The contact center has been broken for decades. The technology to fix it has only just caught up. Pedro Andrade is the person building what comes next.
Pedro Andrade has a simple way of explaining why he took the VP of AI role at Talkdesk. The contact center, he says, has been broken for decades — and everyone in the industry knows it. Agents are navigating ten legacy systems simultaneously. Customers are on hold while repeating information they already submitted online. A department that exists, in most organizations, to absorb complaints rather than prevent them.
What changed his mind about the timing wasn’t a single breakthrough. It was the convergence of two things: a decade of failed deployments that taught the industry what actually matters, and models that have finally caught up to the promises vendors were making years too early.
Andrade splits his time between engineering teams debating RAG pipelines and latency, enterprise CIOs who are simultaneously excited and terrified about AI, and a forward-looking 30% of his calendar dedicated to asking which foundation models are worth betting on twelve months from now.
In this conversation, he talks about what separates real AI from the chatbot era’s false promises, why Klarna’s much-publicized reversal was actually a story about process design rather than technology, and why the most important question in AI-powered customer experience has nothing to do with the model.
Excerpts from the interview;
What drew you to Talkdesk, and why is AI in customer experience such an interesting problem to solve today?
CX has been broken for decades, and everyone knows it. Contact centers are perceived as cost centers. Agents burned out navigating ten legacy systems. Customers are waiting and repeating themselves.
I took this role because Talkdesk sits exactly where this gets fixed. I’m lucky to live in a time when we have “enough technology” genuinely capable of resolving this problem. Not just deflecting calls to save money, but turning the contact center from a complaint department into something proactive and revenue-generating. Making both the agent’s and the customer’s lives meaningfully easier is, honestly, one of the most interesting problems in enterprise tech right now.
Also Read: The Agency-Led E-commerce Model Is Changing
What does your role as VP of AI actually involve? Where do you spend most of your time, and what are you building?
The title sounds futuristic, but the day-to-day is pretty pragmatic.
About 30% of my time is with engineering and data science, translating what new models can do into features that actually work in production to solve real pains (there are always pains somewhere, even in advanced tech). That means architecture reviews for our AI Agents, RAG pipelines that don’t hallucinate, and a deep commitment to latency.
Another 40% is with customers. Enterprise CIOs are excited about AI and terrified of it at the same time. I help them build realistic roadmaps that align with their data constraints and brand. It’s as much change management as it is technology.
The remaining 30% is looking ahead. The landscape shifts fast enough that if I’m not thinking 12 months out, we’re already behind. Which foundation models are worth betting on? How do multimodal, new hybrid models, etc, change the customer interaction model? That’s where our R&D decisions get made.
Contact centers have used AI for years—from chatbots to IVR. What’s fundamentally different about what Talkdesk is building today?
The honest answer is that we finally have technology that’s up to the promise. For years, “AI in the contact center” meant rules-based bots that frustrated customers and sentiment scores nobody acted on. The gap between what vendors claimed and what actually worked was enormous.
That gap hasn’t closed because we got smarter overnight. It closed because a decade of failed deployments taught us what actually matters, and the underlying models finally caught up. You can’t skip that learning curve.
What’s also changed is the scope of what’s possible. Eight years ago, during the chatbot boom, nobody seriously imagined we’d have this level of language understanding, let alone systems that can process images and video and reason about what they mean. That expands the use cases dramatically, and honestly, it creates a whole new set of problems to solve. Which is the part I find most interesting!
When should an AI agent hand a conversation to a human—and who decides where that line is drawn?
The client always owns the decision. Talkdesk provides the engine, but the enterprise dictates the business logic.
Also Read: AI Ads Will Win Only If They Earn Consumer Trust
Every CX platform is making big AI claims. Where does Talkdesk genuinely stand apart in real-world enterprise deployments?
A few things differentiate us, and I’ll be direct about what I think actually matters versus what’s table stakes.
The data advantage is real. Years of interaction data, combined with deep integrations into enterprise systems of record, give our models context that a generic LLM simply doesn’t have. The 360-degree customer view isn’t a marketing line; it’s what makes automation actually work in production and provide truly personalized content.
We’re also not selling a bot. We cover the full contact center surface: self-service, agent assist, routing, workforce management, outbound, analytics, and back-office automation. That end-to-end scope matters because CX problems don’t live in one place.
Industry specialization compounds that. Prebuilt vertical agents and industry-specific retrieval models mean we’re not starting from zero with every customer. That’s what lets us go from pilot to production in days or weeks rather than quarters.
Two things I’m particularly proud of.
- First, AI Gateway lets our AI run on top of third-party or on-prem environments, not just Talkdesk. Many enterprises are nowhere near full cloud migration, and meeting them where they are is a competitive reality.
- Second, differentiation also takes the form of go-to-market motions: we deliberately shifted toward forward-deployed engineers co-building with customers. That changes the quality of what gets shipped, and puts Talkdesk as co-owner of the solution we deliver to customers. And they love it.
And underneath all of it: enterprise-grade trust and safety. Guardrails, governance, human-in-the-loop controls. Our enterprise banking customer taught us early that this isn’t optional for the deals that matter.
Klarna replaced hundreds of support agents with AI, then began hiring humans again. What did the industry misunderstand from that experiment?
Klarna is a great case study, and I say that without judgment. They moved fast, made a bold bet, and learned from it publicly. That takes courage.
What the industry got wrong was treating this as a technology swap rather than a transformation. You can’t just replace humans with AI and declare victory. A lot of what makes customer experience work has nothing to do with the model: it’s data readiness, process design, change management, and organizational culture. Those differ from company to company, and no vendor can package them for you.
The other mistake was skipping the learning cycles. This technology is genuinely new. The right motion is to start small, instrument everything, and treat early deployments as experiments with real feedback loops, not finished products. Companies that went all-in before building that foundation set themselves up for exactly the kind of reversal Klarna experienced.
The ones getting it right are moving incrementally, measuring outcomes honestly, and accepting that some human judgment isn’t a failure of automation. It’s a feature.
Also Read: Your AI Chatbot Knows More Than Your Marketing Team
Some argue that companies are using AI to avoid customers rather than serve them. As someone building these systems, how do you respond?
At Talkdesk, we design AI to remove friction, not to remove connection. If a customer just wants to know where their package is, making them wait 15 minutes on hold for a human is disrespectful of their time; AI solving it in 10 seconds is excellent customer service. But if they need help navigating a missed mortgage payment, forcing them to talk to a bot is cruel. The technology isn’t the problem—the deployment strategy is.
