Measuring the Real Financial Payoff of Enterprise AI

Charles Skamser

Welcome to Inside PX42! I'm Charles Skamser, and I want to jump right in with two rather extraordinary warnings that landed within a couple of days of each other this past week. On October 24, Torsten Slok, the chief economist at Apollo Global Management, published a note arguing that there are no clear signs of an AI payoff outside the technology sector. His evidence is aggregate. Revenue per worker and operating margins have risen sharply among the Magnificent Seven, but they look essentially flat across the other 493 companies in the S&P 500. Then on October 25, Elon Musk posted on X that he expects what he called guppy level death and consolidation among more than 200 specialized AI cloud companies, the so called neoclouds. Reuters and Futurism both covered it on October 26. Those are two different economic questions, but together they tell us this extraordinary AI investment cycle is entering a period of real financial scrutiny.

Charles Skamser

As always, I'm joined by Edward Hamilton and Catherine Spencer. Before we dig in, I want to be upfront about where I'm coming from. I've spent considerable time recently talking with CFOs, Wall Street analysts, and industry analysts about exactly this question, which is whether all this AI spending is going to pay off. And I have a somewhat different spin from the increasingly negative headlines. That view comes from more than thirty years advising C level executives, working with industry and stock market analysts, and delivering advanced technology to Global 500 organizations. I've watched a lot of technology waves, and the same hard question arrives every time: where is the measurable business value? My concern is that we may be measuring the right economic problem at the wrong stage of the transformation.

Edward Hamilton

Let me test that straight away, Charles, because I suspect a fair number of listeners heard the phrase different spin and thought, here comes the vendor. The founder of an AI consultancy tells us the AI skeptics are measuring too early. That is a convenient position for someone selling the transformation. So what would convince me you're not simply optimistic by trade? And the other thing that puzzles me, from the Wall Street side, is that markets keep assigning enormous value to infrastructure and AI suppliers while mainstream companies can't clearly point to an earnings gain. Valuation is a claim about future cash flows. Where are the cash flows?

Charles Skamser

That's a fair challenge, and I welcome it. The burden of proof here is real, and I'd say it sits with people like me. I don't expect anyone to accept my view on authority, and I'm not going to claim PX42 has proven particular return numbers in client deployments, because that isn't what this episode is about. What I'm offering is an argument about where to look and what to measure. And I'd add that the CFOs I speak with are not hostile to AI. They're simply asking for a financial definition of success, which is exactly what they should be doing.

Edward Hamilton

Good, because I'd like to keep the market side honest. Expectations and realized cash flows are different things. A share price can rise on a belief about 2030 while the income statement in 2026 looks unchanged. Slok is essentially pointing at the income statement. So if I'm a board member and I've approved real money, I want to know what has shown up in margins, not what has shown up in a demo.

Catherine Spencer

May I offer a slightly different worry? I think the bubble debate, while legitimate, can become a distraction. Every week there's another argument about whether the whole thing is a bubble, and meanwhile the practical question goes unasked. What are organizations actually doing with this technology, and what has genuinely entered production? In my work with enterprise clients, the interesting answers sit in the unglamorous details. Which workflow changed, who approved it, who trusts it, and who gets called when it goes wrong. A macro headline can't tell you any of that.

Charles Skamser

And that's why I want to separate the two warnings, because they're often blurred together. Slok is talking about buyers, meaning operating companies and whether their revenue per worker and margins have moved. Musk is talking about suppliers, specifically the financing of specialized AI cloud companies. Those are two different balance sheets. A neocloud carries heavy capital spending, debt, GPU depreciation, power constraints, and customer contracts. An enterprise buyer carries operating costs, revenue, working capital, and customer outcomes. One can be stressed while the other creates value, and the reverse can also be true. Musk's count and his forecast are claims, not established outcomes. And his warning is not that AI has no value.

Edward Hamilton

They're not entirely independent, though, and I don't want us to be too tidy about it. If the supplier side goes through a violent shakeout, buyers inherit consequences. Pricing, vendor stability, contract continuity, they all matter to a CFO who has built a process on top of someone's compute. I'd say the two balance sheets are separate but they share a plumbing system.

Charles Skamser

I'll grant that fully. A buyer needs to understand supplier risk, and it's one reason I favor vendor neutral architecture. But the plumbing point supports my larger argument, which is that enterprise leaders can't outsource their economic accountability to the AI boom's general momentum. Whatever happens to the neocloud financing, the buyer still has to answer a very plain question about what changed in the business. That's where I want to spend the rest of the hour.

Edward Hamilton

Then let's start with the evidence, because Slok's note deserves precision. As I read it, he observes that revenue per worker and margins have expanded among the Magnificent Seven, while the other 493 companies in the S&P 500 look flat. He also makes a point about tokens, that the volume of AI usage is not a measure of value. And he stresses that business judgment involves accountability, trust, ethics, and unwritten rules that a model doesn't carry. But we should be careful with what that shows. Is it fair to say zero payoff?

Charles Skamser

No, and I think the caveats matter. The zero financial payoff framing is Slok's interpretation of aggregate data. It isn't proof that no individual company outside technology has benefited. The S&P 493 is also not the same as the non technology economy. And aggregate revenue per worker and margins are moved by many things, such as interest rates, pricing, mix, and demand. They can't prove that AI caused a change or failed to cause one. I take his point seriously as a macro signal. I just wouldn't turn it into a verdict on every enterprise project.

Catherine Spencer

Then let's look at the surveys, which are at least closer to the enterprise. McKinsey's 2026 Global State of AI survey found that roughly 80 percent of respondents say AI improved their individual productivity. That's a strong number. But 37 percent report a positive enterprise level impact on earnings before interest and taxes, usually shortened to EBIT. And only 6 percent meet McKinsey's definition of AI high performers, meaning significant AI value and at least a 5 percent EBIT impact. Those are survey measures, so they're self reported. What I find telling is that the higher performers were much more likely to have redesigned their workflows.

Edward Hamilton

Let me translate that into investor language, because the gap between 80 and 37 is the whole story for a CFO. Eighty percent is individual productivity. Thirty seven percent is a statement about operating profit. In between sit margins, recurring revenue, cash conversion, return on invested capital, and operating leverage. Operating leverage is the key one. If a technology investment doesn't change the relationship between your fixed costs and your growth, it hasn't changed your earnings power. So here's my direct question. After years of investment and enormous attention, is 37 percent not simply a disappointing score?

Charles Skamser

It demands accountability, Edward, and I won't dress it up. If I invested in something for years and fewer than four in ten leaders could point to a positive profit effect, I'd want hard answers. But I'd put two things beside that number. First, it is evidence against a literal claim of universal zero payoff, since more than a third report a positive enterprise impact. Second, the same survey indicates that some firms see financial benefits at the functional level even where they can't yet see them in enterprise EBIT. That is exactly what you'd expect if value is real but locally concentrated.

Edward Hamilton

Or exactly what you'd expect if people are generous in surveys. Self reported EBIT attribution is a notoriously soft thing. A manager who championed a program has every incentive to believe it worked. I'm not dismissing the data, but I won't treat 37 percent as a floor either. It could be inflated as easily as understated.

Charles Skamser

That's right, and I'd accept it could cut either way. It's also why the Federal Reserve Bank of Atlanta's work is useful. In a March 2026 working paper, researchers surveyed nearly 750 corporate executives and found productivity gains that were positive but uneven, with a gap between perceived effects and measured ones. I read that as an evidence problem and a transformation problem, rather than proof of failure. We have weak measurement and uneven execution at the same moment.

Catherine Spencer

The MIT Initiative on the Digital Economy study from September 2026 is the one I'd like to probe. Researchers analyzed S&P 500 filings and classified about 11 percent of firms as deeply integrating AI by the end of 2025, with around 45 percent running pilots. That raises a definitional question for me. What does deep integration mean, and are organizations being judged on the returns of a technology they have barely deployed? Because if only 11 percent have integrated it deeply, then a flat aggregate might say more about adoption than about capability.

Edward Hamilton

It might, but I'd press on the inference. Filings are a blunt instrument. A company can mention AI heavily in a filing for investor relations reasons, and another can use it well and say little. So the classification is imperfect in both directions. And please let's not conclude that the remaining 89 percent derive no benefit. The study doesn't say that.

Catherine Spencer

Agreed. What it does suggest is a familiar pattern, an investment J curve. You spend first on infrastructure, skills, and process changes, and measurable gains arrive later, if they arrive at all. The Stanford AI Index for 2026 makes a related point in its economy chapter, that adoption is broad and agentic deployment is early, and that task level gains are not the same as macroeconomic evidence. The honest conclusion, I think, is that we're probably in the early part of the curve, but nobody should promise that every project reaches the other side.

Charles Skamser

Which is precisely my point, and it brings me to how I'd organize the whole conversation. Individual productivity, departmental financial impact, enterprise EBIT, macroeconomic productivity, and the financial viability of AI infrastructure are all related, but they are different measures. Most of the heated argument in the headlines comes from sliding between them. Someone cites a personal productivity gain to answer a macro skeptic, or cites a macro number to dismiss a functional win. Let me offer a way to keep them apart.

Charles Skamser

This is a PX42 editorial framework, not a validated academic model, so weigh it accordingly. I think of it as a four stage value ladder. Stage one is access to intelligence. That means model subscriptions, copilots, compute, and data services. It's the investment stage, and the benefit is only potential capability. Stage two is personal or departmental productivity. People draft faster, summarize calls, write code snippets, and research quicker. Those gains are real, but they're often local and self reported. Stage three is integrated workflow transformation. AI is connected to enterprise applications and governed data, decisions have defined owners, roles are redesigned, and people and agents can coordinate end to end work reliably. Stage four is enterprise economic performance, where the change shows up in net revenue, margins, working capital, customer retention, risk, or throughput, with measurable attribution and an accountable owner.

Charles Skamser

My spin is this. Many businesses are buying at stage one, celebrating stage two, piloting stage three, and expecting stage four financial results next quarter. That's a mismatch of stage and expectation. The missing work isn't smarter models. It is organizational redesign, integration, delegated authority, economic prioritization, and measurement.

Edward Hamilton

I like the ladder, but let me attack the jump between stage two and stage four, because that's where finance departments get burned. Suppose an assistant makes a process twenty percent faster. Profit does not rise by twenty percent. Maybe the saved time becomes longer coffee breaks, or more internal email, or capacity nobody uses. A CFO can't book saved minutes. Cost reduction is real when you avoid hiring, reduce overtime, retire a contract, or take on more volume with the same team. Revenue is real when it's realized, not when it's an opportunity in a pipeline. So what value can the CFO actually recognize?

Charles Skamser

You're right, and I'd say that's the most underappreciated problem in this whole debate. Saved minutes have no realized financial payoff unless the organization does something with them. It might redeploy capacity to higher value work, increase throughput, improve service levels in a way that retains customers, or avoid a genuine expense. Otherwise the time simply dissolves. And I'd go further. Part of the reason the aggregate data looks flat could be exactly that, local productivity that never gets converted into an economic decision.

Edward Hamilton

Then you've conceded that a good deal of the stage two celebration is hollow.

Charles Skamser

Not hollow, unfinished. There's a difference. A faster analyst is a real improvement in capability. It becomes a financial result only when somebody redesigns the work around it. That's the gap I'm pointing at, and I don't think it's a gap in intelligence.

Catherine Spencer

And the redesign is overwhelmingly a human problem, which is why I push on it so much. When I look at why pilots stall, I rarely find a model that can't do the task. I find a team that doesn't trust the output, a manager who hasn't changed the approval process, a security review that arrives late, or training that happened once and was never reinforced. Someone also has to keep improving the system after the launch party, and that someone needs time, authority, and a budget. MIT Sloan research from September 2026 points in the same direction, suggesting that sustained participation by employees and domain experts, more than raw access to better tools, distinguishes initiatives that reach scale from those that quietly fail.

Edward Hamilton

Let me be a little provocative. If the organizational work is that heavy, then perhaps the honest message to a board is that returns will be slower and more expensive than anyone pitched. The cost of that change management belongs in the denominator. Many business cases I see price the licenses and forget the redesign.

Charles Skamser

I agree completely, and I'd go further. The business case should include implementation and recurring operating costs, including the humans who keep it running. If those are left out, the return looks better than reality, and then the CFO feels misled a year later. I'd rather have a smaller, honest case that survives scrutiny.

Catherine Spencer

That brings us to Slok's other theme, which I think gets too little attention. He emphasizes that business judgment includes accountability, trust, ethics, and unwritten rules. Agents don't learn those from a prompt. In practice, that means we should be talking about bounded autonomy. An agent is given a defined role, defined permissions, and a defined threshold beyond which a person decides. Who approves an action? Who is accountable if a recommendation is wrong? Those questions must have names attached to them before deployment, not after the first incident.

Charles Skamser

Exactly. The goal is bounded autonomy with governance, not irresponsible delegation of executive judgment to a model. And I'd treat agents as capable but imperfect systems that need engineering, verification, authorized delegation, and clear ownership. If we do that, autonomy can expand gradually as evidence accumulates, from human review of every action toward properly bounded action on lower risk decisions. So the real question becomes: if organizations need a different operating model, what should they build, and what should they measure?

Charles Skamser

Let me describe what PX42 proposes as an operating discipline, and I'll say plainly that this is our approach and our point of view, not a set of proven client results. It starts with economic prioritization. Before anyone builds an agent, we assess readiness across the organization, the process, the technology, the data, the integration, the security, and the capacity for change. We call that the AIAgentReadinessIndex. The output isn't a score for its own sake. It's a short list of two or three economically defensible opportunities, with baseline metrics and expected returns written down before the work begins.

Edward Hamilton

Then I'll ask the CFO question about what you just said. Define return on investment for me in a way I'd sign. My definition is incremental financial benefits, minus the full implementation cost and the recurring operating cost, over a defined period, adjusted for risk and attribution. And inside the benefits I want categories kept apart. Booked savings are not capacity released. Revenue opportunity is not revenue realized. Risk avoided counts only against an independently measured baseline. If your framework can't separate those, it's a slide deck, not a business case.

Charles Skamser

I'd sign that definition, Edward. And I'd add that the baseline has to be established before the intervention, because otherwise attribution becomes an argument rather than a measurement. If the program can't say what the number was before, nobody can say what it became because of the agents.

Catherine Spencer

And the second layer is control. Tell me about the architecture for delegation, because that's where my questions are.

Charles Skamser

We call it AIAgentOS. The idea is a vendor neutral architecture that defines agent roles, authority, delegation, permissions, data access, observation, accountability, and audit. Think of it as a control plane. Every agent has a documented scope, a permission matrix, a log of what it did and why, and thresholds that trigger escalation to a human. The progression runs from human review toward bounded autonomous action, and each step is earned by evidence. Around that, PX42 works with an ecosystem. Databricks can serve as a data and AI foundation where appropriate. ServiceNow brings enterprise workflow execution. UBIX contributes analytics. Reliath and its VerifiedTruth capability address verification of evidence and reasoning. And Entrelid concerns architectural authority and controlled delegation. I'd describe those as capabilities that can fit together, not as a claim of a completed customer deployment.

Catherine Spencer

So let me test the human side of that. If an agent recommends something and a manager approves it without really reading it, has accountability actually been preserved? Rubber stamping is a risk. An audit log tells you who clicked, not whether anyone understood. For me, trust has to come from explainability and from the person on the hook having real authority to say no, and real time to evaluate the recommendation.

Charles Skamser

That's a serious point, and I don't have a tidy answer to it. Part of the response is design. High risk actions should require review with the evidence in front of the reviewer, and the review process itself should be measured, for example how often humans override and why. Part of it is culture. People need to feel that questioning the system is part of the job. I'd say the override rate is as informative as the accuracy rate.

Edward Hamilton

Let me bring it down to earth. You have a third idea, I think, on connecting operational signals to financial ones?

Charles Skamser

Yes, EnterpriseBusinessHealth360. The strategic idea is to connect operational and financial signals, anticipate deteriorating business performance, recommend a corrective action, and then measure whether that action improved the outcome. I want to be clear that this is a proposed capability and not a deployed customer proof point. Let me give a few illustrative possibilities, and they are possibilities, not results. Consider a multi rooftop auto dealership group. The system could surface aging inventory and missed lead follow up, recommend bounded interventions, and then track vehicle gross profit, days to sale, conversion in the business development center, and service bay utilization. Or an insurer, where the goal might be to detect claims leakage or service bottlenecks and to measure cycle time, indemnity accuracy, customer outcomes, and risk adjusted savings. Or a manufacturer anticipating supplier delays and tracking service levels, working capital, and expediting costs.

Edward Hamilton

I notice every one of those ends in a metric a CFO already reports. Gross profit per vehicle, claims leakage, working capital. I find that reassuring, I'll admit, because it forces the conversation toward the income statement and the balance sheet. But the hard part is still attribution. If days to sale improve, was it your recommendation, a seasonal shift, or a manufacturer incentive? What's the control?

Charles Skamser

That's the right question, and the answer is design. You use a baseline, comparison groups where you can, staged rollouts, and a measurement owner in finance who isn't the project sponsor. I'd rather a CFO's team own the scorecard than have the technology team grade its own work. Without that, we'd be repeating the same self reporting problem we criticized in the surveys.

Catherine Spencer

I'd add one practical point for the dealership example, since the people matter. Whoever runs the business development center has to help define what a good intervention looks like. If the recommendation arrives from outside their expertise, it will be ignored, and rightly so. The domain experts have to be co authors of the workflow, not recipients of it.

Edward Hamilton

Let me return to the Wall Street lens, since we've been in the enterprise for a while. The skeptic's question is whether we're watching a dot com style misallocation of capital, a long productivity J curve, or some mixture of the two. I don't think a simplistic answer is available. The telecom and fiber buildout of the late nineties involved real technology and real demand eventually, but a lot of capital was destroyed on the way. Time horizons matter, adoption rates matter, the cost of capital matters, and so does whether a given company survives long enough to see the payoff.

Charles Skamser

I think a mixture is the most defensible reading, and I'm cautious about confident forecasts in either direction. Without inventing anything about my conversations, I can tell you how analysts I speak with tend to frame it, which is as several separate questions. Is the capability exciting? Yes. Are vendor revenues growing rapidly? Evidently. Is the funding of AI infrastructure sustainable? That's a question about debt markets, utilization, and energy. And are nontechnology enterprises monetizing AI? That last one is what Slok is probing. These can have different answers. The industry needs cleaner measurement and more operating discipline, and I'd say that's the common thread across all four.

Edward Hamilton

Let me press on Musk's piece, since it is a financing story. Specialized AI cloud companies depend on debt markets and on revenue from AI data centers, and Reuters reported that his warning hit neocloud stocks as debt concerns mounted. His phrase was guppy level death and consolidation, and his count was more than 200. But I'd flag that it's a forecast. The companies are exposed to GPU depreciation schedules, power constraints, and customer concentration, but exposure is not the same as failure.

Charles Skamser

Right, and I'd resist any reading that Musk said AI applications lack economic benefit. He didn't. His concern, as reported, is the financing and possible consolidation of a specific class of infrastructure provider. A shakeout among suppliers is a separate matter from whether a claims process or an inventory process gets better inside a buyer. Consolidation can even be a normal stage in a maturing industry. What I'd tell a buyer is to treat it as a supplier risk to manage, through contract terms, portability, and architecture that doesn't lock the business to a single provider.

Catherine Spencer

That raises a question I find interesting. If the infrastructure wave does suffer a shakeout, could enterprise returns actually improve? Compute might become cheaper, and cheaper inference changes the cost side of every business case. But I'd be careful not to oversell it. Cheaper inference does not integrate your data, redesign your process, or settle who is accountable for a decision. Those costs don't fall because a GPU got cheaper.

Edward Hamilton

Though it would shift the arithmetic on marginal cases, and that's not trivial. A use case that fails a hurdle rate at today's costs might clear it at lower ones. I'd agree with Catherine that cost per token is rarely the dominant cost in a real enterprise program. The dominant costs are people, integration, and change.

Charles Skamser

So we've got a fairly consistent picture. Supplier economics and buyer economics are two different tests. Now let me put some decisions on the table, because I want listeners to leave with something they can use within ninety days. Here are three questions I think every CEO and board should ask about their AI investments heading into 2027. First: which two or three enterprise outcomes will this initiative improve, and how will we measure the change against a credible baseline? Second: what must change in our processes, systems, responsibilities, and people, and not just in our software stack, to capture that value? Third: who owns the result, what authority will we grant the agents, how will we control the risk, and when will we stop, redesign, or scale the investment?

Edward Hamilton

I'd sharpen the investment discipline underneath all three. Ask for cash costs, not just license fees. Ask for unit economics, so you can see whether the value per transaction holds as volume grows. Ask for a range of outcomes with some honest uncertainty rather than a single point estimate, and a timeline for when the value should be recognized. And ask about opportunity cost, because every dollar and every hour of management attention spent here isn't being spent on something else. If a project can't beat the alternatives, it's not a good project, however clever the technology.

Catherine Spencer

And my addition is about participation. Employees, domain experts, and managers should be part of the transformation rather than passive recipients of a new tool. The research from MIT Sloan suggests that sustained involvement is a differentiator, and it matches what I see. If your people don't help shape the workflow, they won't trust it, and if they don't trust it, the economics never arrive. So in the next ninety days, I'd ask every executive to name who in the business is co designing the change, and to give them the time to do it.

Charles Skamser

Let me close with this. I don't think the future of AI will be judged by the number of tokens consumed, the scale of the data centers, or the cleverness of the latest model. Enterprise leaders will judge it by whether it changes the economics and resilience of their business. Slok's warning and Musk's warning are both reminders that the accountability test is arriving, in different forms, across the stack. I'd call that an opportunity as much as a threat, but only for the organizations willing to measure honestly and redesign seriously. Edward, Catherine, thank you both for pushing back. To our listeners, thanks for joining us on Inside PX42. If you'd like to continue the conversation, please subscribe to my LinkedIn newsletter, Intelligent AI Agents Weekly, and feel free to connect with us at PX42.