Why Enterprise AI Stalls Between the Pilot and the P&L | The AI Rabbit Hole 2026 by DVC

Hi everyone.

It's time to talk about AI adoption in enterprise companies. So, for the next half of an hour, we are going to talk about the most I guess hot topic right now in the world of AI adoption. Because everyone seem to have pilots now. Some companies are even having a handful of successful pilot projects. But very, very little of them are seeing some return of investment if any at all. So, today we're having four brilliant guests. Each of them are having different perspective on the topic. So, I suggest we run a short round of introduction and go right into it.

good afternoon everybody. My name is Tarun Raisoni. I am CEO and co-founder of a company called Groove AI. we primarily focus on enterprise AI services to help them with the AI adoption. Focused on security, agentic infrastructure, and the AI infrastructure. Thank you very much and by the way, I love this conference. this is this has got a cool vibe. Thank you.

Sure. Yeah, my name is Raghu Banda. I'm with SAP, the senior director enterprise architect advisor in the AI strategy. I've been with SAP for quite some time.

Various different roles. Last 2-3 years, I've been focused mainly on the business transformations our customers and partners are doing in the space of AI. Apart from that, I also run my own podcast in the evenings and weekends, which keeps me busy.

I'm Kevin Sai. I am in applied engineering, which is what what of Google Cloud, which is the customer-facing arm of Cloud AI engineering. Couple of areas where I focus on where we focus on is around Gemini model quality and also pre-post training.

Hi, my name is Oleg and I'm with JetBrains and I'm leading new business division, which is responsible for building agent systems.

So nice. So Taran, let's start from you. You've seen so many companies helping them adopting AI at a scale. So where do you see AI enterprise efforts usually stall and why?

Yeah, thanks Anastasia. So our experience on the enterprise AI journey has been pilots are very easy to kick off because you can use hyperscalers or frontier models or your own kind of a small set of training and and get started. But when you go at enterprise scale, the first thing you have to think about is really security. And the way you have to think about security is how do you enable scale for all your enterprise users so it is inclusive for all of them keeping the security guardrails of the enterprise in place. The second problem that you run within the enterprise AI adoption journey is the quality of the data. Unfortunately, enterprises struggle with data governance. So you find a lot of duplicate data, poor quality data and the problem is poor quality data leads to poor quality AI where you have a lot more hallucinations.

And the third thing that that we see problem within the enterprise is actually how do you run these AI workloads? Because enterprises have different levels of sensitive data. You know, you could have Claude in your in your enterprise, but if Claude has never seen that algorithm before, you know, do you want Claude to see that algorithm now? So there is like various levels of IP and how do you manage that and how do you scale that up? So So those are the most common challenges I see with enterprises that looking to adopt.

Makes a lot of sense. Do we want to add anything?

Sure, I can add to some of these. We see these challenges on a quite regular basis. I think with SAP as well. We have complex customers with complex landscapes and hybrid landscapes. Starting with a pre-prototype or a POC is pretty easy. We know the data sets. We know what exactly has to be done. These are smaller tasks. But when you get into a realistic or a real-time problem, this is where you have n number of business processes. And we pick up one particular process and you have to understand what exactly is happening there. So this is where things get very complicated like what Tarun was mentioning, the data quality is one of the top things. And to add on top of that, you need to understand the process and what exactly you're addressing there. So this is where many of these things fail when we go into the real productionizing the data and productionizing the actual use.

And third thing is obviously, like you mentioned, governance. And you need to understand what kind of tools you're using and what tools will go in there. Do I know what is going in there? How do I monitor them? So these are some of the things that you'll have to keep in mind. I can go on and on with a whole lot of customers, but I'll stop.

Right, that makes a lot of sense. so it's all about how institutional-wise your organization is actually ready to to be able to adopt, right, AI. That makes a lot of sense. Kevin, from an applied engineering perspective, what do teams most consistently underestimate when they try to operationalize AI at scale?

this kind of dovetails into data, right, quite a bit, right? today we usually are pretty good about scoping out the work it takes to build the AI. But we don't think about the foundation that we have that we're building AI on, right? So give me an example of a not so successful project that we worked on last year. It was a project for building an agentic documentation platform, right? There are two problems we're trying to solve, right? So, in this world of AI, things are going very, very fast. We're actually so so technical writers, we're actually seeing like an increase in bugs faster than they can burn it down, right? For for documentation. That's one. Second thing is the velocity of features getting released are so fast that, you know, I mean, and each release requires new documentation, right? So, there are two problems here.

So, we built a triage agent that looks at the incoming bug, look at existing code, right? And then existing doc, and then look looked at the code and and we're able to the agent's able to give the technical writer a recommendation. So, that allowed us to burn down the really quickly. But, when it came to the the generation agent, we had a lot of problems, because in this new world of, you know, delivering features, you don't have the same level of PRDs. You don't have the same level, you know, of design docs, right? And the human tech tech writer has to go to the engineering team and say, "Hey, what are you actually building?" You know, so they can actually, you know, get the information to write the docs. Now, I mean, the lesson here is if the process is hard for the human, it's hard for the AI, right?

You know, and that's usually something that it that that's usually the part that we kind of usually miss when we design out like a project, right? It's make sure that you have the foundation.

Right, right. Talking about the foundation, Rahul, from your perspective, like enterprise architect perspective, what should be in there at the start of the whole process, so that we make sure that the foundation is in there?

Right, I think it's a great question. I think in the context of enterprise architecture, why I would like to stress like there are like four or five things that we need to focus on. When I talk about this with our customers, because again going back to that, we have complex landscapes, hybrid landscapes and all. Starting with we need to go more about not exactly just about the task productivity, but more on the business outcomes. What kind of an outcome I'm looking into? Am I able to achieve this? And then what kind of tools then I have them. So, that is the first thing. The second thing is what is the business context? Am I able to get the right data for me to attain to this business context? So, what are the different things that are there? So, here I know we have a lot of tools which will bring in the knowledge graph.

The knowledge graphs will build connect all these different points. A data in a particular field, in a table might be different. for example, a normal number like it says 100 or 200. But when you go into the context of this, you need to understand what kind of data it is. What is the context of this data? Is this a purchase order? If this is a purchase order, what happened to this? Where is it being delivered from? What plant location? So, there is a lot of context related to that data. The third thing is you need to go into the integration aspect of it. Where exactly is this being integrated into an end-to-end process? And there are a lot of processes like more than 120 plus business processes. In that we have to go and understand that. The fourth thing is obviously I don't want to go back into that, but governance.

And then finally observability. Observability is where like I need to understand like I have these many number of Now, we talk a lot about agents. I have I'm putting agents in this complete end-to-end process and we are talking about how I can completely automate this process, get into an autonomous domains. But how do I monitor them? So these are some of the things. So these are the five important things I would like to say.

Makes a lot of sense. So it feels like during this complicated process you have to stay on the eagle eye view and zoom out all the time, you know, to to be able to observe what is actually happening and constantly reiterating on that. Oleg, from your perspective, what does it actually take AI to show that real value on the intersection from AI individual productivity gains to the productivity gains that you are able to see on the team or organizational level?

Yeah, thank you for your question. Because I'm not a software engineer, I usually look at this from a like a system engineer perspective. And for me actually it's not about coding, it's about building a software production system. And because I'm also leading like a pretty large engineering organization and most of software engineers are a way smarter than me and they actually try different tools, different actions, but but it was so clear from the beginning it it doesn't lead to to the most successful product organization. So I actually break the system into into three layers, individual layer like developers. then it's it's about skills, it's about habits, it's about trying new new tools and we actually agnostic inside JetBrains. So our developers can use JetBrains tools, they can use I don't know Codex app, you know, Anthropic or whatever whatever they want. And the the approach is freedom, freedom of choice, because developers they like freedom of choice, and we value this.

Second, it's a it's a team level, and at team level you you need to actually bring like a good workflows, and and manage these workflows, and also you need to empower team leads to to understand what's going on in the team, you know, give some some transparency, visibility of agentic impact and agentic workflows within within a team. And at the company level, we are trying to establish some some guardrails to to make sure that the budget is used wisely, that security is not compromised, and there is some some like common enterprise level boring stuff each system requires, because agents in the end of the day is it's a it's it could be perceived as a managed intelligence. This should be traceable, governable, auditable, and and stuff like that. So, if you're success successfully can build this system, and also change your system of work, then I think you can gain much like a way more than just just individual level experiments many companies still running at at this stage.

Makes a lot of sense. it's interesting that we've been discussing that before earlier on, that the narrative right now is primarily focused on individual gains, right? And now we're talking about organizational gains, but Oleg just added an interesting perspective from the team level, right? Because the team is actually the first entity that is actually laying in between individual gains and organizational gains. Tarun, do you also see adoption happening on the team level as well? Do you you know, at all deal with this as the with the sovereignty?

I think in an organization it it usually starts with a team, goes to a department, and scales to the organization. And you know, for enterprise AI to be successful, it has to be very inclusive, right? You cannot make enterprise AI successful if it is just restricted to a few people, right? So, it has to be inclusive, and it has to be very easy to use. So, the user experience of how you build it is super important as well.

May add the one perspective maybe because I think also on this on this on this conference primarily we're talking about like engineering teams, but I think the biggest change I see in my team that actually I renamed technical leads and I I made them product owners. So, because they need to be close to to business results, to customer value. And this change of the mindset and the fact that you can actually bring engineers, product managers, and GTM people into one business context as you mentioned, this is going to be I think the game changer. If you keep these teams separated in a different silos, whatever technology you bring, agents, you know, tools will not change your your your top line and your your time to market. I think this is based on my my experience.

Yeah, I I actually Oleg, I think you you touch a very interesting point, right? With the AI, a good product manager can actually pivot to a engineering and build the alpha. And a good engineer can scale to a product manager and also build the alpha, you know? So, it's very interesting time.

Yeah, that's an interesting dynamics. Rahul, I would want to talk a little bit more about governance and the guardrails. And I wonder when AI systems are actually starting act, not only respond, how should companies think about governance, approval workflows, and accountability overall. What's your perspective on that?

Right. That That's a great question. I think this is the next step, right? I think we are going We are all We know that now. We have a lot of tools involved. We know that teams can use these tools, and then we have the overall organization using these. And there was this one of these presentations, I think, when it happened, then I came to know that like the amount of agents that are being built are now into millions. Like But now you need to know how you have to handle them. We have three to four different things, right? Like you need to understand what kind of data get in gets in there. What kind of operational aspects get into there? And if I'm And what kind of geographical aspects get into there? So, there is a lot of sovereignty-related issues involved there. So, all these kinds of things we definitely have to make sure that they are baked into.

So, I understand like and based out of a particular implementation I'm doing in North America might be different from an implementation I'm doing in Asia Pacific to doing it in Europe. So, this is where things are getting more important. And so, that is how we design systems, how we That is how we design governance and guardrails when you go across these geographies or across the data that you use. So, these are some of the things to get started with.

Makes a lot of sense. Tarun, do you have your own anything you would want to add on top of it?

I think I think I agree with Raghu on on like you know, a lot of these tools that are available. And you know, the one other challenge in in enterprise right now is how do you consolidate all this tooling? Right? Because we have a tool sprawl. We had some before, and then AI just amplified it a lot more. So, so there is a there is an increasing amount of effort at the C-suite on how do you how do you consolidate, you know you optimize it better?

Makes a lot of sense. Kevin, from your perspective, from like

Yeah, I think it's I look at it a little bit differently in terms of process cuz within all the new tooling, it's so easy for people to build as you kind of mentioned, right? You know, PM can do this and just everybody starts doing it, right? You know? And I think that's where I think governance is pretty important, right? You know, we really need to have certain processes. We don't want to stifle creativity, right? But creativity without governance is just slop, right? There's a lot of slop coming out, you know? And then you have different teams building the same thing. Then who who owns Then now you kind of need an authority up on top to kind of say, "Okay, you know, so who has what ownership?" So we I've actually seen we've actually seen pro you know, projects, you know, that you know, are are doing the same thing, right?

You know, when they really should be combined together. And one silo doesn't know the other silo. And that's that's that's another conversation that you had with the different teams, right? Then you have wasted effort, right? You know? Yeah, so there's there's definitely that governance aspect that needs to be part of, you know, the thought process, you know.

Absolutely. And talking about creativity versus quality control and stuff like that, what role do you think a senior technical leader should play in setting that up?

Well, okay. Senior Well, technical leadership there's a lot of things that we have to do, right? But I think in the world of AI, right? I think one of the biggest value that senior leaders can do First of all, senior leaders have authority over a certain business processes, have influence over more, and have visibility over even more, right? Okay? So I think one of the areas that we one of the questions that we should probably looking ask asking is how can I look at a business process and flip from scale through volume scale my work through volume to scale my work through change, right? So give an example, right? So one of the projects that I was very proud I'm very proud of last year was was a bug quality pipeline, right? So we we get the the field files a lot of I mentioned earlier that you know, one of the things we we we work on is gemini quality, right?

So one of the things that you know, that is is that the bugs are filed from the field hundreds of people and it goes to the researchers and the research what we heard from researchers like these bugs aren't usable all right for us. So we built a kind of a bug quality agent workflow and the first time we deployed it, all right? 90% of the bugs were more than 90% were kicked back to the filer saying you're missing this this this and this, right? Then imagine the savings and acceleration that the researchers got out of this. So you don't have a researcher It's not that there's no nobody's reviewing bugs, all right? Our team is was reviewing still reviewing bugs for eval purposes. But instead of the researchers reviewing 3,000 bugs, we'll be reviewing 30 bugs. Right? You know, so you encapsulate that kind of those that logic into the workflow itself and I think that as a senior leader that's where we could we can bring in the values cuz we have we we have visibility over you know, seeing all these different business processes.

That sounds amazing. Well, that's that's the trick I would believe. Cool. Oleg, we've touched that topic a little bit about the scattered tooling and about the fact that each and every engineer right now is becoming their own implementation engineer and also you've touched the topic about the choice and being able to choose. Do you think that that scattered stack is sort of inevitable in this age where we need to still figure out the way of how we scale AI adoption in enterprises or do you feel like there there a way to overcome it.

I think it's it's in in small companies probably probably it's much easier to to standardize across like a few tools and make it more or less like a same same environment. in a bigger companies for example, in in our in my own division, I still okay in like to have fragmentation of tools, different different tools. But they're all and enabled. It means that they they can, you know, developers, product managers, QA can can run different agents. Our position that it's it's okay to have a different tools, but then then if you're building a little layer where you need to govern, you need to control, and you need to run agents as the as a managed services inside inside this this this backbone. So, here you actually cannot afford having like a fragmentation. You need some more or less standard stack. And we also believe that this stack is cannot be closed stack.

It should be quite open and integrated with existing systems. That's why I don't treat it as a platform. I treat it as a rather like a different different layers of technology like for agent orchestration, governance, security, I don't know, code provenance and and stuff like that. And you need to to to if we are talking about the enterprise as a team of teams, then you need some some level there you want to centrally manage some some some stuff. And I don't think company can afford having too many control planes for for software production. Yeah, that it's too too too much. Yeah.

Do you have any your your own perspective on this?

So, I think I'll just go back to like something that Google actually very early coined on in their journey, which I think is very relevant even today. And this was actually a mathematical formula that Sergey Brin came up with, which was essentially 70 20 10. So, 70% of your efforts goes on to the core work. 20% of your efforts goes to make that 70% better. And the 10% is actually completely tangential that you've not even thought about that that takes you further. And I think that is still extremely relevant in the age of AI. So, if you take 70% of the core work of somebody as being a software engineer or a product manager or a salesperson, and that 20% you can augment with the AI, that 70% actually improves a lot faster. And if you spend 10% of your time learning something totally tangential, you can change that 70%.

You you can it can be a game changer. So, I think that's where creativity and how you bring control, governance, etc. should land.

Sounds super reasonable. And I want to close this panel discussion today with a question that I want to address to you all of you. how do you think what in a year from now will differ a company that has stuck on the pilots from a company that has really achieved the AI adoption at scale? Kevin, what do you think?

I I I like the 70 20 10, all right? I I think we I I think in this world in right now there's a lot of there's kind of like a tendency that we do things that's interesting, right? And that 10% can easily, you know, without without governance, right? Become 80%, right? You know? So, I think the the organization that actually is going to do a good job, all right, is going to be able to kind of like do resource management that way, all right, and have a balance between creativity and governance.

Sure, I can add I think a few things here. I think we solved the technology problem here. Now we know that quite a quite a lot of these things you can make it happen. So now you go back to this problem, the question of how do people work with these processes and these technologies? And now it's not just that you are working with these technologies, but now you have a lot of technologies driven by these agents. So how do you make humans work better with these agents? And how can you make sure that whatever is coming out of that, the proper business outcomes that you're focusing on, how is it governed? How can you make it it is trustable for the end user? So that is what is going to be important.

I think that so every enterprise will be challenged by smaller companies who can run faster. And first they actually much more focused on they they're trying to eat your breakfast, you know, I mean, if if every large enterprise currently challenged I think by smaller companies. That's why the the biggest challenge is organizational change. What kind of changes need to be done in the company to to to make it organizationally more adaptable to to the to the current to the current world. And then apply, you know, all possibilities of agenda AI to to to this company. Because if your organization is siloed, people don't talk to people, they don't have like common context, whatever you add to this organization, it will just create like more chaos more frustration and less clarity about direction. If you have, adjust the organization, made it so, you know, I I believe in a hybrid teams like a smaller parts with with a different perspective, business, engineering, GTM growth type of people like even like triangles like like a free person.

I Okay, solo builders are great that but maybe in enterprises I still think, those triangles will win. And then a hybrid team hybrid like agent agent swarms who can actually work, you know, with this team, effectively. And that kind of organizations, I think they will, outperform, their competitors and they can also stand, versus like small startups.

Right. And they're going to be all aligned.

They're going to have the same context so no VP of strategy will ever be needed. Bright future. All good.

so I think if we think about a year from now, I think enterprises should measure themselves on three things. First is growth. so if in the next 12 months with the current momentum, how much revenue growth you have had, how much business growth you have had. The second is operations. Are you able to improve profitability, productivity of your own team? And the third one is retention. Are you able to retain your top talent? And have you improved your customer retention? And all those three metrics actually have to come together. You cannot say like hey, I achieved two or I achieved one. You actually have to achieve all three in this in this next one year because this this year is going to be a a much more much more multiplication of force of AI. So that's that's where I see it.

Thank you so much. I enjoyed that conversation. Thank you.

Thank you.