Generative AI has been making strides of late that have some experts seriously debating an impending dawn of superintelligence. What does that look like in your everyday Word doc or Outlook inbox?
As Microsoft’s new chief product officer of experiences and devices management, Aparna Chennapragada’s job includes translating the latest advances in AI into products and features that offices use every day.
At the HumanX conference in Las Vegas, Tech Brew caught up with Chennapragada about her plans for the new role, what “agent” actually means, and whether AI is coming for your job.
This conversation has been lightly edited for length and clarity.
You started in your current role at Microsoft just a few months ago. What were your priorities coming in?
I would say three things: The most important one and the one that I’m most obsessed about is this vision of, ‘How do you reinvent work, AI first? Productivity at work, AI first.’ The second one is…people talked about this whole consumerization of enterprise in the past, but I think it gets real now. Because I think with AI for the first time, the tech that adapts to us versus the other way, and so you don’t expect a different kind of experience in consumer versus enterprise. So I do think that bringing that consumer sensibility into making products is the second part of the charter. And then the third one: What is the product development process that is AI-forward and model-forward?
A lot of people are very excited about agents right now. But there’s also a feeling that everybody has their own definition of what that means. How does the way you’re approaching agents differ from some of the other enterprise companies?
There are these three characteristics we think of when we think of agents, whether it’s a super agent or these narrow, specialized ones. One is this natural interaction; you should be able to talk to it the way you do to humans. To me, that’s one of the powerful things versus an app. The second is autonomy. Now that’s not a 0-1, binary thing—it’s more of a gradient. Just like, again with humans, you kind of say ‘OK, day one, I’m going to trust you, but verify.’ And then you grow that capability as you train it, and so on. And then the final thing I think about with agents that is probably very unique to Microsoft is we want these agents to be able to work with you, where you work…so that you don’t have to completely rewire and contort yourself just to work with it.
Do you see these advanced reasoning models that are coming out now as something that every company is going to need, or are they going to be more for specialized jobs?
I would say an element of deep reasoning helps even in the broader horizontal task. So to give you an example, if you really wanted to get all the information possible about a specific person or specific company, then taking longer to compute, of course, will give you better information. But I think the additional classes of reasoning to me are where you have a complex task that you want to break down, and you have some element of planning and some element of the output being verifiable. So a great example is coding. Another great example is data analysis. You want to break that down. You want to kind of verify that this path is the right one, and kind of pursue that. But what I am finding, and what we are finding within Microsoft and the industry will see that, too, is that this is not a plug-and-play, just like take-out-of-the-box-type experience. You have the raw horsepower, in some sense, with these deep reasoning models. What companies will have to do—and what we are helping with at Microsoft—is saying, ‘How do you combine them with the both the datasets that the company has so that we can get more intelligence…[and] how do you actually do some of this customization for that particular resource that is data analysis or researching a topic, or coding?’
Keep up with the innovative tech transforming business
Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.
With agents and the adoption curve, are there companies right now that are actually using full-fledged Microsoft agents every day? Or is that something that’s still very early on?
I think across the industry, it’s early. And to me, I think again, comes back to the definition of agents, and this idea being it’s a spectrum. One dimension, is this narrow versus broad? Narrow would be a code-review agent, and broad would be a generalist agent that does everything. And then the other dimension is the level of autonomy or the delegation. Today…we’re seeing agentic features that people are using. For example, I use Copilot and say, ‘Hey, it looks like I’m doing the same complicated prompt every week. Just run that for me.’ It is lowercase-A agentic because it’s doing stuff on my behalf. It’s like saying, ‘Hey, collate all these emails about these customer issues for this product. Just send me an email.’ But that’s narrow, and that’s still very supervised. And so what I’m seeing within Microsoft and from customers is a shift to right and up. Right meaning more broader skills than these specific things, and up meaning more delegation. Say, ‘Go off and run this thing. Send email to A, B, and C and come back.’
And to me, the underlying driver here from narrow to broad is model capability. But from the less autonomous to more, it is about tool chaining. And it is about figuring out how you have checks.
Early on in the ChatGPT era, we were told a lot that AI would not replace people, but that it would be a collaborator and it would create more jobs than it replaced. There’s been so many more advanced models since then. Do you think that’s still true?
I do, and I think the nuance is this: Zoom back all the way out, like, post-industrial revolution; there’s a surplus of physical energy. And of course, there were a gazillion uses of it, and I think about this cognitive revolution. So there is a surplus of cognitive energy. And today, I absolutely believe there will be a million things that because of the scarcity of knowledge work and the scarcity of cognitive energy, we’re not addressing. And so I actually do think that we’ll be very pleasantly surprised by the number of problems that we can solve.
When we think about this replacement, I think about what AI is great for—we are starting with, what are all the mundane tasks that you can take off my plate as a knowledge worker? As part of the customer research, I was talking to a lawyer as part of the deep reasoning product incubation. And he said, ‘Look, I don’t use more than 5% of what I’m trained on in law school because most of the time, I’m foraging and looking for stuff, I’m kind of collapsing this stuff and I’m turning all these gears to get the work done.’ To me, if AI can actually take a whole bunch of those things away—and of course, be the thought partner, too—then you are actually freeing up the customers to do the thing that their 80-year-old selves would be proud to do.