How to Create Real Business and Financial Outcomes at Scale in the Face of the AI ROI Crisis

Charles Skamser

Hello, and welcome back to Inside PX42, the official podcast of PX42 Consulting, where we are discussing how to build the Intelligent Enterprise with AI Agents. I’m Charles Skamser, Co-Founder and CEO of PX42. I'm joined as always by our premier AI strategy consultant, Catherine Spencer, and our lead systems researcher, Edward Hamilton. And today, we are discussing what some are calling the impending AI ROI Crisis. We have to start with the actual numbers. If you look at what Ed Zitron was recently arguing on CNBC, he’s pointing to OpenAI’s projected loss of, what, thirty-eight point five billion dollars against thirteen point zero seven billion in revenue for 2025. And- and he’s saying the math just doesn't work. But from where I sit, advising the Global 500, the issue isn't that AI doesn't work. The issue is that users are treating a massive structural transformation like a simple model-selection exercise. So, we're going to investigate the generative AI ROI crisis, figure out what's true and what's gloom-and-doom hype, and show you how to create real business and financial outcomes at scale based on what we have learned over the past 12 months.

Catherine Spencer

Yes, and it is a massive crisis of confidence, Charles. I mean, from my fifteen years consulting with enterprise clients in London on AI integration and reinforcement learning, I’m seeing boards absolutely freeze up. They look at Alex Karp’s warnings about tokenmaxxing—this absurd, wasteful spending on raw tokens with zero workflow logic—and they think, "Is this all just expensive theater?" It's incredibly classy to see the mainstream financial press finally ask these tough questions, but the answers they're getting are so incomplete.

Edward Hamilton

Exactly, Catherine. It’s Edward here, and looking at this from a pure multi-agent systems research perspective, Zitron’s critique completely misses the architectural shift. He argues that generative AI lacks economies of scale because- because marginal costs rise linearly with compute. Well, yes! If you are building dumb, brute-force pipelines that route every single query to a hundred-billion-parameter model without any orchestration, your costs will be linear, and your margins will be atrocious. But that is a failure of software engineering and multi-agent routing, not a fundamental law of the technology itself.

Charles Skamser

Exactly, it's an engineering and architecture crisis! As we all know from years of hands-on experience, you don't build a high-performance database by letting every single user run unindexed, raw scans across petabytes of data, right? You build indexing, you build query optimization, you build a caching layer. That is what PX42 has created, and we call it our AI AgentOS. It’s an enterprise operating system that treats AI agents, data fabrics, human-in-the-loop controls, and financial KPIs as a single coordinated system. Let's look at Zitron's claim that Oracle is building seven point one gigawatts of capacity for a single client, OpenAI, and that Oracle actually flagged non-payment risk in their annual report. Seven point one gigawatts! That's enough to power a small country. Why? Because the industry is trying to solve reasoning problems with raw hardware brute force instead of elegant agentic orchestration.

Catherine Spencer

It’s a brute-force mania, Charles. The hyperscalers are projected to spend four point eight trillion dollars on capital expenditures between 2026 and 2030, according to Reuters. Four point eight trillion! And yet, McKinsey’s 2025 State of AI survey found that while eighty-eight percent of organizations use AI in some capacity, only thirty-nine percent report any enterprise-level EBIT impact. That gap—that massive, yawning gap between eighty-eight percent adoption and thirty-nine percent financial impact—is the exact definition of AI theater. People are deploying chatbots that summarize Slack channels and calling it a strategy.

Edward Hamilton

And that's why Karp's term "tokenmaxxing" is so spot on. It describes this mindless consumption of input and output tokens without any regard for decision quality, security boundaries, or financial tracking. If you are just throwing tokens at a wall, you aren't building a system. You're just paying OpenAI or Anthropic to train their next models on your enterprise data, which brings us to the IP risk Zitron mentioned. If you surrender your proprietary workflows to a public API, you are literally paying to commoditize your own competitive advantage.

Charles Skamser

Which- which is why we developed the AgentOS framework. We have to move the conversation from "does generative AI work?" to "what is the specific cost, governance, and workflow model required to make it profitable?" It is about moving from disconnected, ungoverned pilots to a structured, twelve-layer production architecture that the CFO can actually audit. Let's break down those layers and show how we actually solve this.

Charles Skamser

Okay, let's look at the actual architecture. When we talk about a twelve-layer AgentOS, we aren't just adding another software wrapper. We are building a system that starts at Layer 1—the Business Outcome Layer. You do not start with the model. You- you- you start with the financial pain. Are we reducing insurance claims leakage? Are we optimizing working capital? What is the exact dollar value of the problem? If you don't baseline that, you cannot measure ROI, period.

Catherine Spencer

And then we move up to Layer 3, the User Experience and Business Portal Layer. This is where our partnership with UBIX.ai becomes so vital. UBIX is fantastic at democratizing data and analytics through a business-user-oriented, no-code SaaS platform. Because- let's be honest—business leaders don't want to write python code or manage vector embeddings. They want to see 360-degree business health in a natural language interface. They want prescriptive recommendations that tell them exactly which invoice is at risk of defaulting, and why.

Edward Hamilton

Yes, and that user experience must be supported by deep, highly secure infrastructure at the lower layers. Let's talk about Layer 7—the Zero-Copy Data Fabric—and Layer 10, the Security and Runtime Isolation Layer. Zitron rightly pointed out the risk of data leakage and intellectual property loss. In a properly designed AgentOS, we do not copy sensitive enterprise data into the vector database of some external model provider. We keep the data where it lives, whether that's Snowflake, Databricks, or a core banking platform, and we use zero-copy retrieval patterns.

Charles Skamser

Exactly, Edward. You- you don't move the data to the agent. You move the context to the agent, under strict access controls, and you execute that agent in a secure, isolated container. Explain the microVM architecture we use at Layer 10, because that is where we stop the "blast radius" of an agent behaving unexpectedly.

Edward Hamilton

Right, so instead of running agents on a shared, open runtime where they can access arbitrary APIs, we execute each agent within a lightweight, hardware-isolated microVM. If an agent is compromised or encounters a hallucination that triggers an unintended function call, the blast radius is strictly confined to that single, transient microVM. It has zero access to the broader network or database unless explicitly permitted by Layer 9—our Governance, Risk, Compliance, and Verified Truth layer. We treat agents as privileged software actors. You wouldn't give an offshore contractor unlimited access to your core ERP without logging and permission limits, so why would you give it to an AI agent?

Catherine Spencer

That Verified Truth framework in Layer 9 is really the crown jewel for regulated industries. If you're in healthcare, insurance, or banking, a model that sounds "plausible" but hallucinates a policy clause is a massive liability. Verified Truth forces the agent to ground every single output in an approved, immutable source of truth, creating a complete, traceable audit trail. If a customer-service agent quotes a policy, Layer 9 validates that the specific clause exists in the active contract before the response is even sent to Layer 3.

Charles Skamser

It- it prevents the hallucination before it ever reaches the user! And- and this is how you turn a risky science project into a production-grade system. If you look at Layer 11, Observability, and Layer 12, Financial Measurement, we are tracking the cost and performance of every single agent call. We can see exactly how many tokens a specific subrogation agent used to process a claim, and whether that decision actually improved the margin. That is how you get the CFO to sign off on scaling the platform.

Edward Hamilton

Let's dive into the economics of this, because Zitron and Karp are right that raw token costs can destroy a business case if left unmanaged. If we look at the current API pricing, OpenAI's GPT-5 is priced at, say, one dollar and twenty-five cents per million input tokens and ten dollars per million output tokens. Meanwhile, their GPT-5 mini is twenty-five cents per million input and two dollars per million output. And GPT-5 nano is a mere five cents per million input and forty cents per million output. Now, if you route every single basic text summarization or classification task to GPT-5 or, heaven forbid, GPT-5.5 at five dollars and thirty dollars respectively, you are literally burning capital.

Charles Skamser

You're- you're lighting money on fire! It's- it's- it's pure madness. Why use a Ferrari to go to the mailbox? That is where Model Routing and our AgentPrune model come into play. If a task only requires a simple classification—like "is this customer email an inquiry about billing or product setup?"—the AgentOS routes that to a highly optimized, low-cost model like GPT-5 nano or a local small language model. It cost pennies compared to routing it to a high-reasoning model.

Catherine Spencer

It’s about matching the cognitive complexity of the task with the economic cost of the model. And then we have Agent Pruning. In a complex, multi-agent workflow, you might have ten specialized agents. If the system realizes at step three that the issue is resolved, it prunes the remaining seven agents from the execution path. It doesn't trigger unnecessary reasoning loops or redundant tool calls. We are optimizing the execution graph in real time.

Edward Hamilton

Exactly, Catherine. We also implement prompt caching and batch processing at Layer 6. If our policy interpretation agent is reading the same five-hundred-page compliance manual over and over again, prompt caching allows us to store that context on the provider's side, reducing our input token costs by up to fifty percent or more. And for non-real-time tasks—like batch processing fifty thousand insurance claims overnight—we route them through batch APIs which offer a fifty percent discount. The math is incredibly compelling when you engineer it properly.

Charles Skamser

And- and don't forget "Rules Before Reasoning." This is a core PX42 consulting principle. If a traditional, deterministic database query or a simple SQL join can answer the business question, you run that first! You do not ask a multi-billion-parameter LLM to calculate a sum or check a date. That is a massive waste of tokens and a source of unnecessary hallucination risk. You use rules and traditional code to do the heavy lifting, and you save the expensive reasoning models for the genuinely ambiguous, complex decisions that require human-like cognitive synthesis.

Catherine Spencer

It sounds so obvious when you lay it out, Charles, but so many companies are doing the exact opposite. They are building systems where the LLM is responsible for everything, including basic arithmetic, and then they wonder why their cloud bill is astronomical and their numbers are wrong.

Charles Skamser

Let's look at some real, concrete mathematical models. Let's make this real. Catherine, walk us through the Customer Service AgentOS model, because the economics of this are absolutely beautiful when you look at the token volumes and the actual capacity savings.

Catherine Spencer

Absolutely. Let's paint a picture of an enterprise handling three hundred thousand customer interactions per month. A traditional, naive GenAI setup might just throw a chatbot at it and hope for the best. In our AgentOS design, we coordinate a society of agents: a customer context agent, a policy agent, a billing agent, a sentiment agent, and a Verified Truth validation agent, with a human adjuster in the loop for exceptions. Now, let's assume each interaction uses an average of nine thousand input tokens and two thousand four hundred output tokens. Across three hundred thousand interactions, that is two point seven billion input tokens and seven hundred and twenty million output tokens per month.

Edward Hamilton

Which, if routed entirely through a premium model, would be a financial disaster. But if we use intelligent model routing and send that high volume to a lower-cost, highly capable model like GPT-5 mini, let's look at the math. At twenty-five cents per million input tokens, two point seven billion input tokens costs just six hundred and seventy-five dollars. And seven hundred and twenty million output tokens at two dollars per million costs one thousand four hundred and forty dollars. That's a total monthly model inference cost of just two thousand one hundred and fifteen dollars.

Charles Skamser

Two thousand dollars! Think- think about that. Even if you add twenty-five thousand to seventy-five thousand dollars per month for integration, hosting, security, the front-end, and support, your total operating cost is under eighty thousand dollars. Now, what is the return? If that system saves just two minutes per customer interaction, and the loaded human-service cost is thirty dollars per hour, that's six hundred thousand minutes saved per month. That is ten thousand hours! At thirty dollars an hour, that is three hundred thousand dollars per month in gross labor-capacity value! That is three point six million dollars a year, from a monthly spend of maybe fifty to seventy-five thousand dollars. It is an absolute no-brainer.

Catherine Spencer

And that's before you account for the reduction in customer churn, faster resolution times, and the elimination of expensive human errors. Now, let's look at an even larger scenario: Insurance Claims Leakage. Say an insurance carrier processes twenty-five thousand claims per month. A claims AgentOS coordinates intake agents, document review, fraud signaling, reserve analysis, and settlement recommendations. This is a much more complex workflow, so let's assume it uses two hundred thousand input tokens and twenty thousand output tokens per claim.

Edward Hamilton

Right, so at twenty-five thousand claims, that is five billion input tokens and five hundred million output tokens per month. Because this requires deep reasoning and contract interpretation, we route this to a premium reasoning model like GPT-5. At one dollar and twenty-five cents per million input, that’s six thousand two hundred and fifty dollars. At ten dollars per million output, that’s five thousand dollars. So our monthly inference cost is eleven thousand two hundred and fifty dollars.

Charles Skamser

Okay, eleven- eleven thousand dollars. Now compare that to the value. If this AI-assisted claims review saves just thirty minutes per claim, at a loaded adjuster cost of forty-five dollars an hour, that is twelve thousand five hundred hours saved per month. That's five hundred and sixty-two thousand five hundred dollars a month in productivity capacity. But here is the real CFO-visible kicker: claims leakage. If that society of agents, using Layer 9 Verified Truth, reduces claims leakage—meaning incorrect payments, missed subrogation opportunities, or unflagged fraud—by just zero point five percent on an annual claims book of one point two billion dollars... that is six million dollars a year in direct bottom-line savings! Combined with the productivity gains, you are looking at over twelve million dollars in annual value. An annual value of twelve million dollars against an inference cost of one hundred and thirty-five thousand dollars a year. That is the difference between tokenmaxxing and true enterprise ROI.

Catherine Spencer

It’s incredible, Charles. And we see the exact same pattern in our third model: Financial Services Fraud and Investigation. A regional bank has one million transactions a month and twenty million dollars in annual fraud exposure. If you routed all one million transactions through an LLM, it would cost a fortune. But our AgentOS uses traditional anomaly detection and scoring rules to triage ninety-five percent of the volume, routing only the complex five percent—the fifty thousand suspect cases—to the agentic investigation society. The inference cost is less than twenty-five hundred dollars a month, but if it reduces fraud losses by just five percent, that's a million dollars a year straight to the bottom line, plus six hundred thousand dollars in manual labor savings. That is how you design for real operating leverage.

Charles Skamser

And this- this is what we mean by "360-Degree Business Health." Most companies today are operating with completely fragmented visibility. The finance team is looking at lagging, historical KPIs. The sales team is looking at an optimistic pipeline. Operations is looking at throughput. It's all siloed. But when you have a multi-agent society monitoring the entire enterprise, they are constantly ingesting these cross-functional signals in real time. They aren't just diagnostic—telling you what happened. They are predictive and prescriptive. They tell you, "We have a supply chain anomaly in sector four that is going to impact our gross margin by two percent next quarter unless we substitute this supplier immediately. Here is the financial impact of that substitution, and here is the drafted action for your approval."

Edward Hamilton

Exactly. It shifts the entire management paradigm from reactive dashboards to proactive, agentic execution. We are moving from descriptive intelligence—"what happened?"—to predictive and prescriptive intelligence—"what is likely to happen next, and what should we do about it?" That is the ultimate promise of the PX42 and UBIX.ai partnership. We provide the architecture, the readiness assessment, and the rigorous governance framework, while UBIX provides the scalable, business-user-oriented decision intelligence platform that puts these capabilities directly into the hands of the executives who own the P&L.

Catherine Spencer

It truly is the end of the era of AI experimentation, and the beginning of the era of AI accountability. Boards and CFOs shouldn't be asking, "How do we build a custom LLM?" They should be asking: "What is our financial baseline? What is our model-routing strategy to control token costs? How are we measuring the ROI of our agent workflows? and how do we ensure our proprietary IP is protected through secure runtime isolation?" Those are the questions that move a company from AI theater to genuine operating leverage.

Charles Skamser

They are the only questions that matter. The skepticism from the financial markets is a healthy, necessary correction. It is weeding out the tourists who are just tokenmaxxing, and it is paving the way for the architects who know how to build durable, governed, and profitable enterprise operating systems. If you're ready to stop running science projects and start driving measurable business health, let's talk. Alright, that's it from me. Great chatting, team, talk soon.

Catherine Spencer

Goodbye, everyone.

Edward Hamilton

Cheers, until next time.