Sponsored by: Capital One Software |Trust at AI Speed: Governance and Security for the Enterprise

Summary

Capital One Software and CDW examine the real day-to-day risks of AI in enterprise environments, including the Gartner projection that 40% of enterprises will decommission AI by 2027 due to governance and security failures.
This video identifies three AI risk categories—control (unrestricted agent access to all data), lineage accountability (tracking agent actions), and inevitable breach (assuming all data will eventually be accessible by third-party agents)—and shows how tokenization and Unity Catalog controls address them.
Unstructured data is highlighted as the new AI risk surface: RAG pipelines and fine-tuning workflows that ingest unstructured documents can expose sensitive information in model outputs, requiring governance controls applied at the data source rather than at the output layer.

Chapters

FAQs

What does Gartner project about enterprise AI decommissioning by 2027?

According to this video, Gartner projects that 40% of enterprises will decommission AI by 2027, driven by failures in governance and security. The presenters argue that traditional governance methods operate at human speed and must be embedded directly into the data platform to keep pace with AI-driven change.

What are the three AI risk categories discussed in this video?

The three risk categories are: control risk (AI agents having unrestricted access to all data an authorized user can see, including from third-party agents); lineage accountability risk (difficulty tracking exactly what actions an agent took and what data it accessed); and inevitable breach risk (treating all enterprise data as potentially accessible by agents at some point).

How does tokenization protect sensitive data in AI workflows?

Tokenization protects sensitive data at the source by replacing real values with tokens before data enters analytics or AI pipelines, allowing models and agents to process the tokenized representation without ever seeing the actual sensitive value. This means even if an AI output is compromised, the underlying sensitive data is never exposed.

Why is unstructured data considered the new AI risk surface?

RAG pipelines and fine-tuning workflows routinely ingest unstructured documents—contracts, emails, PDFs—that may contain sensitive information. If governance is not applied at ingestion, that sensitive content can surface in AI model outputs or be retrieved by agents responding to user queries, creating exposure that traditional structured data governance does not address.

Full transcript

[00:08] Hi everybody. Um really happy to be here at Databricks and to talk to you about trust and AI speed. I'd like to start off by introducing my co-presenter, Jeff Joe. >> Hi everyone. I'm Jeff. I'm the VP of AI and ML at Capital One Software um
[00:22] working on AI products in the data infrastructure space. >> And hi, I'm Rex Washburn. I am the chief architect of the CDW data consulting practice as well as the head of engineering for the practice. Um really happy to be here and talk to
[00:35] you about this. Jeff, you want to kick it off? >> Sure. So, I am part of Capital One Software which might be a little strange and confusing. So, Capital One Software is actually a B2B software arm of Capital One and we are trying to
[00:49] commercialize technology that we built inside and sell it to other businesses. So, uh we are actually a software sub unit of Capital One. >> Not just a credit card. >> Not just credit card, but if you want that, I'm sure I'll get some credit for
[01:02] that. >> have any applications though, I don't think. So. >> That's right. >> What Go ahead. >> Uh so, today we uh have a a fun discussion to talk about AI and
[01:14] security. The high-level takeaways and goals of today are we want to talk about what are we seeing on the real day-to-day risks of AI. Uh what are our best guesses on the risks of AI in the next 5 years? It's
[01:27] kind of a pure guess at this point but how fast is this changing? Who who really knows? Uh and then we hope that we we all share and learn some best practices, some stuff that we've seen internally and externally and share it with you all
[01:40] here. >> So, I mean, why are we here? Everything's been moving at human speed. Our architectures are human speed. The way we think about data is human speed. The way we design data is human speed. The way we consume it has been human
[01:52] speed up until just a very short period of time ago. And now we started to saying, oh well, okay, we can start to figure this out. We have this wonderful AI thing. We're going to start serving up models. We're going to have ML. And it kind of made sense. It was moving fast, but we had an
[02:05] idea. And then it kept evolving faster than we've really ever had to deal with technology, right? So, now we're dealing with a world where all of our data has to be assumed to be accessible by agents at some point
[02:18] in time. Not just the agents you've deployed, not just the agents you've managed through Unity, but agents from third parties. So, we have to start thinking about data in a different way. Uh one of the things I like looking at
[02:30] is looking at your data as if it's the largest attack surface that we are very good at building as data engineers. And I want to think about how do we actually secure that? How do we treat it treat that data with sensitivity knowing
[02:42] that we're in a world where we can't predict what's going to access our data today, and we really have no idea how it's going to be accessed tomorrow. >> Um here we have a couple of statistics that are from Gartner. And so, they kind
[02:55] of paint the picture of where we think the world's going to go. And Gartner basically hypothesizing that organizations are deploying AI agents AI agents faster than they can govern them. And by 2027, 40% of enterprises will demote or
[03:08] decommission AI due to governance failures. Um basically, Gartner's kind of predicting and with the evidence that they're gathering is that the growth and sprawl of AI is becoming a little bit untenable, and likely many projects will
[03:21] shut down or have to be cut back because of this this unstainable growth. Uh a second metric that Gartner uh talks about is that most AI agent deployments will struggle to meet emerging privacy requirements.
[03:33] And by 2029, again, another projection, AI agents deployments will be non-compliant with AI and privacy regulations. I think this is a really interesting point by Gartner. How are laws and privacy regulations
[03:45] going to change in the next few years? This space is so new and so aggressive. Uh there's bound to be changes in in regulation of as we just saw with Fable and Mythos, the government just pulled that back. >> Mhm. >> Who knows what's going to happen
[03:57] literally day-by-day? It's very It's very strange It's a very strange world that we're all in today. Uh >> Well, in talking to my peers, our field CISOs and such as that at CDW, the number of states rolling out new legislation or expanding their
[04:10] legislation around data, data privacy, and AI is expanding on a regular basis. Just in a month, you can see a couple new states showing up with new legislation on the books. And this really leads us to one of the key gaps I'm seeing across the country
[04:24] as I talk to data leaders is the the data governance debt, right? So, we all have we all know we have tech debt of some kind. It's just natural today, right? But the governance debt is real. It is uh slowing everything down and
[04:37] it's increasing risk. And we'll talk a little bit about more about that cuz it's not just your traditional data governance. Want to cover number two? >> Oh, number two. So, managing the
[04:49] autonomous AI risk. So, managing uh the risks of AI are really tricky because there's so much surface area. So, you have to look at the lineage, you have to look at the history, what is it doing, what's the intent. Uh it's a very tricky
[05:02] surface area to to kind of manage. >> Well, and we have to think about this with all this complexity, and we don't have data governance. Now, we have the borderless data state, right? So, we have all of our agents sitting data that
[05:15] we think we've curated. We also have shadow AI. We probably have shadow data. Now, on top of that, we have data that we haven't really thought of before. So, all of our unstructured data, our data that's sitting in SharePoint that folks
[05:27] are building small agents on that they're not really paying attention to. So, that's just within your own domain. Think about your M&A activities. How many companies have you added where you haven't resolved the tech debt or the
[05:40] governance debt? Now, how about those other country those uh you know, M&A activities or those new members to the company that are overseas. And where you have all these other residency requirements, so you have to think about this expanding in a way that
[05:53] we really haven't thought of before. >> That's right. And the last one, protecting the new AI risk surface. The number one thing I hear from both internally and with our customers is they're afraid of what AI
[06:05] what actions the AI is going to take. And then the big risk is the data that it's sending out. And so how do you build a robust system that is protected, making sure that your data is protected in spite of whatever actions your
[06:17] agent's going to take? >> So, Jeff and I are got started talking about this, we realized we're technologists, we really needed a scenario to hang this on. So, something that I'm seeing more and
[06:30] more, especially with agents coming on board and folks wanting to be able to expand the usage of that, is the growth of contact 360s, especially in larger enterprises. So, just the customer 360 that we're probably familiar with, but
[06:43] expanding that beyond just the customer, because we all have multiple roles we play, right? We can be a vendor, we can be a customer, we can be a commercial customer, we can be a retail customer. If you've handled my master data
[06:55] management, you know the headaches that can happen there. Now, imagine giving that sprawl to an agent, doesn't work. Right? So, we're seeing a lot of customers bring all that together. The problem is is when you bring all of
[07:07] your customer data together in one place, and it possibly is it secure, is it not secure, but it's all sitting in that great lake. That is now what I would say the juiciest attack surface I've seen
[07:19] since I had a Captain Crunch whistle. So, it goes back a ways. Um and then you start looking at all the different units. So, we have a fictitious company, think of a multinational retail company that's done M&A across the years. It has PHI data,
[07:33] it has GDPR problems, it has data here in the US, it has different brands across the country, it has EMEA, it has Asia, it has the UK and it has the US. So, when we're talking about that,
[07:47] we need to think about protecting the data up upstream and protecting it downstream. You want to take this one? >> Uh yeah, you want to make sure that the data is protected kind of at the source is one of the most robust ways in the
[07:59] strategies that we've kind of taken internally at Capital One. So, if you can make sure that at the source it's protected, then ideally whatever happens downstream or whatever actions it takes or whatever third-party integrations it has, uh the the data is well protected.
[08:12] >> So, I'll let you keep going. >> So, with the more autonomy that agents have, the governance really matters and there are three big points here. So, one, the most common thing is limiting
[08:24] what agents can access. This is a probably a pretty obvious one, but it can be kind of subtle. You know, it's not just inheriting the permissions and access of the creator of the agent. Maybe it has its own intentions or some purpose, but we're seeing that our
[08:36] people are treating agents like people themselves and they have their own permissions and policies. It gets really subtle and tricky. The second one is managing lineage and accountability. You want to know where is the data coming from? Where How did it get joined? What
[08:49] was the source? Was it from a third party? Was it from an internal? You really have to make sure you understand exactly what your agents are doing, what actions are taking and and where the data is coming from. And the last one is reducing the impact of inevitable errors. So, one kind of
[09:03] thematic strategy that Capital One takes is assume you're hacked. Assume someone's already inside your system. What do you do then, right? What do you do in that scenario when someone's already inside and they're doing whatever they want? How do you stay protected in that case?
[09:16] >> And I think, you know, as we as we were talking about this, we know Unity, we know Databricks brings a lot of this solution to the table. We know we can limit access. We know that we can manage our linear lineage. We can reach into
[09:28] our data sources across the globe. But what else? What about the data that's to the left? What about all those source systems? What about the raw data? What about the data that you have
[09:40] sitting in a lake that you haven't gone and masked where you haven't or you know, somebody does get back past the ABAC rules or bypasses Unity and oopsie happens as I like to say. So, we can say we're going to apply
[09:53] controls consistently across all platforms and that's great for today, but what happens tomorrow when we acquire the new company? What happens when that when that company has, you know, maybe plants all over the place and they have their
[10:07] own tech debt. What happens when we have to go beyond just typical encryption? Isn't it I mean, I've had How many How many people have had leaders just and think about this. Oh, isn't anything encryption going to help us? Can't you just mask it?
[10:20] Well, that sounds great, but what about all the people that have access to the database? What happens if I just exfil the database? Um this is sort of the the place where I get excited. Why is
[10:33] CDW and Capital One software on stage? It's because of the concept of tokenization. Um and that's really a big key here is when you look at this and if all of your
[10:45] data were tokenized at source. Imagine, the data enters your domain, your realm and your first intent is tokenize it. Well, okay, I now I don't have the same
[10:57] problems with exfil. I don't have the same problems with GDPR because if that format is preserved, I can still do analytics. I can still join my data. I can start managing things like data
[11:10] clean rooms across multiple nationalities and multiple data sovereignties, still allow analytics to happen, but always keep that data safe. So, if it's exfilled, it's useless. Right?
[11:22] That's something that we're not talking about a whole lot because we put we create these massive beneficial lake houses, but they're also massive attack surfaces. >> Yeah, one thing I would continue there
[11:35] is there's always a trade-off of security versus freedom, right? So, ultra-secure, yes, you can lock it down, you can hide it, you can redact it, and no one has access, but then you can't do any analytics, you can't use it in AI,
[11:47] you can't even do machine learning, you can't even send it to anyone to do anything, extract any value from the data, right? Every company wants to leverage their data. Then on the other side, if you're too loose, then you might you might risk some of your data
[12:00] leaking out. >> Yep. >> You might have great insights, but hey, guess what? You've just, you know, upset a lot of your customers. And so, that balance is really the tricky thing, and tokenization is really that solution where you can kind of break that that
[12:12] that trade-off. >> Oh, oh, go ahead, sorry. >> So, in in this slide, looking at, you know, AI governance must extend beyond rows and columns, this is really talking about unstructured data. So, of course, tables are very important, but as we know, I think a
[12:24] good chunk of the data just out there in the world is unstructured. Talking about PDFs, old scanned documents, chat text. A lot of the world is met A lot of the data out there is messy, and a lot of
[12:36] companies are just sitting on it trying to understand can I extract value, but maybe they're scared because it might have PII information, sensitive patient information, social security numbers, credit card numbers, etc. And so, the big question is how do you unlock not
[12:49] just structured data, but how do you then leverage your unstructured data where there's a lot of untapped potential. >> Well, and I think I think that's key because we have mass amounts. I think it was uh there's a stat that 80% of an
[13:01] enterprise's data is locked in structured format or unstructured format. Um as soon as we start releasing that data, right? It's sort of safe. I mean, I you know, you you still hear
[13:13] about people exfiltrating docs, but you're not you're not seeing massive amounts of docs exfiltrated. People like to secure it, make the data nicely formatted before they steal it from us. Um but when we start unlocking all of these
[13:25] structured documents, now we're pulling all of this all this data and we're put we're like moving it into our lake houses or we're running local inference on prem and then we're taking all of those embeddings or moving them up to our lake house. No matter what, we're
[13:38] freeing more data than ever before simply because we want to make more data than ever before available to our agents and our AI. And the challenge comes is that's all freewheeling out there and then every
[13:51] shadow AI group that's sitting in your organization, which I mean, we don't have any. I'm sure you don't have any, but you know, I've heard it exists. >> That's right. >> Um but uh they now have access to the data because
[14:04] hey, listen, I happen to have access to lake house. My A back gave me pro access to it. Now I'm going to pull it all down through the handy little, you know, plug-in on Excel and I'm now going to upload it into Claude and I'm going to start hacking on my data and do
[14:16] analysis. And that's all really good, but it you just released all of the PHI. I mean, if you think back to that sort of fictional company, imagine if I'm trying to grow a retail org and I have traditional retail, I've
[14:29] brought on pharmacy, I've brought on grocery, I have different regulations. Now I have to start knitting all of this together and I want to make it available because every one of my acquisitions is also doing chatbots.
[14:42] They're also building AI and everything. I need to do something today, right? Data bricks will take care of me once I get my data into data bricks. But what about to the left?
[14:54] >> Right. >> Right? What happens when that hap when when when that hap when the compromise happens? It's not if, it's when. >> And the uh similar point on this slide, unstructured data is really important in AI, right? For two big use cases,
[15:06] fine-tuning. How do you fine-tune models? It's question answer pairs. It's all loose text. You can't send a table to go train a fine-tune a model. You have to use unstructured text to train your models,
[15:18] right? And then two, rag and context. Pretty much every booth here is talking about context. >> Yes. >> CEO of Databricks talked all about context this morning. What is context really? It's markdown files. What are
[15:30] markdown files? They're just text files. It's endless endless text files. Loose text. The world of AI is all unstructured, right? That is why you really have to think about what is your game plan for leveraging
[15:42] unstructured data cuz it's it it's coming for you. >> Well, and that's the interesting thing is we're talking about tokenizing data. We're talking about leveraging data community catalog for security. But what about the markdown files? What about the context?
[15:55] How are we going to handle that? Make sure that that's secure and still utilizable by programs that essentially are just looking for a nice joint, right? >> Exactly. >> So, we have to basically secure assume
[16:07] the compromise is happening and we have to be able to to be able to handle the outcomes without really having a hiccup in our business, right? >> That's right. >> So, some final thoughts.
[16:19] Tech governance debt is probably the biggest thing hitting us. And I know this is a whirlwind. We only have 20 minutes and this could take a couple hours. I would probably There's a great happy hour later on. Uh you can come to the happy hour.
[16:31] There's some uh great uh cards here. Stop by the booths and find out about that to talk more, but I would say right now is look at data governance differently. Think of governance as activation. And everybody
[16:44] in your org that says, "Oh my gosh, governance. We've tried that before and it's failed." Just remind them that the governance of the old days where we all sat around and we all said, "No, no, no."
[16:56] That has to go away. We have to activate differently. Um anything on governing you want to add? >> Um I think just the big takeaways, you know, what is the strategy that Capital One has taken at scale is tokenization.
[17:09] That was the kind of looking at all the options of things you can do. That's the one that we landed on. That's what we have here at Databricks. And that's really just protecting at the source. Now, there's a lot more you have to do, but step one is protected at the source.
[17:22] And that's kind of the best way, the most secure way to start. And then there are many other things down the road like prompt injection, guardrails, etc. There's a lot of There's a again, another 2-hour talk on what else can you do in the AI world. Um but that's kind
[17:35] of our talk. The last plug for me, I think we already mentioned it, but we do have a happy hour. It's at the top floor of the Marriott at 7:30. Uh free food and drinks. So, if you're around and you want to talk shop, we'll be there. Uh more Capital One folks, if you want
[17:48] to see a demo, we have a booth over there if you want to see >> Go see a demo. You You really want to go see a demo. If you're in the data space and you've ever thought about, "Hey, I have a lake. What happens when the lake gets Something gets compromised?"
[18:01] That's why I got excited. I I don't really do vendor vendor vendor marketecture. Um this is rubber meets the road. And if you really think about it, if you have a Capital One card in your wallet, right?
[18:14] If you trust them with your data, trust them with your data, right? Because they have it. This is not just some vendor that cooked something up. This is something that they live every day that secures my
[18:27] data. So, that's sort of the way I looked at it. Um and it it it's a big thing. One extra comment on securing the borderless data state. We didn't really dive into this. We only got a minute. Um there's a a change I want to call out.
[18:40] And that is data and security folks in your orgs have to work together for this to happen. Right? You can talk about Unity Catalog and all these features, but then what about the security team? What are they
[18:53] doing with I am and how are they playing, right? There has to be a Venn diagram going into the future because when we talk tokenization, there's person with great security knowledge, there's a data person here, and we're figuring out we need to be working
[19:06] hand-in-hand because the attackers now have the same level of models that we do. And the state actors have probably more funding than half of our teams put together quite honestly. If you haven't
[19:18] looked at funding for bad act state actors, you probably should. It'll scare the heck out of you. But it's actually something that I highly recommend. And then come check out the Capital One booth. Definitely should and grab the little tokens on the
[19:31] chair. So >> is. More swag at the booth. So if nothing else, go get some free stuff. That's always the purpose number one purpose of all of this. >> Yes, exactly. So if you have any questions, please come and see us.
[19:43] Really happy to share with some thoughts with you and look forward to learning more sharing more. >> Thanks everyone. Have a good summit.

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.