Agentic vs generative AI in finance, and where CFOs go wrong.

What AI in finance means to a CFO

Alex0:0490% of the people talking to us about what they're trying to achieve talk about AI. We've been doing this for a while. What does AI mean to you when it comes to AI in finance?

Brad0:17A lot of the time, when CFOs or heads of FP&A or finance transformation leads come to us, they are mandated by the board to drive AI efficiencies within the organisation. And they don't actually really know what that is.

Brad0:32To be fair, most CFOs don't really know what that is and how it's going to affect their organisation. It could be a number of things. We can talk about LLMs, we can talk about generative AI, we can talk about agentic AI, and maybe talk through that in a little bit.

Brad0:50Essentially, it's a lever to help drive efficiency within your business. That could be a head count reduction. It could be freeing up people who are overburdened. It could also be giving you insights into data you didn't know was there, to help drive decision-making within your organisation. Keen to get your thoughts on that. Should we talk through the various elements of AI?

Language models, generative AI and agentic AI explained

Alex1:12I was going to say, let's not be like the rest. Let's explain it. Half the time I'm confused by all the acronyms out there for AI. What I do know is that the acronyms being used often aren't what's being built.

Alex1:28If we think about agentic, it's the hot topic right now. It's the buzzword I think everybody's using, a bit like SaaS or EPM or CPM, whatever you want to call it. These were buzzwords everyone was using.

Alex1:45When I think about an LLM, I'm thinking about the engine that predicts the words that come next. It's the foundation that generative AI uses. Say I write an email or create a draft. Or I get it to write a brief or an outline for a presentation. Or I spin up a variation to my forecast. Whatever it is, I'm using generative AI.

Alex2:15And that generative AI is supported by the large language model that sits underneath it.

Brad2:23Yep. True.

Alex2:24It's really simple, right? When we start to talk about agentic, and this is something that frustrates me, and I'd love to hear what frustrates you in the market, when I hear agentic, I see generative.

Brad2:38True.

Alex2:39Agentic is execution. It is executing on what the generative is saying.

Brad2:47True.

Alex2:48What are your thoughts?

Agentic AI performs a task, such as account reconciliations

Brad2:49Well, 1 of the things we're looking at doing within Constancia is building our agentic AI platform. If I look at what we're looking to do there, as an example, it's taking a large language model with some generative AI and feeding data into it.

Brad3:06For example, a set of requirements. That set of requirements could then be turned into design documents or sets of functional requirements. That's agentic AI, and that's what we're going to be trying to do as an organisation.

Alex3:20100%. It's performing a task.

Brad3:22Performing a task, right. If we take that back into the finance world, agentic AI could be automation of your balance sheet reconciliations or your account reconciliations. It could automate 90% of the items coming through there. Then it generates a report that says you need to look at the following 5 reconciliations, because they fall outside of tolerance.

Brad3:45As an example, that's agentic AI. But that could also be process automation. Sometimes agents sit on top of a system, sometimes they sit within a system, or you could talk about automation within a system. So there are various elements people need to clearly understand.

Finance data and the security of frontier models

Alex4:03I'll tell you what. If I'm a CFO, I'm sitting there pretty lost right now. Do you know what I mean? When it comes to AI and finance, there are so many things you just touched on.

Alex4:15Account reconciliations. Account reconciliations take the trial balance. I don't want my trial balance sitting in an AI model.

Alex4:22These frontier models, and when I say frontier model, what I'm really talking about, to put it simply, is things like ChatGPT, Claude and Microsoft Copilot. 1 thing we're working with our clients on heavily at the moment is the infosec, the risk and the security that wraps around that financial data.

Alex4:42That's where I would challenge. I think the technology in the market that is native to AI is the safe investment when it comes to AI, as opposed to developing your own in-house AI for finance data.

Brad5:10Correct. I think a lot of CFOs and IT professionals don't actually understand AI, how it works, how the data gets used within it...

Alex5:19Talk to me about that.

Brad5:20...and how much risk you expose your organisation to. To a large extent, that's why a lot of these SaaS companies are building their little agentic factories, if you want to call it that, that work with modules.

Brad5:33It's a module within their SaaS application. It brings everything together and allows you to work in a secure, controlled environment.

Brad5:40I think you surfaced a really good topic there. There's this big drive in the market, with people all of a sudden thinking they can go and build their own agents on top of all this data and it's going to work just fine. There is so much that goes into it.

Brad5:55If you think about building your own agents, a lot of the time people can build something really quickly through Claude Code. It will work and it will do what you need it to do. But unless you have built it with enterprise rigour in mind, it exposes you to so much risk.

Brad6:12Hackers can hook into those agents so quickly. If you don't have the expertise to secure that, it's something I wouldn't even suggest you undertake. And we're seeing that in the market.

Brad6:26I think this goes back to a point we haven't really covered. 1 of the main concerns CFOs have is how to build the finance teams of the future, and how they use agentic AI in finance.

Building an AI native finance team

Alex6:45AI native FP&A teams.

Brad6:47Yeah. How do they go about enabling that? A lot of the time the easiest wins are the simplest things. We talked about generative AI, and that's developing a board pack, as an example, through a suite of tools. Developing a narrative to support the report, or the contextual information, the supplemental data.

Brad7:12Some of the more process-heavy and task-heavy work that is inefficient in an organisation needs a little more thought. So you need to be able to roadmap out your agentic factory, if you want to call it that, or your process.

Brad7:28Then decide where you use generative and where you use an LLM, and enable your team to use it efficiently and securely. That's particularly true if you're a listed company, as an example. You don't want all that data getting out.

Why Constancia starts with the data model, not the AI

Alex7:36So let's think about why we founded Constancia. We founded Constancia to provide consistency and constancy to the office of finance and operations, both on their transformation and on the value they create from their investment.

Alex7:55I'm going to go back to the simplest thing. We talk about this all the time, Brad. It's the right to left planning of your transformation. And let's talk about AI. AI is technology. What have people been implementing for the last 50 years?

Brad8:12Technology.

Alex8:13Why are people replacing their technology?

Brad8:16To drive efficiency and return on investment.

Alex8:20And also because their data model doesn't support it.

Brad8:23True. True. True.

Alex8:24So Constancia is there to analyse, diagnose and understand the data model that is going to support your AI investment. I speak to CFOs every single day. They're pilots, but let's call them programmes, just to be kind. I'm yet to see one that's been rolled out to production at an enterprise or medium business level.

Alex8:55There are a lot of great MVPs within pockets of finance teams. And I'll say it again. Why are people replacing their EPM technologies or their finance systems? Because there are lots of different pockets of people using different tools.

Brad9:20Correct.

Alex9:21So let's try and avoid the last 20 years. History often dictates the future, right?

AI pilots without a business case, and the cost of tokens

Brad9:27Yeah. I think people are seeing AI as a thing they need to get done really quickly. They're not doing proper business cases to support it. They don't fully understand the costs and implications of implementing AI.

Brad9:42They're really just following the old analogy, throwing something at the wall and hoping it sticks.

Alex9:49Spaghetti.

Brad9:49Spaghetti, right. You've got to understand why you're implementing, and understand the return you're going to get, to justify the spend. There are exorbitant amounts of money being spent on AI at the moment, and enterprises need to really rein that in.

Brad10:10As an example, a lot of these large AI models are token based. They are heavily subsidised through funding, through investors. At some point those investors are going to want a return on the investment.

Brad10:28There's going to be an industry around token-based usage, because it's going to get really expensive. So you are investing in a toolset now that might be 10 times more expensive in 10 years.

Alex10:41So I just did a bit of a search here, right? Simple search. Yes, I used AI.

Brad10:49Generative AI?

Alex10:49Yeah. If we look at AI spend, it's roughly $2.5 trillion in 2026, and it's up 44% year on year. The 1 thing I'll be tuning in for is the answer to how much of that's been realised. That's what I'm going to look at.

Alex11:11Where we're really trying to focus, and I'm going to touch on the spend, Brad, because you talked about that. We work with technology which is AI native. We also work with technologies that aren't AI native but have AI capability.

Alex11:30Look at the licensing model. Look at token cost. Look at the usage allowance. Look at the data retention and the storage of the data that is created by the AI. All of this costs money.

Alex11:48As we look at the commercial and the capital markets today, the money in AI is in the storage of data and the rationalisation of data. Most AI programmes should start with a simplification and clean-up programme. That's being skipped.

Brad12:12Yep. Yep.

Alex12:13And that's a bad thing.

Brad12:15I think sharing of data and knowledge between people and the enterprise is essential.

Most AI pilots fail because they run as tooling

Brad12:21Yes. I want to go back to your point there on the spend. Was it 2.5 trillion?

Alex12:29Something crazy. Who knows? AI generated it.

Brad12:32I think I read a report where something like 80% of these pilots have failed.

Alex12:4070, yeah, 76. The metrics vary. This one here is 79%, but we'll draw an average. We're in finance, right?

Brad12:47That's a crazy amount of money. What we're seeing is people running these programmes as pure tooling, not like a software implementation. It has to be a proper project, run like you would any other large enterprise programme.

Alex13:02Ooh, that's a great topic. Run as tooling instead of enabling. Sorry to cut you off there, Brad.

What a board wants from AI after a private equity buyout

Alex13:07The reason I'm thinking about that is the board pressure the CFO is under. We've been talking to a CFO recently who started about 2 months ago. He's new in the seat. The company has been taken from public to private.

Alex13:24They're starting to move their operating model from an efficiency house and a governance house to a growth house. It's a standard private equity situation, right?

Brad13:34Yep. Yep.

Alex13:35The board is looking for head count reduction. They're looking for time to report and timeliness of insight. They're also looking for strategic business partnering to happen in half the time, with double the insight gained. And they want AI to do that.

Alex14:00How much time do we have, right? Working through that is a situation in itself, because the old people, process and technology adage hasn't gone away.

Brad14:08No, it hasn't.

Alex14:09So I'd love to hear your view as we work through it. Boards want speed. They want cost reduction. They want efficiency and they want better outcomes. And they want timeliness of data.

Alex14:22That's 5 things, and there are a lot more, right? I'm not ignorant to the fact there are a lot more. But from my viewpoint, if I'm a CFO, where does AI fit in this?

Plan an AI programme back from the outcome

Brad14:35You talked about right to left planning, and I think we opened with that. What you've just described is obviously the far right. Or is it left to right planning? Which way round is it?

Alex14:47It depends whether you're in the southern or the northern hemisphere.

Brad14:50So you've started with the outcome they're looking to drive. Then you've got to work your way back from there and understand the roadmap to get there. You have to run it like any other large enterprise programme of work, because you've got strategic initiatives with a clear mandate and very specific outcomes.

Brad15:14Instead of just throwing tooling at it, I think you have to run a proper programme of transformation. It's like any other finance transformation, just with a different toolset, right? I think that's how you've got to see it, and you've got to run it properly.

Alex15:29100%.

Brad15:29We talked about people, process and technology. You've got to analyse the process. You've got to look at the people and the enablement of your people. And then the tooling, and what tooling best suits your organisation.

Brad15:44And you've got to choose a toolset that will be there in 10 years' time, right? Not just some fly-by-night AI startup that might not be there in 10 years.

Alex15:57A lot of vibe coders.

Brad15:59Yes, and you've got to really analyse that and understand the risk that comes with it.

What we covered, and how we help

Alex16:04I agree. Just to wrap this up from our perspective, we touched on a lot of things there. We touched on AI in finance, governance and data quality. We touched on security and information security in finance.

Alex16:20We touched on the different definitions of AI, and by the way, we only did about 20% of them. If you're also struggling to make sense of all this, this is what we're working with people on every day.

Brad16:40Thank you for getting those words out of my mouth. I'm Australian, so I apologise.

Alex16:45Yeah, you suddenly start speaking English.

Brad16:47It's always wrong when there's a South African teaching an Australian.

Alex16:50There we go.