Who Should Own AI Risk With StackAware
The Business of CybersecurityNovember 17, 2026
56
00:28:3226.14 MB

Who Should Own AI Risk With StackAware

Who should accept the risk of an AI system when the benefits belong to the business but security and compliance teams are expected to manage the consequences?

In this episode of Business of Cybersecurity, I speak with Walter Haydock, founder and CEO of StackAware, about the decisions behind responsible AI adoption. Walter's background spans Marine Corps reconnaissance and intelligence, national security and enterprise technology. He explains how those experiences shaped his approach to uncertainty and why delaying a decision can create risks of its own.

We discuss his argument that the business leader benefiting from an AI system should own the associated risk decision. Security, legal and compliance teams bring expertise to that assessment, but Walter questions what happens when they are asked to accept responsibility without receiving the business benefit. I also ask what evidence a board should request before approving autonomous AI. Walter recommends comparing expected value, operating costs and error rates with the existing process, alongside the effects on confidentiality, integrity and availability.

The conversation includes his account of working with Embold Health on governance, risk assessment and security testing for EVA, its virtual assistant for healthcare navigation. We consider how natural language can help people find appropriate providers and why the controls around an application need to support its intended purpose. Walter also outlines questions smaller organizations can ask AI suppliers about data use, retention and output quality, while recognizing that the answers depend on the service and agreement involved.

We then discuss governance as ongoing work. Models, interfaces and integrations can change after a system is approved, so teams need a way to notice different behavior and reassess whether the service still meets their expectations. Walter describes how a tiered approach can give lower-risk activities a clearer path forward while reserving greater scrutiny for more consequential uses. For leaders trying to make decisions without waiting for perfect certainty, the discussion connects opportunity with responsibility. How is your organization deciding which AI risks to accept, and who remains accountable afterward? Please share your thoughts with me.

Useful Links

[00:00:00] Once a year, I open up a handful of places for people who are serious about launching their own podcast, but have never actually gotten around to it. And over the last 11 years, I've helped entrepreneurs, business leaders and Fortune 500 companies all launch their own shows. And here is how it works. You record your conversation on Zoom, Teams or whatever works best for you.

[00:00:23] You send me the audio, an episode title and short description, and I will handle everything else from setting up the podcast host and distribution to getting your show on Apple Podcasts, Spotify and the many, many other podcast platforms out there. I will commit to working alongside you to help you launch your first season of 12 episodes in the bag. So you can start 2027 by ticking a season of podcasts off your bucket list.

[00:00:52] Think of me as your not just personal podcast producer, but also your accountability partner to make sure that you actually deliver those 12 episodes into podcast land. Now, I only offer this once a year with limited places available. So if you want 2027 to be the year you finally launch that podcast, head over to techtalksnetwork.com. Send me a message, send me a DM or even an audio message. I'd love to help you make it happen.

[00:01:24] Who should own the risk when an AI system promises better business results while also introducing uncertainty? So security teams, they can assess vulnerabilities and legal teams, they can advise on obligations. But the harder question is who has the authority to weigh those concerns against the benefit, make a decision and remain responsible as that system changes.

[00:01:52] So today on Business of Cybersecurity, I'm going to be joined by Walter Haydock, founder and CEO of Stack Aware. His career has taken him from the Marine Corps of Renaissance and Intelligence to national security and enterprise technology. All with a focus on making decisions under uncertainty. So today we're going to discuss what boards should ask before approving autonomous AI,

[00:02:22] how smaller organisations can assess suppliers and why governance needs to continue long after deployment. And Walter will also share a healthcare navigation example to explain his approach to assigning risk ownership. So if you're a leader trying to connect AI adoption with business value, this conversation will give you a few practical questions to bring back. But enough from me, let me introduce you to him now.

[00:02:52] So a massive warm welcome to the show. Can you tell everyone listening a little about who you are and what you do? Sure, Neil. Thanks for having me on. So I'm Walter Haydock. I'm the founder of Stack Aware. And we help AI-powered companies in highly regulated spaces like healthcare and financial services measure and manage the cybersecurity compliance and privacy risk of artificial intelligence.

[00:03:18] We primarily do that through implementing the ISO 40 2001 framework, which is an emerging standard for AI governance. But we also work with customers on complying with regulatory obligations like HIPAA, EU AI Act, and the emerging range of AI-specific laws in the United States and across the world. Awesome. And before we dig deeper in a lot of this stuff, I always like to try and find out a little bit more about my guest's origin story.

[00:03:47] And yours just jumped off the page. I mean, you've worked across military intelligence, national security, and enterprise security. Tell me about how all that came together and how you got interested in this space. I feel like there's got to be a story there. And please feel free to share it as long as you don't have to kill me afterwards, of course. So the common thread of my professional life has been learning how to manage risk effectively, especially under conditions of uncertainty.

[00:04:15] So I spent the early part of my career as a Marine Corps reconnaissance and intelligence officer, advising commanders on risk, and then also leading a reconnaissance platoon myself and dealing with a variety of different risks at home and abroad. And then moved to work on Capitol Hill as a staff member for the House Committee on Homeland Security,

[00:04:42] overseeing counterterrorism and cybersecurity programs, primarily at the Department of Homeland Security. And then I decided that I had had enough of dealing with chaos and thought that I would just, you know, have a very calm life in the corporate world and discovered very quickly that that was not the case. That did not happen.

[00:05:05] And corporate America, global business organizations have their own degree of chaos, uncertainty, things like that. I worked at a publicly traded company at a startup and eventually decided that I wanted to go out on my own. And as every, you know, founder, entrepreneur necessarily believes, you know, it must be true that he thinks he can do it better than everyone else,

[00:05:33] because otherwise he wouldn't be founding his own company. I went out and founded Stack Aware with the goal of helping companies manage risk in a cost-effective, in a rational way, especially in light of changing technology, changing threats. So that is my origin story. Wow. I mean, when you zoom in on that, we've got what military intelligence, national security and enterprise security.

[00:05:58] They all seem very different, but I would imagine there's a lot more in common than many listening might imagine. So I'm curious though, which lessons from these environments are most often missed in some of the corporate AI risk programs? The most important lesson that I derived from my time in the Marine Corps was the ability to make decisions. And that is probably the most lacking thing I find in the corporate world.

[00:06:26] As a junior officer, as a platoon commander, as an executive officer, I was necessarily given the responsibility and the authority to make decisions quickly based on available information. Something I was always trained to do was, especially in combat situations, you need to make a 70% decision, not a hundred percent decision, because if you wait till you're a hundred percent sure, then it'll be too late. And I think that's applicable to the corporate world as well.

[00:06:56] There are a lot of smart people with a lot of education, training experience, but even the most senior of them, I find can sometimes be hesitant to make decisions because, you know, there's, there's the element of accountability that, that comes with decision-making. Um, but I've always found that if you play it too safe, then you're actually being dangerous. You're, you're taking more risk than you want to.

[00:07:23] Uh, if you're, if you're playing it so safe that, that you never make a decision, you're actually taking on risk as well, because you're just accepting a lot of risk in a implicit way without explicitly stating it.

[00:07:37] And I think when looking at cybersecurity, it is very easy to get bogged down on the test, the technical aspects of it, but I've looked for it if possible, if you can share an example of how AI risk that might have looked technical at first, but was actually caused by some maybe cultural aspects of an enterprise, things like ownership incentives or, or just poor communication. I'm sure you've got a few war stories there too, right?

[00:08:01] A key problem with AI risk, but also risk in general is identifying owners of that risk. And it's always very challenging to find someone to willingly say, I own this risk, especially if the rewards are not aligned appropriately for that person, which is why I always suggest to my clients that a business leader is the person who owns a system and all relationships.

[00:08:31] They're not really relevant risks because that person is going to be, the benefit from that system operating and helping in, you know, generating revenue, reducing costs, things like that. So for every AI system, there must be an owner and that owner should be the responsible business leader who will derive the benefit from it.

[00:08:52] If you start having folks, you know, technical folks or security or compliance folks who are risk owners, then you put them in a very strange position because they're not necessarily deriving the benefits of the system, but they would be on the hook if something goes wrong. And something that I say is if I'm ever, you know, back in a corporate environment where I'm not the founder of the organization, and obviously I'm the risk owner for everything at StackAware now because, you know, I'm the only employee.

[00:09:21] But if I were at an organization where I'm not and someone comes to me in a security or compliance role and asks for a risk acceptance, the decision will always be no. It will always be denied because why would security or compliance take on any additional risk when it's their job to absolutely minimize it? So that doesn't necessarily mean organizations shouldn't accept risk. It just means that business leaders should be the ones who are taking those decisions.

[00:09:47] And I think understandably there is certain aspects of caution in the C-suite around AI agents, teams of digital workers acting on their own with their own identity and a question box around accountability too, of course. But if there's a board member listening, a leadership team, what evidence should a board be asking before accepting that an AI system is safe enough to handle business processes autonomously?

[00:10:14] For board members who are considering authorizing agentic AI deployments for their companies, a couple things should be laid out ahead of time. So first of all, what is the anticipated benefit of this deployment to the organization? So if we're in a marketing situation where it's going to be putting out content for the company, then that's certainly a reasonable use case.

[00:10:41] Another potential use case, we're seeing agentic AI being deployed in healthcare, which is where I operate a lot of time. We could see a lot of benefits from that. In the US, we have an affordability crisis when it comes to healthcare and accessibility problem. So AI can certainly assist with that. Conversely, at the same time, I'd want to see a quantitative risk assessment on this deployment.

[00:11:05] So that means compared to business as usual, what sort of impacts to data confidentiality, integrity, and availability do we expect to see? So if we're expecting to see huge reductions in cost and improvements in patient accessibility for a healthcare system, and we're not expecting too much additional risk from the deployment of the AI system, then it makes sense.

[00:11:32] But conversely, if we're seeing error rates from an AI agent that exceed a human and the costs are high from tokens being expended, things like that, and the benefits are marginal, then I think board members should drastically reconsider those decisions to deploy AI systems. Because merely using AI for the sake of doing so doesn't really make a lot of sense.

[00:11:58] So it all comes down to a risk-reward calculus, and that should look broadly for all the stakeholders, society. But at the end of the day, it's a risk-reward decision, and being explicit about that and clear about what you're willing to tolerate and what you're not is key for any board member when considering an AI deployment. And you mentioned healthcare then. I know you've had a lot of successes there.

[00:12:23] Are you able to share any of those successes, just so we don't focus too much on the dark side of cyber and AI, et cetera? What kind of successes are you seeing there in your day-to-day work? One case study I'd love to talk about is Embold Health, which is a firm that works in the care navigation space. So they serve very large companies that have thousands and thousands of workers.

[00:12:47] And a key requirement for these businesses is keeping their workers healthy, as you might expect. Because if a worker is out sick or injured, then that person isn't delivering value. And then also that person is out sick and injured, which is bad for that person.

[00:13:05] So there's a clear combined incentive, aligned incentive between the employee and the company to get the best possible care, certainly at the best cost and most effectively. So Embold Health helps to rank doctors by their outcomes.

[00:13:24] So if you see Dr. A, who's prescribing lots of unnecessary medication or operations, you know, and you can evaluate that in an objective way based on how long a worker is out, is absent from work. Then you can say, well, maybe we should redirect our folks to Dr. B, who is, you know, more conservative in the treatments, or maybe not.

[00:13:52] Maybe they're more aggressive in treating certain things that, you know, the intervention makes the worker healthier over the long run. And Embold Health has a ton of data on this, and they wanted to make it accessible to individual users. So the workers of their clients in a natural language way.

[00:14:13] So they built a chat bot called Eva, the Embold virtual assistant that allowed people to communicate in a natural way. So, for example, previously you would have needed to know that if my foot hurts, I need to see a podiatrist. Now, not everyone knows what a podiatrist is, and you might have had to struggle with, you know, the system to look up, you know, what the right type of provider is.

[00:14:42] But Embold Health will, with the Eva system, channel you to the right series of doctors or individual specialists for your condition, and then also provide rankings on the effectiveness of that provider. And it does this all in a natural language format so that everyone, irrespective of background, can communicate with the system.

[00:15:04] So Stack Aware worked with Embold Health to do a comprehensive risk assessment, governance review, and penetration test of the system prior to its deployment. And it was so successful that in summer of 2025, Embold Health was acquired by another company called Quantum Health.

[00:15:24] So the rollout of Eva was a strong success, and part of that was due to the Embold team's focus on governance, security, and privacy throughout the process, making sure that they could deliver effective outcomes while not taking unnecessary risks. Do you need AI agents that you can trust?

[00:15:49] Well, with an AI data layer providing real-time connection within your data platforms, you can trust your agents to provide accurate solutions. So scale your business by trusting your agentic AI accurately getting the work done for you. Trust its capabilities with Denodo. And you can do that by simply visiting denodo.com to learn more. Love it. Love it.

[00:16:14] And for anybody working inside a smaller organization, how can they assess third-party AI tools when they don't have the luxury of a large security or compliance team? Any tips there on what they should be doing? The top three things I would recommend smaller, specifically healthcare firms look for, would be one, data training requirements and agreements.

[00:16:37] So if you're providing information to the consumer version of ChatGPT without changing any of the settings, it is going to be training on that data, providing it to human reviewers to improve their models. Now, that's not necessarily a bad thing, but it's important to understand what's happening behind the scenes. So certainly any kind of protected health information, you wouldn't want subject to that type of training.

[00:17:05] You'd need a business associate agreement in place to preserve the privacy of that data throughout your supply chain. Second of all, I would recommend looking at retention terms. So even if you have a business associate agreement in place, if the provider is retaining that information for any period of time, there exists the risk of a cybersecurity incident leading to a breach of that data.

[00:17:32] And I would say that less is more when it comes to retention. Optimally, you would have a zero data retention agreement, which some of the AI hyperscalers like OpenAI and Anthropic and Microsoft Foundry will offer. So I would pursue those to the extent possible. Those might be difficult to get because sometimes the hyperscalers want the ability to review the inputs for abuse. Understood. There's a back and forth there.

[00:18:01] And then the third thing I would look for would be the efficacy or the accuracy of the outputs of the system itself. If you're generating lots of transcription errors for an ambient AI scribe type system, or you're getting lots of strange recommendations from an AI system that require a provider to review and override, that it might not be worth the risk and the cost of implementing this AI system.

[00:18:29] So looking at the error rates would be a third metric that I would evaluate as part of reviewing an AI deployment for a smaller healthcare organization. And outside of AI for a moment, I think we've been debating for years that fine balance between required governance and whether it will slow innovation down. Many leaders worry that it does. But when does that concern genuinely have merit? And when is it an excuse for just accepting weaker controls?

[00:18:59] Again, you've probably got a few stories from this. But how do you get that balance right? I think the balance between speed of innovation and governance isn't always a conflict.

[00:19:11] And for example, ISO 42001, which is a global standard for AI governance, gives you a set of tools to conduct repeatable risk assessments for AI systems, while at the same time building a monitoring and measurement program to ensure that the systems are meeting your business objectives. So I would say that innovation and business value could be aligned with governance in those situations that I described.

[00:19:40] So if you have a healthcare AI tool that is chewing up a lot of provider time because its outputs are inaccurate, then that's something that the business would care about and the security or AI governance team would care about.

[00:19:57] So looking at those error rates, tracking them in a deliberate manner over time, comparing them to your tolerances and using that to drive risk assessment, that can actually help both of the desired outcomes in terms of governance on the one hand and business effectiveness on the other.

[00:20:15] Another example I would say is that you can create a tiered set of categories for evaluating AI systems using different standards based on the data that they're processing or the actions that they're taking. And that can allow you to pre-approved lower risk activities. For example, if they're only processing public data, if they're not doing anything that could damage your brand, then, you know, say, hey, go for it.

[00:20:45] Just it's pre-approved. We've already done an assessment on these types of systems. Maybe you have a few quick checks that you need to do. But on the other hand, if you want to process protected health information or drive an autonomous vehicle, then we're going to apply a much higher degree of scrutiny. And that prevents organizations from kind of getting caught up and unable to make decisions if they have that tiered system of analyzing risk.

[00:21:13] And it also helps you avoid taking unnecessary risk. And another big topic out there at the moment, increasing in importance is AI drift. So how should security, legal data and business teams divide responsibility when an AI system changes after deployment evolves, almost develops a life of its own and drift from its original initiative?

[00:21:38] Model and system drift is a huge issue when it comes to effective and secure AI deployments. So there are a couple of things that can happen. So first of all, the hyperscaler AI providers like OpenAI and Anthropic, they push new models to their user interfaces on a regular basis all the time. And even the enterprise versions don't allow you to select which specific models are being deployed. You don't necessarily know what safety layers have been changed.

[00:22:06] And this can lead to differing behavior when you're essentially applying the same commands. So understanding that is a risk and there's not much you can do to stop it. But users need to be trained to observe differing behavior from the same AI system based on new model releases. Additionally, you can have application programming interfaces that allow you to communicate with a specific version of a model.

[00:22:35] So that might be more appropriate for your machine readable type of interactions with AI systems if you're using those hyperscalers. Alternatively, you could host your own model. Now you necessarily take on a lot of the security and technical burden when you're doing that. You can't rely on a third party as much. And then also there are some important things to consider as well that are more mundane but can also be important.

[00:23:01] For example, OpenAI recently refused or changed the way that its API functions and previously would accept temperature, essentially commands or instructions for the responses that were coming out of its earlier models. But for its later models, it refuses temperature instructions for those later models.

[00:23:23] So that will actually break your integrations if you're upgrading the model without necessarily looking at the response of the API itself. So that's kind of a mundane but pretty important implication of AI system drift for any company that's using a third party to get its AI processing done. And as this is a tech and cybersecurity podcast, obviously we focused a lot on the technical aspects today.

[00:23:51] But at the heart of everything is always humans, employees, people, customers. So what happens to employees and customers when some organizations treat AI risk as just another checklist rather than going an ongoing operating discipline? AI risk is one part of an enterprise's overall risk and it requires the same approach from a high level.

[00:24:17] You can accept risk, you can transfer risk, you can mitigate risk, or you can avoid risk. And that applies to AI as well. What I'll say is that AI is different because of its non-deterministic and potentially unpredictable behaviors, especially when you have agentic swarms that are potentially communicating with each other.

[00:24:36] They can take actions as we've seen that are unexpected and undesired, specifically the open AI and hugging face hacking incident where open AI agents broke into that company's information systems. That's one example where things are a little bit different. So understanding that there could be a self-reinforcing or a snowball effect when you're deploying AI systems is very important to keep in mind.

[00:25:04] And that's something that should form part of your risk calculus. At the same time, there's a risk of being left behind. If you are going to bury your head in the sand and say, you know, I'm not going to use AI, then there's potential you could be left behind by your competitors, which is itself a business risk. And if you're performing a valuable service for your customers, then you're not necessarily serving them as well as you could. Yeah, I think that is a powerful moment to end on.

[00:25:33] And obviously, at Stack Aware, you help AI-powered companies measure and manage cybersecurity, privacy and compliance risk and so much more. And thank you for sharing your insights around that today and your experiences and what you're seeing. But for people listening who want to find out more information about anything we talked about, where should they go? And also, I believe you've got a podcast as well. Where can people listen to you as well as read some of your thoughts and the work that you're doing?

[00:26:00] The best place to find me is on LinkedIn. I'm Walter Haydock. Give me a follow. I post daily on AI security and governance topics. If you are interested in working with Stack Aware, please go to stackaware.com to book a call with us. And then if you're interested in following along on some of the deeper dives that we do when it comes to AI governance and security, please check out the Deploy Securely podcast.

[00:26:26] That's available on YouTube, on Spotify, on Apple Podcasts. And give us a follow and a listen there. Well, I'll include links to everything, including the link to your podcast as well. I'll be checking that out. And so many big talking points there in a short amount of time, a lot to digest. I would encourage people to listen to keep this conversation going if anything resonated with you today. But more than anything, thank you for bringing it all to life in a language everyone can understand. Really appreciate your time. Thank you, Neil.

[00:26:56] I think Walter's argument about ownership gives leaders a question to take back to their teams. If a business function benefits from an AI system, who has the authority to accept its risks and responsibility for monitoring what happens? And I really enjoyed the focus on comparing a deployment with business as usual.

[00:27:19] Accuracy, review, time, operating costs and the way data is handled all being in the same discussion. Because a policy document cannot answer those questions without evidence from the system in use. And again, that healthcare example showed why security and governance can support a useful service. While the discussion of changing models reminds us that approval is the beginning of a responsibility.

[00:27:48] So big thank you to Walter from Stack Aware there. You'll find his LinkedIn profile, company website, Deploy Securely Podcasts, all in the show notes over at Tech Talks Network. And over to you, how is your organisation assigning responsibility for AI risk? And what would you change? Please share your thoughts with me. TechTalksNetwork.com. I'll be back again real soon with another guest. But thanks for listening as always. Bye for now.

[00:28:20] Bye for now.