The AI-Enabled CFO: Governance, Data Integrity, Risk Oversight and Cybersecurity

Speaker A: M welcome to the Financial Executives Edge, a production of the Financial Executives Journal. Here, finance meets bold leadership. Join us for sharp insights and practical strategies to elevate your thinking, drive change, grow your impact and empower your career. This isn't just insight. It's your edge. The Financial Executive's Edge. This episode is sponsored by AI Dentified. The nation's top RIAs and wirehouses use AI Dentified to turn relationship data and money in motion events into organic growth without adding headcount. It's AI powered intelligence built for modern finance leaders. Learn more@aidentified.com welcome to the Financial Executive's Edge.

Speaker B: Today's podcast centers around the AI enabled CFO covering governance, data integrity, risk oversight and cybersecurity. I'm, um, Lynn Gargano, your host and moderator with the Financial Executives Networking Group and editor for the Financial Executives Journal. As we all know, AI is transforming finance at unprecedented speed. But with greater intelligence also comes greater responsibility. And in this episode we explore the evolving role of the CFO in ensuring that AI is deployed responsibly, securely and strategically. Because in the age of intelligent finance, insight may come from AI, but ultimately the responsibility belongs to the cfo. We're joined today by Shane Reed, CEO of USC Data North America, the business memory for AI readiness and governance, as well as John Hetherington, founder CEO of we deliver your vision. He's a technology strategist that helps fintech leaders make better decisions faster. So thank you gentlemen for joining us today.

Speaker C: Thanks for having us, Lynn.

Speaker D: For sure. Glad to be here, of course.

Speaker B: So I think we're going to start at the highest level, this discussion and determine what, what does really effective AI governance look like within an organization and in particular within the finance function. So I'm going to toss this out to Shane to get us started.

Speaker C: Well, I was having this conversation with CFOs and you know, and CIOs where we're seeing this in the businesses and I, you know, I gotta say I think governance is gone way beyond of just being a policy document in those businesses. You know, it still goes back to uh, and we're in the data world, right? So I'll come from my aspect, I'd say it's data discipline. I look at it. So, you know, when we're having conversations with CFOs, it's like what are, you know, the ones that are doing it really well can answer a couple of key questions and that's, you know, where, you know, where does the data come from, who owns it and, and I think the benchmarking, the most important bit is like, what does wrong look like? We actually don't know what right looks like and what wrong looks like. Then how do we benchmark that? And I think without those answers, you know, your governing outputs, you can't actually validate. So that's probably where we come from, from, from more of an AI governance piece is actually being able to understand it and being able, uh, to put your finger on it if there was something wrong where that, where that went wrong, especially from auditing perspective.

Speaker B: Okay, John, any thoughts on that?

Speaker D: Yeah, I totally agree with Shane. It's not a policy document, it's a culture. And what we're seeing talking to CFOs controllers in the Pacific Northwest, because we just completed a survey of finance leaders where she found that only 23% were actively using AI. And a lot of them were just firefighting the day jobs, getting the job done for reporting board updates, managing their teams. So although they're thinking about AI governance, some of them are not even there yet. So similar to what Shane's saying is our conversations are really around what data are you using for what decisions? And then how does AI play a part of that?

Speaker B: What advice would you give CFOs in evaluating the AI governance structure or even readiness to even jump into AI adoption?

Speaker D: So building on what Shane said, it's similar to what is wrong, like find out the problems with your data. Because we also always talk about garbage in, garbage out. But with AI, it's now garbage in, exponential garbage out. AI doesn't sleep, doesn't get tired, doesn't

Speaker C: stop, and 24, 700 times worse.

Speaker D: Exactly. So it's really now about how do we make sure we're making good decisions better? Uh, which really comes down to the process. I know people talk about people process technology, but what we find is that without a good process, your best people won't perform. And that's so, so important in things such as the AP process, the financial close process. Just making sure you're getting your basics done well before you start automating and accelerating with AI.

Speaker C: Yeah, and Lyn, we do like we almost mandate discovery with clients now because understanding, yes, the people to process technology and to be fair, this day and age, technology will always work. It's just a horrible saying. It's just a question of time and money to make the tech work. Um, you know, but we're talking about a governance framework and we're talking about that discovery. Like we really, really dial it down and go, let's Build a data accountability map. And that's just a fancy way of going, let's follow the bouncy ball of, you know, documenting which systems feed which decisions. So we always understand what has driven that. So, you know, if, and if the governance, you know, I think governance collapses when no one can trace the number back to its source.

Speaker B: Right.

Speaker C: So we can build that and then we understand the process. And to be fair, and I'm sure John sees it all the time, is a lot of people are trying to digitize and put AI on a legacy process that was 17 stages. And we go, guys, if we just connected this better and did better things to the data, it would be six, six areas that we' would be managing now, not 17 in there. So absolutely understanding that process and working, working backwards.

Speaker D: We're seeing a lot of boards just say to CFOs, go do AI. You need to get on the AI train. And, and to Shane's point, that puts a lot of CFOs and accountants deer in the headlights. Where do we start? So building on what Shane said, it's, it is literally about, okay, what do we have now? What works, what doesn't work? Fix that foundation, which might be just removing some approvals from your process flow to streamline before you start applying new technology

Speaker C: or should you be doing it in the first place? You know, sometimes you go, we need AI. I suppose the first really bad question ask is why. Yes, like who is it? Is it helping the client, is it helping the business, is it helping the employees? And is it going to help the bottom line? And if you can't measure that and answer that straight up, I go, you're not ready.

Speaker B: Yep.

Speaker C: Cause you don't know what you want.

Speaker B: So many organizations treat AI adoption as a technology project, which it goes so far beyond that. You've both focused as well on the criticality of data integrity. So what are the biggest data quality risks that CFOs underestimate?

Speaker C: Multiple systems and you know, people are bought more over. I think duplication, I mean, I think duplication and even definition drift in the business because you've got the CFO who's, let's just think about, you got two departments, they're using the same metric, but they calculate it differently and it's fed from different sources. So, you know, finance still calls it revenue, the sales area and the CRO still calls it revenue, but they're off by 15%. So an AI, it will ingest both, uh, will ingest both data sets and it learns that wrong pattern and it's very difficult when we keep feeding it this information to roll back when it's built. So. And the CFO doesn't know that until they present to the board and everything goes sideways on that. So I think once we go in, we think about, it's an old acronym, rot, you know, redundant, obsolete and trivial data. You know, if we can reduce that out of the business and remove the duplication out of the business, at least we know we've got a clean data set that we can baseline in the business and understand. That's my style. So John's different.

Speaker D: Yeah, and echoing that as well, like I said mentioned earlier, bad process, automated is still a bad process. So the inconsistencies come from human input. And much as we want to automate and have AI or any automation streamline the process, remove the data quality issues, unless there's a human checking that, then you're still going to have a bad result. And we often see inconsistencies around the data across different entities. Then there's an Excel bridge which extracts data from your ERP and puts it into a Power BI report. Each one of those transition points is a position of data error. So looking at those transition points, how can we shore up the quality, consistency and accuracy of that data? So you know, your reports are uh, consistent, accurate and reliable. And then as a result you're not suffering bad data. Which I saw an interesting statistic from IBM around data quality problems costing the US $3.1 trillion. And that's not AI related, that's just bad data. Uh, finding the data, fixing the data, uh, correcting mistakes. 3.1 trillion. So anything we can do to help that is going to help the bottom line.

Speaker C: Yeah. And literally on average the businesses, when we go in and we do a data cleanup, at best 25 to 30% of the data is actually usable. It's the real data set. We're talking at least 70% of the data that exists in any business that you go into could be just removed from the entire scenario of a business that goes back to cost as well. Think about storage and management and security all around that additional data. You do not need to have it. And that's what we're trying to pull out.

Speaker B: No, no, absolutely. I know from my experience as a CFO and overseeing um, automation projects, that resonates well. And that's a, that's a critical part. So maybe a gentleman could potentially give some examples of the consequences of having, you know, bad data when you overlook these risks. So flawed data going in have you had clients or worked on projects where you've seen the, the consequence or the outcome of that?

Speaker D: Yes, for sure. So we work a lot with medium sized businesses, 50 to 500 people in the 50 to $100 million range. So these folks are starting with AI, they're starting to get used to AI, they're dipping their toes in the water. And what we found is the first bit is just a chat, just a search, just to understand how you can use AI. But that very quickly jumps into a, okay, let's produce a report to the client on billing for example or a status on the uh, stage of the financial reports. So the AI, once we know that problem to solve, the AI is then searching a specific directory or folder, uh, for the information. To your point around data quality, it's not only the data quality within that folder, it's uh, also the data available within that folder. So we found during testing, and this is why testing is so, so important. It's is that when we were running the queries to produce the client update report, it was only looking in one folder.

Speaker B: Ah.

Speaker D: All the other client information was stored across many folders all over the place. So pulling all that data together can be a bigger exercise than the actual project itself. So it's not necessarily just about the quality. It's making sure the AI has access to the right data, uh, to pull out the right information to create the right result. Which in this case was a client report.

Speaker B: Yep, disparate data, a big problem, especially in larger organizations. Right.

Speaker C: Well we, we've got to add context.

Speaker A: Right.

Speaker C: So we, and that's our biggest job when we're going in with our clients is, you know, to John's point about that unstructured data that's hiding in all of these folders is we try and change their minds. Again, we need to remove the folder structure. So we actually, we contextually now add structured and unstructured data together and we're labeling and doing data classification. So like okay, give me all of the invoices to make decisions for an AR or an ap, you know, and when people do it wrong gives you a point about what have we seen go bad. I mean we've seen people make vendor decisions and headcount decisions and pricing decisions based on AI outputs weren't necessarily checked or benchmarked. Um, that's when we, because underlying data was bad and AI doesn't necessarily, isn't always wrong. I know everyone knows the term hallucination, but that goes Back to there's 17 copies on average of Data in a business. And it goes, I'm just going to take number six and we're. Well number six is from three years ago and it related to another project that shouldn't be in this baseline. And that's when we go back to context with data.

Speaker B: Absolutely. So you start with your AI governance and you stress test for AI readiness to say and understand why you want to adopt AI. You know, at an enterprise level and particularly in the finance function, you make sure you're, you have integral data, right? Consistent, non disparate data, usable data. And then you have to think about, you know, integrating all this information into your financial systems and ultimately there's going to be aspects of sensitive data and security concerns. When you're, whenever you're implementing new AI tools, what do you see, uh, as ah, new vulnerabilities emerge when integrating AI tools, what is the biggest risk?

Speaker D: So my advice for Anyone, not just CFOs, anyone adopting AI is you treat AI like an intern, like a new hire, someone new to your business, new to the company. Because like someone new to the company, you wouldn't give them access to every single system. You wouldn't give them access to all the sensitive data. You give them an initial set of uh, work to do so they can work on it, prove themselves out, check for mistakes and then they build competence. And that's the same for AI. So when we're training AI on the models or the task or this uh, procedure they want it to complete, it will make mistakes. But that's when you test it out, train it, give it more information. Because at the end of the day the security is all about how you, you work with that intern or that AI to prompt it in the right way, to guide it, to help it make better decisions.

Speaker B: Have you seen this go astray or gone wrong where there is a data breach when trying to use AI tools?

Speaker C: Yeah, I think, I mean it's a good point on the intern, um, knowledge, but I mean more I get back to the data point, you know, we're seeing, um, we're back to the, where we're seeing data sprawl, right? So teams start using the AI tools, you know, even, even approved ones, right? But they start moving data so they're exporting to CSVs, they're pasting into prompts, they're connecting third party apps, uh, you know, and suddenly sensitive financial data that they think, oh, it's a, you know, it's a private data model, we're cool. But it's now sitting in five places that it was never meant to, you know, in there again. So that attack surface expands and there's nothing's malicious. No one did anything deliberate. But, um, you know, no one's necessarily making, you know, deliberate decisions to expand it. But it is naturally occurring again across the business. We're finding sensitive data. So when we, when we do, like our PII or sensitive data scans, you know, we are picking up things on local drives and down in your downloads folders because they've pulled something down and we're like, it's, it. It's back to being unmanaged.

Speaker D: And that initial protection is so, so important. We're doing a lot of experimentation with Claude Cowork, where it runs agents 247 for you. And there's a lot of hype right now around how amazing it is and how it's going to revolutionize work. But it has to start with that security model, that constriction of what data are you allowing access to, because it will do anything and everything. So you want to be very careful in terms of test it out on one or two folders of some sample data, but kept within your environment, ideally on a separate machine that's not linked to your network or your own personal device. Then you can test it out, start to build the confidence about how it's delivering results, checking the results, and ultimately, with anything we've talked about today, having one person accountable for that result.

Speaker C: I think that's what's missing too, John. You see it in there. You go in there and you start talking to the client. There's a lot of people responsible for certain areas of the process, but no one is accountable. I mean, luckily I'm seeing really good CFOs being accountable for it, um, rather than, you know, blaming Johnny Ponytail in it, who's just trying to try to keep people out of the systems. But all of a sudden he's, he's to blame for the quality and where, where it goes wrong. Uh, he's got enough of it.

Speaker D: I used to be Johnny Ponytail, by the way, in my younger days.

Speaker C: I was in my 20s, John.

Speaker B: Yeah, but it all loops back to governance, Right?

Speaker A: Right.

Speaker B: And part of governance is accountability and making sure that infrastructure in place. Because at the end of the day, yes, the ultimate responsibility is with the cfo, but there has to be accountability throughout the organization. Everyone owns an effective governance strategy. Um, now we talked about kind of security and sensitive data, Right. And assessing that internally. But let's think about external parties, like vendors. Right. So how should CFOs approach this vendor risk. Yeah.

Speaker C: Well, I'm going to take it back to where this is where I think you take a step back in the cfo. We're going to go back to the legal and procurement level as well, the amount of clients I'm working with. And I go, let's have a look at the contracts. And the contracts are so out of date that, you know, you're not having vendor, vendor contracts. Now answering the question which all of them need to be able to answer is where is my data stored? Is it used to train your model? And who can see it in the organization? Because we forget that so much nearly everything in the business, you know, whether it's cloud or like everyone's doing the same thing, everyone's building their own, their models m based on client data or everything. And whether or not they should or shouldn't be. Um, now I think most vendors have the answers, but I still see that most buyers are not asking those questions of the vendors. And that's very, very important because if you go back to cyber insurance and all the issue around there is like, who's really holding the cards at the end of the day if client data or citizen data goes missing? Was it the vendor, was it the organization? So I think that that baseline needs to be done and everyone sort of understand, you know, where is the data and you know, what are they using it for?

Speaker D: Yeah. And I would ask any vendor three key questions. So the first one is, where does my data go? So where is it being stored? Who has access to it? Both internally in our organization, but also in your vendor organization. And the key one is it's not necessarily about the vendor, but the vendor's vendor. Because typically a lot of AI models these days are built on third party large language models. Not a huge amount of people are actually building their own AI. They're borrowing models. So by inference you're giving your data to a vendor, who then giving it to another vendor. So that chain has to be understood and secured. So what we've done is actually created a scale of everything from free, private, sorry, free ChatGPT, public Gemini, which is very unsecure, that's public, that's Internet, all the way up to your openclaw and your custom large language models. So generally with enterprise, if you're focusing on the copilot enterprise with Microsoft or your Gemini's enterprise workspace, it's contained in your tenant, your environment, and that's generally with the right controls, generally the most secure place to put it. So we actually have a scale of this which we can share with your listeners if they wanted to comment, uh, framework in the chat for this or comment framework on my LinkedIn profile, I'll send them that scale for free, no sign up and they can see where they land on that scale of privacy and security for AI, you know, and

Speaker B: just beyond the whole data protection, there's the broader risk of cybersecurity. So from your perspective, how do you think AI both increases or mitigates cybersecurity risk?

Speaker C: Well, I think it's always a two horse race, right. We use an American been living here now three years, I'll use the American sort of football. Uh, I think there's an offense and a defense team on both sides of this. So AI is a threat multiplier on both sides, uh, of the fence. So on the offense, the attackers are using AI to craft more convincing phishing exercises and more targeted fraud and faster exploit discovery. And then on the defense side, AI is generally really, really good, um, and has been for years for anomaly detection, you know, so flagging unusual access patterns or transactions, you know, behaviors that a human would miss traditionally. So I think the, let's go about the CFO's job. I think it's just to make sure and convince the board and the business that the defensive investment is keeping up pace with the, with you know, the offensive threat and very often go on the businesses and thankfully now budgets are starting to invest in this, but they weren't necessarily prioritizing that even a couple of years ago.

Speaker D: Yeah, and even just some of the basics which seem obvious but most people don't do it around the two factor authentication for when you're signing in to your ChatGPT or your copilot or just making sure you have those two factor security levels of a code authenticator code or a thumbprint face print just to give that extra level of security. And then in the AI environment that you've built is having someone or idea with a backup who can ultimately pull the kill switch. So if you find that something's going wrong, something's maybe might just be the AI starting to hallucinate and go off in the wrong direction too fast, too quickly and you need to shut things down quickly. You need that one person with a backup who has that kill switch to just stop, say let's reset, let's understand what's happening. So it's that protect, detect and then more importantly react. Most of the disaster recovery business continuity plans we talk about in business when something bad is happening is not so much the detection you Find out pretty quickly it's happening, it's how do you react? What's your playbook for making sure you can respond quickly, tell the right people what's going on and ultimately fix it.

Speaker B: Are there any cyber threats specifically that target finance AI systems that a CFO should be aware of since a lot of sensitive data tends to reside in the accounting finance function.

Speaker C: John, I don't know you've been watching the media lately but there's been some doozies that have uh, been done. Like even like let's go back to sort of the co pilots of the war and I think this is where this is being ignored. And you watch the McKinsey issue that occurred a few days ago. So I wrote about that. Um, it's prompt injection.

Speaker A: Mhm.

Speaker C: So a model, basically people are getting in there and poisoning the apps, the models themselves. So this is bad actors manipulating data into a training pipeline that they'll skew outputs. And prompt injection is basically when someone's getting in there and they craft an input. It tricks the AI system into leaking data or bypassing controls. So you get on the inside and we've got private data models and it's our own LLMs and stuff. Stuff.

Speaker D: Okay.

Speaker C: But if someone gets in there and they can access all of the training data that's being put into that through prompting, um, you know that's pretty bad. You're bypassing controls and you know this isn't theoretical like this has happened, this happened. Microsoft had this in the last couple of months where co pilot had been done, uh, and businesses financial data was being pulled back out of the system through that engineering. So I think finance is a very, very high value target, you know and AI systems, you know, that kind of need to be treated like infrastructure and you know, not just software tools. Now I think maybe we're getting our heads around that maybe.

Speaker D: Yeah, the financial information is definitely very valuable. What we're seeing it's more around the personal information, particularly medical information where you have name, address, history because they're then using that to spoof the call centers, the email spam filters. So it really comes back to having very good protection around phishing emails where they clicking on the link. So training your teams with regular phishing tests where you create a spoof phish email with a link to click and they're getting very, very sophisticated. That's what AI is doing now is creating volume of attack because it never sleeps, keeps going. But also the sophistication around how to personalize emails to individual people before it used to be one person sending out 5,000 emails. Now it's one AI sending out 5,000 different emails to different people. So it knows what you've done, it knows some of your history. It's gone your social media profiles. It's then triggering some events in the email which might think, oh, I wanted a discount on that clothing store I just visited or with the company. We're now doing discounts for clothing through the merchandise. And it gets very, very sophisticated. So just having those phishing tests to help people spot the email, the first thing to look for is the email address. Where did that come from? Because they'll often have a weird email with one or two different typos. So instead of Microsoft, it might be Micro Sift. And you're just looking for those little bits of anomalies, uh, that'll help you protect against. Get them sharing your username, password and everything else with. Attack the attackers.

Speaker C: And you see it. I mean, I think, John, I think still business email compromise is still one of the biggest items still exist, especially in the, in the, uh, in the finance space. I mean, you've seen examples now, Lyn, where there's been, you know, people have been the cfo, and this is a famous one for last year. CFO was in a teams meeting. Video had been cloned, the audio had been cloned, the language had been cloned because they had four, four or five months of data that they had been sitting there collecting. So they've done a transfer, uh, thinking they were in a meeting with the executives, effectively sign off. The bad actors knew exactly what the delegation authority was. They've transferred $250,000 out of the business and then talking about it in a meeting later. And they're like, what are you talking about? This meeting didn't exist. The three other people on the call were all spurts. It's very scary. So we tell our clients now to do things like, uh, and this is definitely in wealth management, especially wealth management. Very, very high target Financial services World is if you're being asked to do something over the phone by a client, train them to SMS something onto yourself, you know, at the same time only something that you would know and then like basically read the number back to me, you know, so it's just adding that two factor even in. Now, I know that's scary to say that we can't even trust each other on a phone call or a video call these days. But that's an extra layer that you probably should be bringing at, uh, a governance level when we're talking about items that could be very, very detrimental to the business. Where it was a wrong decision made.

Speaker B: Absolutely. So before we wrap up today, we talked about AI governance, data integrity, the risks lurking everywhere from data protection to cybersecurity and to informing the board. This raises a very important question, and it's a leadership question. If AI, uh, increases speed and complexity, what principles must remain constant in financial leadership? Final thoughts, John?

Speaker D: Judgment. So we're seeing a big shift not just CFOs, all leaders. Because the experience you built up as a CFO for the last 20, 30 years can now be summarized for 20 bucks a month in a chatbot. So it's creating a shift to say, as CFOs, where's the value I add to the business in addition to my expertise and experience, which a lot of people can get access to? So it comes back to that one word, judgment. What decisions are we making? Why are we making those decisions? And how do we make the best decisions for the business based on the risk profile and what we want to achieve for our business goals? That's what I would say. Judgment, Shane.

Speaker C: I'd be very similar sentiment. I think accountability and explainability for everything going on. So every AI driven financial decision still needs a human who owns it and can explain it. I know speed is a competitive advantage, but until something goes wrong, uh, then auditability is everything. Uh, the CFOs who will lead this era are, uh, the ones who demand that AI makes their teams more accountable, not less than what they're doing. So, so that's what we're seeing. I'm seeing some really good CFOs, surprisingly, who are 612 months ahead of this. They're really making this a foundation and they've brought everyone on board. They are accountable in their areas. They could with a blindfold or know how that decision was made with the tools because they can trace it back to the system, they can trace it back to the data, the baseline.

Speaker D: Yeah, trust.

Speaker C: And it's cool to see.

Speaker D: Yeah, trust, verify.

Speaker C: That's it. You're right. Trust. Um, yeah, trust. I mean, I think when people trust it and that's fed up to the board level, when the board can trust, the tools are starting to work. And this is not trying to reinvent the wheel in the business like Sprint in a little, a little area like, dude, pick off little bits in the business at a time and that trust will start, uh, running itself across the business.

Speaker B: Well said, gentlemen. And that brings us to the end of our AI enabled cfo. Governance Data Integrity, Risk Oversight and Cybersecurity Podcast. A huge thank you to both you, John and Shane for imparting your insights. It's clear from today's discussion that, uh, the AI enabled CFO is really not simply just adopting new technology, it's ensuring that AI operates within governance integrity and as you said, Shane and John, accountability. And while AI can predict outcomes and certainly drive insights, leadership and accountability will always define the CFO role. So if you've enjoyed today's conversation, please be on the lookout for the AI Enabled CFO and Financial Executive Workshop series where we continue the discussion covering the full financial landscape with practical use cases and candid conversations around the CFO's evolving role in AI adoption. Thank you for joining us. And until next time, stay informed.

Speaker A: This episode is sponsored by AI Dentified, the nation's top RIAs and wirehouses, and use AI Dentified to turn relationship data and money in motion events into organic growth without adding headcount. It's AI powered intelligence built for modern finance leaders. Learn more@aidentified.com thanks for listening to the Financial Executive's Edge. If today's episode sparked new ideas or helped sharpen your perspective, be sure to follow and review us on your favorite podcast platform. You can also visit financialexecutivesjournal.com for more insights, articles and upcoming episodes. Until next time, stay sharp, stay strategic and maintain your edge. The Financial Executive's Edge.