Cisco has announced Dialog, a new layer for the AI agents that talk to customers on its Webex platform. Webex has offered AI agents for customer chats since 2024. What Dialog adds is that an agent keeps going after the conversation closes, working with staff, other agents and company systems until the customer's problem is solved. Cisco expects a beta in the first quarter of 2027, so none of this is in customers' hands yet.
A case is a chain of steps
Normally, when a support chat ends, the customer becomes the project manager. They call back, repeat the story, and ask who owns the next step. Handing that follow-up to an agent is a real gain for the customer, and it is the part of the announcement worth wanting.
But look at what the agent now has to do. For a damaged sofa it checks the order, reviews the photos, reads the refund policy, books a replacement, emails the warehouse, updates the account, writes to the customer, waits for the tracking number, confirms delivery and closes the case. That is easily twenty steps, spread over days, across systems that were built by different people. A human team does this with a lot of lost handoffs. An agent that owns the case does the handoffs itself, and that is the appeal.
Why a small error rate becomes a big one
If an agent does each step right 95% of the time, the chance that all twenty are right is 95% multiplied by itself twenty times, which is about 36%. At 99% per step it is about 82%. For a shop with 1,000 such cases a month, that is roughly 640 cases with a mistake somewhere at 95%, and roughly 180 at 99%.
That is the worst case, where every step is equally risky and nobody catches anything. Many mistakes will be small, like a clumsy email. But one of them will be a replacement sent to the wrong address or a refund paid twice, and the customer finds out before the company does. Better models help, but they move the number slowly. Going from 95% to 99% per step still leaves about one case in five with a problem.
Where the person belongs
The way out is shorter unchecked chains, not a perfect agent. A person who looks at the work at step five stops an early mistake from being carried through the other fifteen. In the sofa case the agent can run most steps alone, and a person approves the two that matter: the replacement or refund, and the final message to the customer. That is perhaps two minutes of a person's time on a case that used to take twenty across a week of chasing, and it removes most of the damage a wrong step can do.
Cisco is also planning a dashboard built with Splunk, its data analysis product, that shows agent health and escalation patterns, with release planned for December 2026. That is useful for spotting a bad week. It shows a mistake after it has happened, which a checkpoint prevents.
Short chains are safe now
The same arithmetic explains why the other half of the announcement is easier to trust. Webex users can also invite an agent into a meeting, where it can pull usage numbers and update a PowerPoint deck. That is two or three steps, and the person who asked is looking at the slide before anyone else sees it. At 95% per step, three steps still come out right about 86% of the time, and the person catches the rest.
So the rule is plain. A short chain with a person watching can run now, and a long chain that nobody watches cannot. When a vendor shows you an agent that owns a case from start to finish, ask which steps a person approves, and how many steps happen between two of those approvals.