How AI Projects Fail: Women in Data Share What Kills Enterprise AI

Summary

Databricks experts Robin Sutara, Maria Zervou, and Holly Smith share the patterns that most reliably cause enterprise AI projects to fail, with poor data quality and skipping people-and-process work at the top of the list.
This video covers practical approaches to scaling AI adoption including business user hackathons to build psychological safety, framing AI as an intern-as-a-service to set realistic expectations, and using Unity AI Gateway for centralized budget controls and security policies.
Teams are guided to shift from token maxing to token efficiency and to invest in data foundations before attempting to scale AI initiatives beyond proof of concept.

How AI Projects Fail: Women in Data Share What Kills Enterprise AI

Most AI projects fail because companies skip the foundation: data governance and team readiness. Holly Smith, Maria Zervou, and Robin Sutara from Databricks share what they have learned from working with hundreds of enterprise teams. From poor data quality to skipping people and process, they reveal the patterns that guarantee failure and how leading enterprises avoid them.
Learn how to build AI adoption at scale by starting with data foundations, creating psychological safety through a business user hackathon, framing AI as an 'intern as a service,' and managing risk early through Unity AI Gateway budget controls and security policies. Discover how to shift from token maxing to token efficiency, balance innovation speed with governance, and address the real barriers teams face when adopting AI.
🤝

Chapters

FAQs

Why do most enterprise AI projects fail?

According to the panelists, most AI projects fail because companies skip the foundation — specifically data governance and team readiness. Poor data quality, ignoring people-and-process work, and trying to scale before proving foundational value are the most common patterns that guarantee failure.

What is the intern-as-a-service framing for AI adoption?

Framing AI as an intern-as-a-service helps teams set realistic expectations about what AI can and cannot do autonomously. It encourages users to treat the AI as a capable but supervised assistant, reducing both over-reliance and the anxiety that can block wider adoption.

How can teams create psychological safety for AI experimentation?

One approach discussed is running a business user hackathon, which gives non-technical teams a structured environment to experiment with AI without fear of making mistakes. This helps broaden adoption beyond data scientists to the wider organization.

How does Unity AI Gateway help manage AI risk in enterprises?

Unity AI Gateway provides centralized controls covering both cost and security, including budget limits to prevent runaway token spend and policies to block unauthorized tool use or PII leakage. This lets teams move quickly on AI innovation while maintaining the guardrails required for enterprise governance.

Full transcript

[00:21] Hello and welcome back to Data and AI Summit. And we have a bit of a treat for you this afternoon. We were asked to do a women in data segment about generic kind of women in data talking about our careers. We thought that might be a little bit dull candidly because actually we've got a range of experts available to us and so I instead
[00:36] thought, "Hey, why don't we talk about a lot of our AI expertise and something a little bit different that we have not seen before." You've had a lot of guests here today tell you about what they have been doing, uh what's worked really well, you've heard a lot of positive stories. But with me I have two absolutely fantastic guests who work
[00:53] with customers all day every day as part of their day jobs. Been spending a lot of time this week talking to people as well. I wanted to spin it on its head a little bit because it is very useful to talk about the good things. It's also useful to talk about the things that don't work so well as well. And we have
[01:08] a broad range of a diverse perspective here. Uh we've got Robin who works very much at the high level kind of like big enterprise level. Maria who works like a little bit closer to home and then I'm just a lowly practitioner. So I want to bring my perspective as well. But with that in mind I would like to welcome uh
[01:25] Maria and Robin. Hello. Hello. Hello Holly. Good. Thank you. Very busy. Uh now you both have those kind of jobs which candidly I don't really understand. So Robin, if we start with you, can you just tell the people at home what is your job and how do you help customers? Absolutely. So thank you again so much
[01:42] Holly for having me and Maria. We're super excited to share with you. Uh so Robin Sutaria, I'm currently the AI enablement leader at Data Bricks covering all of the Americas. So essentially I own helping the organization as we think about what is the next evolution of our workforce. How
[01:58] do we make sure that nobody gets left behind as we think about how do we ensure that everybody knows how to use AI in their day job and how that might potentially change the workflows, the processes, the existing systems that they have in place as the technology continues to move. I think we would say
[02:14] relatively quickly and recently. Um so a lot of it's focused I actually own enablement inside Databricks. So all of our go-to-market roles or account executives, our solution architects, how do I make sure that they have the tools and systems and they know how to
[02:30] use it. Customer enablement, so how do I make sure all of our phenomenal customers here know how to use our products effectively. And then partner enablement because we have such a huge amazing ecosystem of partners. We want to make sure they can use the products and and can help us really drive this
[02:45] next transformation of the workforce. So enough to keep you busy then, yes. Uh uh Maria, you have a very interesting role as well. Tell people about it. Yeah, hello everybody. Maria Chief AI Officer for EMEA. Yes, my role is very interesting. Yesterday we had a brain
[03:00] date actually discussing what does this role mean and what how is it going to change, you know, and just bouncing different ideas. Yep. But what I do for Databricks is I'm working with different customers across the region on what are the AI products we need to build and how to build them
[03:15] in production long term. They have to stay there. It's not about testing technology now, it's about actually delivering value. And what does it mean? How do we execute? How to bring all the organization with us so it can be technology plus people as well, you know. So it sounds to me, especially if
[03:32] someone is like really getting started at the beginning of the journey, they might come to you and be like hey, we've got some ideas. Where where do we get started? Like where are some good places to get started and where are some maybe not so great places to get started? You do not get start. You don't get
[03:48] start unless you have some the foundation in place. Maybe that place more than your as well, right? So, it always goes back to the data. Yep. So, now, because the data set can be so big, we try to limit the space and say, "Okay, do you have at least some data that good enough that we can work on
[04:03] top?" Okay. All right. And so, we isolate the problem and say, "Okay, what data do you have? What state are they? And let's figure out what we can build on top of this now." So, what is a common ambition that maybe lacks the data for people to do that they maybe not don't realize? I think they don't understand what good
[04:18] data means, right? Okay. Because you may have a lot of assets, but unless someone is owning it, unless someone is expressing it, unless it's being consumed by other people or agents at this point. Okay. So, you old-fashioned data lake where you just like throw all the data in and don't curate any of it.
[04:34] You got no idea what it is. You have to curate. Useless. Add metadata. Yeah. Think now about the whole ontology and business semantics that we have to build on top. Yeah. All of this defines like a good data foundation for AI. Okay, then. Okay, then. Interesting. Uh and Robin, I want to go to you, as
[04:49] well, because uh and you mentioned about kind of enabling, frankly, not a trivial number of people, but I think one of the interesting things about your role is you have such a broad range of consumer when it comes to to AI and kind of the things that people can do with it. Like, you've got, you know, our field
[05:05] engineering team, who are like super techy and keen, and then you've got people who are less inclined. We love your account executives. Uh we love their relationship building. That's what they're great at. Uh they're all right. They isolate them, sure. They
[05:21] they do. They're all super passionate about this, as well. So, tell us again, you know, from a getting started perspective, if you have a bunch of people that maybe haven't taken the first steps yet, like, what works? What doesn't work? Yeah, I think we've learned a lot of things in going through our own transformation inside of Databricks that
[05:36] we're now sharing with customers. It's what I spent all day yesterday talking about is really uh that how do you get started? Once you have that first use case or that first uh sort of pilot or POC, it will not roll out at scale in production unless you think about the workforce behind it, who's actually going to use it. So, internally, we did
[05:53] a lot of things. We built up communities. We've tried to really figure out how do we establish the right personas so that the organization can start to see how do I use AI in my job. I think that was the biggest barrier for many non-technical roles inside of Databricks. I mean, we have a
[06:09] phenomenal technical capability within the company, but it really was like how do we how do we create space that they have the time to think about what are the things they do. So, we did things like an actual business user hackathon where we essentially pulled them out of their
[06:25] day-to-day job for one day and said, "What is one tactical problem that you are trying to solve for?" Whether it's for a customer or for yourself. That seemed to be okay. Although, we were still talking Claude in a terminal, so half of them didn't know how to launch a terminal or install
[06:40] Claude or Isaac, right? So, then we said, "Okay, let's really think about this." We know that the workforce will likely not increase in size as fast as it has previously for Databricks, right? We've grown really, really large as a company. We will not be able to hire as many people as we have in the past. So,
[06:57] what we've been telling sort of account executives is, "I'm going to get you hire somebody. It's an intern as a service. I want you to sit and tell this intern exactly what do you want them to do, check their work, give them feedback." And so, we're finding that that is just opening up a whole another
[07:12] level of people starting to realize that's how I think about an agent, that's how I think about AI. The possibilities are limitless on what I can do. So, I do really want to come back to that framing because I think originally, gosh, especially around January time, this is when we have our company alignment, and it was really, really
[07:29] intensive in terms of like, "Hey, this is the new thing. We haven't quite figured out like what is the way to do this." One of the framings that really helped me is like we're not expecting you to be an AI expert, we're expecting you to be a caveman and bang rocks together and see what happens. And I think that's the maybe less polished way
[07:46] of saying "He does an intern." Which is a nice way of talking about it. But um I think for me it's like this like it's okay if you fail a little bit with this kind of stuff as well. And I don't know about you but we have a executive assistant team. They obviously work really really hard. There's a lot
[08:02] for them to do. But I can also see this kind of like right of passage as like they're setting up their agents and then like once in a while like a message will pop up in like the general Slack channel to every single employee saying "Testing my code setup." And it's like good for you, you know? So there is an element of failure in there. I mean I guess that
[08:18] there's very much difference in terms of failure in terms of, you know, oh whoopsie daisy I sent a Slack message I I shouldn't have done versus oh sorry I, you know, generated a contract and now we're giving away our services for free. I mean like when it comes to managing
[08:33] risk like what works, what doesn't? Well well either. Yeah, so I think yeah oh yeah go ahead. I'll take it. We're both getting answered because we can't answer it. Yeah yeah okay right. When it comes to in terms of risk in AI we need to control it early, right? So
[08:50] we need to apply or understand the risk in two different areas. Yeah. One is what data am I fetching to the LLM to be used? Yes. But also what are the tools the the model this agent needs to fetch to answer user question? So it's what data do I have access but what are all the
[09:06] other estates that the AI can access Mhm. to give an answer back to the user. Yeah. How am I controlling risk across all of this? Well We have a solution on that on Databricks, right? So on the whole AI we have the Unity AI Gateway. Yes. Budget controls, policies, routing to
[09:23] models, everything is controlled outside. And on the data side for us AI and data are the same thing, right? So we can apply permissions from user to data. Very nice to control the risk of are we exposing the data that we shouldn't be exposing to. So, data and AI risk management is
[09:40] kind of connected for us on in my world. Yes. And and Robin. Yeah, I think that I would also add to that. That's the technical, I think, parts of it, absolutely. But, I think for many organizations, they're still trying to rethink the process, the operating model, the governance and controls and policies, because the
[09:56] policy that existed in your organization 3, 5, 7 years ago may not apply today because the technology is so much better than it used to be. Yeah. So, your chief legal officer your chief compliance or risk officer more than likely said, AI, oh my gosh, stay away.
[10:12] We don't write complete risk aversion that they did not want to take any sort of risk, which then inhibits the pace of innovation that the business can do. And so, what I'm seeing is a lot of executive boards go back to revisit to say, do we still need the same risk and controls in place? How much can we now solve with
[10:29] the technical the technology and the platform, which is now phenomenal, right, by the way? And then, how much do we revisit our internal processes so that we can mitigate that so that we can shift the balance of the risk versus pace of innovation? Yeah, I think and I was a bit of a nut. I did read through the data risk policy
[10:45] on AI usage. Bravo. And it is it is very interesting to see the kind of attitude of like, yes, we do expect you to use these tools and that is okay, but you are accountable for the risk that you create, especially from a personal usage perspective. Now, obviously, if you're doing like a
[11:00] massive project that takes, you know, a lot to set up and that is a team effort to go ahead and think through. It's not something you do right at the end, either. Um, I think one of the other things that I've seen interesting I First of all, I did not expect to hear the word token maxing this week as much as I have.
[11:17] Got any thoughts on token maxing? And like, what have we learned from this over the last couple of weeks? For me, we've learned that we've even passed the stage, that now we're talking about token efficiency, right? So, token maxing is, I need to learn, I'm going to burn all the tokens, I'm going to do do do do do do stuff. Now we're like, what
[11:34] is the value I'm getting? Yes. Show me. And this is where we are now. We learn a lot on how much are we consuming, how we do even understand the value from these systems. It was a big learning and now we are trying to abstract the tokens and the models away
[11:50] and try to optimize the architecture behind the systems. Yeah. So that we optimize the token that we need before we even hit the model. Yeah. You know, we're trying to abstract a little bit the model layer and how much we are paying for example to all the different labs.
[12:05] All right. And so Robin, leader board of token maxes, target Oh my gosh, yeah. As a people manager, I got to do it every day. I can see exactly who's using how much. Where do I fall in that? And oh my gosh, please don't put me at the bottom of my own team. How do I make sure I'm doing that well? But I think there is a big balance
[12:21] there, isn't it? I think there is actually a not too much, but if people are using like next to nothing, it's like are they even making use That's also a huge red flag as well, right? But I agree with Maria, there has to be some articulated value that we pre-create at the beginning. Like for me, it's very much about how do I make the team more efficient? Knowing I'm not going to be
[12:37] able to double the size of my team this year, I need to double the size of the business. How do we do that? It's going to be via that agentic usage and how do we make sure that we're using the tokens in the right way to do that? Yeah. And again, we're very lucky at Databricks in that there's a lot of effort sits on the infrastructure so that we can measure all of these things
[12:53] as well. Obviously, we've got I don't want to say a plethora of tools, we have quite a variety at work. We've got a quite the choice. I mean, how much is it worth investing in this to understand how much your team is using on a day-to-day basis? Absolutely. Well, and it's definitely helping me think about how how do I
[13:09] think about helping the team? If you're not using it, why are you not using it? What's the what's the prohibiting thing? Is it you're not comfortable, you don't understand, you don't have the right capabilities? And I think that's just helping us think about the next evolution of the workforce and how we actually help people grow.
[13:24] And how is user-friendly this, right? Because as you said, when we're doing these works with the business, if you open up the the closed terminal, and that is not easy to interact, right? So, how do we make it as simple to people to use? All right, thank you so much. I really appreciate you coming on here today.
[13:40] This has been excellent fun. I'm sure we're going to carry on chatting after this.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.