Cisco’s four-part playbook for making agentic AI work

By Nathan Brooks

Cisco’s $15 billion services organization employs 20,000 people; Liz Centoni says one early AI pilot merely exposed five hidden engineer handoffs to customers faster.

Centoni, Cisco’s executive vice president and chief customer experience officer, leads professional services, customer success, technical support, and the renewals engine for software and services. The 26-year Cisco veteran led the company’s $28 billion Splunk acquisition in 2024. Her remit now spans internal operations and customer delivery as Cisco pushes to become AI-native.

Cisco is combining agentic AI—systems that can act on tasks or recommend next steps—with older predictive models and rules-based automation. Centoni’s governing position is blunt: good revenue and margins do not make the status quo safe, and leadership, not technology availability, limits progress.

Her four operating principles are:

• Redesign any workflow constrained by human availability.
• Keep improving processes and offerings because customer expectations keep rising.
• Treat current quality as a floor, not a ceiling, while AI capability changes quickly.
• Equip managers to support leaders beneath them who can bring teams along.

Those principles lead to four practical moves.

What should an agent actually change?

“We didn’t just start bolting on AI. We looked at the underlying workflows to understand how we could redo work,” Centoni said.

When Centoni assumed her current role two years earlier, customers complained about finding the right technical expert. The route often passed through repeated engineer handoffs, with each engineer asking the customer to explain the problem again.

Cisco’s first response—an agent that found the right engineer faster—made the underlying design worse. “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. We just bolted AI onto the wrong thing and said, ‘automation,’” Centoni said.

The corrective starts with the desired outcome, not the existing process map. Some steps were designed decades ago around problems an agent may not be able to cut through. Subject-matter experts and users must define the new workflow; teams without close contact with the work should not design it.

Cisco must also set explicit boundaries between autonomous action and human handoff. Centoni asks where automation and AI can work at 90% accuracy or better, while keeping experts in the loop for consequential judgment and customer conversations.

The economics come from Cisco’s history. Over more than 40 years, the company solved similar problems tens of thousands of times, yet often treated each customer as a bespoke engagement requiring more experts.

That experience now sits in a knowledge base and playbooks for agents and people. Using customer context, including project KPIs, Cisco can customize value realization with less expert labor while improving consistency, security, and time to project completion. That capacity also gives experts more time to discuss iteration and demonstrate value beyond what customers paid for.

Takeaway: Automate the redesigned service, not the legacy handoff chain.

How can AI prevent customer outages?

Cisco traditionally optimized for faster root-cause analysis, aiming to reduce a three-hour diagnosis to 30 minutes. Centoni now wants agents to act earlier by finding recurring clusters in case histories and monitoring customer environments for warning signals.

The agents can recommend next steps before an issue becomes critical, using Cisco’s accumulated knowledge of failures. In a 12-month target rather than a reported result, Centoni wants to tell a customer it would have had 10 outages last year and that Cisco prevented eight.

Takeaway: Move the unit of value from faster recovery to fewer incidents.

Which model should a team choose?

After a year focused on AI adoption and maximizing token use, boards increasingly demand return on investment, or ROI—the measurable business return from an investment. Centoni’s sequence starts by asking whether rules-based automation can complete the task; AI becomes necessary when several signals must be combined to make a prediction.

Many Cisco use cases mix old predictive models, automation, and newer large language models, or LLMs. Model choice must follow use-case constraints, including data sovereignty rules that require data to remain on premises, rather than a designer’s preference.

Cisco’s renewals assistant demonstrates the rule. It combines renewals, sentiment, risk, and install-base agents. Centoni chose Mistral because it was one of the few vendors able to provide a customized private, on-premises model with the required capabilities, and the system also uses automation.

Takeaway: Select the least complex tool that satisfies the task and its constraints.

How should teams measure an agent?

Adoption can produce what Centoni calls compliance theater: employees mark an AI tool as used, then return to familiar processes. Centoni instead checks whether workflows changed and whether the renewals team gained capacity after deployment.

Agents also need different completion criteria from conventional software. Centoni says conventional software can reach 100% feature completeness after build, test, and deployment, whereas an agent’s first iteration may be only 60% usable. Users become part of the technology stack, so a feedback channel is essential to iteration and scale.

Cisco’s internal support agent runs inside the support workflow. When a customer opens a case, it retrieves relevant history and context, sparing the support worker the first 30 minutes of repeated questions, and it recommends a next action based on the customer’s environment.

Takeaway: Measure changed work and added capacity, then improve the agent with users.

As product teams move from pilots to operations, the durable pattern will be workflow ownership: find the constraint, redesign with the people doing the work, choose tools to fit, and prove the result through capacity and customer outcomes.