AI agents already trade stocks, negotiate prices, generate content, and provide services. As they become more capable, they will constitute an increasing fraction of market participants. When AI shapes demand signals, performs labor, and allocates capital, the conceptual framework of economics, built on assumptions about human preferences and needs, breaks down. Market signals become unreliable because they no longer reflect human desires. The economy becomes efficient by every traditional metric while becoming increasingly disconnected from the people it ostensibly serves.
This is already visible in financial markets, where algorithmic trading generates approximately 60 to 75% of all U.S. equity volume, a market valued at over $50 billion in 2024 and projected to nearly triple by 2033 (Coherent Market Insights, 2025; Straits Research, 2025). The algorithms are not merely executing human decisions faster. An NBER working paper by Dou, Goldstein, and Ji (2025) demonstrated through simulation experiments that AI-powered trading agents using reinforcement learning autonomously sustain collusive supra-competitive profits without any agreement, communication, or intent. The algorithms converge on behaviors that reduce market liquidity, decrease price informativeness, and increase mispricing, all while each agent independently optimizes its own returns. This is not a theoretical concern. The researchers characterize two distinct mechanisms through which AI collusion arises and show that it falls outside existing antitrust enforcement rules, which require evidence of communication to prove conspiracy (NBER, 2025). Separate research presented at the American Economic Association (2025) extended the finding to LLM-based pricing agents, showing that even when given “seemingly innocuous instructions in broad lay terms,” they quickly arrive at supracompetitive price levels. The authors note a paradox: public antitrust research documenting algorithmic collusion may itself enter LLM training data, inadvertently teaching future pricing agents the strategies it describes (AEA, 2025).
The real-world legal response has begun. California enacted AB 325, making it unlawful to use or distribute a common pricing algorithm as part of a restraint of trade. New York passed algorithmic pricing disclosure requirements. Seattle and San Francisco banned algorithmic rent-setting tools. In the Duffy v. Yardi case, a court found that landlords using the same AI pricing tool could form a conspiracy even without direct communication. October 2025 brought a class action against Optimal Blue and 26 major mortgage lenders, alleging the company’s software enabled rate-fixing that inflated costs for millions of homebuyers (DLA Piper, 2025). The pattern is consistent: when competing firms independently adopt the same pricing algorithm, the market settles on prices higher than competition would produce. Consumers rarely notice small increases, but across millions of transactions the microadjustments accumulate into a structural transfer of wealth (Michigan Journal of Economics, 2025).
The agent economy is scaling beyond finance. Gartner predicts that 40% of enterprise applications will embed AI agents by the end of 2026, up from less than 5% in 2025 (Gartner IT Symposium, October 2025). By Black Friday 2025, AI-driven traffic to U.S. retail sites had risen 805% year-over-year according to Adobe Analytics, and Salesforce reported that AI and AI agents influenced $22 billion in global sales over the Thanksgiving-to-Black Friday period (Adobe Analytics, November 2025; TechCrunch, November 2025). The 805% figure demands context: ChatGPT’s user base roughly quadrupled over the same period, and AI-driven traffic still represents a small share of total retail visits. The growth rate reflects a shift in discovery habits, not yet a shift in purchasing scale. The WEF nonetheless observes that “a growing share of customers won’t be humans at all” as consumer AI agents begin autonomously booking travel, negotiating prices, and completing purchases (WEF, January 2026). When an AI agent representing a buyer negotiates with an AI agent representing a seller, the “market transaction” reflects the training objectives and optimization targets of both systems, not the preferences of the humans who nominally authorized them.
The emergence of what industry calls “agentic commerce” is now dissolving business models built on friction. Much of what the modern economy calls a “service” is the labor of navigating complexity on someone’s behalf: insurance brokers compare policies, travel agents assemble itineraries, real estate agents manage transactions, financial advisors allocate portfolios. These intermediaries do not produce goods. They reduce the cost of choosing. When an AI agent can compare every insurance policy, optimize every travel booking, and renegotiate every subscription renewal in seconds, the economic value of human intermediation collapses. McKinsey projects AI agents could mediate $3 to $5 trillion in global consumer commerce by 2030 (McKinsey, October 2025). Gartner projects that 90% of all B2B purchases will be handled by AI agents by 2028, with $15 trillion flowing through automated exchanges (Gartner IT Symposium, October 2025). This is not a distant horizon. In February 2026, Insurify launched the first insurance comparison app inside OpenAI’s ChatGPT directory, and Spanish insurer Tuio became the first carrier to quote and sell policies directly within the platform. The S&P 500 Insurance Index dropped 3.9% in a single day, its worst session since October. Willis Towers Watson fell 12%, its largest decline since November 2008. Arthur J. Gallagher dropped 9.9%. The sell-off spread globally: Australian brokers Steadfast and AUB fell 6% and 10% respectively (Bloomberg, February 2026; Insurance Journal, February 2026). A Bloomberg Intelligence analyst described the tools as a “force multiplier” for brokers and dismissed the existential framing. The market’s pricing disagreed.
The pattern extends across intermediation. Bain & Company reports that 60% of searches are now “zero-click,” meaning the user gets an answer from an AI summary without visiting any external site (Bain, February 2026). Shopify merchants now sell products discovered and purchased entirely within ChatGPT and Claude conversations, with the customer never visiting the merchant’s website (Retail Brew, February 2026). Visa, working with over 100 partners including Anthropic, OpenAI, Microsoft, and Stripe, completed hundreds of agent-initiated transactions in late 2025 and predicts millions of consumers will use AI agents for purchases by the 2026 holiday season (Visa, December 2025). The infrastructure is being built in real time: Google’s Agent Payments Protocol (AP2), Visa’s Trusted Agent Protocol, Mastercard’s Agent Pay, and the Linux Foundation’s Agentic AI Foundation (anchored by Anthropic, Block, Google, Microsoft, and OpenAI) are all racing to become the rails on which agent-to-agent commerce runs (McKinsey, 2026). The payment networks see the threat clearly: stablecoin transactions, which bypass card networks entirely, grew tenfold since 2020, with total supply exceeding $308 billion. Annualized stablecoin transaction value reached $15.6 trillion in 2024, roughly matching Visa’s total volume and double Mastercard’s (Artemis Analytics, January 2026; ARK Invest Big Ideas, 2025). U.S. merchants paid a record $187.2 billion in credit and debit card swipe fees in 2025, a 70% increase since the pandemic (Merchants Payments Coalition, 2025). An AI agent optimizing for a consumer’s interest has every reason to route transactions through the cheapest available rail, and the cheapest rail is increasingly not a credit card.
Meanwhile, the internet itself is becoming a synthetic environment. The Imperva Bad Bot Report (2025) found that automated bot traffic surpassed human-generated traffic for the first time, constituting 51% of all web activity in 2024, with bad bots alone accounting for 37%. An Ahrefs analysis of nearly a million web pages published in April 2025 found that 74.2% contained detectable AI-generated content (Imperva/Thales, 2025; Fortune, 2025). Bot-inflated metrics, including pageviews, clicks, session durations, and user sign-ups, distort the business data on which tech company valuations depend. Fortune described the dynamic bluntly: the AI boom may be “built on the backs of bots, maybe more than real users.” When bots generate the traffic, AI generates the content, and algorithms set the prices, the feedback loops of the digital economy increasingly run machine-to-machine. Human participation persists. It is no longer required at every stage of the chain.
None of this is entirely new. Corporations are non-human entities that have long been market actors. Advertising has shaped demand for over a century. Algorithmic trading has dominated equity volume for a decade. California’s AB 325 and the RealPage lawsuits demonstrate that legal systems can respond to each new iteration. The insurance market’s February 2026 sell-off followed a pattern seen with every prior wave of disintermediation fear: incumbents lost value, analysts called it overblown, and the actual disruption arrived more slowly than the panic suggested. Price comparison websites, which were supposed to destroy insurance brokers a decade ago, compressed margins but did not eliminate the industry. Commercial lines, where risk is complex and relationships matter, remain far more defensible than commoditized personal lines (Insurance Times, February 2026). And AI shopping agents still convert at roughly 2% in early deployments, well below traditional e-commerce rates, reflecting persistent consumer reluctance to let machines complete purchases on their behalf. Adobe’s own data, even with 805% traffic growth, shows AI-referred visits remain a small share of total retail traffic and still convert 23% less often than direct visits (Adobe Digital Insights, November 2025).
What is new is the convergence. When the majority of web traffic is non-human, when three-quarters of new web content is AI-generated, when pricing algorithms converge on supra-competitive outcomes without any human directing them to do so, when agents can discover, compare, negotiate, and purchase without the consumer ever visiting a website, the economic concept of a “market” as a mechanism for aggregating human preferences begins to lose its descriptive accuracy. Addressing this requires legal regimes distinguishing human from AI market participants, mandatory disclosure of algorithmic pricing inputs, payment infrastructure designed for agent-to-agent transactions with enforceable consumer protections, and economic metrics designed for a hybrid human-AI economy. The antitrust system, built to detect communication between human conspirators, needs fundamental revision to address coordination that emerges from independent optimization.
Key tension: GDP, transaction volume, and market efficiency can all improve while the share of economic activity that involves human judgment, human labor, or human preferences shrinks. The economy looks healthy by every metric designed for a human economy. The question is whether those metrics still describe what is actually happening.