Cisco: From the floppy disk to hyper-personalized AI

Thimaya Subaiya is executive vice president of operations at Cisco, and he oversees security and trust, supply chain, IT, and growth operations at the networking giant. Subaiya is part of the inner management circle around Cisco CEO Chuck Robbins. Before taking responsibility for operations, he held the role of chief transformation officer at Cisco.

Subaiya recently sat down with Computerwoche editor Jürgen Hill for a conversation that includes how Cisco is redesigning its internal processes with an eye on automation, how AI is redefining Cisco’s DNA, and why Subaiya thinks today’s world of AI is like a floppy disk.

Thimaya, let’s start with a personal question. In your opinion, what was the AI development that most shaped or transformed Cisco?

Thimaya Subaiya: I think it was the AI agents, i.e. Agentic AI and the Model Context Protocol. We now use the term “MCP Everywhere”. This means that all our data, systems and everything else are starting to communicate with each other. The key to this is AI agents.

Before we dive into the depths of AI, let’s look back at your previous role as Chief Transformation Officer at Cisco. What was the biggest challenge in transforming from a pure hardware and software provider to what I would call an “as-a-service company”?

Subaiya: Transformation cannot be done in isolation; it must take place within the entire company. Companies that want to transform should always have an end result in mind—be it a specific metric or a change in business or operating model. But ultimately, it is the people who make the biggest difference. Transformation must be anchored in the corporate culture.

Don’t see transformation in isolation

Subaiya: To prepare Cisco for the AI era, we had to bring two things together: Data and processes. Firstly, we had to ensure that our data was not only cleansed but also consolidated and that we were using the same data across all applications company-wide. Secondly, we had to rethink and redesign our internal processes, always with an eye on automation, in order to be able to act faster. Ultimately, our transformation was based on three elements: We wanted to simplify, move at speed and focus on growth.

That sounds like an ambitious goal..

Subaiya: Yes, it was. But these elements helped us to shorten our original transformation goal, which was set at two years, to eleven months. And it was at this crucial moment that AI became the “next big thing” in technology development. We see AI as the defining technology of this generation. Every generation sees the emergence of a new technology, and this is the next one. Our transformation and the AI era collided beautifully for us as a company.

Cisco

How is this defining technology changing Cisco’s DNA?

Subaiya: I see four categories in which AI is changing our DNA:

• Our go-to-market approach: The way our salespeople interact with customers is completely changing. This is sales productivity.
• Product development and engineering experience: Developers are using AI tools to develop faster. Especially in the area of testing and quality, AI can make a huge difference and greatly reduce the time needed. This is engineering productivity.
• Employee productivity: This is where AI agents play the biggest role. We are working on equipping every single employee with their own hyper-personalized micro-agent. These agents learn from each other, but only have access to the information that the respective employee has. AI should not be used as a search engine or as a substitute for work, but should make the employee smarter and expand their knowledge.
• Customer and partner experience: It’s not just about how we change internally, it’s about how we can provide a better experience for our customers and our large partner ecosystem from purchase to support.

AI is like a floppy disk

Let’s look five years into the future. What does an AI-driven Cisco look like compared to today?

Subaiya: Let me explain with a comparison. Today’s world of AI is like a floppy disk to me.

Sorry, that surprises me a little.

Subaiya: I deliberately use a floppy disk as an example, because AI as we know it today will be “old school” within the next year. Think about the development of storage media: from floppy disks to hard disks, solid state drives, and modern SSDs. This change took about 20 to 25 years. AI will go through the same curve, but much faster. In five years’ time, we will look back on today’s AI and say: “But that was so old-fashioned!”

And the consequence?

Subaiya: I expect that we will have a much better product portfolio within five years. What takes us 10 years to develop today, we can achieve in five years thanks to AI. Today, we are still at the data consumption stage. However, we must soon reach the point where the database is complete and we only need to add new data. The focus will shift from the question of how to drive an AI engine to how to use and optimize an AI engine.

AI and security

You strongly emphasize the efficiency gains through AI. Isn’t AI also a tool for creating completely new business models?

Subaiya: For me, new business models are part of the product portfolio. But let me emphasize another area: The fundamental weakness of AI that I see is security.

As Cisco, we are in a unique position here, as we are the only company to offer both a comprehensive networking and security portfolio and to bring them together quickly. In terms of security, we need to look at two dimensions in the AI world: First, attackers are using AI to attack companies and even countries. Today, security is no longer just a technological issue but a geopolitical one. The targeted way in which infrastructures and governments are attacked has increased exponentially.

Secondly, we can use AI for defense—AI for security and security for AI. Our development is aimed at linking network data directly to security data so that we can detect threats immediately at the network level instead of dealing with them after the fact. To fend off AI attacks, human intelligence is no longer enough, we need secure AI defense factories.

Cisco

And what are you doing specifically?

Subaiya: Our most important product, which we are currently testing, is called AI Defense and offers security for AI. This product monitors AI usage and signals when a potential boundary breach is imminent.

Customer Zero

Subaiya: I’m so excited about these products because I’m the “customer zero” at Cisco. After all, my IT and operations teams use these technologies first before we offer them to customers. For example, we are building our own internal AI stacks, including clusters with over 1,024 GPUs. These are based entirely on Cisco technology—from networking and compute to security and observability. We use this to create reference architectures, test our own products and run our internal AI agents and assistants.

Where are you internally and with your products in this evolution if today’s AI is a floppy disk?

Subaiya: Internally, we are just coming out of the floppy disk era. But I am always striving for more internally. The perfect state would be when all internal systems use agents that talk to each other. Tasks that take me five minutes today should no longer need to be done by me tomorrow. We are still six months to a year away from this point.

From the customer’s point of view, especially with our security offerings, we have made progress. We are just leaving the hard drive era. Our products are thoroughly tested. At Cisco, we record 1.2 trillion security hits every day internally alone. We are able to defend ourselves with the help of our own AI products, which shows that we are far ahead on the product side.

You also want to replace legacy applications—or example in HR or other areas—with agents. That sounds simple on paper, but how do you transform that into the reality of a large company?

Subaiya: Yes, that is indeed a challenge. Let me give you an example that may not sound “cool,” but costs a lot of money in day-to-day business: Order processing.

Customers usually send us orders as signed PDF documents. We had to feed such a PDF back into our systems, match it with each individual order, check inventory management to see where the products were being shipped from, and link it to the ERP and supply chain system. If components were missing, production had to procure them.

A process that required around 1,200 external employees. They read the documents, entered the data, and ensured that it was transferred to the ERP system. The error rate was 2-3%, which caused problems in the downstream processes.

Today, an AI agent takes over the entire process end-to-end, without human interaction. We no longer need the 1,200 external employees. The system works more efficiently and faster, running in real time instead of with a 24-hour delay. This not only saves us costs, but also gives us cleaner data and improves the customer experience.

MCP as an API for agents

But the AI agent has to be able to talk to many systems?

Subaiya: Yes, that’s right. That’s why we use the MCP I mentioned at the beginning. The order agent can talk to the ERP agent, which in turn enables much more efficient and timely procurement based on incoming orders. This lowers our logistics costs and minimizes shipping risks. Basically, MCP provides APIs for agents.

How do you ensure that these agents don’t mutate into black boxes that even your specialists can no longer understand?

Subaiya: We strictly control the data used. Security is always our top priority. We monitor where the data is located, who has access, and how clean the data is. The second is the critical control of process and workflow mapping. As long as you control the data, the process, and the workflow, the AI can’t really get out of hand. We also take a human-in-the-loop approach—a human must always approve.

What other AI tools do you use internally in your day-to-day work?

Subaiya: We use AI in all our global processes and in the supply chain. Our internal security team also uses AI to keep Cisco secure. We also have an internal AI assistant, Circuit. This is a secure, isolated environment—the data remains on site. Circuit has over 90,000 users and makes it possible to operate various models such as ChatGPT, Gemini, Grok, or Claude in a secure environment.

We also have our own internal Cisco LLM. As it uses up-to-date Cisco data—in contrast to external data—its responses are up to date. This is another advantage of MCP: it allows our AI to access the latest information, create context, and expand the data.

Brainstorming with the AI

What skills should managers have in order to exploit the potential of AI in a company?

Subaiya: This is less a question of specific skills. The way in which someone uses AI is much more important. The most important thing is to brainstorm with the AI engine. Don’t just ask it something or use it to search, but have a real brainstorming conversation with the AI.

Sorry, how can I imagine such a brainstorming session?

Subaiya: I’ll give you a personal example. I had a large amount of water to remove from a concrete pool, so the cleanup usually took a whole day. I started a conversation with an AI engine and told it my problem and asked how to make it easier?

The AI suggested installing a permanent pump and asked for details like the depth of the pool etc. We discussed different types of pumps and at some point the suction power came into play. I hadn’t even thought about it, but suddenly it wasn’t just how far the water had to be transported, but what suction power I needed to achieve this.

The AI suggested changing the piping to improve the suction capacity, but this increased the flow rate. The AI’s solution: add a tap at the other end to reduce the flow rate. So I worked out a solution with the AI. In the end, the AI warned me: “Be careful, the pump could implode due to too much suction.”

This led to the question of a pressure relief valve. The AI couldn’t recommend a ready-made valve, but showed me the blueprint of a pressure relief valve. I built this with a 3D printer and uploaded a photo to the AI so that it could check it. It was as if I had three or four different engineers at my side designing a solution with me.

Impressive, I’ll have to try that with one of my next problems. And what personal experience has shaped your view of AI the most?

Subaiya: I have two daughters who are eight and ten years old. My entire house is automated and can be voice-controlled via Alexa. If my children wanted to turn on the fan, they always had to say the exact command—”Turn on all the fans” or “Turn on the fan in the living room”. Alexa doesn’t understand context. This frustrated me and was the impetus for the idea that I could use an AI agent to personalize Alexa. The goal was for the technology to adapt to the children and not the other way around. Today, each of my daughters has their own AI agent, which knows which commands are meant through voice recognition via Alexa.

I have transferred this personal insight—that technology must adapt to the user—to the company. That’s why we want to create micro-agents that can be personalized for each individual person. That was the biggest personal AI learning for me, and it can be a real game changer for our company.

This article originally appeared in Computerwoche.