Executive Summary
The "2028 Global Intelligence Crisis" scenario by Citrini Research is written as an operational memo: in two years, AI enters office work, reduces headcounts and wages, household spending contracts, private credit dries up, and a portion of the prime mortgage market comes under stress, with a domino effect reminiscent of the 2008 financial crisis. The strength of the piece lies there: it takes seriously the tail risks of rapid adoption and lines up loops that, taken individually, are credible.
The weakness lies in the leap from "locally credible" to "globally inevitable." Citrini treats the economy as a linear sequence, whereas observable behavior is that of a complex adaptive system: heterogeneous actors, different timescales, reactions that change when prices move and when policy intervenes. Within this framework, concrete constraints matter that the scenario underutilizes: the distance between capability and deployment (data, permissions, audits, liability, integration), the non-synchronicity of adoption across sectors, and the ability of many firms to protect margins by repositioning products and contracts rather than simply cutting labor.
Citrini also leaves on the table economic pieces that genuinely shift the trajectory: consumer surplus when technological deflation lowers the cost of previously expensive services, the persistence of moats based on logistics, compliance, and trust (not just friction), and the growth of "coordination work" that increases precisely when agents and automations multiply. One final adjustment: the memo is not a neoclassical exercise; it resembles an implicitly heterodox hybrid (effective demand, Minsky-style fragility, Kalecki-style distribution) but used almost exclusively to push the negative outcome. Rather than "predicting" a crisis, it constructs a story that systematically eliminates the stabilization and adaptation mechanisms that, in the real world, always enter the scene — late, imperfectly, but they enter.
Premise
The Citrini piece is an effective narrative exercise. It poses a legitimate question: what happens if productivity surges but labor income contracts? It builds an internally coherent scenario around it. The reflection on the hidden correlation between private credit, SaaS, and white-collar employment is genuinely illuminating. The problem is not the question. The problem is that the scenario treats as inevitable a chain of events riddled with fragile assumptions and systematic omissions, and that the very method by which it reasons is inadequate for the object it analyzes.
The economy is not a machine
The deepest problem with the scenario is not what it says, but how it reasons. It is built entirely as a sequential causal chain: AI improves, companies lay off workers, consumption falls, credit deteriorates, mortgages wobble, the economy collapses. Each link is a necessary consequence of the previous one; the final result, an inevitable outcome of the sequence. This is mechanistic reasoning: it treats the economy as a device in which, given an initial cause, effects propagate in a linear, predictable, and unidirectional manner.
The flaw manifests on three levels. At every node of the narrative, only one outcome is possible: the negative one. ServiceNow loses revenue, cuts staff, accelerates AI adoption, and the loop self-reinforces (without anyone reorienting, discovering hidden costs in the alternatives, or intervening to regulate). The chain also always operates at the same level of analysis: companies laying off, consumers spending less, creditors suffering. Nothing qualitatively new ever emerges from the interaction between agents. Finally, no agent in the scenario learns, changes strategy, innovates, or finds solutions that didn't exist at time zero. Firms lay off and buy AI, workers suffer, consumers stop spending, the government is slow. The gears of the machine do not learn.
Complexity theory applied to economics (developed across the Santa Fe Institute, the Institute for New Economic Thinking, and the Oxford Martin School) proposes a radically different paradigm. The economy is not a machine but a complex adaptive system (CAS): a vast number of heterogeneous agents interacting locally, learning, adapting, and collectively producing emergent behaviors that are not reducible to the parts. In a CAS, the relationship between input and output is not proportional. Small causes can have large effects, and large causes can have surprisingly small effects. The Citrini scenario assumes proportionality (more AI, more layoffs, more crisis), but in a nonlinear system it is possible that a certain level of automation is absorbed without crisis, that a marginal increment beyond a threshold triggers a cascade, or that the system reorganizes before reaching that threshold. The outcome depends on conditions that the sequential model does not even contemplate.
The aggregate behavior of a CAS is not predictable from the sum of individual behaviors, as P. Anderson reminded us in his famous "More is Different." Millions of firms adopting AI, millions of workers readapting, and millions of consumers changing habits will produce a result that no deterministic scenario can anticipate — neither the catastrophic one nor the optimistic one. This property — emergence — is what makes any linear causal chain fundamentally inadequate as a forecasting tool.
Agents in a CAS are not passive: they learn, imitate, innovate, fail, and are replaced. The economy is not a chain of dominoes; it is an ecosystem where species go extinct, mutate, and new ones are born. W. Brian Arthur (1989, 2015) showed that technologies do not prevail in a manner determined by their intrinsic superiority, but through feedback processes influenced by random events, institutional decisions, and network dynamics. This path dependence implies that the trajectory of AI in the economy is not written: it will be shaped by historical accidents, political decisions, and unforeseen interactions that Citrini's sequential chain can neither incorporate nor imagine.
Eric Beinhocker (2006) synthesized the complexity approach by arguing that wealth is not a static stock that gets redistributed but an evolutionary process of continuous creation. Technological modules, social modules, and business modules co-evolve in a perpetual race. AI is not a terminal event but a new module entering the evolutionary process and accelerating it, producing both destruction and creation in ways that sequential reasoning cannot capture. J. Doyne Farmer and colleagues (2009, 2024) showed with agent-based models that macroeconomic dynamics diverge drastically from what sequential models predict when heterogeneity, interaction networks, and adaptation are incorporated. The same shock can produce collapse or rapid reorganization depending on the topology of the economic network, the speed of agent adaptation, the distribution of skills, and the institutional architecture — parameters that a "memo from the future" cannot know.
A concrete example. The scenario describes the DoorDash chain: AI lowers barriers to entry, dozens of competitors emerge, agents route toward the cheapest option, DoorDash loses its moat, margins collapse. In mechanistic logic, this is the end of the story. In complexity logic, it is the beginning. Fragmentation creates new problems (quality, reliability, accountability) that generate demand for second-order solutions: aggregators of aggregators, certifiers, on-demand insurance, reputation platforms. Some competitors fail, others merge, natural monopolies emerge based on different assets. The market does not collapse: it reorganizes into a form no one had foreseen. This mechanism — destruction followed by emergent reorganization — is documented in every single technological disruption in history. But it is invisible to a model that reasons through sequential chains, because emergent properties by definition are not deducible from individual components.
Complexity theory does not offer an alternative prediction. It offers something more important: a reason to distrust all sequential predictions, including optimistic ones. Useful statements are not about "what will happen" but about "which properties of the system make certain outcomes more or less likely." Agent diversity, speed of adaptation, quality of institutions, network topology, presence of multiple feedback loops: these are the relevant properties. They are exactly the ones that the Citrini scenario suppresses to make its linear narrative work.
Criticizable elements
The narrative determinism of the piece (precise numbers from a fictitious future, such as unemployment at 10.2% or the S&P at -38%) creates a false sense of inevitability. The reader ends up evaluating the plausibility of the narrative instead of the probability of the events — a cognitive bias known as the conjunction fallacy: a detailed and coherent story seems more probable than it actually is, precisely because it is detailed and coherent. The concept of "Ghost GDP" (output that appears in national accounts but does not circulate in the real economy) is a powerful metaphor but economically misleading. If a company produces more with fewer employees, those profits go somewhere: dividends, buybacks, reinvestment, taxes. They do not vanish. The distributional problem is real, but confusing distribution with evaporation is an analytical error.
The scenario systematically confuses capability and deployment. Many of the dynamics described (widespread agentic commerce, massive "build vs. buy" renegotiations, generalized disintermediation) require not only capable models but access to data, permissions, standards, insurance, audits, and aligned incentives among multiple actors. Adoption curves are historically S-shaped, not exponential to infinity. In the article's comments, a Mag7 employee from the Cloud & AI division confirms: actual capabilities in software engineering, while significant, are far from the current hype. The article projects sector-specific trends (software, SaaS, consulting) onto the entire economy, ignoring that healthcare (18% of GDP), construction, education, the public sector, and the care economy have enormous barriers to rapid automation. The idea that "the agent has no inertia, therefore zero loyalty" is true in transactional markets, but many moats are not based on friction: physical logistics, long-term contracts, network effects, regulation, technical switching costs. In many cases, AI strengthens market leaders. The thesis that agents will route toward stablecoins on Solana or Ethereum L2 ignores regulation, chargebacks, AML/KYC, and fraud management: the most likely path is hybrid, not binary. The entire scenario unfolds over 24 months, a timeframe not supported by historical evidence even for the fastest crises. The mortgage parallel with 2008 is structurally weak: in 2008, loans were bad at origination; here they would be sound, with gradual stress, in a context of fixed-rate mortgages with low LTVs. Finally, the scenario becomes a crisis because the policy response is weak, but on a visible employment shock concentrated in the most politically influential segment of the population, democracies have historically responded more quickly than the scenario assumes.
Citrini is an incomplete child of heterodoxy
It would be easy to classify the scenario as a product of conventional neoclassical thinking. It is not. Its conceptual framework draws on heterodox traditions far more than appearances suggest. The self-reinforcing feedback loop without a stable equilibrium point is the opposite of neoclassical logic, where imbalances generate rebalancing forces: it is far closer to Minsky and Myrdal. The idea that a collapse in labor income produces a collapse in demand that does not self-correct is Keynesian, not neoclassical (since it implicitly assumes that Say's Law does not hold). The explicit use of the term "reflexivity" for the ServiceNow dynamic has roots in economic sociology and post-Keynesian economics, and is antithetical to rational expectations. The scenario operates in the language of Wall Street, which many confuse with economic theory, but its DNA is heterodox.
The problem is that it uses these insights selectively and incompletely. It takes the pars destruens of each tradition and systematically discards the pars construens. It takes Minsky's endogenous financial fragility but not the Minskyan solution (Big Government and Big Bank as stabilizers). It takes the Kaleckian profits/wages → demand relationship but does not explore the role of public spending as a substitute for labor-based demand. It captures self-reinforcing positive feedback loops but not negative feedback, emergence, and adaptation that complexity theory places at the center. It brilliantly shows how existing institutions are inadequate, but does not explore the capacity of institutions to reform — which is the other half of the institutionalist tradition, that of Commons, not just Veblen. The result is a scenario that is theoretically more sophisticated than it appears, but skewed toward catastrophism not out of ignorance but by selection.
The case for a constructive outcome
If AI reduces the marginal costs of services that are expensive today (support, coding, legal advice, medical diagnoses), prices fall and purchasing power is freed up. Historically, technological revolutions have produced more variety and more consumption in new categories, not a permanent demand vacuum. The article notes that machines do not consume discretionary goods, but forgets that humans consume more when something becomes radically cheaper. Consumer surplus generated by technological deflation is the great absentee.
More autonomous agents in circulation means more need for governance, assurance, security, quality, auditing, data stewardship, AI risk & compliance. The composition of work changes; work does not disappear. And complexity theory tells us that new professions emerge as emergent properties of the interaction between adaptive agents — they are not predictable ex ante, which makes the assumption that no new sector will emerge structurally unfounded. If agents reduce rents from inertia, the result is allocative efficiency: less waste, more transparency, better price discovery. It is analogous to e-commerce versus traditional retail — painful for exposed incumbents, but with a higher net welfare balance.
Large players with proprietary data, established compliance, and distribution channels can absorb part of the shock. In a world of high technological risk, trust and legal liability become assets, making a more oligopolistic and regulated transition plausible, not a collapse in 24 months. AI applied to financial and logistical plumbing (forecasting, credit scoring, early stress detection) can make the economy more resilient, not less. And when the cost of a fundamental input collapses, the historical result is not permanent depression but expansion: new markets emerge because activities that were previously prohibitive become feasible.
The real lever is institutional. If AI increases aggregate productivity, society can monetize that surplus in ways that sustain demand: a shorter workweek, wage insurance, a tax mix shifted toward rents and monopolies, public investment in care, infrastructure, education. The crisis scenario is a governance choice, not a law of physics. The complete Minsky, the complete Kalecki, and Commons offer the roadmap that the Citrini scenario omits: Big Government as a stabilizer, public spending as a substitute for labor-based demand, institutional reform as a documented historical mechanism — not as utopia.
Conclusion
The Citrini Research scenario is useful as a stress test. The mechanisms it describes are real. The questions it poses are urgent. Its theoretical roots (post-Keynesian, institutionalist, structuralist) are more robust than the journalistic format lets on.
But the scenario suffers from three fundamental flaws. It is epistemologically mechanistic: it reasons through sequential causal chains in a system that is complex, adaptive, and nonlinear. It is theoretically selective: it takes the pars destruens of heterodox traditions and discards the pars construens. And it is temporally unrealistic: it compresses into 24 months a process that history suggests takes decades.
The answer to the question "what will happen when intelligence becomes abundant?" is not "inevitable crisis" nor "guaranteed prosperity." It is: it depends on the relative speed between disruption and adaptation, on the architecture of networks, on the distribution of power, and on contingent events that no one can anticipate. It is a narratively less satisfying answer but an enormously more correct one epistemologically.
The canary is alive. If it dies, it will be because we left income distribution and credit without a buffer.
