Ask George Murnane how to separate real artificial intelligence from a marketing slogan, and he gives you one question: what does the model predict, and what is its loss function?
George Peter Murnane has spent more than three decades running asset-intensive aviation businesses, 14 of those years as a chief operating officer, a chief financial officer, or both at once. He is now chief executive of Jet.AI Inc. (NASDAQ: JTAI) and a director and CFO of AI Infrastructure Acquisition Corp., the blank-check company that closed an upsized $138 million IPO in October 2025. That résumé sits at the exact junction where capital, operations, and AI claims collide. It also makes him unusually hard to sell to.
The George Murnane filter: name the prediction, name the loss function
The test is deliberately unglamorous. If a company can tell you precisely what its model forecasts and what error it is trained to minimize, the AI is probably real. If the best it can offer is that the technology “makes the experience smarter,” the label is doing work the software cannot.
That distinction matters more in aviation than in almost any other industry, because the cost base is high and the margins are thin enough that a small efficiency gain compounds into real money. Murnane’s filter is a way of routing scarce capital toward the few applications that move those numbers, and away from the many that only move a pitch deck.
Where AI is real in aviation: the AOG math
Start with predictive maintenance, the application Murnane considers genuinely valuable. The economics are not subtle. A single aircraft-on-ground event can cost an operator between $10,000 and $150,000 per hour of downtime, once you add lost revenue, crew rest and overtime, passenger re-accommodation, and the scramble to source a replacement part.
Predictive maintenance attacks that cost directly. By reading sensor data, flight history, and maintenance records, the models flag a deteriorating component before it fails, which lets an operator move an unplanned repair into a scheduled window. A 2022 Deloitte analysis cited by Radome Technologies estimated that predictive maintenance, properly implemented, can cut maintenance costs by up to 30% and reduce AOG events by more than half. Delta cut unscheduled maintenance by more than 30% using predictive engine monitoring.
The use case is specific enough to survive Murnane’s question. The model predicts a component failure. Its loss function penalizes false negatives, the missed failures that ground an aircraft, and false positives, the needless part swaps that waste a maintenance slot. There is a number on both sides of the ledger. Investors have noticed the same thing: the predictive airplane maintenance market is projected to grow from $5.35 billion in 2026 to $18.87 billion by 2034, a compound annual rate above 17%.
Predictive maintenance is not the only application that clears the bar. Crew scheduling optimized against duty-time limits has a defined objective and a hard constraint set written into federal regulation. Dynamic pricing run against forward booking curves predicts demand and optimizes yield. Document automation in SEC filings and merger diligence has a measurable output and a measurable error rate. Each of these can be described in a sentence that names what is being predicted. That is the tell.
The failure modes George Murnane watches for
The opposite of a loss function is an adjective. Murnane’s interviews return repeatedly to two ways companies dress up old or absent technology as AI.
The first is rebranding. A regression model that has been forecasting demand or pricing risk for 30 years gets relabeled “AI” because the term raises a valuation. The math is the same forecast it always was, repackaged under a more valuable label.
The second failure mode is more current and more expensive. A company bolts a large language model onto a workflow without redesigning the workflow underneath it. The result is a chatbot marginally more eloquent than the FAQ page it replaced, sold as a transformation. The model is real, but the value is not, because no one re-engineered the process the model was supposed to improve.
Murnane’s caution here is partly reputational arithmetic. As he has put it, the credibility cost of overclaiming compounds faster than the marketing benefit. For a public company whose name carries the letters “AI,” that is not an abstract risk. Overstate what the software does, and the first product failure under pressure becomes the story.
How Jet.AI uses AI where the value is measurable
Jet.AI gives Murnane a place to apply his own test in public. The company’s software, built when it operated as a private-aviation platform, concentrated AI on functions with a number attached: booking optimization, matching customers to the right operator, and customer communication. Its CharterGPT app and the Ava agentic booking model used natural-language processing to compress a booking process that once ran on phone calls and faxes.
Those tools handle a deliberately narrow set of jobs. Flying the airplane, vetting an operator’s safety record, and resolving a mechanical failure at midnight stay with humans and regulators. The AI sits where its prediction is cheap to measure and its errors are cheap to correct, which is exactly where Murnane argues it belongs.
From booking software to AI data center infrastructure
The most telling application of the loss-function test is the one Jet.AI is now living through. The company has moved away from running aircraft and toward building AI data center infrastructure, describing itself as a technology company focused on data center development across North America, with projects spanning more than a gigawatt of planned capacity.
The reason behind the pivot reads like a case study in Murnane’s own discipline. Jet.AI built genuine AI products, including a large language model agent for private aviation, then ran into a constraint that no amount of marketing could fix: unreliable uptime for the computational resources those products depended on, which occasionally slowed the company’s ability to serve customers. The bottleneck sat below the algorithm, in the power, land, and compute the products ran on.
So the company went after the bottleneck. Based in Las Vegas, with access to land, solar power, and natural gas, Jet.AI signed a letter of intent for a 50-megawatt project on a 120-acre campus with room to scale toward a full gigawatt. The framing Executive Chairman Mike Winston used was almost a rebuke of the category’s usual rhetoric: the move “isn’t a flashy move, but it’s a smart one,” because data centers are “the bedrock of the AI economy” and create value that is “tangible, stable, and meaningful.”
That is the loss-function test pointed at infrastructure rather than software. The prediction is straightforward: compute demand keeps climbing, and the assets that supply it earn against it. The error is measurable in megawatts delivered and uptime maintained. There is a number on both sides.
Why the loss-function test travels
Murnane’s filter works because it is industry-agnostic. It ignores whether a technology looks impressive and asks instead whether anyone can state what the system is optimizing and check the result against reality.
That discipline is the through-line of his career, from pricing aircraft assets at global carriers to evaluating a data center SPAC. The same question that exposes a relabeled regression model also exposes an overhyped acquisition target. In both cases, the executive who can describe the objective function is operating from evidence. The one reaching for “smarter” and “seamless” is operating from hope.
For a sector where roughly every company now claims an AI strategy, the value of a one-question screen is that it is fast and hard to fake. Name the prediction. Name the loss function. If those two answers are specific, the technology is likely doing real work. If they dissolve into adjectives, the only thing being optimized is the marketing.
