Cisco AI solves, humans decide – the future of network troubleshooting?

After Cisco has talked a lot about AI so far, but has shown few concrete solutions, there is now an interesting approach with AI Canvas. The keynote at the live in-house exhibition featured an AI agent for automated network troubleshooting, which analyzed the root cause of a network problem using various data sources, suggested solutions and implemented them after human approval and subsequently validated the effectiveness of the solution.

Behind AI Canvas is a web-based dashboard with an AI assistant that the network manager can use to hold a conversation and act as an agent. It guides them through the diagnosis and coordinates decisions and solutions with the tool user based on live telemetry and the experience of network experts.

The basis for troubleshooting is provided by an LLM called Cisco Deep Network Model, which is specialized for the specific application. According to Cisco, it has been trained with over 40 million tokens and experts have contributed over 3,000 reasoning traces and carefully commented on and validated each logic level. Specifically, the manufacturer draws on findings from the Technical Assist Center (TAC) to learn from solved problems.

Data from the network, the cloud, security tools, observability, and collaboration platforms is to be used for the dashboard to obtain as complete a picture as possible. Cisco categorizes the interaction between the various agents that obtain telemetry data, identify potential network errors and suggest solutions under the keyword AgenticOps. “The age of agent-based AI is just around the corner,” said DJ Sampath, Vice President for Products, AI Software and Platform at Cisco.

AI Canvas in a practical example

As an example, Cisco showed an open ServiceNow ticket that was inserted with the ticket number in the AI Assistant of AI Canvas. AI Canvas then pulled the contents of the ticket from ServiceNow to record the specific problem. It then pulled data from the Meraki, ThousandEyes and Splunk cloud platforms and analyzed it using the Deep Network Model. AI Canvas then began gradually to generate a dashboard with various widgets and graphs, from which it displayed relevant insights from different platforms.

The Meraki dashboard showed an increased packet loss from a Meraki MX firewall, Splunk provided a report on increased TCP retransmissions and the path visualization from ThousandEyes showed the specific point where the packet loss occurred, namely an overloaded WAN link. To avoid the bottleneck for a mission-critical cloud application, the tool suggested implementing a QoS policy in the SD-WAN on the firewall. The human had to approve this, AI Canvas carried out the change independently against the Meraki platform via API and subsequently validated the quality via ThousandEyes.

The seamless process without switching between the dashboards of the different solutions, the involvement of the system manager in decisions by the assistant and the speed of the solution were convincing. Cisco also showed how other people can be involved in collaborative troubleshooting – and pointed out that troubleshooting is a team sport rather than a finger-pointing exercise, which some system managers tend to overlook.

Control by people

Cisco emphasizes that AI Canvas is not intended to replace IT operating personnel, but only to support them. Humans should always be part of the process to ensure that all processes are compliant. Cisco emphasizes transparency at every stage, for example by providing traceable information on the origin of the data used. The central idea: while AI Canvas takes over repetitive and time-consuming tasks, the final decision-making responsibility remains with the human being.

In the long term, it looks as if critical processes can benefit greatly from automation through agent-based systems to enable people to rectify errors more quickly in critical situations. Cisco expects that in the future, employees will have to carry out inspection and approval steps before autonomous AI agents are given full decision-making powers. In an interview with iX, the manufacturer stated that it has an approach for delegated privileges for agents based on its in-house Duo solution to control the corresponding authentications for agent identities, their authorizations and to monitor the activities they perform.

(vbr)