Customer Focus & Team Excellence: Shashi Upadhyay’s Fast-Paced Method at Zendesk

At Zendesk, agentive AI now shapes the product strategy. Shashi Upadhyay, as the President of Product, Engineering and AI, represents a unique profile with experiences in both startup environments and big tech companies. Currently, he leads a significant division of 2,000 people. Following Zendesk’s recent announcement of new developments in agentive AI for customer experience, he explains how this transition is being practically implemented on a large scale, leveraging his extensive experience.

From founder to corporate leader, how have your experiences shaped your product development approach?

I’ll start with the startup experience. I founded Lattice Engines, which was later acquired by Dun & Bradstreet, and ultimately went public. I’ve been through the full journey from inception to IPO.

In a startup, the first lesson is emotional: the constant fear of failure.

In a startup, the first lesson you learn is emotional: the constant fear of failure. A misstep could mean the end of the company, affecting everyone who has invested in us. This fear, mixed with the optimism of building something great, becomes a part of you.

To cope, we do three things: listen very closely to customers, recruit the best possible people, and act quickly. These are deeply ingrained in me now.

From a big tech company, I learned how to operate at scale.

From a big tech company, I learned how to operate at scale. You can’t inspect everything, and you don’t know everyone. You need to focus the organization on a few key priorities while still maintaining that same paranoia about clients, talent, and speed.

For me, Zendesk is the perfect mix. The shift towards AI feels like a startup, with a huge opportunity that starts modestly, yet with the resources of a large company. It’s not as massive as big tech, but it’s enough to act decisively and not so large as to become slow.

So again: focus on customers, team excellence, and speed. That’s my philosophy.

How do you align product, engineering, and AI towards a unified goal of impact and simplicity?

Yes, I oversee approximately 2,000 people, 300 of whom work solely on AI. When I joined seven months ago, my priority was to demonstrate how crucial AI would be. Not everyone was convinced initially. Zendesk was a traditional, high-performing SaaS company, but with significant room for improvement in AI. So, we began evolving. First, by reallocating human resources: we went from a few individuals to several hundred working full-time on AI.

What matters is the outcomes for customers, not the number of licenses.

Secondly, in a large organization, you need a short list of priorities and to repeat them until you are tired of them. That’s when others start to get it. I repeat: AI is key. We are transitioning from a SaaS company to an AI enterprise. What matters is the outcomes for customers, not the number of licenses.

Thirdly: re-skilling. We are investing heavily to make our teams AI-oriented. Fourthly: extreme focus on the customer. Fifthly: targeting markets we already have at hand, such as employee services, which I’ve made a priority.

We’ve also made acquisitions, like Local Measure, to strengthen our CCaaS platform.

All this rests on the constant repetition of five or six messages. Either people align, or we recruit those who believe in this mission.

LLMs evolve faster than traditional development cycles. How do you design your product roadmap in this context?

This is the fastest-moving technology I’ve ever seen. Internally, we anticipated that certain agent-to-agent capabilities would take 18 months… They arrived almost immediately. The model layer evolves the fastest. OpenAI, Anthropic, Google, DeepSeek… each leads in different areas. These companies can afford to continue for ten years if needed.

We let customers choose the best model at any given time.

So, we abstract that layer. We let customers choose the best model at any given time. We focus on use cases. Which ones create the most value? Then, we select the model that works best for that case, at that time.

We have established a robust evaluation pipeline. We do backtesting, we tailor the model to the task, as quickly as possible. If we get too attached to a single model, we make a mistake. So, we are model-agnostic, focused on usage.

Regarding work methodologies: it’s like moving from Waterfall to Agile. Now we are moving to something else. We encourage small teams to test new approaches. We give them tools like GitHub Copilot, Kodium, etc., without imposing anything. This freedom facilitates the transition.

How do you distinguish a good idea from a product that deserves to be industrialized?

When it’s a genuinely good idea, customers practically snatch the product from your hands. Even if the product isn’t finalized, they grab it and adapt it.

We had a clear example at Zendesk: our product for internal services was born this way. Customers used our CX product for HR or IT cases. They modified it themselves. We observed this, then built around it.

It’s crucial to identify what customers are trying to do on their own.

It’s crucial to identify what customers are trying to do on their own: hacks, workarounds, in-house developments… If everyone is creating a chatbot themselves, there’s a massive need. It’s up to us to respond. So: stay very close to the customers. The signal is often clear.

AI is often seen as a tool for efficiency. Can it also help differentiate in a saturated market?

Yes, without a doubt. The real power of AI is not just efficiency; it’s making customers more satisfied. Simple examples include 24/7 support, instant responses, consistent tone, no mood swings… And eliminating a classic frustration: having to repeat your issue to every new person you speak to.

The goal is for the customer to no longer dread contacting a company.

AI can remove this friction. Even in highly competitive markets, there is room for improvement. The ideal scenario is AI managing the entire experience, and as soon as it reaches its limits, a human takes over immediately. The goal is for the customer to no longer dread contacting a company. We are far from this today, and AI can truly make a difference.

What is your stance on bias, transparency, and accountability in AI systems?

The provider must be responsible. Hallucinations, absurd responses… These are known issues. When you build your own solution, you face these very quickly. A provider must be an expert in the domain.

We promise well-trained AI from day one.

We focus solely on customer service. No distractions with marketing or sales. We train our AI exclusively in this area. It won’t go off on tangents. We promise well-trained AI from day one. You can improve it with your data, but the foundation is already solid.

So yes, the responsibility rests with the provider. That’s why we say: don’t do it yourself! Let us handle it. It’s our job.

How do you maintain a high technical level without slowing down speed or experimentation?

It’s not a choice; it’s a matter of context. In established areas (infrastructure, latency), we use structured approaches. But when building for humans, where tastes and behaviors change, you need to iterate quickly.

I don’t want iteration on security. But I do want a lot on AI/human interactions: tone, responsiveness, etc. So at our company, some teams are very process-driven, others are experimental labs.

Have you noticed a change in product design expectations, in terms of managing automation, customization, and control?

Yes, and AI makes these balances easier than ever before. Previously, customization and scaling were in conflict: customization required humans, scaling demanded uniformity. This compromise is disappearing. AI can detect intent, personalize at scale, and act. We can do both simultaneously, without compromise.

What are the signs that AI is well integrated into a product? And conversely, how can you spot a failed integration?

A well-integrated AI is invisible. You don’t notice it. It just works. Conversely, a poorly integrated AI is obvious: it’s clumsy, disconnected. It’s like adding a medieval piece to a Roman ruin. Many AI experiences today look like this.

Tomorrow, applications will respond in natural language. “Show me the last 20 red tickets,” and it will appear. It’s already feasible. So: invisible = well integrated.

For a CX company starting with AI products, what advice would you give? And what mistake should be avoided?

First, determine which indicator you want to improve. In customer support, the fundamentals don’t change: solve the problem, quickly, in a good experience, and get a recommendation. The rest, like deviation rates, is secondary. Start from these objectives, measure your baseline, then see if AI improves these indicators.

Many products fail because they don’t know what they’re trying to achieve. Or they choose the wrong problem. For example: focusing only on efficiency. You can be very efficient… by doing nothing. But that’s not the goal. The goal is efficiency with customer satisfaction. If you start from there, everything else is just implementation.