Most companies will never train a frontier model. Their AI work happens somewhere less visible: the last mile, where new models meet a decade of internal systems, data permissions, and org charts.
That is where Utkarsh Shrivastava works. He leads engineering for enterprise AI at Netflix, where his organization builds enterprise search and the company's agentic suite. His team built a centralized, paved path for agentic work, including a marketplace where teams publish their own MCPs, skills, and plugins, and an enterprise search that had to honor every existing security permission. Their tools now have been widely adopted across the company.
Ahead of his talk at AIE Code Summit, we talked about how that rollout happened, why the hardest part turned out to be ownership, and why the next 12 to 18 months come down to proving ROI.
This interview has been edited for length and clarity.
Give me a high-level view of your background and what you're working on now.
I lead an organization here at Netflix focused on workforce AI, so I'm responsible for enabling GenAI for Netflix. Of course, I'm not the only one doing it here, but I sit in a platform org, and what I do has company-wide implications.
In my role, I work with frontier labs and I constantly evaluate cutting-edge technology to see whether we should invest in it. Before this, I wore many different hats. I led global customer service for Netflix once upon a time. I did interesting work at the intersection of software and data, including building a data lakehouse for the company. And before Netflix, I was at Yahoo doing B2C, working on Yahoo Mail's flagship products.
What's your relationship to AI?
That's an interesting question. I haven't been asked that before. If I have to truly explain it, I would say right now it feels a lot like a cat-and-mouse game. Not all companies are at the frontier of research with LLMs. So my job is to make sure I'm aware of the latest and greatest technology out there, and to truly evaluate it in a way that can scale to the company's needs.
That's also a focus of my talk. I'll be talking about how frontier labs are only the tip of the iceberg. Companies that have been around for a decade have a lot of historical context to implement. So I'm the guy working on the last-mile integrations with GenAI.
You rolled out one shared assistant to all of Netflix. Walk me through it: the starting point, what launch actually looked like, and where it is now.
That's also the basis of my talk, so I'll be sharing a bit more there. When AI was a new thing, each team was trying to do its own stuff, and that looked wildly different. We wanted to encourage that. However, we needed a centralized place and a paved path for doing agentic work at the company.
So I worked with several frontier labs to evaluate which technology we would want to integrate, and I worked across the various enterprise pillars to make sure we had a centralized way of doing things. That meant coming up with a strategy for MCPs. These teams own their MCPs, skills, and plugins, but they needed a marketplace to publish them, for example. They needed guidance: if two adjacent areas are advertising the same set of data, who wins? Which one should take priority? Or should we not open up everything to everyone? Those are the kinds of questions I had to navigate, and we had to do it ASAP. We did it on a very shortened timeline.
As for where it is today, we have company-wide adoption. Adoption is broad, and it’s become the way people work day to day. It has gotten to the point where it's the new norm. Day to day, this is the tooling you use.
Who were the early heavy users inside Netflix? Was it who you expected?
This was also an interesting journey. It was a mix, and it followed where leaders were most eager to try new ways of working. As the industry moved quickly, teams across the company wanted to try the new tools coming out. Before anyone could, we had to make sure those tools met our security and data standards. As options multiplied, it became hard for teams to keep up, so we took on that work centrally and gave them a standard, supported way to get started.
The teams themselves are the experts in their domains. We provide the platform, and they decide how it fits their work. That split let each team focus on what it knows best while we handled the common foundation.
Did teams like marketing, finance, or engineering adopt faster than others, or was it about the same?
No, they were not the same. A lot depends on what their leadership wants them to do. Adoption followed where teams were most eager to try new ways of working. Largely, the company direction was: yes, we are moving, even though individual timelines may vary.
This is where I learned that some leaders and individuals within those pillars were very technology-savvy and wanted to move fast, so we saw wider adoption there. In fact, I can call out one enterprise pillar where the leader was really passionate about not only adopting newer technology but spreading it into their day-to-day workflow. Those early partnership wins helped us prove the value. Because one of the key questions in our industry is this: what is the productivity measurement? That's an industry-wide question we are all trying to answer, and these early adopters helped us make progress there.
Are you mostly building internally, or building and buying?
It’s a mix. We evaluate the best of what’s available, from frontier labs to open-source models, and build the last mile that fits how Netflix works
Speaking of the last mile and security: how do you handle permissions, what people can access, and how they find things through enterprise search?
A good way to look at this is before and after the advent of GenAI. Before, everything sat behind its own walls, and each team only had access to its own data. It was very clean that way. We built enterprise search to respect existing access permissions from day one. We worked on enterprise search as well. Before GenAI, each team’s data sat behind its own walls. We built enterprise search to respect existing access permissions from day one. Then some business units came to us and said, ‘We never got access to this kind of information because of the earlier architecture. Is there a way you can enable that for us?’ So we systematically took some of those use cases and enabled them more broadly.
All of this needed a methodical approach. It was all very new. Everyone was trying to wrap their heads around what it does and what's coming. It had so much potential. We went through a lot of interesting times, especially because the current workforce setup was before genAI arrived. Things were done a lot differently. We wanted to make sure this was setting us up for the future. And that's where all the leaders wanted my opinion, because they wanted to know how to think about some of these things ahead of time.
You mentioned departments getting access to information they never had before. What's an example of something that couldn't have been done before and is possible now?
I want to answer with two examples, both things I'm very excited about. First, it is faster and cheaper to build than before. Earlier, if a business user wanted access to data, they would reach out to their data analyst or data scientist, who would mine that information and curate a graph for them. With AI, all of that is possible, so they can access it very fast. That has truly been game-changing for us. The business can finally roll up its sleeves and mine a lot of information that otherwise wasn't available.
The other part of your question was what's possible today that was never possible before. I wouldn't say there's anything that wasn't possible, but the cost of building something like it was so steep that we would never do it. We have one agentic suite where all the MCPs related to all the enterprise pillars are integrated. Think of a database-like situation where you want to index on a primary key and get all the information related to it. That's just a click away now. We can modify permissions, the datasets start pulling in data, and then you apply LLMs on top of it. That does the magic. None of this was possible before. We would have had to have large areas of investment to enable something like this.
A rollout at this scale is never seamless. What did you learn that you'd tell another company to watch out for?
I would answer that with one word: ownership. Depending on the architecture, who owns which part of the data and the experience changes, and that needed careful alignment. That was the hardest thing to navigate: getting every team clear on its role.
Netflix's culture has a very clear reputation. How did that shape getting everyone to move in the same direction?
Because Netflix has that transparent culture, I would say it was easier for us to navigate this transition. People had good questions up front, which is why they were asking questions. I got questions like yours a lot, throughout the year and beyond. But once they understood how we were going to roll it out and what our intention was, they put their full weight behind us.
And that matters, because they're the ones who have to adopt it, use it day to day, and get value out of it. The culture of freedom and responsibility, and context, not control: those are the elements that really helped us get to where we are.
Where are you headed with these tools over the next 12 to 18 months?
Over the next 12 to 18 months we’re focused on measuring how these tools improve the way teams work, so we invest where it helps most. But we also want to power our use cases and the work we're doing to entertain our users, which is our primary goal.
Is there anything we haven't covered, or something from your talk you want to dig into?
In my talk, I conclude by saying there is no one-size-fits-all, and you really need to build the culture. Building an AI-forward culture in your company, and explicitly calling it out top-down, is a huge win. What I've seen is that your board of directors and your C-suite have a lot of desire to do this. It's really the next level down that has to take the baton and implement it for the company. That layer has to work absolutely hand in hand with the industry, and they should have a goal of transforming the workforce with an AI-enabled toolkit.