Cisco's Liz Centoni: Why bolting AI onto broken workflows fails

Cisco's $15 billion services business learned its first agentic AI lesson the hard way: an agent built to speed up customer handoffs made customers angrier, not happier. Liz Centoni, EVP and chief customer experience officer, disclosed the failure as part of a four-point playbook for agentic AI that she says now guides a 20,000-employee organization covering professional services, customer success, technical support, and renewals.

Centoni, a 26-year Cisco veteran who led the company's $28 billion Splunk acquisition in 2024, is steering the networking giant toward becoming an AI-native company in both internal operations and customer offerings. Her governing principles are blunt. Every workflow constrained by human availability is a workflow for redesign. The status quo is not an option. Today's quality is the floor, not the ceiling. And leadership, not technology, is the limiting factor.

Why did the first AI agent backfire?

When Centoni took her current role two years ago, customer feedback pointed to one recurring irritation: the path to the right expert required repeated handoffs, with the problem explained in detail each time. Her team's first instinct was to build an agent that would find the right engineer faster.

"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 correct question, she argued, was why the handoffs existed at all. Mapping a workflow and automating it is not enough, because steps designed decades ago may address constraints an agent no longer has. "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."

She also urged leaders to draw careful boundaries between where an agent can act autonomously and where a human must take over, asking where automation and AI can operate at 90%-plus accuracy.

How does expertise become scalable?

For most of Cisco's history, complex customer engagements meant adding more experts, and engineer availability throttled speed. "When I came into this business, we treated every customer we had like a bespoke model," Centoni said. "But over the last 40-plus years, we've solved these problems tens of thousands of times."

Cisco has converted that history into a knowledge base and playbooks that AI agents and human experts share. An "expert agent" now customizes deployments using customer data and project KPIs without layering on many human specialists. "The outcome for the customer is consistency, security, and better time to project completion," she said. Human experts remain in the loop, with more bandwidth for customer conversations.

Can AI prevent problems rather than fix them?

Cisco is shifting from faster root-cause analysis — compressing three hours to 30 minutes, in Centoni's example — to prediction. Agents now monitor customer environments continuously and recommend next best steps before issues turn critical. Her stated goal for the next 12 months: "You would've had 10 outages last year and we prevented eight of them."

How does Cisco choose models?

With boards increasingly demanding AI ROI after a year of adoption-focused spending, Centoni applies a simple test. "Before we even go down the AI path, we ask, 'Can automation do this?'" she said. Rules-based work gets automation; predictions that combine multiple signals get AI; often both run together with older predictive models and newer LLMs.

Model selection follows the use case, not designer preference. Cisco's AI assistant for renewals — combining a renewals agent, a sentiment agent, a risk agent, and an install base agent — runs on a customized Mistral model in a private instance, chosen because few vendors offered an on-prem model with full capabilities for data sovereignty requirements.

What metric replaces adoption?

Centoni rejects adoption as a success measure, calling checkbox usage "compliance theater." Instead she checks whether workflows actually changed and whether the renewals team gained capacity since the assistant's deployment.

Agents also demand a different build mindset. "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." Cisco's internal support agent illustrates the payoff: it pulls a customer's full history the moment a case opens, saving the first 30 minutes of fact-gathering and recommending next best actions based on the customer's environment. As Cisco's agents iterate with user feedback, the open question is whether prevented outages and freed expert capacity will show up in measurable returns.