Corporate Finance Explained: AI in Corporate Finance

[00:00:00 – 00:19:35]
Right now, in active production, JP Morgan is running over 2000 different artificial intelligence models. 2000. Yeah, 2000. And they are backing this up with a $17 billion annual technology budget. Which is just, I mean, it’s a staggering number. It really is. And if you think that scale is purely, you know, a Wall Street phenomenon, look at Walmart. They are processing over 2.5 petabytes of data per hour through their AI-driven systems. Every hour. That’s hard to even conceptualize. To put it in perspective, that’s roughly equivalent to analyzing 50 million tall filing cabinets of text every 60 minutes. Well, it’s a massive amount of computing power. But, you know, what’s truly significant here isn’t just the sheer volume of data or even those crazy budget sizes. What is it then? It’s the fundamental shift in how businesses are actually treating this technology. Because, I mean, AI in corporate finance has officially crossed a threshold. Right. It is no longer this cute, isolated pilot project run by some siloed innovation team. It has really become the core operational infrastructure of the enterprise. Okay, let’s unpack this. Because if you’re a finance professional listening to this deep dive right now, those mega cap numbers can feel completely detached from your daily reality. Oh, absolutely. They seem like sci-fi. Exactly. So our mission today is to bridge that gap. We’re diving into a highly strategic resource from the Corporate Finance Institute, specifically looking at generative AI in finance, use cases, and applications. It’s a great breakdown. Yeah. And the goal here is to shortcut your path to being a really well-informed, highly productive corporate finance pro by just slicing right through all that vendor hype and looking at the tactical reality. Because the truth is the capabilities we’re talking about, they aren’t just locked up in GP Morgan server rooms anymore. The entire landscape has been democratized. I mean, we are seeing tools like Microsoft Co-Pilot natively embedded right into the software we all use every single day. Excel, Teams, Outlook, all of it. Exactly. So this means the kind of predictive analysis and data synthesis that was required a billion-dollar R&D budget three years ago. Yeah. It is now literally sitting on the desktop of an FP&A analyst at a mid-sized manufacturing company on a random Tuesday morning. That’s a huge shift. It is. So the strategic conversation for you or your CFO has entirely shifted. The question is no longer whether you’re going to deploy AI. Right. It’s how you prioritize its rollout and, crucially, how you govern it. So you don’t inadvertently lose millions of dollars. Okay. So if this technology is already sitting on my desktop, what am I actually supposed to be doing with it when the rubber meets the road? Good question. Because conceptually, we all know AI can write a poem or generate an image. Yeah. But in corporate finance, what is the actual mechanism driving value? Well, the most immediate application and really the absolute baseline where everyone needs to start is task automation. Okay. Specifically, I’m talking about the extraction and reconciliation of unstructured data. Give me an example. Think about the friction of dealing with hundreds of messy, totally uniquely formatted PDF supplier invoices or bank statements. The worst. Right. Historically, an analyst had to manually key that in, or they relied on very fragile optical character recognition software. Yeah. The stuff that broke if a supplier moved their logo an inch to the left. Exactly. The old Control-F keyword matching approach. The system was essentially blind to context. But generative AI actually understands context. So when you pair it with automation tools like Power Query, it can look at a totally scrambled PDF and just understand that total due, amount owed, and final balance all mean the exact same thing. Oh, wow. Yeah. It extracts the line items, cleans the data, and reconciles it against the ledger. It’s doing the intellectual heavy lifting of categorization, not just scanning pixels. So task automation buys you time. But time to do what? You aren’t just going to sit around enjoying a two-hour lunch break. I mean, we’d like to, but no. Right. You’re going to use that freed-up data to actually look into the future, which leads right into the predictive side of this. Yes. Because we’re talking about moving way beyond your standard Excel what-if scenarios here. Far beyond. Finance teams are now piping that newly planned data into platforms like Azure Machine Learning to run highly sophisticated predictive models. Okay.

But where the CFI resource highlights a truly transformative shift is when you blend that predictive modeling with narrative generation. Wait. How does that work? Okay. So let’s say an FP&A team is looking at a massive variance report, and the European division has completely blown past its Q3 operating expense budget. A nightmare scenario. So normally, an analyst has to stare at a blank screen, cross-reference a dozen different spreadsheets, and try to write up some coherent explanation for the executive team. Right, which takes hours. But instead, they feed the raw data and the predicted outputs into a Gen AI tool like ChatGPT. Then it just writes it. It drafts the initial variance analysis narrative. It instantly synthesizes all those data points into a readable, grammatically perfect first draft that explains the root causes of the overrun. That is wild. It’s like having a brilliant but incredibly hyperactive intern. That’s a great way to put it. Right. They can process a mountain of information in seconds and give you a surprisingly solid rough draft. But, and this is a massive but, you absolutely would not let that intern send the final report to the board of directors without you checking every single line. The analogy hits the nail on the head. The human is elevated from a data gatherer to an editor and validator. Which is a much higher value role. Exactly. And this exact dynamic is transforming risk assessment and stakeholder communication too. How so? Well, Treasury teams use these systems to monitor global macroeconomic indicators and transaction history simultaneously. Right. Looking for trouble. Yes. Flagging potential counterparty risks or foreign exchange exposure. And once the AI identifies the risk, it can instantly translate that dense data set into three completely different formats. Really? Like what? A high-level executive summary for the C-suite, a granular operational report for floor managers, and a specific narrative for the investor relations team. Three entirely different formats from one single source of truth. Let me pause you there, though, and push back a little bit. Sure. Go ahead. If I am a CFO and I hear that my team’s major competitive advantage relies on basic off-the-shelf tools like ChatGPT and Power Query, I’m going to be a little concerned. Because it sounds too accessible. Exactly. Isn’t that too generic? If a company wants to actually beat their competitors, shouldn’t they be hiring engineers to build a custom proprietary large language model just for their own data? You know, that is the instinct for a lot of ambitious executives. But the CFI resource issues a very explicit warning about this. Oh really? Yeah. Companies that leap straight into trying to build their own custom LLMs almost universally fail. Universally? Yes. They incinerate massive amounts of capital and time, and usually end up with a product that is already obsolete by the time it actually launches. Why does it fail so consistently though? Is it just too technically difficult? It’s technically difficult. It’s incredibly expensive to maintain, and it’s highly prone to hallucination. Right. The secret to actual productivity and finance isn’t owning a proprietary model. The secret is what industry insiders call the stack. The stack. Okay. You don’t need to reinvent the engine. You just need to stack existing powerful tools together efficiently. They’re chaining them together. Exactly. You take basic automation like Power Query. You feed that into an established predictive model. You top it off with a commercial Gen AI for the narrative and you wrap the entire thing in human review. Got it. The value isn’t in the AI itself. It’s in the specific frictionless workflow you build with it. Okay. That clarifies a lot, but I mean it also raises a huge red flag for me. What’s that? If the stack is just combining readily available Lego blocks, essentially, why am I not seeing this highly efficient workflow at every mid-cap company? What is the actual bottleneck stopping everyone from operating like JP Morgan? The bottleneck is the single most critical and easily the most ignored aspect of this entire technological shift. Which is? You cannot layer brilliant artificial intelligence on top of a broken data foundation. Ah, the unsexy stuff. Exactly. This is the tedious work of data governance. If your underlying data is a mess, your AI will simply generate terrible decisions at the speed of light. Now, if you’re an FP&A analyst listening to this, you probably don’t control your company’s data architecture. But this right here explains exactly why that new, incredibly expensive AI tool your boss just bought isn’t working the way the vendor promised it would. Right. It’s not the tool’s fault. So, break down what a broken data foundation actually looks like in practice. It typically fails across three critical layers. Okay. The first is data integration. Let’s say your company operates with deeply siloed systems. Which is pretty much every company. True. So your enterprise resource planning software doesn’t talk to your HR system, and neither of them connects to your customer relationship management software. Total silos. Right. So if an AI is asked to analyze the profitability of a new sales initiative, but it can only see the CRM and is completely blind to the HR labor costs, its analysis is functionally useless. You have to unify that data. Which is why we hear so much about cloud data platforms like Snowflake or Databricks. Exactly. They aren’t just massive hard drives, right? They act like a centralized, instantly searchable nervous system for the entire company. They break down those silos. But even if you do integrate the systems, you hit the second layer, which is data quality. Garbage in, garbage out. The classic paradigm, but amplified. If your system is full of missing values or inconsistent product classifications, like one department calling a product so UA and another calling it alpha product or stale pricing data. It’s just a mess. The AI will confidently synthesize all those errors into a completely wrong, yet highly persuasive, financial model. And because the AI output looks so polished and professional, it’s actually much harder for a human to spot the underlying data error. Exactly. It looks right, but it’s fundamentally flawed. Wow. Okay. But what about the third layer? You mentioned governance and risk. Yeah. The third layer is data access controls. And this is where companies face severe legal exposure. How so? Well, generative AI tools are incredibly efficient at searching absolutely everything they have access to. Right. So imagine your company has an internal AI tool and it crawls a shared drive, reading a highly sensitive, unreleased quarterly revenue forecast that the CFO’s office is still finalizing. Oh no. I see where this is going. Right. Then an analyst in a completely different department, say marketing, asks the AI to help draft a public-facing blog post about the company’s momentum. The AI just hands it over. The AI trying to be helpful inadvertently includes that sensitive, unreleased financial data to prove the point. Here’s where it gets really interesting. That is a catastrophic SCC disclosure violation waiting to happen. Absolutely catastrophic. You’ve literally just leaked your own earnings to the public because your access controls were flat. So the companies that are actually winning at AI in 2026 aren’t the ones with the flashiest tech teams or the best vendor pitches. No, not at all. They are simply the companies who suffered through all that boring, unsexy data governance work back in 2020 and 2021. That is the harsh reality. The organizations that skipped that foundational step are now realizing their ambitious 12-month AI deployment timeline is actually a 36-month timeline. Ouch. Because they have to pause everything, go back and physically clean their data first. Let’s look at the compounding advantage of actually doing that unsexy work. I want to circle back to those behemoths we mentioned at the start. JP Morgan and Walmart. Right. Because when you look at their use cases, they aren’t using AI to launch crazy revolutionary new consumer products. They’re using it to relentlessly compound incremental operational advantages. Yes, JP Morgan’s trajectory is a perfect case study for this. One of their most famous early deployments was a project called Coin Mine or Contract Intelligence, which they rolled out way back in 2016. They had this massive bottleneck with commercial loan agreements. And a commercial loan agreement is essentially a densely packed legal novel. You can’t just run a simple search on it. Precisely. It’s too complex. So they used machine learning that could actually understand the semantic meaning of the legal clauses. Wow. It extracted the critical data points, covenants, interest rates, termination clauses from these complex documents. Saving them how much time? Well, they took a manual process that previously required 360,000 hours of legal and financial review every single year. Oh my God. And reduced it to a process that executes in mere seconds. 360,000 hours. That one single project likely funded their next five years of AI research? Most likely. But it’s not just that they had a smart idea. It’s how they manage it. It is their entire philosophy. First, JP Morgan treats AI as a core operating expense. It is essential infrastructure, not some discretionary budget line that gets slashed through the moment the market dips. That makes a huge difference. Huge. Second, they invested early and heavily in that unified data foundation we just discussed. And third, they apply rigorous credit risk style governance to their models. What does that mean in practice? It means when they deploy a system like LOXM for algorithmic equity trading, or when they roll out an internal large language model to 60,000 employees, they’re treating model risk with the exact same gravity as they treat financial market risk. Amazing. Now let’s pivot from Wall Street to the factory floor, or in this case, operational retail. Walmart. Right. The data giant. 2.5 petabytes of data an hour. How does a retailer translate that much compute into actual operating income? By monitoring a mind-boggling array of variables simultaneously, Walmart systems are ingesting local weather patterns, real-time social media sentiment, granular competitor pricing changes. All at once. All at once. Plus, global supplier capacity constraints. They process all of that to drive hyper-local demand forecasting, which then dictates per-scale inventory replenishment and incredibly precise labor scheduling at the store level. So the secret sauce isn’t inventing some revolutionary new engine. It is simply using AI to reduce the friction in a thousand tiny gears across the entire operation. That’s exactly it. Because I mean, improving an inventory turn by a fraction of a percent or lowering a stockout rate by a millimeter, those seem like rounding errors on their own. On their own. They are rounding errors. But when you compound those fractional friction-reducing improvements across Walmart’s $680 billion in annual revenue, they generate billions of dollars in pure operating income that simply would not exist without the AI. Incredible. And just like JP Morgan, Walmart isn’t only focused on deep special-aid systems. They also deploy broad, everyday productivity tools. Like the co-pilot stuff we talked about. Exactly. They have an internal generative AI tool called My Assistant that has rolled out to 50,000 corporate employees to help with drafting, summarizing, and basic analysis. So to win, the AI has to be deep in your specialized proprietary systems, but also broad across the daily workflows of your regular employees. Yes, both ends of the spectrum. Well, we’ve seen the magic of doing it right. But let’s look at the dark side. What happens when a company ignores these rules? It’s not pretty. What happens when you ignore the messy data foundation, or worse, when you ignore the human-in-the-loop principle? Because as the source material shows, the result is usually a catastrophic financial write-down. The cautionary tales are arguably more instructive than the success stories. And in corporate finance, the most glaring operational failure of the AI era is that Zillow offers.

Ah, yes. Let’s walk through this crash, because it was spectacular. Zillow launched its iBuyer business to systematically purchase homes directly from sellers, do some light renovations, and flip them. Right. And the entire business model relied on their proprietary Zestimate algorithm to generate the offer price. But in 2021, the wheels completely fell off. The algorithm systematically overestimated home values across the country. It was autonomously buying thousands of properties at prices significantly higher than the market would actually bear. And the financial fallout was brutal. Zillow took a $304 million write-down on inventory, suffered another $240 million in wind-down losses, laid off 25% of their workforce, and completely shuttered the business line. Massive destruction of capital. Over half a billion dollars vanished. But let me challenge the narrative here for a second. Okay. Wasn’t Zillow’s algorithm just incredibly unlucky? I mean, the 2021 housing market driven by COVID was completely unprecedented. Nobody could have predicted that specific surge and subsequent volatility. Your challenge actually highlights one of the most vital rules in algorithmic finance. The AI wasn’t unlucky. The failure was mathematically inevitable. Really? Why? Because of a concept called regime change. Unpack that for us. What is a regime change? An AI model is entirely trained on historical data. It learns patterns based on the past. Makes sense. Therefore, when the underlying economic environment structurally shifts, when the regime changes from, say, a predictable pre-COVID market to a highly volatile stimulus-driven COVID market, the AI becomes blind. It doesn’t know the rules have changed? Exactly. It continues extrapolating trends from a reality that no longer exists. We call this model drift. Zillow’s AI lacked any margin of safety for this inherent drift. And they compounded that error by violating the human in the loop rule, right? Absolutely. For consequential capital allocation decisions, an AI should only ever recommend an action, while a human expert must approve it. And Zillow didn’t do that? Zillow inverted that entirely. The AI was making multi-hundred-thousand-dollar purchasing decisions autonomously, and the human teams were just doing spot checks after the capital was already deployed. That’s terrifying. Furthermore, their feedback loops were painfully slow. It took quarters for the executive team to realize the model was bleeding cash instead of days. So Zillow is the ultimate operational failure, trusting an algorithm too much in a shifting environment. But they’re also massive strategic failures. Very true. Let’s look at IBM Watson Health. Starting around 2015, IBM went on a $5 billion acquisition spree. Huge investment. Yeah, they bought up major healthcare data companies like Fitell, Explorys, and Merge Healthcare. And the grand vision was to take Watson, the AI famous for winning Jeopardy, and force it into the medical sector as an advanced diagnostic decision support tool. It was supposed to revolutionize medicine, but instead, the strategic thesis completely collapsed. What happened? In 2022, IBM sold the entire Watson Health portfolio to a private equity firm for roughly $1 billion.

[00:19:36 – 00:23:30]
They incinerated $4 billion in enterprise value. Why did a company with that much technical pedigree fail so badly? Was the AI just not smart enough to understand medicine? It wasn’t an intelligence failure. It was a data foundation and workflow failure. IBM fundamentally misunderstood the mechanism of healthcare data. How so? They underestimated how incredibly messy electronic health records actually are. Messy in what way? Like, disorganized? Exactly. Unlike structured financial ledgers, health records are full of localized doctors’ shorthand, subjective margin notes, incomplete histories, and PDFs formatted completely differently from one hospital to the next. So the AI just couldn’t read it? It couldn’t parse it reliably. Furthermore, IBM failed on clinical integration. Meaning the workflow. Right. The AI required doctors to completely change how they input data during their already frantic daily rounds. It couldn’t naturally slot into the existing workflow. They ignored the stack. Exactly. IBM ignored the Unsexy Data Foundation and tried to force a revolutionary breakthrough rather than compounding incremental wins. Okay, this brings us to the ultimate synthesis. We’ve seen the incredible heights of JP Morgan and the catastrophic depths of Zillow and IBM. How should the listener be operating in this landscape right now? The operational rules drawn from these sources are exceptionally clear. First, you must start with specific use cases that have a clear, measurable return on investment. Like automating invoice extraction. Exactly. Do not try to solve everything at once. Right. Second, you have to fix your Unsexy Data Foundation before you deploy the shiny new tool. Now, silence. Right. Third, ensure your feedback loops are rapid enough to catch model drift before it costs you money. And fourth, every single workflow must be designed with human-in-the-loop oversight. The AI does the heavy lifting. The human makes the final call. Perfectly said. Those are the organizational rules. But let’s bring this down to the individual. If you’re sitting at your desk tomorrow morning, what does this mean for you personally? It requires a harsh career reality check for 2026. Your biggest career risk is not that an artificial intelligence is going to replace your job. The machine isn’t coming from my desk. No. Your biggest career risk is being massively outproduced by the human analyst sitting at the desk right next to you. Wow. If your colleague is utilizing these democratized AI tools to instantly draft their variant narratives, automatically clean their messy data sets, and generate complex scenario sensitivities, they are going to produce three to five times the output that you can generate manually. So the battle isn’t human versus AI. The battle is AI-augmented finance professionals versus non-AI-augmented finance professionals. And the gap between those two groups is compounding so rapidly that if you don’t start now, it will be mathematically impossible to catch up. Precisely. The imperative is to pick one single workflow this week. Just one favorite. Figure out how to apply an AI tool to it and begin building your own compounding advantage. Which leaves us with one final deeply provocative thought to ponder based on everything we’ve heard today about Zillow, Model Drift, and those dangerous regime changes. Think about those thousands of AI models running at JP Morgan and the massive scale of Walmart’s algorithms. If artificial intelligence models inevitably drift over time, and if they inherently become blind and stop reflecting reality, the exact moment market cycles experience a structural shift. Which they always do. Right. Then the most valuable, irreplaceable skill for a human finance professional in the future won’t actually be the technical ability to operate the AI. No. Technical operation will be fully commoditized. The ultimate human superpower will be possessing the deep, fundamental market intuition required to look at a highly confident, beautifully formatted, grammatically perfect AI output, and instantly recognize that the algorithm is dead wrong.