Scaling AI momentum into enterprise execution

So I am up here. Joining me is Taj and Karina from CoreWeave. Thank you very much. And we're gonna be having a conversation today about probably one of the most challenging aspects of this whole AI journey that all of us are on, which is maintaining momentum through the adoption journey.

So we are all under real world pressure from our leaders to drive adoption of AI. We have to balance kinda top down governance, with grassroots adoption, and we have to build an operating model that turns all of the tools that we're given into transformative business capabilities.

So I'm joined with, by... With Taj, on the technical and operations side, at CoreWeave, and Karina from the people in organizational side.

And these two have just developed an incredible partnership that we're gonna delve into today. And, Taj, based on our conversation yesterday, it feels like it's that partnership that's kept this whole thing moving when things could have stalled.

Absolutely.

So Taj, you see this across the entire market, not just inside your own walls. Let's start out by talking about how you actually categorize where companies are right now in terms of AI adoption.

And specifically, where does CoreWeave fall into that picture?

Absolutely. And thank you, first of all, for having us here and allowing us to share our story here. It's been an experiment of sorts that Karina had started eighteen months ago, and and we and we are seeing the fruits of our labor, but also being very adaptable and flexible in how we're sort of approaching it. So to your question where companies are, all of you are from different sort of, you know, parts of the industry.

One of the way I see it as, you know, some people are ahead of the curve, some people are behind the curve for different reasons.

I categorize this into three or four categories of their greenfield companies, AI native companies like ourselves, who have a little bit of an unfair advantage of, you know, we're starting fresh so we can hire differently. We also have... And particularly in our case, which is another double unfair advantage that we are an AI, know, AI company itself. We offer compute.

So we do have the privilege of having a lot more AI vendors in the ecosystem. So it gives us a little bit of, I would say, head start and advantage of AI adoption enablement. So that's one. I call it the greenfield.

The other are traditional companies which are behind the curve for obvious reasons. They're slow to move. They're also traditional company, and reminds me of my wife when she's in Oracle, part of it was moving fast, which was the OCI part. But part of it was moving slow, which is the traditional part of Oracle.

And then I also look at the fourth is the founder run companies. Founders are particularly vested in companies' growth, and they do...

And regardless of where the company is, founder run companies are also, you know, more sort of AI native. I'm also cognizant of the fact, besides these four categories, some have conservative budgets, budget constraints, etcetera. So you have, you know, different sort of constraints around the adoption. Those are the dimensions I see.

CoreWeave, of course, falls under two things. It's a founder run as well as it is AI native company. And what we've done here is we've created an incredible framework. Again, we've learned it over time.

One thing we do good at CoreWeave, which we've learned is experiment. And we've got three layers. We call it literacy, fluency, and ROI. Those are three stages.

Last year or maybe the first twelve months was really around literacy, which is about prompts and, you know, teaching people how to prompt. And second one is fluency, which is really about building agents. ROI is when those agents are getting massively used across the company.

So we did the first twelve months literacy. We are now doing fluency. We also started doing ROI, but next year will be the year of ROI.

Yeah.

Okay. So literal... Literacy, fluency, and ROI. That makes a lot of sense. And that's something that was developed within IT as an idea.

But you've also had great partners like Carina. So, Carina, what does real organizational change require that tool adoption metrics or things like fluency and literacy alone misses? And how does that end up shaping the model that CoreWeave actually took?

I I I think back to the beginning of our partnership and some of the early conversations we had, we're really trying to understand that AI is...

It's not just a it's not just a technology change that is coming for us. But when you start to implement these agents, they start to, like, work. You've got digital workers and humans working side by side. It's fundamentally changing everything about your workforce.

What the role designs look like, what responsibilities look like, operating models have to change. New roles have to get created. Some roles might go away, but then there's new ones that are gonna get...

You know, come out in order to support the AI that you're delivering. So it's pretty clear to me early on that we needed to have a proper evolution of not just like get people to adopt it, but now how are we changing the way that we work, the way that we think about work, and how we hire and continue to build our talent pipeline? That lens, I think, is sort of the difference that's gonna make a difference in whether a company is actually gonna be able to evolve versus just put a bunch of technology in place.

Yeah.

To follow-up a little bit, you mentioned, you know, some roles might go away, but some new ones are being created. Would love to just like dive in on like, what's an example of one of these awesome new roles that are being created?

This is my...

This is probably our...

This is one of our kismet moments, think, when Tash came to talk to me about this idea he had with like, manager of agents. This is gonna be a thing, Karina. We gotta figure this out, which put us on a journey of thinking about once you deploy all these agents and they're doing this work, who is performance managing. Right?

Who is managing these agents? Who's looking at how well they're performing? What the accountability decision rights are looking like? Evolving them because your business isn't static, your business is constantly changing.

So it's not like you can just deploy something. Just like any product, you have to continue to evolve it. What is that role going to be? What are the core capabilities?

And so for us, that was where we first started. Think now we're experimenting in that. We have people that are starting to sit in these new roles.

They may not be the right ones. They may not be where we end in the end, like, at at the end of the day....

Change in six months....

Candidates that are in these roles, we let them know upfront.

Yeah. I mean... Breed of person to feel comfortable.

Yes. And my hypothesis is that early career folks, which is the group I think that is most concerned about AI and what is happening, I think are potentially like we have a hypothesis that we think that they are right for these roles. Yeah. And we think of them almost more as like this operation sort of engineer. It's not just like an observability role. There's much more to it than that because you do have to understand all the business context.

We kind of feel pretty strongly that it's gonna be critical in order for us to actually pull this off that these roles need to grow and evolve.

And just to add to that, I mean, part of what Karina is saying is these are agent managers that we'd literally talked about eighteen months ago. Of course, we are launching that now. And I think what's most exciting to me is this is the first time in the industry IT is actually legitimizing shadow IT because these will be technical roles in the business. There'll be technical role in a...

On the HR side. That said, we did start about eight months ago on AI engineer role, which is sort of ITFDEs, if you may. And we already have about 20 of them who are churning out agents in in IT. Yeah.

So those are new roles. I know rest of people are saying fretting about employment, but these are early career people that we are hiring. Yeah. And they're they're...

The rules are being created and more and more are needed next year.

And you heard it here. You're allowed to go create your own shadow IT organizations now. So go tell your CIO that. Heard it from Tosh.

How many CIOs are in this room right now?

So set the stage for us on the mechanics. Like, this doesn't feel scalable upfront. So what are some of the things that you do behind the scenes to make this scalable?

I can start with the operating model because it doesn't work without partnership and, I I would say, thought leadership of folks like Arena and leaders like Arena.

The operating model. We have a simple operating model. Again, this was an experiment which has booked wildly for us and now we're expanding it. It's worked with people, supply chain, with marketing, and one more group I'm forgetting.

But essentially, it's four critical roles. One is a leader themselves. You have to have the leader leading in. And we've got amazing stats to show when leaders don't lean in.

We have some departments that have about 30% lower adoption. And we measure adoption with a simple metric. A lot of people look at metric.

I had hypothesized this last year, prompts per user. It's a very basic. It's a lazy number, but start with that. And it becomes interactions per user if you start adding cursor and codecs and others.

But, yeah, back to the operating model. You got the leader. You also need a PMO ninja of sorts. And we have one in the audience sale reports to Karina who's incredible.

And we have four of these for the four different orgs. So you need the leader. You need the PMO ninja, AI ninja in my mind. And then on the...

And that's the spoke. I call it the hub and spoke model. On the spoke side, have IT. Have a leader who does, you know, AI enablement and everything else.

And then we have AI engineers. Maybe there's a fifth part, which is the grassroots developers inside HR. I think if you have five of those activated Yeah. The result is magic.

If any of those are missing We've seen literacy lower, fluency lower, certainly ROI lower.

The ninja that you mentioned. That sounds like a really cool role. What's what what is a typical profile of this ninja?

Karina, do you want it? Because you have a ninja in your teeth.

I do. And and I I I don't wanna make the world sad that I... You're not sure if you'll find yours, but I'm gonna encourage you to go out. I think one of the first things, this is not traditional project management and it's not even a silly like traditional product management in a lot of ways.

You're looking for somebody that has genuine curiosity, that they just look at the world still with wonder and are willing to ask all of the questions that other people in the room don't want.

So a five year old's mindset.

I mean, you can find...

Yep. Exactly. But but with the maturity of a grown adult who can have the role. I I just...

I do think it is a mindset. And I think that's a lot of what we are trying to build. It's definitely a product mindset where you're able to think through impossible challenges. You're willing to run through walls.

You're like, we don't know how to solve this, but we will find a way, someone who's willing to iterate. So a lot of the folks that traditionally work in product will have a lot of this thinking. And so those are good profiles to start with. But you really wanna pressure test because what you don't want is a a pure, like, project manager.

I think there's a little bit of rigidity that can come with that, which is really useful in some areas. But this is a very unknown space, And you're gonna get met...

This person is gonna be met with all kinds of hurdles. They are also the person who will help you navigate your organization. There are a ton of blockers in companies, especially for some of the larger existing companies, that you are gonna run into anytime you try to change, but especially with AI. So you need someone who's not going to be discouraged.

They're resilient, persistent. They really need to understand that product thinking mindset so that they can iterate and, like, get their way in the door and not get stopped.

Yeah.

That makes a lot of sense.

I'm lucky to have a few of those people working with me. So I feel very, very...

And one thing I'll add incidentally, I was just thinking of the four leaders.

Karina certainly is leading that pack in terms of AI thinking, you know, with their Ninja partner. Those four have four Ninjas. They were not hired. They were from the inside. Some are technical, some are nontechnical, but I think it's that product thinking. Yeah. That customer obsession, that curiosity.

Those are critical attributes.

Yeah.

So that's kind of the organizational view in terms of what the people need to be to drive scalability. Karina, anything else in terms of, like, what does redesigning work around AI mean in practice?

Yeah. One of the areas I'm spending a lot more of my energy on right now is thinking about organizational design.

I'm looking at things like manager roles and the way manager roles exist today. And I'm trying to really think about what are gonna be the core capabilities of this role as we start to expand.

What does succession planning look like when I am looking at a lot of the sort of level one, level two work is very ripe for agentic solutions. It just it just is by the nature of the type of work that it is. So how do I think about building that next group of talent? What what is the definition of succession planning?

What is the level and balance of that? There's just a lot of rethinking on the traditional models of org design, scale, right, how you build your bench. All of that has to be reimagined. Within the within the HR group itself, not everybody within the group was experienced around AI when...

With the folks that we hired.

And in order to get them to engage and reimagine the models that they're used to working in, we also have to get them to think about AI differently. So we've we've started using different language to try to simplify things and make it easier. And once you engage in that thinking of being, okay, if this is a role I have today and this is happening and now I'm putting this AI in place and it's doing it instead, How do I use all the design, the patterns, all the things that I know from my background and apply it to reimagine what is the new operating model for these humans? What is performance going to look like?

So you you have to retrain. There's gonna be...

You're gonna hear us say this a lot. This is a mindset shift. You need a new mental model. You need to break things down into small chunks, and talk about them in a very simple way so that people don't think it's a technology problem alone, which could be a barrier to them engaging in the process.

Yeah. The change management is just critical.

It it is absolutely critical. But maybe not in the traditional sense.

Yeah. Yeah. And and that's one of the reasons, you know, I also think Karina's here. Don't have my, you know, CMO here and so the others.

I do feel, and people don't understand the gravity of the organizational chain that will happen in the next three years, it has to be done with HR very closely. It reminded me of Jensen saying, you know, court, IT is the new HR. I feel, yes, it was about agents, but IT needs to work with HR. And in some cases, we're just talking about some companies have blended them.

To me, that's the way to do it. Roblox is the one I I know. Atlassian is another. They've actually blended IT and HR.

And I think we have to force that thinking because as these agents are becoming sort of digital workers, we also wanna make sure the human angle is sort of retained.

Yeah.

The, you know, the human judgment, all those parts. That... To me, this relationship is really critical to show how it really works. And we are happy to... That's why we're here to share that because we are learning ourselves how it's gonna evolve.

So on the topic of agents, yesterday we were talking and you had mentioned that you have this entire process of, you know, registering agents and making sure that agents can graduate from, you know, the the the mad scientist in the basement of CoreWeave designing it to actually being deployed in a more scalable fashion.

And that of the 50 graduated agents at CoreWeave, only five clear, like, this transformative bar Yep. Because they actually rethink a process.

And from what I understand, two of them came from your work together.

So I would love to understand, Karina, from you, kind of the real before and after. If you can describe for us what the agent was, what it's done, and then like the before and the after process.

Start with ask HR. That one feels...

Yeah. Ask HR and also the other one.

I feel... And I feel like Yeah. Copilot one's pretty interesting too. The greenhouse that I'm working. Actually, let's start with that one because that's a little bit more of an interesting story. Right.

The greenhouse... Because you had started that idea which Did. Which my engineers came So and my engineers came and said, our head of HR is going nuts and she's asking us to do something which you've never done.

And I was like, is that...

Was that... Is that the actual words that we're...

Literally. Literally. Yes. So they came to me for like...

It's fair.

Please say it last...... To anyone. Never call your head of HR nuts just for the record. That's...

And and and so my initial reaction was what was it? So I heard it was outlandish idea. And I'm like, do it. You're gonna spend three weeks.

I asked them how much. Do it and let's see where this goes. And now we have a story that we are telling the world, which I spoke about in another conference a month ago. Please.

It feels really good to have people come back later and be like, oh my god. You were right. This is so cool. I'm not gonna lie, I did love that moment. I wanted to take a picture. Okay. For those of you that do work in HR, work in any company that has these tools, you may or may not understand this pain.

We have a tool called Greenhouse that is what people use for recruiting. We have this tool called Workday, which is what we use as our HRIS.

These two tools need to talk to each other in order for us to hire and onboard people. Most companies have this problem even if they have different tool sets. And there is an integration that you can build between these two systems. If you have ever done this, you know that those integrations are always flawed. They're never actually that great. Error proofing, they are better than nothing. They're better than doing manual work.

But they are not awesome. They are still quite ineffective and challenging. So my first challenge to the team as they were talking through all the complexity that they were running into, for example, there was no visibility when errors happened. So somebody was still going to have to run around and have a bunch of meetings and ask a bunch of questions to try to figure out why this thing failed. That felt like a flawed thinking approach. I was like, why would I go live with this? Why would we do this?

Have we thought about what role AI can play in this journey? It is amazing at anomaly detection. That is something that it does very well. It is very good at things that are go to point a, go to point b, like why are we not thinking about leveraging AI in the same way that I have a human today that will check an email, will go to this system, and it will enter the data into this system, and it will ask all these questions, why can AI not do that job instead? And it turns out that it can.

It is not just an agentic solution. It has a mix of things in it, which I think is where some of the pushback was coming originally.

But that integration, our ability to do that, has changed now the way that we think about integrating tools.

Any integration that we're even considering in the future, I'm gonna pause and I'm gonna think, is this the right thing to do? Now we have this level of visibility where when errors go wrong, it sends the error right back to the person that has the information so that they can really quickly fix it, remedy it before it even gets into Workday. So we have fixed the root cause of the issue where it belongs, which is also now changing the recruiting team's process. So now they're changing the way they work to try to improve the inputs into the system.

So it's creating It's like a feedback loop.

It is amazing feedback loop. And there's, like, multiple along the way. It's also now improving our data integrity because now your chances of getting something bad into my system slim to none. It's gonna be very, very hard for you to get any garbage into our our HRIS, which is the core of where we keep all of our data.

And this is something that's going to be able to grow and evolve with us. Let's say next year, I'm sick of Greenhouse. If anyone's here from Greenhouse, this is just a hype...

This is just a hypothesis. I wanna get rid of it. I don't want it anymore. Most companies, when you try to change out a tool, they're like, ugh.

This is gonna be twelve months. All this change management is gonna be so hard. It's gonna take so much time. I'm not dealing with that situation anymore.

Now, the change from one tool to a next is gonna be so much more lightweight. I'm not rebuilding an integration. We're just going to have the model and what we've already built adapt to that new tool. So now I have business flexibility.

I have tool flexibility. I'm designing better processes because I'm focusing on the root cause of the issue.

That's technical debt.

It Totally.

It is gonna grow with us. And I think, again, now that it's built, I think there's a lot of excitement around it because now the thought is, where else can we do this? Where else should we be doing this?

Yeah. And then part of the story that was remarkable was, of course, there were specifics like start date was wrong, employee experience improved, cost centers, which I, you know, deal with in my team. Nightmare. Became accurate through this agent. But the most important part was we were not using a SaaS product in between, which we typically use.

So...

Okay. That got the attention of our board and our c and our c suite, my god. This is...

You know, we hear about, you know, SaaS eclipse is actually happening. Yeah. To us, this was a really good experiment which has become a pattern now to your point of graduating agents. But in...

You know, you might have heard the word graduation. So one of the things that ninjas do, they inventory agents, registry. They measure the risk profile. This is all done in our spokes, not in IT, but the governance model.

Then they graduate agents. Something like this, hey, IT. Can you help? Scale it. This is a concept that we've tried, we've prototyped, and then eventually it's testing, which I'll admit we haven't even gotten there.

But those are four things that Ninja do, and this is part of the graduation thing that happened.

That's just a great example of of something that worked.

We know that not every initiative that you probably have tried to launch has worked.

So what are some patterns that you've noticed that differentiate the success rate between spokes? And how do you hold people accountable? How do you hold yourself accountable in those situations?

I actually love that question. I'll go ahead, Karina, please feel free to add. I love that question because it's about... You know, when we did the literacy, it was all about let's just get people to learn prompting. And it was amazing doing those literacy campaigns and seeing, you know, prompts per user literally go from 60% to 90% in marketing when we did very intentional three hour boot camps with them.

But I think the question is what's really sticking when it comes to fluency for agents. It's not great. I'll admit that because we wanted to go for volumes, and I've seen failures between 30 to 50% in in places, which is also fine. It's by design. Some of the patterns that we are seeing are what you sort of heard from Karina, data not being clean.

And who in the audience agrees you don't have clean data? Probably everybody here. Right?

And so what happened is...

And I'll talk about one of the agents that we had produced. It was most high profile agent. We are a data center company. Things break down like transformers, chillers, etcetera.

And we had a hypothesis an agent could...

When they break down, it's a lot of communication going into the database, finding the vendor, placing an order, getting the RFQ. We identified it. Look. Beautiful demo.

It was shown in the company all hands. This was a year ago or maybe nine months ago, and it was, the best agents in sliced bread. Turned out it failed because we didn't...

Our vendor database was off. We have 250 vendors and the parts number were not there. So it morphed into another agent. How do we clean the data?

With human intervention.

Human in the loop. So think that became sort of a pattern. A lot of agents we realized we didn't have data, and they became data cleanser And maybe there's a whole generation of data cleanser agents before you actually have the high ROI agents. Yeah.

So imagine just being able to pivot. You could never do that in traditional software. Being able to pivot quickly Absolutely. Then be able to to do that. It's just amazing. Okay. So let's shift gears slightly.

We all now have an adoption mandate.

You all must drive your teams to use AI, but you don't have this model. You don't have the hub and spoke model that you've talked about. You don't have the change management in place.

What's the first thing you're gonna do on Monday knowing that you've got this mandate but not the model? We'll start with you first, Taj.

I think start small. Start small. Inventory what you have. Because you'd be surprised some of your...

You don't have to go with the Frontier Labs. Some of your incumbents have. Asana has AI capability. If you have Glean, it has AI capability.

In my previous stint, I was pushing Glean more than I was pushing...

I think Claude wasn't even there than OpenAI because it had the similar capability. So I would say start small, baby steps.

One thing I will say is put your measurement in place. Whatever tool that you have, start with that. You don't have to go out and get expensive, you know, token capability. Again, I'm thinking of non dev. For dev, do need for development.

And have your measurement in place. Be experimented with prompts per user. I think it's a lazy but a beautiful metric that you can rely on, you can stand on because that to me is the foundation of everything. And measure it for each department. Measure it for people.

Talk to people. Don't be punitive. Encourage them. Help them. Because then you can have targeted enablement. I think once that gets going, then you get to the next level of fluency. And so I'll just start small, have a measurement in place.

Karina, would love to hear from you.

Yeah. I I think you just... You you need to start finding the places in your business where it will matter.

You definitely need to start small.

But you're you're building trust and you're kind of... You're building a bit of a brand for what it's going to mean in your company. And so some small experiments are not gonna be as interesting. You're not gonna learn as much from them because they're not gonna push the boundaries enough. So I think you really need to think about, like, in your company today, where are some of those high friction areas, noisy areas where you hear a lot of complaints or you know this is maybe slowing the business down or it's creating. Go dig in in one of those areas to try to find your starting place.

Because if you can find a way to look at a space that is not running that great today and maybe people are only thinking about a SaaS tool to solve it, and you can even come up with one or two small agentic solutions that will move that forward, even if it doesn't solve everything, but it lets the people around it. They will become your coalition of the willing. That will be an RO's eye story that is interesting. You will learn more even if it fails.

The learnings will be way more valuable. It will expedite what you do next. I just...

I wouldn't experiment in the safest of places. And I know that might sound...

Very intuitive. I know. And I and I I get it. But...

But if you're not...

Learning. Yes. And it needs to matter. People need to give it... I was about to say a bad word. Sorry. People need to care.

People need to care. Otherwise, it's just gonna start to sound like what they hear on TikTok or YouTube like AI, AI this, AI that. Everybody's just talking about AI. If you say AI one more time, I'm gonna throw you out the window.

You've got to show them. It takes the fear away. It makes them feel like they can be a part of it and then the company can see the value. So just pick your pick your place wisely.

So start small, leverage AI where it already exists, identify metrics, identify the friction areas, find a champion.

These are all, like, basic things that I think we all kinda miss sometimes. So...

Maybe just one more thing.

This is just because I want to encourage this because we need more awesome ideas. It is very easy to just think about efficiency when we talk about AI because that's the easiest thing to grab on to. People talk about bots. Like, here's this thing. You can do this. It's very easy, like, repeatable.

And that's great. Do all of those things. That is not as interesting as what else might be possible. So make sure you're expanding your thinking. Think about how AI might help you personalize experiences at scale.

Think about, things that maybe you could only do if if you wanted white glove service and you wanted it to be adaptable to the human that's on the other side of that interaction, but you would have to hire tons of people that could hold people's hands. How can AI engage in that area? Think about how you create your your sensing layer across your company. You've got all of this amazing data. We don't have the compute in our own brains to be able to really understand how to process it. How can you use AI to give you better signals about where you should spend your time, where the friction exists?

Try not to just stay in the efficiency lane. That's gonna limit your thinking, and it's gonna... It's just gonna limit what might be possible. So make sure you're thinking bigger than just how do I make this faster? How do I make this more efficient?

Love it. Okay. As we wrap up, we're gonna do some rapid fire. You ready? Okay.

Go for it.

These are not supposed to be gotcha question, but be honest. Okay. Taj, what is the biggest myth about AI adoption?

I think one thing which I used to think is easy is, I would say, it's not easy. Going from that fluency to ROI, it's not easy. Like, when you build an agent, how do you productize it? None of these frontier labs have solved it, so we had to solve it our own way. It was... It wasn't easy. Like, the journey is not easy. I mean, we're giving you a lot of framework, literacy, and AI, etcetera, etcetera.

It's hard.

But it... We do acknowledge it wasn't easy. It did require a village. It required different sort of thinking.

Yeah.

It required being self critical. But I would say it's not it's not been easy, the journey. Yeah. We we might make it look easy, but... Yeah.

Karina, biggest myth about AI adoption.

I just... That any change is going to happen just by having it. I just think that, like, we talk about how this is gonna change everything. Just having the tool isn't really gonna change much of anything. I think that is just such a myth. Like, so just creating an...

Build it they will come situation It with AI.

It's just not.

Yeah.

Humans will always want to... It's like a natural psychological instinct to kind of move to that status quo. We wanna optimize for the status quo. It's just this innate thing that happens to us.

So people will find a way to get right back to where they started from almost always if they can. It's just the nature of how we operate. So biggest myth is that just by adopting a tool, you're gonna actually create change.

Alright.

So today and what you've just heard isn't just a story about tooling, it's a story about a partnership.

Two people who refuse to let we don't know stop them from building a plane and the runway at the same time.

I have to assume at times that was probably a little uncomfortable, and that's okay. So if you leave with one thing, let it be this, stop waiting for the framework to arrive before you start.

You both did not have a framework. You built one along the way. So thank you so much for your candor in joining me up here, and thank you all for being here.

Thank you for having us.