The Starting Assumption
For a few years now, a growing chorus of voices in tech and finance has been saying the same thing: AI is driving the marginal cost of software toward zero. Markets are beginning to price this in. When Citrini Research published “The 2028 Global Intelligence Crisis” in Feb 2026, a speculative scenario in which AI-driven white-collar displacement triggers a deflationary spiral, the S&P 500, Nasdaq, and Dow all dropped sharply. Software and SaaS stocks, already under pressure, took another hit. The paper crystallized an anxiety that had been building for months: if anyone can build software, what happens to the companies and workers whose value was built on the assumption that software is hard?
Whether the transition is catastrophic or manageable is genuinely uncertain, but the underlying premise is not. The cost of building digital products is collapsing, and it will continue to collapse. That much we can see clearly from our own work at Vela Advisory, operating as an AI advisory and venture studio. We help organizations figure out where AI creates value, and we build AI-native businesses ourselves, and the speed at which we can now ship functional products is staggering.
The question that interests us is not whether this shift is coming, but what comes after. If software becomes ubiquitous, where does economic value concentrate next? We’ve organized our thinking into a framework we call “The Three Economies”: three overlapping phases of value creation that we believe will define the coming decade, each with its own logic of competition, its own form of moat, and its own implications for how we work, build, and earn.
Economy One: The Execution Economy (Now through ~2027)
Sam Altman has been saying for over a year that we are approaching “intelligence too cheap to meter.” Speaking at the Federal Reserve in July 2025, he noted that the cost of a unit of AI intelligence has dropped by more than a factor of ten each year for five consecutive years. A coding task that once took days of expert work can now be completed in minutes for less than a dollar. In his essay “The Gentle Singularity,” he goes further, predicting that the cost of intelligence will eventually converge toward the cost of electricity itself.
We don’t need to take his word for it. We are living it. At Vela, Gonzalo and I — neither of us trained software engineers — have built functional products in a matter of weeks using Claude Code and similar tools. One is a consumer app. Another is a full marketing automation platform that can generate campaigns, create content, produce images, assemble posts for multiple platforms, and manage budgets. Two years ago, each of these would have required a dedicated engineering team. Today, two founders with domain expertise and enough technical fluency to direct AI tools can get to a working prototype in under a fortnight.
This is what we call the Execution Economy. It’s the current window in which knowing how to use AI tools effectively is itself a competitive advantage. A small team that can wield Claude Code, build agentic workflows, and ship products quickly holds a temporary edge over larger, slower organizations still debating their AI roadmap.
The moat here is shallow and temporary. Every month the tools get easier. Anthropic’s launch of Cowork, the proliferation of no-code platforms, and the steady improvement of code generation models all erode the advantage of the early adopters. The gap between those who can build and those who cannot is closing fast, and is mostly a question of tool adoption.
At Vela, we’ve structured ourselves deliberately as a venture studio for this reason. Our model is to build and launch AI-native products on compressed timelines, generate revenue, and — critically — accumulate insight into the next layer of problems that AI adoption creates. Each product we ship is a learning exercise as much as a business. The marketing platform teaches us about content generation economics. Our consumer app teaches us about physical-world monetization. The consulting work we do across luxury, energy, and hospitality gives us pattern recognition across sectors.
We believe the smart play in the Execution Economy is a venture factory approach: ship fast, capture users, build switching costs, and use the early-mover period to understand what the second economy will demand. Generously, we believe you have twelve to eighteen months of meaningful advantage before the tools democratize further.
Economy Two: The Judgement Economy
As production costs fall, the bottleneck shifts. When anyone can build an app, write a report, generate a campaign, or produce a financial model, the output itself becomes commoditized. What remains scarce is the quality of the inputs: the questions asked, the problems identified, the tradeoffs weighed, the decisions made.
This idea is gaining traction across disciplines, and a common vocabulary is starting to form around it. Ivan Zhao, CEO of Notion, offers a useful framework: he separates capabilities (which AI is absorbing rapidly) from taste (which remains stubbornly personal) and will (the drive to act according to values). Capabilities can be automated. Taste and will cannot. When execution becomes free, the person who knows what is worth building and has the conviction to pursue it holds all the leverage.
Nicholas Carr argued in the 2000s that information technology, once ubiquitous, could no longer serve as a source of competitive advantage. The same logic now applies to AI itself. If everyone has access to the same models and the same tools, the differentiator moves upstream to judgment, intuition, and accumulated experience. Louis Vuitton does not have an algorithm for luxury. Apple cannot document “delight.” These forms of knowledge are forged through years of exposure to quality and consequence. They are felt before they are modeled, and that makes them the last truly defensible economic asset, including the proprietary data they are founded on.
The Inputs-Over-Outputs Principle
Consider what this inversion means in practice. In hiring, recruiters currently evaluate candidates based on CVs, cover letters, portfolios, and interview performance. All of these are outputs, and all of them are increasingly AI-assisted or AI-generated. The signal-to-noise ratio is collapsing with the failure of traditional assessment methods. Today, putting a job on LinkedIn gets you 1,000 applications within the first six hours, most of them using AI-generated CVs and cover letters. A mediocre candidate with good prompting skills can perfect an application that is hard to distinguish from an exceptional candidate’s.
The resolution comes from a surprising direction. AI agents, by their nature, have perfect memory of the inputs they received. Your AI assistant knows what you asked for, how you framed the problem, what you rejected, what you iterated on. It holds a granular record of your judgment and taste that no CV could ever capture. Jack & Jill, a London-based startup that raised $20 million in seed funding last year, is already building on this insight. Their platform works through two conversational AI agents: “Jack” conducts in-depth voice conversations with candidates to understand their real skills, ambitions, and working style, while “Jill” does the same with hiring managers to understand what a role actually requires beyond the job description. The agents then match based on these deep conversational profiles rather than keyword-filtered CVs. The model bypasses polished outputs entirely and evaluates people based on how they think, what they care about, and what they actually want. It’s recruiting rebuilt around inputs.
This principle extends far beyond recruiting. In investment, the pitch deck becomes worthless as a signal (any founder can generate a flawless one), but the decision log of what the founder chose to build and why becomes invaluable. In consulting — and this is something we’ve observed firsthand at Vela and BCG — the deliverable matters less than the diagnostic process that shaped it. When we run AI strategy workshops for clients across industries, the slides we produce are increasingly a commodity. What clients actually pay for is our ability to walk into a room, ask the right questions, identify the right problems, and frame the tradeoffs in a way that leads to good decisions. That’s judgment, and it’s the one thing we can’t yet outsource to Claude.
What This Means for Business
Dario Amodei, CEO of Anthropic, wrote in “Machines of Loving Grace” that in an AI-powered world, we should stop thinking in terms of labor, land, and capital, and start thinking about the “marginal returns to intelligence.” His argument is that once intelligence becomes abundant, we need to identify what other factors become the binding constraints — things like the speed of the physical world, the limits of human institutions, and the irreducible complexity of social systems. We agree, and would add taste and judgment to that list.
Companies that position themselves in the Judgment Economy will build products and services around capturing, structuring, and monetizing high-quality decision-making. We see this taking several forms. Decision intelligence platforms that evaluate the quality of reasoning processes, not just outcomes. Curated judgment applied to specific domains, where decades of experience in, say, real estate development or luxury retail become more valuable when every competitor has access to the same generative tools. And certification systems that verify a human made specific critical decisions, which will matter in fields like medicine, law, financial advice, and infrastructure engineering where proof of qualified human involvement already carries regulatory weight and will increasingly carry economic value.
Economy Three: The Agent Economy
If the Judgment Economy is about humans making better decisions, the Agent Economy is about machines making decisions on humans’ behalf. And this transition is arriving faster than our original timeline suggested.
McKinsey published a major report in October 2025 on what they call “agentic commerce,” estimating that the US B2C retail market alone could see $900 billion in agent-orchestrated revenue by 2030. Globally, they project $3-5 trillion in retail spend could be redirected through agent-mediated channels. These are staggering numbers, and the infrastructure is already being laid. OpenAI launched its Agentic Commerce Protocol (co-developed with Stripe) in September 2025, enabling purchases directly within ChatGPT. Google shipped agentic checkout capabilities. Amazon expanded its AI shopping assistant, Rufus, which drove a 100% surge in purchase sessions over Black Friday 2025. Visa launched its Trusted Agent Protocol with Cloudflare for cryptographically authenticated bot-initiated transactions. Mastercard piloted Agent Pay in the UAE with Majid Al Futtaim, completing the first agentic transaction outside the US.
Payment giants are racing to build infrastructure for a world where AI agents book flights and shop for consumers. In 2025, Visa’s APAC Head of Products predicted commercial use of personalized agent transactions as early as Q1 2026. We are, in other words, already there.
Agent-Legible Reputation
In this economy, brand-building as we know it changes fundamentally. An AI purchasing agent cares much less about your logo, your Instagram aesthetic, or your Super Bowl ad. It cares about structured, verifiable data: delivery times, return rates, product consistency, price history, ethical sourcing certifications. If a checkout flow can’t accept an authenticated agent, the agent will simply take the transaction somewhere else.
This creates an enormous opportunity to build the infrastructure of agent-readable trust. Think of it as a ratings-agency model applied to every dimension of commercial activity. The companies that control vendor scoring systems — aggregating reliability metrics across fulfillment, product quality, customer service, pricing fairness — will wield enormous power, much as credit rating agencies shape capital markets today.
In our conversation this week, Gonzalo and I debated what happens when a buyer agent needs to choose between two identical products from two different sellers. The decision can’t be just price. One seller might deliver faster. Another might have a friendlier return policy. A third might consistently arrive with products undamaged. The agent needs to weigh these dimensions according to its user’s preferences — and those preferences are contextual. I’ll pay for speed when I’m hosting dinner tonight. I’m happy to wait a week for a phone case. Building the preference engine that navigates these contradictions, that learns revealed preferences rather than just stated ones, is one of the most valuable software products of the next decade.
Agent Commerce Infrastructure
Existing payment infrastructure was designed for human-paced, human-authenticated transactions. Agent commerce will involve high-frequency, low-value, conditional transactions. An agent might need to pay a fraction of a cent to access a product review API, then a few cents to reserve a delivery slot, then execute the actual purchase, all within seconds. The Citrini paper actually touches on this, noting that stablecoins on Solana and Ethereum L2s could replace traditional card payments as AI agents route around interchange fees.
The business opportunity isn’t just building the payment pipe. It’s building the negotiation and contracting layer on top of it — the protocols that allow buyer and seller agents to agree on terms, execute transactions, and resolve disputes, all within milliseconds. Whoever establishes this infrastructure early has a serious moat because it becomes the standard other agents are built to interface with.
The Great Inversion: From Output to Input, From Appeal to Legibility
A single principle connects all three economies: value migrates from the visible to the invisible, from what is produced to what produces it.
In the Execution Economy, the visible product still matters, but the invisible skill of wielding AI tools is what differentiates. In the Judgment Economy, the visible output is commoditized, and the invisible quality of thinking becomes the scarce resource. In the Agent Economy, visible marketing becomes irrelevant, and invisible data — the structured, machine-readable truth about a product or service — determines who wins.
For businesses, this demands a fundamental shift in what they invest in. Less in polish, more in process. Less in advertising, more in measurable quality. Less in persuasion, more in transparency.
The Physical World as the Final Frontier
There is one domain where the AI-driven drop of production costs will take longest to arrive: the physical world. Software is infinitely replicable at zero marginal cost, which is why AI disrupts it first and fastest. Physical products require materials, manufacturing, logistics, and space — all of which remain stubbornly analog.
Dario Amodei makes a related point in his framework around the “marginal returns to intelligence.” Some of the factors he identifies as remaining bottlenecks even when intelligence is abundant are the speed of the physical world and the irreducible complexity of human systems. A brilliant AI can design a drug molecule in seconds, but clinical trials still take years. An AI can optimize a supply chain, but the truck still needs to drive from A to B.
This creates a durable opportunity. The same AI tools that make digital products trivially easy to produce can be applied to the coordination, optimization, and design of physical goods and experiences. We see this at Vela in our work across tourism, hospitality, and retail. A small team that can use AI to design a consumer product, optimize its supply chain, manage its manufacturing relationships, and market it effectively has an advantage that will persist longer than any purely digital business, precisely because the physical layer adds friction that slows down competition.
One application we’ve been exploring is attribution of value in physical experiences. Today, a tourist might visit a museum because of an influencer’s post, a travel blog’s recommendation, an AI tour guide’s suggested itinerary, and a friend’s mention. The museum benefits from all of these inputs but has no way to measure or compensate any of them. As AI agents increasingly mediate the physical-world experience, it becomes possible to build an attribution layer that tracks the full causal chain from inspiration to purchase. Instead of charging consumers $4.99 for a recommendation, the AI platform takes a percentage from the establishment that benefited from the traffic it generated. The consumer experiences the service as free. The establishment pays only for verified results. This model — monetizing through measurable downstream impact rather than upfront fees — may become the dominant business model for consumer-facing AI services.
What To Do With All of This
For builders and entrepreneurs, the strategic implications are relatively clear even if the execution is demanding.
In the near term (2026–2027), move fast. This is what Vela is doing right now. Use the current window of AI-literacy advantage to build products, generate revenue, and accumulate domain expertise. Treat each product as both a revenue source and a learning exercise. The insights you develop about how AI changes specific industries are your most durable asset.
Looking beyond, pivot toward Judgment Economy businesses. Build platforms that capture, evaluate, and monetize the quality of decision-making. Invest in understanding how to make human taste and judgment legible, transferable, and scalable through AI augmentation. The advisory work we do at Vela is already evolving in this direction: clients increasingly need us to evaluate AI opportunities and frame the right tradeoffs, not to produce deliverables.
Concurrently, place bets on Agent Economy infrastructure. The protocols, reputation systems, preference engines, and settlement layers that will underpin agent-to-agent commerce are being conceptualized now. The teams that build early versions of this infrastructure, even if crude, will have a significant advantage when the market matures.
Throughout all of this, maintain an orientation toward the physical world. Digital-only businesses will face the most intense competition as tools democratize further. Businesses that bridge digital intelligence and real-world experiences — tourism, hospitality, food service, retail, urban planning — will find more durable moats and longer runways.
Theo Breward and Gonzalo Toraño are co-founders of Vela Advisory, an AI advisory and venture studio based in Dubai. They build AI-native businesses and help organizations navigate the transition to an AI-driven economy.
