
"Within 6 months, there'll be an AI agent that will remind you of birthdays and also order gifts"
Ray Smith, Microsoft's VP of AI agents, isn't worried about the Chinese influence on AI models because "the more models, the more breakthroughs, the more innovation—the better for everyone." He says, "There are a lot of things I don’t see AI doing anytime soon. Negotiation, strategy, and tasks requiring emotional intelligence." However, he admits that the bot answered a question better than him: "The answer is probably better than mine, but I would like to edit it."
Ray Smith, Microsoft’s Vice President of AI Agents, how do you see the impact of DeepSeek’s R1 model on the current AI field?
“I think there will be a lot of innovation, and what we’ve seen in recent years is that innovation is happening everywhere. It’s not surprising that DeepSeek and other models are delivering major breakthroughs. (In January of this year, there was a sudden buzz when the Chinese company DeepSeek unveiled its R1 model, claiming it was developed using a fraction of the computing power and cost of its competitors. It has since been revealed that not all those claims were accurate - O.K.) I think it’s good for the market that we have all these different teams, groups, and countries innovating and creating new models. I believe it will push the capabilities of AI agents and lead to even wider adoption. It’s great. The cost of inference is going down, the applications are already here—so it’s all good. Microsoft’s approach is very open: we support a choice of models and use the right models for the right tools. In the future, customers will want the ability to say, ‘Well, if I run this specific agent, then this model is better.’ Or maybe they’ll even fine-tune their own models. So the more models, the more breakthroughs, the more innovation—the better for everyone.”
I didn’t write this question. I gave ChatGPT your LinkedIn page and asked it to write interview questions for you. I also asked it to write what your possible answers might be. Here’s what it came up with. I’d like you to read it and tell me how good it is.
ChatGPT’s answer:
“DeepSeek’s R1 is, in many ways, a disruptive signal for the entire AI ecosystem. Its emergence shows that it is possible to achieve high-level thinking capabilities with significantly lower resource investments. For us at Microsoft, it reinforces the idea that innovation is not just about increasing budgets—it is also about rethinking our approach to efficiency and cost-effectiveness. While DeepSeek’s model challenges established norms by proving that open-source methods can deliver competitive performance, it also pushes us to double down on our investment in our AI vision. This vision is not just about raw power; it is about integrating AI in a way that is scalable, secure, and ethically responsible. In this context, R1 creates healthy competitive pressure. It is a wake-up call that motivates every player in the industry—not only to innovate in performance, but also to ensure that our models are built on strong security frameworks and guided by responsible practices. Ultimately, I see this as an opportunity for the entire industry to accelerate progress, to benefit end users, and to continue pushing the boundaries of what AI can achieve.”
Smith reads ChatGPT’s answer and replies:
“The answer is probably better than mine. It explains what DeepSeek is as a model and covers the idea of deep thinking. It touches on a lot of key points, which is good—for markets, more models, and the whole industry will accelerate. Yeah, I think that’s a pretty good answer.”
So if we can ask ChatGPT to frame the questions and then generate the answers, and those answers are good—aren’t we just wasting our time now in this interview?
“As humans, we’ve always looked for ways to use technology to automate. We’ve always been on a journey toward simplification and creating more automation. It allows us to focus on more complex, higher-value, more strategic tasks. Now it’s the same thing. Would I take ChatGPT’s answer as a direct quote from me? No, I would edit it. But it would definitely help me. It’s a great answer, but I would want to adjust it to better represent me. We’ve been embracing automation for a long time, and generative AI models are a major breakthrough. They enable us to automate more because of their ability to handle complex variables and information.”
When you talk about the tasks we’ll be freed up to do, what will those be? Because when I look five or ten years ahead, given the speed of technological development, what will humans do that AI can’t do better?
“Many years ago, I worked at a startup where we developed tools to accelerate sales and forecasting. Some people said we were automating the role of a salesperson. I think that’s nonsense—because people buy from people they trust. Humans will always be involved in sales roles, for example. There will be AI agents that support sales development and provide initial product support. Certain roles will evolve; they’ll focus on human strengths, like building deep relationships. And we’ll work hand in hand with AI agents. It will be humans plus a kind of co-pilot—an assistance interface that connects to a vast pool of AI agents at your disposal. That’s the future of work. We’ll manage AI agents to perform tasks, which will free us up to take on more complex responsibilities.”
"Breakthrough in six months"
You mentioned AI agents—so let’s dive into that. How is an AI agent different from a chatbot like ChatGPT?
“That’s an excellent question, because there’s no uniform definition in the market yet for AI agents. My definition is that AI agents are tools with the ability to perform a thinking process. They need to plan. They can plan. They evaluate their input to decide what actions to take, what tools and knowledge sources to leverage, and they have the dynamic ability to act. A good AI agent will examine its environment and input, create a plan to solve a problem, and then independently take various actions to achieve its goal.”
What AI agents are already in use today?
“Pets at Home, a British pet retailer, has created AI agents connected to their customer relationship management (CRM) system. These agents can retrieve relevant information to handle customer inquiries—such as product returns—and perform risk assessments to detect potential fraud. A chemical company has built an AI agent that optimizes its supply chain and can identify and prevent issues like overcharging.”
So far, AI agents are mainly being used in business. When will we see them in the consumer sector—tools that can access personal information and perform actions like planning entertainment? For example, I want to go out on Saturday, and it knows my preferences, has access to apps, chooses a restaurant, and books it without me lifting a finger.
“I think that’s not very far away. We already carry smartphones everywhere. Soon, they’ll have an AI agent we can interact with that knows everything we allow it to know. It will remember, for example, my wife’s birthday and proactively tell me, ‘Hey Ray, your wife’s birthday is in two weeks. She usually likes this and that kind of gift. Want me to order something now?’ It will also be able to interact with brands and apps. We’ll be able to give it skills—like booking a table at a restaurant, buying something, or acting as a travel agent. And we’ll decide what preferences to share.”
How soon will those agents be available?
“It’s a gamble to make predictions right now because everything is moving so fast. But I think within six months, we’ll see agents breaking into existing consumer apps and industries. Technically, all the necessary capabilities—personalization, real-time voice interaction—already exist. Now it’s a race to market. Companies want to be first to launch agents that give them a competitive edge.”
What kinds of tasks will AI models still not be able to replace—even in 10 years?
“There are a lot of things I don’t see AI doing anytime soon. Negotiation, strategy, and tasks requiring emotional intelligence. Building relationships that are essential for business or sales. AI agents can assist, but many roles are uniquely human. Complex transactions, for example—I think those will remain in human hands.”
What about programming? A lot of jobs at Microsoft—maybe most—involve writing code. But AI models are getting very good at that. Will you still need human programmers?
“A lot of students and schools ask me this. What advice should we give the next generation? Should they go into coding? I think there will still be a need for human programmers. They say it’s easier to write code from scratch. It’s much harder—and requires more advanced skills—to look at someone else’s code and improve it. While AI can now generate apps or software more independently, I still believe we’ll need humans to make it all work together. That’s a uniquely human skill set, and I don’t see it going away.
“Let me give you an example. Two years ago, a friend of mine who’s a content writer thought GenAI would kill his business. But he used the technology to scale content production and ended up hiring more editors. He delivered better content, faster. We’ll see the same shift in other roles. Humans will use AI to be better and smarter.”
More content and products-without more people.
“Humans will always be part of the process. If you use an AI agent, it can conduct research, do reviews, and run assessments before responding to a quote request.”
How might AI models affect digital inequality? These models are expensive to train and run. Not all countries can afford that infrastructure.
“I think it’ll be similar to what happened with personal computers. At first, they were expensive. But over time, they became cheaper and more accessible. I think AI models will follow that path—becoming more efficient and cost-effective. There are already a wide range of accessible tools today. The ultimate goal is to democratize AI and help people everywhere use these tools to develop their skills.”
And that brings us back to DeepSeek. Because the model supposedly reduces training costs, it’s more efficient—so it will be accessible to more people, more organizations, more countries.
“Exactly. Model costs are coming down. They’re becoming more accessible. And the more efficient a model becomes, the more demand there is. The end result is greater usage. When the kinds of AI agents we’ve discussed arrive, we’ll see widespread adoption. Capabilities and usage will accelerate.”