Amid market fears of an AI bubble, the ‘boring’ non-tech firms like manufacturers, healthcare, finance, and entertainment giants stand to capture trillions in value by evolving from AI adopters deploying basic tools to fundamentally transform the economics of their operations with agentic systems. New Carnegie Mellon and Stanford University research validates this potential by showing AI agents complete some realistic workflows 88.3% faster at 90.4-96.2% lower cost than humans. That creates the kind of discontinuity that could yield 6x returns, even if the company isn’t named Tesla. Of course, with caveats.
The AI valuation disconnect could not have been any starker last week.
On one side, Tesla shareholders ratified a compensation structure tied to sextupling the company’s valuation over the next decade through aggressive bets on AI, robotics, and automation. More than just a vote of confidence in Musk, it was an endorsement of the exponential economic potential of these technologies to transform a singular tech leader.
On the other side, the Nasdaq remains volatile as it hemorrhages and then adds hundreds of billions in market cap from one day to the next amid AI bubble fears. Analysts questioned whether AI hype had finally peaked. Michael Burry of “The Big Short” fame placed massive bets against technology stocks Nvidia and Palantir. That triggered a war of words with Palantir CEO Alex Karp, but it was also followed by SoftBank’s surprising move to sell its $5.8B stake in Nvidia. Observers pointed to these moves as a verdict that AI’s potential has been overhyped.
What if both narratives miss the real story?
What if the magnitude of value creation possible through AI is just fundamentally misunderstood? What if the companies positioned to capture trillion-dollar returns aren’t the ones Wall Street is watching?
While markets obsess over the tech industry’s frenzied pursuit of AI superiority, they’re overlooking where much of the massive value creation will occur in the transformation of the 423 “non-tech” companies in the S&P 500 that are just starting to tap agentic AI’s potential. These are sectors like health care, energy, logistics, consumer goods, and industry. It’s companies like Caterpillar, Walmart, Pepsico, JPMorganChase, and Disney.
They don’t dominate the headlines and hype when it comes to AI. But a study released last week allows us to begin to grasp the massive scope of the looming value of migration in these sleeping giants.
Research from Carnegie Mellon and Stanford University compares how humans and AI agents complete identical work across data analysis, engineering, computation, writing, and design. These are the actual workflows that constitute modern operations, and the results reveal something material: agents finish tasks 88.3% faster at costs that range from 90.4% to 96.2% lower than those of human professionals, highlighting the potential to enable efficient collaboration by delegating easily programmable tasks to agents.
The question every executive leader and investor should now be asking: What is the estimated economic impact of delegating easily programmable tasks to agents in enterprise workflows?
The Carnegie-Stanford research allows us to start grasping the potential for delegating these “readily programmable” tasks, per the paper’s terminology, that the authors estimate to be 82.5% of the occupation of knowledge work in tech and non-tech companies.
This isn’t a one-time boost. It’s a permanent step-change in earnings power. Multiply this across those 423 non-tech companies on the S&P 500 over a decade, and you are describing the first steps towards one of the largest value creation events in economic history.
As I’ve written extensively, we’re entering the Agentic Era where AI systems don’t merely assist humans but orchestrate entire workflows autonomously. The inference economy emerging from this transformation will allow synthetic colleagues to multiply organizational output without proportional cost increases. I had defined a synthetic colleague to be “distributed networks of specialized agents that share persistent memory, coordinate through structured communication protocols, and pursue decomposed goals under orchestrated supervision”.
Of course, not all companies have the potential to 6X their valuations (perhaps not even Tesla!). There are many layers within this big slice of the S&P 500 cake. A company that makes heavy machinery will still be bound by different physical constraints than a bank, but can benefit from the value creation in the same way, turning those same physical assets into foundations for their moat in an agentic world.
So, let’s refine this question: How do other companies think about multiplying their value by 2X or 3X? I’ll break that answer down into three parts:
The Margin Expansion Mathematics
The Revenue Multiplier
Timeline (aka, patience!)
Given the swirling debates over AI hype and bubbles, the temptation is to sit on the sidelines and see how it all plays out. Whether you are an investor or executive, that is a recipe for missing what may prove to be one of the century’s most asymmetric investment opportunities.
Below, I try to provide guidelines on where I think “pockets” of value exist, and what they entail.
Margin Expansion Mathematics
The cost structure implications require careful analysis. As the research indicates, not all work is equally amenable to agentic replacement.
If we analyze the 11 official categories of the S&P 500, the 423 “non-tech” companies are found in Materials, Real Estate, Utilities, Energy, Consumer Staples, Health Care, Industrials, Communications Services, Consumer, and Financials (with some exceptions in the last four). Whatever the core business, to some degree, each of these businesses relies heavily on knowledge work to operate.
The Carnegie-Stanford research identifies three categories of programmability that determine cost compression potential:
Readily programmable work includes deterministic tasks solvable through code execution: data analysis, system configuration, computational tasks, and structured writing. These tasks are estimated to represent between 20% and up to 50% of total tasks. For this category, agents achieve the cost reductions stated supra. The workflows are redesigned entirely around programmatic approaches, with humans in the loop and agents completing, under human supervision, these programmatic tasks. This applies, as suggested in the paper, to both tech and non-tech environments.
Partially programmable work includes tasks theoretically amenable to code but requiring workflow redesign or human judgment at key decision points: complex engineering design, strategic planning, content creation, and operational optimization.
Minimally programmable work includes tasks requiring visual perception, aesthetic judgment, or contextual understanding where current agents struggle, such as relationship management, negotiation, strategic oversight, and quality assessment requiring domain expertise. For this category and the latter, we take the hypothesis that, given the state of agentic AI, agentic replacement of human workflows is not yet possible.
Now let’s apply the Carnegie-Stanford programmability categories to JPMorganChase.
The finance giant has about 317,000 employees and about $96B in annual non-interest expenses, based on extrapolation of Q3 25 figures. We can estimate this to include $48B in labor costs, i.e., 50%, using average banking & finance labor ratios.
According to the Carnegie-Stanford research, programming is required in 82.5% of relevant occupations. So, we would want to analyze the tasks that fall within about $39.6B of that labor.
(Note: While the research estimates that programming skills are required in 82.5% of occupations (based on U.S. Department of Labor ONET data), this figure reflects broad occupational requirements rather than the precise proportion of daily tasks that are readily programmable or automatable by current AI agents)
Assuming that 30% of tasks are readily programmable (e.g., data analysis, risk computations), this indicates about $11.9B could be redirected to agents. In an optimistic scenario, the agents then achieve cost savings close to c.93% or about $11.1B.
Taking $11B in gains as a base case (per the above), JPMorgan’s net profit margins could expand from 32% to 38% based on $ $180.6 billion revenue in 2024. (Note: for calculation purposes, JPM reported 2024 net income was $58.5 billion on $180.6 billion revenue, which implies net margin ~32.4%).
Of course, that implementation is not just about labor reductions. It allows for the reallocation of human time. In a recent interview with CNN, JPMorgan CEO Jamie Dimon played down the impact of AI, saying: “We always redeploy” employees and that “there will be jobs eliminated by AI... [but] it will also create jobs.”
But I think he is underestimating the true financial impact of agent-related savings. I believe the 6% margin gain calculated supra., well ~50% of that would ultimately be a net gain in JPMorgan’s P&L, leaving us with 3% margin improvement in our example.