How does a nightclub owner in Kathmandu end up becoming one of enterprise tech’s foremost AI engineers? By demonstrating a facility for solving fiendishly difficult problems, argues Yashodha Bhavnani, one of Nepal’s most successful after-dark impresarios and the new head of AI at the content platform Box. Straight out of college, Bhavnani spotted a gap in the market for local-friendly after-hours leisure spots and set up a venue designed exclusively for Kathmaundis. All it took to get from there to Box, says Bhavnani, was an instinct for “breaking new ground [and] finding a problem, targeting and solving it.”
That’s what happened at MITRE where, after leaving college, Bhavnani worked on complex data dissemination challenges for the US Census Bureau. After that came an 11-year stint at Google, where she helped shape its Assistant, Bard and Gemini products and helped build a measurement system for LLMs to bridge the models’ understanding of perception and truth. Then Box came calling. “I knew the intersection of AI and content in the enterprise would be the first area where we will find the most value,” recalls Bhavnani. “That’s what attracted me [to the role].”
In the following interview, edited for length and clarity, Yashodha Bhavnani discusses the barriers to successful AI adoption and why AI agents are the next revolutionary workplace tool.

You clearly see huge potential for AI and enterprise content, and you’ve previously said generative AI means the latter will start working for the former, but what does this mean exactly?
This keeps me up at night because I’m really excited about it. Let me explain. For example, I very often need to compile a report that requires research across my content. A few of my PMs [product managers] and engineers have sent me proposals and I need to report to the board on our product strategy. Imagine I have an agent that can access my content and do what I tell it to – scan my content and write a draft for me. This emergence has already happened. This is the next play for making knowledge work so much easier for the end user. It’s all about saving time. It’s all about taking away burdensome, annoying tasks.
The other intersection I find exciting is [how] AI can structure content. During audits, financial institutions were getting people to take hours out of their day to categorise yearly statements, etc. Now they can use data extraction AI to save many hours and do the job with greater accuracy. AI can read those documents and pull out the precise information, unlocking your data.
In your experience, when trying to adopt and integrate these types of technologies into their processes, what is the biggest mistake companies make?
One is philosophical. There’s a worry, an apprehension, about this technology, about how innately secure it is or that it’s going to take away jobs, or it’s going to have the wrong outcome. This philosophical barrier is real and something we should talk about. What I always tell customers is to think of AI as the next tool that has come along, just like the calculator or the internet. It’s very important to start to embrace it now, because otherwise all your competitors will start becoming more efficient, more accurate than you – all because they’re using a better calculator.
Another is the perception it’s a fad. AI is very much here to stay; the technology has crossed the tipping point. It’s not only getting smarter, but we see a trajectory where, in certain areas, AI will cross human knowledge. And it’s getting cheaper. Both of which equate to it becoming more accessible.
I have also seen people use AI for something that’s not a real problem just to test it. And then it has no value. You must think of AI as a tool, and you must use it to solve either a problem that’s really annoying, burdensome or just requires automation, such as data extraction. Or find the hardest problem you are spending a lot of time on and see how AI can solve it for you.
Box is preparing for AI agents to join the workforce, but many complain a standard definition of an AI agent doesn’t exist yet. How would you define them, and are we in need of an official definition?
The way I’ve been looking at it is, AI agents are basically autonomous workers. Very simple. An autonomous worker highly skilled in a certain topic which can deploy tools, meaning it can use an API, call its knowledge source, etcetera. And it can interact with other systems, including other agents. And when you put this very simple definition together, what you get is a helpful tool. That’s the way I think of it. It’s a very helpful tool that can interact in a natural language format with humans.
I’m confident agents will be a part of the workforce. Today, we work with ourselves, and we have tools and machines. Tomorrow we will work with other autonomous agents. And I think some AI agents will surpass what humans can do. You can pass some of the burdensome work to an agent and become more creative.
In terms of a definition, what usually happens is that different companies launch different agents. There are some similarities, and then the definition sort of becomes cohesive over time. And I think the cohesion is now on a conceptual level, at least in Silicon Valley.
The automation of low-level tasks or entry-level roles through agentic AI, as you have explained, is likely to change the way future professionals are trained. How should people and employers prepare for that?
The way to prepare for it is to just use it as a tool. When the calculator came out, we were like, ‘Oh, here’s a math problem, let me try the calculator.’ It’s exactly the same. Whether it’s a low-level data entry role or a high-level executive role, all professions will be impacted in different ways. Not essentially replaced but impacted.
The boring stuff can be done by an agent. What can you now do with your newfound time? I firmly believe this is what even entry-level folks are craving for, which is, ‘How can I add more value?’ Nobody wants to go through a thousand-word document. It’s probably going to be a relief to a lot of people.
What do you think is the next step in this product’s evolution?
I personally think the next evolution is going to be exciting, which is the ability for agents not only to do things for you, but interact with other agents. I think this is really important because there’s no point in a very intelligent system if it can’t interact with other things. I don’t know any customer who has all their files and all their knowledge in one place. So, if you think about systems of systems, the ability for different agents to interact on your behalf, to get the right insight, to help you with the workflow process, that is just insanely valuable. And I think this is the next leap that will happen.