Cisco builds Webex agents that keep working after the customer hangs up

Engadget's picture of the new Webex is a presenter in the middle of a pitch who asks an agent to pull up statistics and update the PowerPoint on the fly ^[3]. Engadget wrote that this "could be extremely useful, assuming it works" ^[3].

Cisco's larger claim is about the work that happens after the customer hangs up. Dialog is a layer beneath the Webex AI Agent platform that ties a customer's history to enterprise systems, knowledge and policies, SiliconANGLE reported ^[4]. An agent running on it coordinates with employees, other agents and backend systems until the request is resolved, and it keeps its context across channels ^[4]. "We are moving away from a world where we build agents to where we train and onboard agents like digital co-workers," said Vinod Muthukrishnan, who leads Webex customer experience operations ^[11]. Cisco says teams will not have to build individual conversation flows, and that Dialog will run third-party agents alongside Cisco's own ^[5].

Cisco's executives talk about the agents in the present tense. "Agents are entering the workforce at an unprecedented pace, both to help you and operating by themselves," said Snorre Kjesbu, Cisco's senior vice president and general manager of collaboration ^[15]. Most of the new features and products are planned for early 2027 ^[9]. What a contact center can run now is narrower. Webex AI Agent is generally available for Unified Contact Center Enterprise, with voice support for Contact Center Express planned this month ^[14]. Neither report includes pricing.

The only measured performance in either source comes from outside Cisco. Engadget, citing multiple studies, put agent failure rates on multi-step tasks at 60 to 70 percent ^[18]. It also noted that an agent with 95 percent reliability per step completes a 20-step workflow about 36 percent of the time ^[17]. That is 0.95 multiplied by itself 20 times, or 0.358 ^[20]. At 99 percent per step, the same 20-step job still fails about 18 percent of the time ^[21]. Cisco says Dialog "continuously refines agent performance with every interaction" ^[8]. Engadget answered that "that's kind of the promise with all AI agents" ^[19].

Cisco's executives say the oversight job grows with the scope. "We are talking about securing, observing and governing an entire execution chain that starts with the conversation and goes well into the enterprise," Muthukrishnan said ^[12]. The AI Agent 360 control plane is meant to connect agent behavior to escalation patterns and business impact ^[6]. A count of agent sessions would only measure activity. For a pilot, I'd ask for two figures: the share of post-call runs that closed with no human involved, and how long escalated customers waited.

For a 2027 platform decision, I'd sort each candidate task by two things: how many steps it takes, and whether a person can cheaply undo a wrong one. Rebooking a callback is short and easy to reverse, so it can run after the call with sampled review. A long but reversible task belongs in the Dialog beta, with every escalation logged. Refunds and account changes are short and hard to undo; the agent drafts and a human approves. Anything long and hard to undo stays with a person through 2027. I'd sign the meeting and calling platform on what is generally available today and buy the agent layer as a separate pilot. The cost of that caution is a mid-contract renegotiation if Dialog leaves beta and its resolution rates hold up.