In a world where AI agents and human employees have to work together, there needs to be a place where they can meet. If Cisco has its way, AI Canvas will be that place. What kind of new interface is this? And how does it work under the hood? We took a closer look and talked to Anand Raghavan, VP Products, AI at Cisco.
With AI Canvas, Cisco wants to put an end to the fragmented experience that IT teams have known for years in AIOps and SaaS environments. Instead of yet another central dashboard that has to contain all the data, Cisco is focusing on a distributed architecture with MCP servers per domain and an agentic orchestration layer that sends questions posed in natural language directly to the right systems.
AI Canvas is designed as an open ecosystem and is going to work with both Cisco and non-Cisco products. It continuously learns, including through human feedback. With this, Cisco wants to go beyond previous promises from the market, with an approach that is both multi-vendor and proactive. AI Canvas is not GA yet, but a very important focus product for Cisco.
The dilemma of SaaS dashboards
To better understand what Cisco wants to achieve, it is important to take a step back and look at how things work in many organizations. Many have jumped on the SaaS bandwagon over the past twenty years. This has undoubtedly brought the necessary benefits. However, each SaaS solution also comes with its own environment and dashboard. In itself, that’s fine, but there is “no easy way to get a unified experience across all these dashboards,” Raghavan outlines the problem.
With the increasing number of SaaS dashboards, this problem is not getting any smaller, Raghavan continues: “Whether it’s Cisco environments or not, whether I’m a NetOps employee, SecOps employee, or someone involved in observability, each domain has its own dashboards.” Having a separate dashboard for each sub-area has simply been”the SaaS way”for the past twenty years, he concludes.
Raghavan is quite knowledgeable when it comes to dashboards, by the way. He was one of the first employees at ThoughtSpot. That company marketed itself as the Google for BI and analytics (a slogan that Raghavan can take partial credit for). The then CEO of ThoughtSpot, Sudheesh Nair, had a fairly strong opinion about dashboards. “Dashboards must die,” we heard him say several times. You could say that ThoughtSpot already had a similar vision to what Cisco wants to achieve with AI Canvas at the time and wanted to generate output based on the input in natural language from users.
Changes and trends around dashboards
However, dashboards are also subject to change. First of all, there is the way employees interact with them. This now takes place on a large scale via natural language interfaces, i.e., spoken or written human language. That in itself is nothing new; it has been around for years. The BI world in which Raghavan has been active is a good example of this. However, it is much more widespread now. More and more people are coming into contact with it.
When it comes to AI Canvas, two other trends are more interesting, as far as we are concerned. The first is that larger organizations in particular need finely tuned models and applications that function well in their own environment. This has implications for how you make insights, telemetry, and other data available, via dashboards or other means.
The second relevant trend that Raghavan identifies is that there is increasing demand for the consolidation of dashboards. People no longer want to constantly access each of the dashboards. In the past, we have seen attempts to solve this problem by digging a data lake and throwing all the data into it. However, this did not usually deliver what organizations wanted or expected. Here, too, you run into the trend mentioned above, namely that not everyone wants to send all their data to a central location (often in the cloud).
According to Raghavan, there is an increasing demand for simplicity when it comes to gathering insights and the like. “Can’t I just ask a question in natural language and get an answer?” as he puts it.
A familiar story, but Cisco puts a new spin to it with AI Canvas
Up until now, the AI Canvas story sounds very familiar. That is, we’ve heard these kinds of stories many times before. The aforementioned concept of the data lake is a good example. That didn’t really work out. Now there are data lake houses on the shores of data lakes to bring more structure to the data, but it’s still quite a job to set something like that up.
We should also consider companies such as ServiceNow, which sees itself as part of the infrastructure where everything comes together and can be linked. It does this quite successfully, as evidenced by the company’s success in the market. However, this is still just a single dashboard, Raghavan points out. It certainly does not contain everything you might need. You still have to go to other environments and dashboards to retrieve additional information if you are looking for something specific or if you are troubleshooting.
The big problem that even successful players such as ServiceNow have is that the fairly static dashboards where all the information comes together often only answer the first question you ask, according to Raghavan. “When you’re troubleshooting, however, it’s never about the first question, but about the second, third, and fourth. That’s where existing solutions often fall short, precisely because of the static dashboards,” Raghavan summarizes. AI Canvas takes a fundamentally different approach. That’s what we’ll be discussing in the rest of this article.
AI Canvas
AI Canvas is an example of what Cisco calls AgenticOps. We wrote a more detailed article about this concept during Cisco Live last June, in which we also discussed the role of the Deep Network Model, an AI model developed by Cisco specifically for networks. AgenticOps is intended to be the successor to AIOps. The idea is that AIOps has become too limited.
The most important component of the name AI Canvas is not actually the first part (AI), but the second part. It is first and foremost a canvas. In fact, it is basically an empty canvas. That may sound a bit strange, but it is very important to understand. Such an empty canvas implies a certain dynamism. After all, it is empty at the beginning, but is filled every time you want to troubleshoot a problem. This filling is done by AI Agents, which even determine the design of the widgets that will appear on the canvas. People work together with the agents to arrive at good answers as quickly as possible.
Dynamism and autonomy
The dynamic aspect of AgenticOps and AI Canvas is what makes it different from AIOps, Raghavan tells us. A lot of AIOps consists of static playbooks. These always do the same thing when they encounter an issue or problem. Within AI Canvas, there is not only more dynamism, but also more autonomy for AI (Agents). The latter follows more or less logically from the former. If there is more freedom and dynamism for agents to do their work, they must also be able to make autonomous decisions more often.
Autonomy sounds nice, but it also comes with responsibilities. What happens if agents take a wrong turn somewhere? What will you see on your canvas? To address this, the AI Agents have access to the reasoning traces associated with these types of agentic processes. This means they can see what steps they have taken to get to where they are at that moment. Importantly, they can also take one or more steps back in this line. To do this, AI Canvas relies heavily on the Deep Network Model.
This enables AI Agents to reason through multiple scenarios and options before actually showing anything to a human employee. “As AI Agents discover more, they also understand more,” says Raghavan. This statement implies that AI Canvas works with some kind of memory via the Deep Network Model. To remember things, AI Canvas uses the reasoning traces it stores in Splunk. These traces are not much larger than firewall logs, says Raghavan, when we ask how big such a memory is. So it will not entail enormous additional costs.
Agentic layer on top of MCP servers
Now that we have a clear understanding of the theory behind AI Canvas, let’s look at how it works. In essence, AI Canvas is an orchestration and agentic reasoning layer on top of a collection of MCP servers. It is therefore explicitly not a Cisco-only product. “The future is multi-vendor, and it looks like it will be MCP-powered,” says Raghavan. Communication between silos, systems, and components will run through these types of servers when AI Agents are deployed.
To enable AI Canvas, Cisco initially set up MCP servers for all components of its own portfolio that it wanted to link. Raghavan gives the example of an MCP server for Meraki. That server has all the tools and APIs on board to work with Meraki. But every MCP server within Cisco’s portfolio now also has indirect access to all those APIs, and vice versa.
Decentralized vs. centralized
The above architecture is decentralized. That is a very important part of the story of AI Canvas. Raghavan: “This solves a fundamentally big problem. Pre-MCP, you had a single intent orchestration engine, which worked fine if you had 40 or 50 APIs. However, when you’re talking about thousands of APIs, the number of false positives became extremely high.”
If we look at the architecture, there is an agentic orchestration layer on top. All it has to do is classify the intent of a specific flow or process. Based on that, this agentic layer can determine which underlying MCP servers it can or must turn to.
If an employee cannot log in to Webex, for example, the MCP servers from Meraki and ThousandEyes also spring into action. The results of troubleshooting the problem with logging in to Webex are displayed in AI Canvas. A dashboard is put together for this specific report. If it had been a different question, the dashboard would have looked different.
Open system with broad possibilities
AI Canvas is essentially an open system because of the decision to use MCP as the common denominator. All sources, systems, and platforms for which an MCP server exists can be connected to it. Based on what we have seen since Anthropic made MCP available to everyone last November, that number is increasing steeply. There does not seem to be much debate about which standard to choose. In that respect, we expect that a lot of data from many sources will be able to flow into AI Canvas.
AI Canvas is also well suited to the requirements of organizations that want to maintain control over their own data, for example because it must always be on-premises or even air gapped. The underlying Deep Network Model can only be trained on the data that an organization has. For Cisco, this is also a great way to get their AI PODs into customers’ data centers, of course. These have been developed to run smaller, more specialized models locally.
Ultimately, the success of an initiative such as AI Canvas, which relies heavily on the presence of MCP servers, is highly dependent on the availability of those servers. No MCP servers, no AI Canvas, to put it bluntly. That’s not entirely true, Raghavan immediately points out, because in principle you can also send all kinds of data to Splunk or ThousandEyes via APIs and then access everything there. However, we don’t think that’s the most cost-effective option. What could also happen is that customers start demanding the availability of an MCP server from their supplier. Ultimately, that is always the goal of open systems. With the already considerable dominance of MCP, this latter scenario is quite realistic.
All in all, AgenticOps as Cisco envisions it is already much more than an interesting concept. AI Canvas is coming relatively soon, as it is seen as a priority internally at Cisco. We understand that the alpha version of AI Canvas should be ready by the end of September. We will try to get a full demo before then and will definitely share our findings here.