing multiple signals together to drive a prediction, then AI," she said. "In many cases, it's not binary. We have a combination of things where we use old predictive AI models, automation, and some of the newer LLMs."

Tools should fit the use case, not the designer's preference. Cisco's renewals AI assistant — which combines a renewals agent, a sentiment agent, a risk agent, and an install base agent — runs on Mistral. Centoni said Mistral was one of the few vendors offering an on-prem model with full capabilities, which mattered for data sovereignty requirements.

What does success actually look like?

Centoni rejects adoption metrics. Users check the box, then finish the work with old tools. She calls this "compliance theater." Instead, she checks whether workflows changed and whether the renewals team gained capacity since the AI assistant launched.

Agents also behave differently from software. "With software, you build, test, deploy it, and it's 100% feature complete," she said. "With an agent, it's 60% of the way there in its first iteration in terms of usability, and the user is part of the tech stack that gives us feedback." A good feedback channel is what makes an agent usable at scale.

Cisco's internal support agent shows the payoff. When a customer opens a case, the agent pulls their full history and context instantly — sparing the support worker the first 30 minutes of gathering details. It also suggests the next best action based on what's known about the customer's environment.

"That's a lot of time and frustration saved up front," Centoni said.