Closing the AI Governance Reality Gap With AI Score
Tech Talks DailyNovember 15, 2026
3788
22:0017.77 MB

Closing the AI Governance Reality Gap With AI Score

How do you govern AI when the technology, its users, and the work it performs keep changing after you approve it?

On Tech Talks Daily, I'm joined by Alex Harland, co-founder and CEO of AI Score, to discuss the distance between an AI policy and the reality of everyday operations. A company may approve a platform and complete its initial review, yet employees discover new applications, vendors introduce features, and agents begin working with different information. Alex argues that governance has to keep following those changes if it is going to remain useful.

What interested me was his emphasis on individual use cases. An inventory of product names can tell you what a business has purchased, but it may tell you very little about what those products are allowed to do. We discuss why an organization needs to connect its AI assets with approved activities, named owners, and continuing evidence of how they operate. The same platform can create very different risks depending on the information it handles and the decisions it supports.

We also look at vendor assurance. Adding AI questions to a supplier questionnaire is a reasonable start, but the answers need evidence behind them. Alex describes AI Score research involving approximately 250 organizations using AI. He says 92% were beyond pilots, while only 16% had an inventory. These are company-reported findings, and the conversation does not establish the survey methodology. They provide a starting point for a practical question rather than a universal measure of enterprise readiness. Can your supplier explain what AI it uses, for which activities, and how it knows when that picture changes?

AI agents add another challenge. Existing identity, permission, and network controls remain relevant, while autonomous behavior creates a need to notice changes after deployment. Alex discusses monitoring whether an agent starts accessing external information when its approved use was limited to internal sources. We talk about the people responsible for those systems and how operational accountability should connect with board oversight.

There is also a commercial argument for governance. Alex sees it as a way to pursue business objectives within an understood risk appetite. Controls still need to justify the work they create, and the interview does not provide a measured customer return or a product demonstration. The useful lesson for listeners is to ask whether governance is helping a business make better decisions as its AI usage changes.

We close with the company's growth following the $5.4 million seed round referenced in the supplied guest notes and our conversation. Alex describes the challenge of keeping pace with customer needs and a rapidly changing market. For listeners responsible for AI adoption, security, procurement, or board reporting, the episode offers a practical way to connect written rules with observable activity.

I'd like to hear your thoughts. What would help your organization keep its AI approvals, ownership, and day-to-day use aligned as the technology changes?

Useful Links

[00:00:00] It's not your fault that your agentic AI systems are acting outside of compliance. There are just too many data sets to guardrail them all. But with Denodo, you can now organize your hundreds of data sources into one layer and govern your agents with a single approach. Try it now with Denodo by visiting denodo.com to learn more.

[00:00:28] Picture the scene, it's 6am. You're on vacation in Mexico. A day of maybe cocktails and taking it easy. Taking a much needed break. But then I come along and ask you to come on my podcast. Would you take a break from the vacation with your wife to come and speak with me? I suspect most of you would say no.

[00:00:52] But my guest today, Alex Harland, co-founder and CEO of AI Score, he said yes. And today, instead of drinking cocktails, we're going to discuss why AI governance needs to follow what happens after deployment. And understand how to assign responsibility to real people. And explore why a list of approved products tells you very little about their actual use.

[00:01:21] And if you're trying to balance AI adoption with sensible oversight, I think today's conversation will offer you a useful starting point. So a few questions for you before we begin. How much do you really know about the AI that is already working inside your business? You might have approved a platform. Or you may have written a policy. Maybe sent everyone the training slides. Meanwhile, the software has changed. Employees, they've found new uses for it.

[00:01:51] And an agent could be doing something that nobody has actually signed off. So keeping this picture current is becoming a business problem with consequences for customers, costs and trust. And everything I've just talked about here are just a few of the things myself and Alex are going to cover today. But enough from me. Let's get him onto the show now. So a massive warm welcome to the show. Can you tell everyone listening a little about who you are and what you do?

[00:02:21] Yeah, of course. And thanks for having me. My name is Alex Harland. I'm the co-founder and CEO of AI Score. AI Score helps organizations safely, securely, but effectively adopt AI. And I was doing a little research on you. I'm always fascinated by my guest's origin story. And with your background at the NCSC and now as a founder, where do you see the biggest gap to how organizations describe AI risk and how they actually manage it? I would imagine there's somewhat of a disconnect there. Yeah, of course there is.

[00:02:51] I mean, for me, I think I come back to the simple fact that I still believe too many organizations are treating AI as a tool that's being rolled out rather than treating it as a shift in sort of a technological shift in the fabric of how they are operating as an organization. You know, introducing new product, changing existing product, changing the ways almost every workflow within a business works.

[00:03:19] That requires more of a thought than just a simple rollout of a tool. And I think that shift is definitely the biggest gap between the reality of AI and the way that organizations see it. So I'm curious, from what you're seeing out there, what does a useful AI risk assessment capture that maybe a general security or data protection review often tend to miss? Because it's almost like a new playing field here. So I'm curious what you're seeing. Yeah, it is a new playing field. And I think it comes back to that previous point, doesn't it?

[00:03:49] Yeah. If we're treating AI like a tool, then the existing controls, the existing reviews will still stand true. The reality is that they don't. And sure, we still need all of those pre-existing things. All of those things are still absolutely relevant. But we also need to understand how much of a monumental change this becomes. And so it's not just, sure, there are models to think about, there are tools, there are skills, there are agents.

[00:04:17] And all of these kind of components are operating in their own unique ways that kind of outstrip what we've previously seen. And it's all happening at a scale and a speed and an unstructured data complexity that outstrips everything before. So I guess what I'm saying is it really has to be a complete strategic shift in how we view this problem. And for that reason, an AI risk assessment has to be holistic. It has to be complete. And it has to be done, you know, not just in isolation as a sort of strategic item.

[00:04:47] And at AI Score, you describe yourselves as the intelligence governance layer connecting your AI and your ambitions. And to bring to life some of what we're talking about here, are you able to share an example of an AI system that looked low risk in pilot, but then became much harder to govern when it reached real users in that live environment? You don't have to name any names, but any stories spring to mind there? Yeah, plenty.

[00:05:12] I mean, most, particularly if we look back at the last few years, I think most AI deployments started in some degree as a pilot. They started small in an isolated use case and then over time have grown and scaled through the organization, both in complexity, perhaps in verticals and horizontals. And I think you can take any number of those, or you could just very simply look at the sort of approval deployment of a big,

[00:05:40] one of your big hyperscaler AI platforms. I won't name them, but, you know, there are three or four. Yeah. Take the deployment of one of those. I see this time and time again with organizations who choose to deploy one of these within their organization and then almost turn their back on it and think that the job's done. The reality is that these capabilities are shifting and the capabilities are changing rapidly.

[00:06:05] But not only that, the way that users are using them is changing rapidly too as they begin experimenting, begin understanding more and more about how they can maximize the capability to achieve some sort of effect. The growth of those platforms completely shifts and changes. And, you know, beta features are added that fall under different licensing agreements and things are turned on in the background. And we've seen a lot of this over time.

[00:06:31] And so I guess what I'm saying is every AI deployment tends to typically start small or start as a pilot and grows rapidly, but then it changes rapidly too. And so really it's about making sure that you're, it's a continuous governance cycle around all of these capabilities. It's not just to sort of turn that thing on and turn around. Yeah. And I suspect we will have a few leaders listening and you're setting off a few light bulb moments right now.

[00:06:57] So for those people listening, how should they be deciding which AI use cases need that more formal assurance and which can still be handled by maybe lighter controls? I think the key there is the point on use cases is really valuable because I get that question quite a lot. But I get it from a user base, you know, which department can I let my technology team run free? Because they potentially have more understanding of what's happening. But I think that misses the point. I think the point really is on the use case.

[00:07:27] And so there's two elements there for me. There's a real deep requirement to understand your use cases. And when you're managing those assets, when you have your AI inventory and all of the capabilities that fulfill your objectives as an organization, it should really, the granularity of that should go down to the use case level.

[00:07:47] So that you can explicitly say, you know, this platform or this SaaS product or this agent can only do this sort of use case or this parameter of work. And then being able to monitor and manage that going forward so that you can detect if it does or doesn't happen. And so I guess when it comes into which of those use cases potentially are higher risk or should be monitored more effectively, I think that will change based on the organization.

[00:08:14] There'll be some organizations out there that deploy AI to work with customer-facing financial models, perhaps. That to them will be incredibly important because there'll be FCA regulated. And it'll be something that they need to tightly monitor. Others may be, you know, that may be completely irrelevant and they may be in a completely different sector. So really, I think sort of a politician's answer to your question. But I think the use cases is critical.

[00:08:42] It's about monitoring those and managing those going forwards. And then those use cases, the risk profile of them will differ for all organizations. There are obviously then regulations and law coming into effect that stipulate specific use cases, either around the UAI Act, like automated hiring decisions and things like that. But that will come through in those risk assessments. Yeah, I completely agree. And I think also over the last 10 to 15 years,

[00:09:09] many organizations are possibly guilty of outsourcing risk or believing they can. And as a result, they just rely on questionnaires from their vendors, from their third parties that they're working with. But again, what evidence should they be requesting to test whether those answers actually reflect the system in practice? Because again, another disconnect there. Yeah, good question. I think it's something I've noticed recently, actually, which is a good thing.

[00:09:36] But the vendor question is that they have a whole new section now when you're filling it out. And yeah, you tick yes to I use AI, which I think pretty much everyone ticks nowadays. There's then a whole plethora of questions around how you're using it, how you're governing it, how you're securing it, and to what capacity. I think that's all good stuff. But you're right. You have to then be able to evidence it. And we recently conducted a piece of research into 250 organizations who were utilizing AI.

[00:10:07] All of them pretty much were fully deployed with AI. We found, I think it was 92% were well beyond pilots. And the reality was that only 16% had an inventory and could tell you at any point in time exactly what was being utilized, in which capacity, where, and why, and by who. And I think if you're in that position and you're answering one of those diligence questionnaires, or you've got somebody answering a diligence questionnaire for you,

[00:10:32] and you don't think you can even answer that very basic question of what AI is being used, for which use cases, the last point, and can I monitor that going forwards, you probably don't have that evidence. So I think that's where I would be leaning in. I'd be asking more specific questions about how. Well, I want to know how you're governing your AI, because if it's still on a spreadsheet, then you've got it wrong, and it's not going to work for you. Yeah, we are coming towards the end of 2026.

[00:11:02] And I think the big dominating theme this year has been the explosion of agents, not just individual agents. We're now talking about teams of AI agents, digital workers, that could be thousands of them talking to each other, not just throughout the organization, but beyond, and to those third parties and vendors too. So there's so many big questions around identity management, but how can teams monitor models and agents after their launch,

[00:11:27] when their behavior, data sources, or integrations may drift or indeed change? That's a great point as well, and highlighting the fact that, you know, identity access management is still super important. Yeah. Well, I mean, still super important. It was never not important, but it's core to agents too, as our network controls, as our permissions, and all of the traditional control mechanisms that we've always deployed. I think, for me, the real shift with agents is,

[00:11:57] and we see this from the big hyperscaler incidents that have happened over summer, the real big shift here is that you can have all of these things in place. You can have a sandbox in place and you can do a risk assessment, reduce some guardrails on a model and allow it to try and solve a problem for you. But if you're not anticipating the unexpected, i.e. that this agent might go and compromise a third-party organization,

[00:12:26] then, and you don't have the additional layers to monitor that, then you're going to fall into real big problems pretty quickly. Because, to your point, agents are autonomous, they're moving quickly, they're making decisions in an automated capacity, and they're spreading through our organizations. And so, yes, all of those traditional controls are important. And when we do our risk assessment and deployment, we need to think about how we deploy, how we're layering those controls in. But I think more so, we need a live inventory of these assets. We need to be monitoring their behavior continuously,

[00:12:55] and we need to be checking for things like drifting behavior. Has an agent started doing something completely different to what it was doing in the previous six weeks? Is it all of a sudden accessing external data sources, where we only proved it for internal data sources? Is it taking some sort of behavior that was unexpected, that we need to detect, to then respond to rapidly? I don't think many organizations are in that space yet.

[00:13:23] I think they're probably still in the space of the traditional controls around identity access management and network controls, as opposed to the more nuanced, unstructured analysis of how those agents were operating. And as the pace of technological change continues to ramp up, and FOMO possibly ramps up even more, many businesses are just scared of being left behind. And there's somewhat of a myth out there that assurance will just delay deployment

[00:13:51] and keep them further behind, which is completely wrong. But I'd love to tackle that point from a different angle here. I mean, what is the long-term cost of moving quickly without being able to explain an audit or AI decision further on down the line? Yeah, I mean, I think this is the question of the decade for me. At AI Score, we believe governance is about moving towards your objective, whilst under the boundaries of control and within your risk appetite.

[00:14:21] It's absolutely not. If your sole focus is on the control element, you're missing the point. It's control for control's sake, and it's ineffective. So I think it is a fair concern for organizations to say, well, if I'm focused too heavily on control, do I miss the competitive advantage? But actually, I think AI governance is a competitive advantage in itself because it allows you to move towards that objective more quickly without losing control.

[00:14:50] And I always use the, I'm not into F1 racing, but I always use an analogy, which is probably incorrect. But I always say on an F1 car, governance is your traction control, right? If you didn't have it, you could, sure, you could floor the accelerator if they have accelerators, I'm not sure, but you could floor it. And you'd come off the track at the first corner. You certainly wouldn't make it the whole way around. With the right application of the traction control and the brakes,

[00:15:20] you're able to go as fast as you possibly can around that course. And I think it's the same thing. And for me, you know, governance is that there are lots of risks wrapped into that. There's cost risks, there's cyber risks, there's reputation and regulation. And as long as you can keep moving forward quickly, whilst understanding those moving risks, that's the key to the success. And again, in that same research we did recently, 84% of organizations agree with me when I say that,

[00:15:48] that governance over AI is a competitive advantage. And I'll always try and give people listening a valuable takeaway. And if we do have people listening and they have that knowledge and they admit that their AI systems affect a customer, a citizen or an employee or someone in their teams there, what should meaningful accountability look like beyond just publishing your policy? Because again, this is a massive question right now. Everyone's rushing and putting so much out there,

[00:16:15] but when it goes wrong or an unexpected result comes from an agent or AI in general, who's responsible, who's accountable? So what should that look like? Yeah. I mean, it's the first, pretty much the first line in any governance. Yeah. A pack that you take, isn't it? Whether it's NIST or ISO or whatever you're following as a framework, it's the first piece. And for me, named accountability is critical.

[00:16:42] So, and I think you just mean that in the paper policy. Sure, you have a paper policy. That's really important for audit and regulation. But we need to have that in a living system. And that living system of your AI within your enterprise, it does need, it does have to start with accountability. It has to start with the person that's technically or business from a business perspective accountable and responsible for that asset. But then it's about a live monitoring and a live inventory of those assets as they're moving, as they're changing.

[00:17:11] But at a deeper level, what are they approved for? What are they not approved for? It's more nuanced than just saying, here's a logo. This is approved by Jane, that it needs to go deeper. And then I guess from that kind of operational ecosystem, accountability has to go up to the board. So we have to be reporting these risks and these opportunities to the board to make sure that we've got that full governance from the very, very top all the way through the operational structures.

[00:17:38] And then when I say move away from policies, from paper policy into a platform that runs this, it's about then carrying that all the way down to a tactical level. Fantastic advice. And back to you for a moment here. You've had an incredible year. I think AI score, I read, raised $5.4 million in seed funding round. So tell me about this year. What's been keeping you up and what's keeping you busy? And also, what can we expect next year from you guys? Any teasers you can leave us with? Thank you. Thank you. Yeah, it's been a crazy year, a busy year.

[00:18:09] But I mean, it's a testament to the times. I think the next 10 years are about one of the overarching conversations is going to be AI governance. And it's going to be the organizations that do it well, that reap the benefits through these next 10 years. So we're seeing that in the funding that we're able to raise and the support that investors are willing to give us. But we're seeing it with the customers we work with. And we're super excited about that. If there's one thing that keeps me up, it's just that we want to move faster. We want to do more.

[00:18:40] The environment is changing and moving so quickly that actually keeping pace with it in itself is a challenge, a great challenge, but certainly a challenge. Teasers for the next year? I'm not too sure. So hopefully you'll see us continue to grow on the path that we're growing and working with bigger and more complex and forward-leaning organizations as we go. Well, it sounds like we need to get you back on in 2027. Just saying that out loud is a scary thought.

[00:19:10] But for anybody listening that is getting serious about AI governance and they want to understand what matters, continuous decision-grade insights measured against their strategic objectives, where can they find out more information about all things AI Score and connect with you or your team? Where should they go? Yeah, they can find us at aiscore.ai. They're very, very, very welcome to reach out to me on LinkedIn. My name's Alex Harland. I think I'm in there somewhere.

[00:19:40] One of the Alex Harlands, you'll see me. Link to AI Score. Or my co-founder, Benita Tibb. Feel free to reach out to any of us or go through the website. But I'd love to catch you up. And even if it's just a conversation about AI governance and how everything that I've said is wrong, I'm very, very willing to listen and be in here more than anything. Well, to save anyone searching for you on LinkedIn, I will include links in the show notes. So please, as I said, AI Score is billed

[00:20:07] as the intelligence governance layer connecting your AI, most importantly, and your ambitions. So please, check out the links in the show notes. But more than anything, just thank you for taking this call and speaking with me today, even while you're on holiday. That is commitment right there. But thanks for joining me. Thanks, Neil. Thanks for your time. One of the many things that will stay with me from that conversation is just how quickly an approval can become an outdated picture of reality. A tool gets a new feature.

[00:20:35] Somebody connects another data source, and suddenly the original assessment describes a very different situation. Keeping up takes ongoing attention, clear ownership, and evidence of what people and agents are actually doing. So my takeaway here is to choose one that AI can use in your business, and then ask three questions. Who owns it? What is it approved to do? And how would we know if any of that changed?

[00:21:04] The answers to those three tiny questions could tell you where the next conversation needs to happen, without turning the entire company into another committee. So a big thank you to Alex from AI Score for joining me, taking a time out from his vacation as well. Kudos to him for doing that. You'll find the company website and his suggested contact destinations in the show notes. But please share your thoughts. How closely does your AI governance reflect

[00:21:33] the way your business actually uses AI today? What are the answers to those three questions I just gave you? Please hop over to techtalksnetwork.com. Send me an audio message. Let me know. But that's it for today. So a big thank you for joining me. I'll speak to you all again real soon. Bye for now.