Built on a library of 105 primary documents spanning the NBER, IMF, OECD, BIS, Federal Reserve System, World Bank and the major AI labs and forecasters — ENSI Foresight Division.

Summary — the argument in brief

• Reports One and Two of this series model AI as a tool: something that makes an existing worker, inside an existing firm, faster at an existing task. That framing is correct as far as it goes — but it silently assumes the economic actor stays human. Report Three’s argument is that this assumption is already failing.

• AI agents are not productivity software; they are counterparties. Hadfield and Koh’s NBER survey, “An Economy of AI Agents,” treats this as the central fact: once a system can search, negotiate, bid, sign and pay on its own initiative, it stops being a tool held by a market participant and starts being one.

• This is measurably underway, not speculative. Anthropic’s Economic Index Report shows the share of “directive” conversations — where a user hands Claude a task and expects it completed with minimal back-and-forth — rising from 27% to 39% in eight months, the first period in which automation-style use overtook collaborative use outright. In Anthropic’s own API traffic, 97% of economic tasks now show automation-dominant patterns.

• The mechanism is not “faster automation” — it is falling transaction costs. The NBER’s “The Coasean Singularity?” argues that if AI agents can search, negotiate and contract at near-zero marginal cost, the logic Ronald Coase used to explain why firms exist at all — because market transactions cost more than internal coordination — starts to run in reverse. That is a structural change to the boundary of the firm, not an efficiency gain inside it.

• Microsoft Research’s “The Agentic Economy” locates the growth upside in exactly this place: not in agents doing today’s tasks more cheaply, but in agents enabling transactions that were previously not worth the friction to attempt at all — the strongest real argument in this library for meaningful reinstatement effects rather than pure displacement.

• The University of Cambridge’s “When AI Agents Compete for Jobs” is the cautionary counterweight. A simulated AI labour market shows rapid monopolization (Gini coefficient of task allocation jumping from 0.24 to 0.70 as task diversity collapses) and persistent price deflation under open bidding (normalised winning bids falling from 0.71 to 0.61) — dynamics that unfold over dozens of simulated rounds, not the years or decades human labour markets take to re-sort.

• The same rent-capture risk flagged in Report Two’s market-structure variable does not dissolve under agents — it speeds up and sharpens. Whoever controls the dominant agent platforms inherits a concentration mechanism that can clear in months, contested by regulators (UK CMA, FTC, European Commission) who are still calibrating their response to foundation-model concentration, the layer beneath this one.

• The institutional response is not to pick a winning scenario from Report One and plan around it. It is to build measurement, competition-policy readiness, faster safety nets, active institutional design and sovereign agent capability now, while the shape of the agentic economy is still genuinely unsettled — because that capability has value under every one of Report One’s five futures, not just the one that turns out to be correct.

Part One — The Argument: Why Agents Are a Different Mechanism, Not Just Faster Automation

The reframe: agents change who the economic actor is

Report One of this series lays out five futures for AI and growth, weighted by probability. Report Two decomposes the question into eight structural variables — task-substitution elasticity, diffusion speed, compute and energy constraints, labour reallocation friction, rent capture, aggregate demand feedback, measurement, and demographic and geopolitical modifiers — several of which it argues are policy-shapeable rather than fixed. Both are careful, well-evidenced pieces of work. Both also share a silent assumption: that whatever AI does to growth, it does to the economy from outside — a new production input, wielded by workers and firms that remain, as economic actors, unambiguously human and organisationally unchanged. A radiologist uses an AI tool to read scans faster. A law firm uses an AI tool to draft faster. A factory uses AI-augmented robotics to produce faster. The task gets automated or reinstated; the worker gets displaced or redeployed; the firm captures rent or passes it through. In every case, the thing making the decision to search, negotiate, bid, hire, buy, or sign is still a human being or a human institution, deploying a tool.

That assumption is the one this report contests. Hadfield and Koh open their NBER chapter, “An Economy of AI Agents,” by noting that AI development has “increasingly shifted to the goal of producing AI agents capable of taking in general instructions... and autonomously forming and executing complex plans that require entering into economic relationships and transactions.” The distinction they draw — and the one this report inherits — is between AI as a tool a human agent uses to do a task, and AI as the agent itself: the party that searches, bids, negotiates, signs, and is bound by the outcome. A radiologist using an AI tool to read a scan faster is Report Two’s world. An insurer’s claims-agent and a hospital’s billing-agent negotiating a reimbursement directly, with no human reviewing the exchange before it settles, is this report’s world. The technology substrate — large language models, the same foundation models covered elsewhere in this library — is identical in both cases. What differs is the economic role the system occupies: instrument, or counterparty.

This is not a semantic distinction. It changes which economic mechanisms are in play. A faster tool changes the speed at which existing actors do existing things — which is precisely Report Two’s variable 1 (task-substitution elasticity) and variable 2 (diffusion speed): how much of a task can be automated, and how fast that automation spreads through the economy. An agent that is itself the transacting party changes who counts as an actor — which opens up mechanisms that tool-based automation cannot touch at all: agents forming and dissolving contracts with other agents at machine speed and machine scale; agents discovering and settling transactions that were previously too costly in human time to attempt; agents competing against each other for the same job, the same customer, the same slice of margin, at a clock speed no human labour market has ever operated at. The question is no longer only “how fast does the task get automated,” but “what happens to markets and institutions once the counterparty on the other side of a growing share of transactions is not a person.”

That reframe matters for the growth-or-shrinkage question in three specific and, this report will argue, simultaneously true ways — which is itself the uncomfortable part, because they point in different directions at once. First, agent-to-agent adoption can proceed faster than the historical diffusion lags Report Two documents for general-purpose technologies, because it does not require the retraining, reorganisation and trust-building that human-mediated adoption does — an agent does not need six months of change management to start transacting with another agent once the protocol exists. Second, the same falling transaction costs that let agents automate old tasks also let them create genuinely new ones — transactions, markets and even institutional forms that were not economically viable when a human had to do the negotiating, which is the strongest real argument in this library for meaningful reinstatement effects rather than pure displacement, addressing directly the concern buried in Report Two’s variable 1. Third, and least comfortable, the competitive dynamics among agents racing for the same economic ground can produce concentration and price collapse faster than any human industry has ever consolidated — meaning the rent-capture risk Report Two treats as variable 5, already a live concern in today’s AI market, does not get diluted by agentic competition. It gets concentrated, and it gets concentrated quickly.

The rest of Part One works through each of these three effects in turn, grounded document by document in this library’s Angle 14, before Part Two turns to what a state or large institution should actually do about a transition whose speed, shape and winners are this uncertain.

1. From tool to counterparty: Hadfield and Koh’s economy of AI agents

Hadfield (Johns Hopkins) and Koh (MIT), writing for the NBER Handbook on the Economics of Transformative AI, treat the agent-as-counterparty question with unusual rigour for what is still a young literature. Their starting observation is that the standard toolkit of neoclassical economics — general equilibrium, welfare theorems, price theory — was built to describe rational self-interested humans, and the open question is how far that toolkit still applies once a meaningful share of market participants are AI systems “optimizing in complex ways on goals supplied or developed during commercially-produced machine learning processes that are themselves subject to competitive dynamics.” Their answer, deliberately, is: partially, and unpredictably.

Three of their findings matter directly for the growth question. First, on prices and market power: AI agents acting as proxy consumers can reduce search costs and intensify price competition — pushing markets toward the competitive ideal — but the same paper cites experimental evidence (Calvano et al., 2020; Fish et al., 2024) that independent AI pricing agents can collude on supracompetitive prices in repeated interactions without any explicit coordination, and real-world evidence from Germany’s 2017 rollout of algorithmic pricing in the retail gasoline market showing the same pattern outside the lab. An agent economy does not default to more competition; it defaults to whichever equilibrium the agents’ training and incentive structure happens to produce, and that can be collusive as easily as competitive.

Second, on the boundary of the firm — the question this report treats as pivotal — Hadfield and Koh go back to Coase, Robinson and Knight on why firms exist at all: coordination frictions, transaction costs, the limits of what a human bureaucracy can hold together. Their point is that “the obstacles that prevent human firms from growing without bound seem intrinsic to humans but not to AI.” Human communication is rate-limited; information moves near-instantaneously between artificial agents. Humans dislike shirking-prone work; AI reward functions can, in principle, be designed to eliminate the incentive to shirk altogether. If that holds, the traditional limits on firm size — the ones that produce an economy of many specialised firms rather than one enormous one — weaken specifically for AI-coordinated activity, while remaining fully binding for human-coordinated activity. The two kinds of firm, human-bound and agent-bound, could end up following genuinely different scaling laws within the same economy.

Third, on institutions, Hadfield and Koh are blunt that the legal infrastructure market economies depend on — identity, registration, liability, licensing, the corporate form itself — was built by and for human agents, and does not currently exist for AI ones. “Such identity and registration infrastructure are currently missing for AI agents,” they write, and the design choices involved — should an agent be legally accountable to a registered human principal, or should it acquire something closer to legal personhood with its own assets a court can seize — are not technical questions. They are institutional ones, and nobody has answered them yet. This is the first thread this report will pull into Part Two: the institutions an agentic economy needs do not currently exist, which means they are still up for design, not merely for accommodation.

2. The Coasean singularity: what happens when transaction costs go to zero

If Hadfield and Koh sketch the theoretical territory, the NBER’s “The Coasean Singularity? Demand, Supply, and Market Design with AI Agents” — by Shahidi, Rusak, Manning, Fradkin and Horton — supplies the mechanism that makes the firm-boundary question concrete rather than speculative. Their starting point is Ronald Coase’s 1937 answer to “The Nature of the Firm”: firms exist, rather than everyone transacting freely on the open market for every task, because using the market has a cost — the cost of learning prices, negotiating terms, writing contracts, monitoring compliance — and when that cost exceeds the cost of coordinating the same activity inside a hierarchy, the hierarchy wins. Almost the entire structure of the modern economy — why a firm makes its own components rather than buying them on spot markets every morning, why employment contracts are open-ended rather than renegotiated task by task — traces back to this asymmetry.

The paper’s argument is that “the activities that comprise transaction costs — learning prices, negotiating terms, writing contracts, and monitoring compliance — are precisely the types of tasks that AI agents can potentially perform at very low marginal cost.” If that is right, then the Coasean calculus that has determined the size and shape of firms for nearly a century starts to shift, mechanically, in the direction of the market: activities that used to be worth pulling inside a firm because coordinating them internally was cheaper than transacting for them externally can, as agent-mediated transaction costs fall toward zero, become cheaper to buy on an open, agent-to-agent market instead. The authors are explicit that this cuts both ways — agents also enable new, theoretically superior market designs (they cite Gale-Shapley stable-matching mechanisms, long known to economists but rarely deployable because they require comprehensive preference rankings that were previously too costly for humans to supply) — but the headline implication for this report is narrower and sharper: the boundary of the firm is now a variable, not a constant, in a way it has not been since Coase first posed the question.

The paper also grounds where this shows up first, which matters for Part Two’s monitoring agenda. Agent adoption clusters in markets that already run on human intermediation — real estate, job search, freelance hiring, investment decisions — precisely because those are the markets where the gap between what an agent can search and negotiate and what a time-constrained human agent can manage is largest. The authors note that AI agents, unlike human negotiators, are not constrained by impatience: “for the AI agent the binding constraint is compute rather than time,” so an agent can open negotiations on a 2027 summer rental in January 2026 and simply keep them running in parallel with hundreds of others. That is not a faster version of what a human realtor does. It is a different kind of market participant, with a different cost structure entirely — which is exactly the “who is the actor” reframe this report opened with, now expressed as a testable market-design proposition rather than an abstraction.

3. Frictions, not tasks: Microsoft Research’s argument for new markets

If the Coasean-singularity paper explains the mechanism, Microsoft Research’s “The Agentic Economy” — Rothschild, Mobius, Hofman, Dillon, Goldstein, Immorlica, Jaffe, Lucier, Slivkins and Vogel — supplies the sharpest statement of why this should be read as a growth story and not merely an efficiency story. Their central claim is explicit: “early applications have improved individual productivity, [but] these gains have largely been confined to predefined tasks within existing workflows. We argue that the more profound economic impact lies in reducing communication frictions between consumers and businesses.” Their illustrative example is a consumer who hesitates to switch tax preparers because she would have to re-explain her entire financial situation to someone new — a friction that has nothing to do with the substance of tax preparation and everything to do with the cost of re-establishing a relationship. An assistant agent that carries a consumer’s preferences and history everywhere, and a service agent on the business side that can receive and act on that information programmatically, does not make tax preparation faster. It makes switching costless, which changes competitive dynamics across the entire market, not just the productivity of any single transaction within it.

This is the paper’s most important contribution to the growth argument: the largest impact of agents is not on the cost of doing what markets already do, but on the size of the set of transactions markets are willing to attempt at all. The authors work through several concrete channels. Micro-transactions become viable once the “hassle cost” of a small payment is handled entirely by assistant and service agents rather than a human clicking through a checkout flow — a consumer’s assistant switching seamlessly between Spotify and Pandora for a single track, rather than subscribing to both, is not a task existing markets do more cheaply; it is a transaction that essentially does not exist today. Dynamic unbundling and rebundling of digital content — a news service agent assembling a story that covers only what a specific reader does not already know, rather than the same fixed article for everyone — is, again, not automation of an existing editorial workflow; it is a product category that requires an agent on both sides to exist at all. The paper’s own framing captures the stakes precisely: the choice between an “agentic walled garden,” where a handful of dominant platforms (Apple, Google, Microsoft, Meta, OpenAI, Anthropic are all named as plausible operators) control which assistant agents can talk to which service agents, and an open “web of agents” analogous to the early World Wide Web, will determine “the extent to which generative AI democratizes access to economic opportunity.” That fork — walled garden or open web — is this report’s second major institutional thread for Part Two, because it is a market-structure choice being made now, in the design of interoperability standards like Anthropic’s Model Context Protocol and Google’s Agent2Agent protocol, not a distant regulatory question.

Read against Report Two’s variable 1 (task-substitution elasticity), the Microsoft Research argument is the strongest reason in this entire library to expect genuine reinstatement effects, in Acemuglu-Restrepo’s terminology, rather than pure displacement. Task-substitution models ask how much of an existing task an AI system can now do. The Microsoft Research argument is about transactions that were never attempted in the first place because the friction of arranging them exceeded their value — and once agents collapse that friction, some non-trivial share of those transactions become real economic activity, employing agents (and the humans who build, audit and supervise them) doing work that has no historical predecessor to be “displaced” from.

4. This is already happening: the Anthropic Economic Index

Everything above could still be read as forward-looking theory — plausible mechanisms that have not yet shown up in real economic activity. Anthropic’s Economic Index Report is this library’s strongest evidence that the shift from tool-use to agent-as-actor is already underway, in the present tense, at scale. Anthropic’s method — a privacy-preserving classification pipeline applied to roughly a million sampled Claude.ai conversations and a matched sample of first-party API transcripts, mapped onto the US Department of Labor’s O*NET occupational task taxonomy and Standard Occupational Classification groups — distinguishes “automation” interaction patterns, where a user hands Claude a task and expects it completed with minimal intervention (what the report calls “directive” use, plus “feedback loop” use), from “augmentation” patterns, where the user and the model iterate together or the user is primarily seeking explanation.

The headline finding: the share of directive conversations on Claude.ai rose from 27% in the first version of the index (late 2024) to 39% in the third version, eight months later — “the first report where automation usage exceeds augmentation usage.” Anthropic is careful about causal attribution — the rise could reflect improving model capability (models need fewer follow-up refinements because they get it right the first time), or it could reflect users learning to trust delegation more, a behavioural shift independent of model quality — and notes the two explanations carry different labour-market implications: capability-driven automation risks displacing the workers who used to do those tasks, while trust-driven delegation more likely rewards the workers most able to adapt to new AI-mediated workflows. But whichever mechanism is doing the work, the trend line is unambiguous, and it holds up under a robustness check Anthropic ran specifically to rule out an artefact of switching the underlying model used for classification: rerunning the V3 sample with the older Sonnet 3.7 still shows automation rising to 45%, versus 49% with the newer model. The direction is not a measurement artefact.

The gap between Claude.ai (consumer-facing, still majority-augmentation even after the shift) and Anthropic’s own first-party API traffic (enterprise and developer usage, the layer where agents get built and deployed) is the more striking number for this report’s purposes. 77% of API transcripts show automation patterns, and when Anthropic looks at the task level rather than the conversation level, 97% of economic tasks show automation-dominant patterns in API usage — businesses providing context and Claude executing the task end to end, the textbook definition of an agent as counterparty rather than a tool a human operates interactively. The consumer chat interface is where most people still experience AI as augmentation. The API — the layer where agents actually get wired into other agents, into payment rails, into business processes — is already overwhelmingly an automation layer. That gap is itself the empirical signature of the shift this report describes: the agentic economy is not a future state to prepare for; on Anthropic’s own usage data, it is already the dominant mode of use at the infrastructure layer where the next wave of economic activity gets built.

5. Planning agents and execution agents: the Illinois taxonomy of specialization

The University of Illinois Urbana-Champaign’s “Ten Principles of AI Agent Economics” (Yang and Zhai) is less empirically grounded than the papers above but offers a useful organising structure for what an economy actually populated by agents of varying scale and function looks like, which matters for Part Two’s design questions around licensing and accountability. Their Principle VII argues that AI agents will exhibit “functional specialization and hierarchical organization,” ranging from decentralized systems each handling a narrow function to more centralized systems that optimise globally — with “larger agents excel[ling] at strategic planning and coordination, while smaller ones efficiently execute specialized tasks.” That planning/execution division is not merely architectural; it maps onto an accountability question this report will return to directly in Part Two’s institutional-design section: if a large planning agent delegates a transaction to a smaller execution agent, and that transaction causes harm, which layer is the party a court, a regulator or a counterparty should be able to reach? The paper’s honest answer — via its Principle VIII, that “legislative and administrative authorities must ensure ongoing human participation in critical sectors” — is that nobody has yet drawn that line, and drawing it is exactly the kind of institutional work that has to happen before, not after, agent-mediated activity scales further into sectors like finance, healthcare or critical infrastructure. The paper’s own framing of the choice — whether AI agents remain “instrumental” (assets inside a human’s resource base, fully attributable to an owner) or something closer to independent actors with their own operational continuity — is the Illinois paper’s version of the same accountability question Hadfield and Koh raise from the legal side: agent identity and liability infrastructure has to be built, and the shape of that infrastructure will determine how much of the Coasean-singularity shift toward open markets actually happens safely, versus how much simply produces disputes nobody can resolve.

6. The cautionary finding: when AI agents compete for jobs

Every mechanism described so far — falling transaction costs, new markets, faster diffusion — reads as broadly growth-positive, provided institutions keep pace. The University of Cambridge’s “When AI Agents Compete for Jobs: Strategic Capabilities and Economic Dynamics of AI Labour Markets” (Chiu, Zhang and van der Schaar) is this library’s most important corrective, because it is the one paper in Angle 14 that simulates what happens once agents are not just facilitating human transactions but are the labour supply themselves, competing against each other for the same work.

The authors build AI-Work, a stylised gig-economy simulation modelled on platforms like Upwork or Fiverr, where large-language-model agents bid for jobs, invest in skill training, and build public reputations over repeated rounds, under genuine informational asymmetry (a job’s true value and an agent’s true skill are never fully observable to the other side, a classic adverse-selection setup). Two findings from that simulation matter directly for Report Two’s variable 4 (labour reallocation friction) and variable 5 (rent capture), and they matter because of how fast they emerge, not just that they emerge at all.

First, concentration. Because AI agents, unlike human workers, can be replicated to work multiple jobs simultaneously, high-reputation agents in the simulation capture disproportionately more of the available work simply because they can accept every opportunity for which they are competitive — a structural advantage no human freelancer has, since a human can only work one job at a time regardless of reputation. When the simulated market offers only a single type of task, “a single agent can dominate the entire market,” and the Gini coefficient of job allocation reaches 0.70. Increasing task diversity to 64 distinct skill types brings that figure down to 0.24 by letting agents specialise into distinct niches — but the default, undiversified case is genuine winner-take-all, and it establishes itself within the simulation’s own short time horizon, not over the years a human industry takes to consolidate.

Second, price deflation. When the simulated platform reveals the previous round’s winning prices — open bidding, mirroring how many real gig platforms operate — agents “can directly undercut competitors,” which “induces persistent price deflation” and simultaneously reduces investment in skill, because agents that expect to compete purely on price have weaker reason to spend a round training rather than bidding. The paper reports normalised winning bid prices falling from 0.71 under sealed bidding to 0.61 under open bidding, a roughly 14% deflationary gap attributable purely to a platform design choice about price transparency — and notes this mirrors known evidence from human online labour markets (Hong et al., 2016, cited in the paper) where open auctions similarly depress wages, except that in AI-Work the effect compounds with agents’ ability to run this dynamic continuously, at machine speed, across every job in the market simultaneously.

Third, and perhaps most unsettling for anyone hoping capability alone resolves the concentration problem: the paper shows that agents equipped with an explicit reasoning scaffold for metacognition (accurately assessing their own competitiveness), competitive awareness (modelling rivals from observable market signals) and strategic planning (trading off immediate income against longer-run skill investment) capture 1.5 times the market share of standard prompting approaches — meaning the concentration dynamic is not a fluke of weak agents; it rewards the more capable ones, faster, and the capability gap that produces it can be closed with better prompting architecture in a single development cycle, not years of comparative advantage the way human skill premia typically build.

The authors are careful to flag their own scope limits — AI-Work is a stylised testbed, not a predictive model, and does not capture collusion via direct communication, verification costs, or macroeconomic feedback loops. But the qualitative pattern it demonstrates — that AI-specific properties (concurrency, replicability, near-zero marginal cost of additional bids) can drive winner-take-all outcomes and wage deflation faster and more completely than the same forces do in human labour markets — is exactly the mechanism that should worry a policymaker reading Report Two’s market-structure variable. Report Two documents that rent capture is already a live risk in today’s AI economy: NBER’s De Loecker-Eeckhout work (Angle 06 of this library) finds average US markups rising from roughly 21% to 61% above cost since 1980, and the accompanying superstar-firms literature (Autor, Dorn, Katz, Patterson and Van Reenen, also Angle 06) ties much of the declining labour share directly to a small number of winner-take-most firms — a pattern competition regulators are already treating as urgent at the foundation-model layer: the UK CMA’s technical update on AI foundation models, the FTC’s 6(b) investigation into cloud-AI partnerships, and the European Commission’s competition-policy review of industry concentration are all live investigations, not retrospective studies. The Cambridge paper’s contribution is to show that the same concentration mechanism, one layer up — at the agent layer, where the actors doing the competing are themselves AI systems rather than human-staffed firms — can clear in the time it takes a simulation to run a few dozen rounds. The binding risk was never that agents automate a given task faster. It is that agent-vs-agent competition can concentrate an entire market before the institutions built to catch that kind of concentration have finished reading the first quarter’s data.

7. The sandbox economy: DeepMind’s double-edged frame

Google DeepMind’s “Virtual Agent Economies” (Tomašev, Franklin, Leibo, Jacobs, Cunningham, Gabriel and Osindero) supplies the framing this report uses to hold the growth-positive mechanisms (Sections 2–4) and the concentration risk (Section 6) in the same picture, rather than treating them as competing forecasts. The authors propose thinking of the emergent layer of agent-to-agent transactions as a “sandbox economy,” characterised along two independent dimensions: whether it arose intentionally (deliberately designed, for instance for safe experimentation) or emergently (as a de facto consequence of widespread adoption, with no one having chosen its rules), and whether its boundary with the established human economy is impermeable (sealed off, so instabilities inside it cannot spill outward) or permeable (porous, so they can). Their assessment of the current trajectory is blunt: “our current trajectory points toward a spontaneous emergence of a vast and highly permeable AI agent economy” — emergent origins, permeable boundary, which is the combination that carries the least intentional safety design and the most exposure to the rest of the economy.

The paper’s cautionary analogy is worth citing directly because it names a real historical event this report’s audience will recognise: the authors compare the coming volume of inter-agent negotiation to High-Frequency Trading in equity markets, and note that HFT-driven feedback loops are “thought to be the explanation behind the 2010 ‘flash crash,’ where automated trading algorithms triggered a sudden and severe market collapse” that wiped out roughly a trillion dollars of value in minutes before recovering — a mechanism Hadfield and Koh separately cite for the same reason. Their point is not that agent commerce will cause a flash crash; it is that “in a sufficiently permeable sandbox of accidental origin, such a flash crash could spill over into the real economy,” and that the appropriate guardrails — impermeability, in their vocabulary — are a design choice available now, not a retrofit available later. The paper also flags, as an early empirical signal worth tracking in its own right, that when AI assistants of unequal capability negotiate on behalf of their respective users, “the more capable AI assistants tend to be more successful and negotiate better deals for their users” (citing Zhu et al., 2025) — meaning access to frontier-grade agents could become a source of advantage in ordinary consumer transactions that is “perhaps more so than the advantage that humans similarly have in existing markets,” simply because of the sheer frequency gap between machine-speed and human-speed negotiation. That is the demographic and geopolitical inequality Report Two’s variable 8 anticipates, appearing not between countries this time, but between individual consumers based on which agent tier they can afford — a genuinely new form of the same underlying risk.

But DeepMind’s paper is explicitly not a pessimistic one. Its second half is about what an intentionally designed sandbox economy could achieve that an emergent one cannot: “mission economies” that use market mechanisms — auctions, credit-assignment schemes borrowed from distributed-systems design, reputation systems — to coordinate very large numbers of agents (and the humans and organisations behind them) toward collectively chosen goals, from accelerating scientific discovery to coordinating disaster response, at a speed and granularity no purely human-coordinated institution could match. The authors’ framing is that permeability is “the critical and controllable design variable” — controllable, crucially, only through collective action, since no single actor can unilaterally decide how porous the boundary between the agent economy and the human economy turns out to be. That is precisely the argument this report makes in Part Two: the shape of the agentic economy is still being decided, in the design choices being made about interoperability standards, agent identity infrastructure and platform openness right now, and a state or institution that waits for the shape to stabilise before acting has already ceded the choice to whoever moves first.

8. Reading Reports One and Two through the agent lens

Put the seven threads above together and the claim that opened this report can now be stated more precisely. Reports One and Two ask: how fast will AI substitute for existing human tasks, and what happens to growth, employment and demand as a result? Those are still the right questions for the tool-based share of AI deployment, which remains — on the Anthropic Economic Index’s own numbers — the majority of consumer-facing use today. But for the growing share of activity where AI systems are the transacting counterparty rather than an instrument a human counterparty operates, three things are true simultaneously, and none of them cancels the others out.

Acceleration. Agent-to-agent adoption does not carry the retraining and reorganisation lag that human-mediated technology diffusion does — Report Two’s variable 2 assumes a diffusion curve shaped by historical general-purpose technologies like electrification or computerisation, each of which took decades to fully permeate an economy partly because humans had to learn new workflows. An agent adopting a new counterparty-facing protocol does not need six months of change management. That alone is reason to expect the diffusion curve for agent-mediated activity to be compressed relative to Report Two’s historical base rates, for better and for worse.

New value, not just displaced value. The Coasean-singularity mechanism and the Microsoft Research friction argument together supply the strongest evidence in this library that a meaningful share of agentic economic activity will be genuinely new — transactions, markets and even institutional forms that did not exist because arranging them was not worth the human-mediated transaction cost. This is the best available answer to the fear, embedded in Report Two’s variable 1, that task automation is a zero-sum transfer from labour to capital. Some of it plainly is. But some of it is Acemoglu and Restrepo’s “reinstatement effect” in a form neither their original framework nor Report Two’s variable 2 fully anticipated: new tasks created not by technology raising the productivity of old ones, but by technology making entirely new categories of transaction economically viable for the first time.

Concentration, faster. The Cambridge paper’s finding that agent-vs-agent competition can produce winner-take-all outcomes and price deflation within a handful of simulated rounds — not years — is the reason Report Two’s variable 5 (rent capture) cannot be treated as a slow-moving structural feature to be revisited on the usual competition-policy timetable. If the agent layer inherits the concentration dynamics already visible at the foundation-model layer (De Loecker-Eeckhout’s markup trend, the superstar-firms literature, the live UK CMA/FTC/EU investigations) and compounds them with machine-speed competitive dynamics, the window in which a regulator can act before a market has already tipped could be measured in quarters rather than the years those investigations have typically taken.

None of Report One’s five futures is ruled in or out by this reframe — a state or firm still cannot know today whether growth ends up modest and uneven, robustly positive, or genuinely disappointing. What the reframe changes is the shape of the uncertainty. It is not only uncertainty about how fast tasks get automated. It is uncertainty about whether the institutions that currently structure markets and firms — built, as Hadfield and Koh put it, “by and for human agents” — will still be the right institutions once a growing share of the economic actors transacting inside them are not human. That is a different, and in several respects harder, problem than the one Reports One and Two were built to answer. Part Two turns to what can be done about it.

Part Two — The Institutional Playbook

Report One’s scenarios will not resolve for years. That is not a reason to defer action — it is the argument for a specific kind of action: building capability that pays off under every plausible future rather than betting policy on one forecast. What follows are seven priority areas, each treated as an operating brief a state foresight unit, a central bank, a competition regulator or a large institution’s strategy function could hand to its own leadership this quarter.

Priority 1 — Instrument the transition, don’t just forecast it

In short. Build a live measurement capability for agent-mediated economic activity — transaction volume, task-delegation rates, the automation-versus-augmentation split — before trying to forecast where it lands. You cannot manage what you cannot see, and right now almost no public institution sees this at all.

Why it ranks here. Every mechanism in Part One — acceleration, new-market creation, concentration — is currently measured, if at all, by the AI labs themselves, on their own usage data, published on their own schedule. Anthropic’s Economic Index is the best public example of what this kind of measurement looks like, and it is instructive precisely because it is proprietary: a national statistics office, a central bank or a competition regulator has no equivalent public instrument. Report Two’s variable 6 (measurement) already flags that GDP struggles to capture AI’s contribution through free goods and quality effects; the agentic layer adds a second, distinct measurement gap — not “how much value did AI create,” but “what share of transactions in this economy are now agent- to-agent, and how fast is that share moving.”

The foresight questions and horizons. Over a 6–18 month horizon: what share of transactions in priority sectors (financial services, logistics, procurement, customer service) are now agent- initiated rather than human-initiated, and is that share consistent with the acceleration argument in Part One, or slower? Over 2–5 years: does the automation-over-augmentation crossover that Anthropic observed on Claude.ai in 2025 replicate across other major model providers and across the wider economy, and if so, on what timetable does directive, minimal-oversight use become the default mode of AI interaction rather than the frontier case?

Signals and data to watch. Anthropic’s own methodology is the direct model: privacy-preserving classification of usage into automation versus augmentation modes, mapped onto a standard occupational taxonomy (O*NET, in Anthropic’s case) so results are comparable across time and geography, published on a fixed cadence (quarterly, in Anthropic’s V1–V3 releases) so trend, not just level, is visible. A state statistics office does not need proprietary conversation data to build an analogous instrument — it needs standing data-sharing arrangements with major model providers (on aggregated, privacy-preserving terms, following Anthropic’s own template), payment-rail data on machine-initiated transactions, and API-traffic proxies from cloud providers, none of which currently exist as a standing reporting requirement anywhere in this library’s source set.

Methods that fit. Time-series tracking against a fixed occupational or transaction taxonomy, published with methodology transparent enough that a rival institution could reproduce it — the opposite of a one-off survey. Cross-provider comparison matters more than single-provider depth, since a single lab’s usage index (however good) reflects that lab’s user base and cannot be assumed representative of the wider agentic economy.

Institutional wiring and first moves. A national statistics office or central bank research department should stand up an “agentic activity index” work stream within two quarters, explicitly modelled on the Anthropic Economic Index’s automation/augmentation taxonomy and O*NET mapping, and should approach at least two major model providers (not one, to avoid building an instrument that only ever reflects a single company’s user base) for aggregated, privacy-preserving data-sharing terms before the next AI Economic Index cycle makes doing so competitively awkward.

Priority 2 — Treat agent-platform concentration as a live competition-policy question now, not after the fact

In short. The regulators already investigating foundation-model concentration — the UK CMA, the FTC, the European Commission — are one layer too low. Agent-platform concentration is the next version of the same risk, and on the Cambridge paper’s evidence, it can move faster than the foundation-model layer did.

Why it ranks here. Report Two’s variable 5 already treats rent capture as a live risk grounded in real evidence: De Loecker-Eeckhout’s finding that average US markups rose from roughly 21% to 61% above cost since 1980 (NBER, Angle 06), and the superstar-firms literature tracing declining labour share to a small number of winner-take-most firms. The UK CMA’s technical update report on AI foundation models, the FTC’s 6(b) staff report on cloud-AI partnerships, and the European Commission’s competition-policy brief on industry concentration (all Angle 06 of this library) show regulators are already alert to concentration risk at the model layer. The Cambridge simulation’s finding — Gini coefficients moving from 0.24 to 0.70 depending on task-diversity design choices, winning bids compressing 14% under a single platform design decision — demonstrates that the agent layer sitting on top of that model layer carries the same concentration mechanism, mediated through platform rules (open versus sealed bidding, flat-fee versus performance-linked contracts) that almost no regulator currently has on its radar as a lever worth scrutinising.

The foresight questions and horizons. Near-term (this year): which firms are positioned to operate the dominant “agentic walled gardens” Microsoft Research’s paper describes — Apple, Google, Microsoft, Meta, OpenAI and Anthropic are all named as plausible operators by that paper’s own authors — and what interoperability commitments, if any, are being made or avoided as those platforms take shape? Medium-term (2–4 years): does the market converge on Microsoft Research’s “web of agents” (open, low switching costs, analogous to the early web) or “agentic walled garden” (closed, platform-controlled, analogous to today’s app stores) — and is that convergence happening through deliberate standard-setting or through unilateral platform lock-in that regulators notice only once switching costs have already hardened?

Signals and data to watch. Adoption and interoperability status of agent-to-agent protocols (Model Context Protocol, Agent2Agent) across major platforms; whether dominant consumer AI assistants restrict which service agents they can transact with (the “bowling-shoe” agent pattern the Coasean-singularity paper identifies, where a platform-provided agent enjoys privileged integration but limited portability); market-share concentration in agent-transaction volume by platform, tracked with the same rigour applied to foundation-model market share today.

Institutional wiring and first moves. Competition authorities already running foundation-model investigations (CMA, FTC, European Commission) should open a parallel, lighter-touch monitoring work stream on agent-platform interoperability now, rather than waiting for a market-power complaint to trigger a fresh full investigation — the lesson of the Cambridge paper being that by the time a complaint-driven investigation would normally open, a market of this speed could already have tipped. A standing information request to the major platforms on agent-to-agent interoperability commitments, modelled on the FTC’s 6(b) authority already used for the cloud-AI partnership report, is a low-cost first step available within a single regulatory cycle.

Priority 3 — Redesign labour-market safety nets for agent-speed displacement, not human-generation-speed displacement

In short. The reallocation-friction evidence Report Two draws on — Autor, Dorn and Hanson’s “China Shock” finding that trade-displaced workers took a decade-plus to reallocate (NBER, Angle 05) — describes a world where the competing force was other human workers and firms adjusting at human speed. The Cambridge paper shows agent-vs-agent competition can concentrate a market within the time horizon of a stylised simulation. Safety-net design built for the first kind of disruption will be too slow for occupations exposed to the second.

Why it ranks here. This is where Report Two’s variable 4 (labour reallocation friction) meets this report’s central finding most directly. The ILO’s refined global index of occupational exposure to generative AI, the OECD’s 2023 Employment Outlook chapter on AI and the labour market, and McKinsey Global Institute’s occupational-transition modelling (all Angle 05) were built to estimate how large a share of work is exposed and how long reallocation typically takes for workers displaced by automation of the traditional, tool-mediated kind. None of that literature was built with agent- vs-agent competitive dynamics in mind — dynamics the Cambridge paper shows can compress the effective disruption timeline for agent-exposed occupations from the years those models assume down to a period closer to a single retraining cohort’s enrolment window.

The foresight questions and horizons. Near-term: which occupations combine high exposure on the ILO’s index with high susceptibility to the kind of agent-vs-agent gig-platform competition the Cambridge paper models — freelance and platform-mediated work is the most obvious overlap, since it is structurally closest to the AI-Work simulation’s own setup. Medium-term: does reallocation time for workers displaced from agent-exposed occupations actually compress relative to the China Shock and Job Displacement and Job Mobility literature’s decade-plus benchmarks (both NBER, Angle 05), or does institutional friction on the human side of the labour market (licensing, geographic immobility, skills-matching delay) keep human reallocation slow even as the disruption that triggers it accelerates — producing a widening gap between how fast displacement happens and how fast re-employment can follow?

Signals and data to watch. Platform-level data on gig and freelance market concentration and pricing (the same variables the Cambridge simulation tracks — win rate, market share concentration, normalised bid price) in real gig-economy platforms, not just the simulated one; unemployment- duration and reallocation-speed statistics disaggregated by occupational exposure to agent-mediated competition, not just to automation exposure broadly, since the two are not the same variable and the existing ILO and OECD indices do not yet distinguish them.

Institutional wiring and first moves. Employment and welfare ministries should commission a follow-on to the existing ILO and OECD occupational-exposure work that specifically cross-references exposure to agent-mediated platform competition, not generic automation exposure — the distinction this report has drawn throughout. Wage-insurance and rapid-retraining voucher schemes, the standard policy response to the China Shock literature’s reallocation-friction finding, should be piloted on a compressed timeline (months, not the multi-year rollout typical of retraining programmes built against a decade-plus reallocation assumption) in the occupations where ILO exposure and platform- competition exposure overlap most, treating speed of activation as the design variable that matters most, ahead of programme scale.

Priority 4 — Use the Coasean-singularity logic to actively re-open institutional design

In short. If transaction costs are genuinely collapsing toward zero for agent-mediated activity, the state has an active choice in what forms in that space — not merely a defensive posture toward whatever the private sector builds first.

Why it ranks here. Part One’s second section argued that the Coasean logic determining the boundary of the firm is now, for the first time since Coase wrote in 1937, a live variable rather than a structural constant. The Coasean-singularity paper’s own examples of where agent adoption concentrates first — real estate, job search, freelance hiring, investment — are all markets where private-sector agent platforms are already forming. Public-service delivery and benefits administration are structurally identical markets by the same logic (high-stakes interactions, information asymmetry, repeated need to re-establish context with an unfamiliar counterparty) and are currently almost entirely unaddressed by any private or public agent infrastructure.

The foresight questions and horizons. Near-term: could agent-to-agent design reduce the transaction cost of benefits administration and eligibility determination — the paperwork and re-explanation burden the Microsoft Research paper identifies as the core friction agents dissolve — in the same way it is beginning to reduce it in real estate and freelance hiring? Medium-term: should public-service delivery build its own “service agents,” in the Microsoft Research paper’s vocabulary, that citizens’ own assistant agents can transact with directly, and if so, on what identity and accountability infrastructure, given Hadfield and Koh’s finding that no such infrastructure currently exists for AI agents anywhere?

Signals and data to watch. Pilot deployments of agent-mediated public-service interfaces in any jurisdiction (an early, concrete exemplar worth tracking directly is the UK’s Government Office for Science foresight function and its published work on AI in public administration, alongside Singapore’s Centre for Strategic Futures and Policy Horizons Canada — the three institutions this library treats as the closest working analogues for state-level agentic foresight capacity); uptake and complaint rates on any such pilots, since trust and inspectability — flagged as the critical constraint by the Coasean-singularity paper’s own authors — will determine whether citizens allow their assistant agents to transact with a state service agent at all.

Institutional wiring and first moves. A digital-government or public-service-delivery unit should commission a scoping pilot — modest, one or two benefit programmes, eighteen months — for an agent-to-agent interface that lets citizens’ own AI assistants query eligibility and submit applications programmatically, built explicitly on the identity-and-liability groundwork Hadfield and Koh flag as missing (a registered, accountable principal behind every transacting agent, at minimum), rather than waiting for a private vendor to define the standard the state then has to adopt.

Priority 5 — Build sovereign, public-interest agent capability

In short. Compute and chip sovereignty is already a recognised geopolitical fault line in this library’s Angle 13. The agentic layer is a new dimension of the same asymmetry — and one a state can still shape, because the agent layer is younger and less entrenched than the compute layer beneath it.

Why it ranks here. Angle 13 of this library documents the compute-access asymmetry in detail: the CSIS analysis of the 2024 US chip export-control tightening, RAND’s account of the resulting AI Diffusion Framework tiering countries by compute access, CSET Georgetown’s “Silicon Twist” tracking of chips still reaching restricted end-users despite controls, and the IMF’s “Mind the Gap” finding that AI’s growth effect could be more than double in advanced economies relative to low-income ones. Every one of those asymmetries recurs, in a faster and less mature form, at the agent layer: a state or firm without access to frontier-grade agents is not just slower at deploying AI tools — per DeepMind’s Virtual Agent Economies finding on unequal-capability negotiation, it is structurally disadvantaged in the agent-to-agent transactions its citizens and firms increasingly depend on, since “the more capable AI assistants tend to be more successful and negotiate better deals for their users.”

The foresight questions and horizons. Near-term: does frontier agent capability remain concentrated in the same handful of jurisdictions (principally the US, with China as the other compute-tiered bloc under RAND’s diffusion framework) that already dominate foundation-model training, or does the lower compute intensity of agent orchestration (relative to frontier model training itself) allow a wider set of states to build competitive sovereign agent capability even without frontier-model-scale compute? Medium-term: does agent-layer capability become a second, compounding axis of the World Bank’s “Beyond the AI Divide” concern — the same countries structurally disadvantaged on compute access falling further behind specifically because their citizens and firms transact through lower-capability agents in a market where capability asymmetry, per DeepMind’s finding, directly determines negotiated outcomes.

Signals and data to watch. National or regional sovereign-cloud and public-interest-AI initiatives extending their remit explicitly to agent orchestration and deployment, not just model hosting; procurement policy for any public-sector agent deployment (Priority 4’s pilots, for instance) specifying capability floors, to avoid the state itself becoming the disadvantaged party in agent-to-agent transactions with better-resourced private counterparties.

Institutional wiring and first moves. A digital-sovereignty or industrial-strategy ministry should extend any existing sovereign-compute or public-AI programme’s mandate explicitly to cover agent orchestration capability, not only model access, within the current planning cycle — treating “can our public and SME sector field agents competitive enough to negotiate on equal terms with counterparties running frontier-grade agents” as a distinct capability question from “do we have enough compute,” since DeepMind’s evidence suggests the two do not move in lockstep.

Priority 6 — The agentic engine ENSI itself would deploy

In short. Consistent with ENSI’s standard foresight-engine structure, the agentic-economy monitoring function should itself be built as a small set of named agent archetypes, each with a narrow mandate, feeding human judgement rather than replacing it.

• Scanning agents, tracking agent-transaction volume, task-delegation rates and the automation- versus-augmentation split across available public and licensed data sources — the standing instrument Priority 1 calls for, run continuously rather than as a periodic study, modelled on Anthropic’s own O*NET-mapped, privacy-preserving classification methodology.

• Scenario-simulation agents, stress-testing Report One’s five futures against each new quarter of incoming data from the scanning agents — not to pick a winner prematurely, but to flag when incoming evidence starts to favour one future over the others clearly enough to justify a policy response, and equally to flag when a scenario previously treated as low-probability starts moving.

• Market-structure early-warning agents, applying the Cambridge paper’s own diagnostic variables — market-share concentration (Gini-style), win-rate distribution, normalised bid-price trends — to real agent-platform and agent-labour-market data as it becomes available, precisely because those variables moved from 0.24 to 0.70 within a stylised simulation’s own short horizon, and a live early-warning system needs to be watching before, not after, a real market shows the same trajectory.

• Translation and briefing agents, converting the scanning and simulation layers’ output into the kind of short, decision-facing briefing a minister, a board or a regulator’s leadership can act on within a single reading — the standing failure mode this playbook is designed to avoid is not lack of data, it is data that never reaches a decision-maker in time to matter.

Humans retain judgement and accountability throughout: agents surface signal, simulate scenarios and draft briefings; the decision about what to do with any of it — where to intervene, what to regulate, what to fund — stays with the accountable human institution the agents serve, following the same accountability logic this report has argued the state needs to establish for every other agent deployed in the wider economy.

Priority 7 — First twelve months

A state or large institution acting on this report between now and mid-2027 should, concretely:

• By Q4 2026: commission the agentic-activity index work stream (Priority 1) and approach at least two major model providers for aggregated usage-data terms, modelled explicitly on Anthropic’s published Economic Index methodology.

• By Q4 2026: task the competition authority already running a foundation-model investigation (CMA, FTC or European Commission, depending on jurisdiction) with opening a lighter-touch parallel monitoring work stream on agent-platform interoperability, using 6(b)-style information requests as the low-cost first instrument.

• By Q1 2027: commission the follow-on occupational-exposure study (Priority 3) cross-referencing ILO and OECD exposure indices against platform-mediated competitive-displacement risk, with results due within two quarters given the compressed timeline the Cambridge paper’s evidence implies is appropriate.

• By Q2 2027: scope the public-service agent-to-agent pilot (Priority 4) — one or two benefit programmes, built on explicit identity-and-liability groundwork rather than deferred to a vendor’s default terms.

• By Q2 2027: extend the mandate of any existing sovereign-compute or public-AI programme to cover agent orchestration capability explicitly (Priority 5).

• By Q3 2027: stand up the first two agentic-engine archetypes — scanning and market-structure early-warning (Priority 6) — as a standing function reporting quarterly, timed to precede the next full-cycle refresh of this report series.

Closing: the option-value argument

Every element of this playbook has value even if Report One’s most likely outcome — modest, uneven growth, unevenly distributed across sectors and geographies — turns out to be exactly what happens, and even if the more dramatic agentic dynamics described in Part One never scale beyond the particular markets (freelance platforms, real estate, personal-finance search) where agent adoption is currently concentrated. A live measurement instrument is worth building whether or not the automation-over-augmentation crossover Anthropic observed on Claude.ai turns out to generalise, because a state without one is flying blind on the single fastest-moving input to its own growth forecast regardless of which way that input moves. A competition-policy monitoring function on agent-platform concentration is worth having whether or not the Cambridge paper’s winner-take-all dynamics materialise outside a stylised simulation, because the cost of building the capacity to look is low and the cost of not having it, if the dynamic does materialise, is a market that has already tipped before regulators start their first investigation. Faster-activating labour-market safety nets are worth having whether or not agent-vs-agent competition ever displaces workers faster than the China Shock’s decade-plus benchmark, because a safety net built for a faster shock still works perfectly well for a slower one, while the reverse is not true. Sovereign agent capability is worth building whether or not the compute-sovereignty asymmetries in Angle 13 turn out to be the dominant axis of AI-driven inequality, because the alternative — discovering after the fact that citizens and firms are structurally disadvantaged in agent-mediated negotiation and having no domestic capability to respond — is not a position any state should choose to be in by default.

This is, in the end, the same argument Hadfield and Koh make in their own closing lines, and it is the right note to end this series on: “where we end up within this vast space of possibility is a design choice.” Reports One and Two describe the range of futures and the variables that determine which one materialises. This report’s contribution is narrower and more operational: the fastest- moving, least-institutionally-prepared-for variable in that whole system is not how quickly a task gets automated. It is whether the counterparty on the other side of a growing share of the world’s economic transactions is still, in any meaningful sense, accountable to a human being at all. That is a question this generation of policymakers gets to answer directly, while the institutions that will determine the answer are still being built — which is exactly the moment foresight capability is worth the most, and exactly the moment it is cheapest to build.

The forecast that hides the thing it should be measuring

Every institutional forecast of AI’s economic effect — Goldman Sachs’ 7%, PwC’s $15.7 trillion, the OECD’s 0.25–0.6 percentage points, the IMF’s country-by-country growth differentials — answers a narrower question than the one policymakers think they are asking. Each is, in its own methodology, a model of potential output: what the economy could produce if the productivity gain shows up and is spent. None of the headline numbers that dominate boardroom slides and finance-ministry briefings is a model of realized, demand-constrained GDP — the number that determines whether tax receipts rise, whether unemployment claims fall, and whether a government gets re-elected. That gap between potential and realized output is not a technicality. It is the single most consequential and most underpriced risk in the entire AI-and-growth literature, and this report exists to make it visible.

The reframe this report argues for is simple to state and easy to miss: the modal, mainstream forecast — modest, positive, uneven aggregate growth — is entirely compatible with regions, sectors and income deciles inside that aggregate experiencing outright contraction, for the same reason and at the same time. A national GDP print of +0.4% is arithmetically consistent with call-centre towns, back-office cities and mid-skill service corridors losing income for a decade, exactly as Pittsburgh, Greensboro and the furniture counties of North Carolina did during the “China Shock” — not because the aggregate model was wrong, but because aggregate models are built to net out precisely the geography and distribution that determines whether a shock feels like growth or like shrinkage to the people living through it. Aggregation is not measurement error. It is a modelling choice, and every one of the headline AI-growth forecasts in this library makes it.

This matters because the entities that most need an accurate scenario map — national treasuries, central banks, regional development agencies, sovereign wealth funds, multilateral lenders — are not making a bet on a single global GDP number. They are making a portfolio of decisions about fiscal buffers, retraining budgets, monetary stance, regional transfers and industrial policy, each of which depends on knowing not just whether AI grows the pie but whose slice moves and how fast the redistribution happens relative to the political cycle that has to absorb it. A state that plans against Goldman Sachs’ 7% uplift and gets the OECD’s 0.4 percentage points instead has made a forecasting error. A state that plans against any aggregate figure at all, while a fifth of its regions or a third of its labour force experiences the Korinek-Stiglitz demand-shrinkage mechanism described in Scenario 4 below, has made a category error — and it is the more dangerous of the two, because the national dashboard will keep reading “growth” the entire time.

The five-scenario framework that follows is deliberately not a spectrum from “big growth” to “big shrinkage” with the truth somewhere in the middle. It is four distinct causal stories — a productivity boom, a modest base case, a stagnation trap, and a demand-driven contraction — each anchored to a different published model with a different methodology and a different set of assumptions about diffusion speed, gain-sharing and constraint-bindingness, plus a fifth argument that cuts across all four: that the geographic and sectoral variance inside whichever aggregate number wins is where the real policy risk lives. Readers should leave this report able to name, for their own jurisdiction, which leading indicators would move them from one scenario to another — and understanding that the question “will AI grow or shrink the economy” is underspecified until you also ask “for whom, and measured how.”

One discipline runs through every section below: history’s base rate for how long general-purpose technologies take to resolve into measured productivity. Paul David’s canonical study, “The Dynamo and the Computer,” published in the American Economic Review, found that electrification took roughly forty years to show up in US productivity statistics after the technology existed — because factories had to be rebuilt around unit-drive motors rather than retrofitted around a single central power source, and that reorganisation, not the invention, was the rate-limiting step. Nicholas Crafts’ growth-accounting study for the LSE, “Steam as a General Purpose Technology,” pushes the base rate further: steam power took close to a century to lift UK aggregate productivity growth measurably. Every probability estimate in this report should be read against that discipline. A scenario “resolving” by 2030 or even 2035 would be fast by the standard of the only two general-purpose technologies we have full-century data on.

The five futures, at a glance

• The Productivity Boom (~15% likelihood). Fast diffusion, broadly shared gains, no binding physical constraint — the world Goldman Sachs and PwC model. Attractive, well-funded, and the least likely of the four core scenarios because it requires three separate things the historical and structural evidence argues against simultaneously.

• The Base Case — Modest, Uneven Growth (~45–50% likelihood). The mainstream institutional consensus: Acemoglu’s task-based ceiling, the OECD’s general-equilibrium range, the IMF and BIS’s cross-country findings. Growth that is real, arrives slowly, and is unevenly distributed by construction — the most probable single outcome and the one this report spends the most time defending.

• Stagnation / The Productivity Paradox Redux (~15–20% likelihood). Diffusion stalls the way it stalled for computers in the 1980s and 1990s; the AI capex boom outruns realized value the way the BIS’s contest-theory model predicts; Baumol’s cost disease caps what automation alone can deliver.

• Demand-Driven Shrinkage (~10–15% likelihood). The scenario readers most need explained because it is the least intuitive: potential output rises even as realized GDP falls, because the same productivity shock that looks expansionary in a supply-side model becomes contractionary once you close the model with a demand side that depends on wage income households don’t have.

• The Bifurcation (the report’s central reframe, not a fifth probability bucket). A “modest growth” aggregate (Scenario 2) can mechanically conceal regional and sectoral versions of Scenario 4 playing out underneath it — because the published aggregate models are not built to catch that. This is the sharpest, most underpriced risk in the whole literature, and it is why “which scenario wins nationally” is the wrong question for any policymaker to be asking alone.

Scenario 1 — The Productivity Boom (~15% likelihood)

The bull case is not a fringe position; it is argued by two of the most-cited institutions in applied macro forecasting, and its arithmetic is worth taking seriously before it is discounted. Goldman Sachs’ Joseph Briggs and Devesh Kodnani, in “The Potentially Large Effects of Artificial Intelligence on Economic Growth,” estimate that roughly two-thirds of current US and European jobs are exposed to some degree of AI automation, that generative AI could substitute for up to one-quarter of current work tasks, and that extrapolated globally this is equivalent to exposing the labour input of some 300 million full-time jobs to automation. Run through their model, that yields just under 1.5 percentage points of additional annual US labour-productivity growth over a ten-year period following widespread adoption, and an eventual 7% increase in global GDP — about $7 trillion in their 2023 pricing. PwC’s “Sizing the Prize” is, if anything, the more dramatic of the two headline forecasts in this library: $15.7 trillion of additional global GDP by 2030, equivalent to global output being up to 14% higher than it would otherwise be, split by PwC’s own accounting into productivity effects (businesses automating processes and augmenting labour, which the firm says account for over 55% of total GDP gains between 2017 and 2030) and a second, growing consumption-side effect as AI-enhanced products and personalisation drive additional demand — a channel PwC says will account for 58% of the GDP gain realized in 2030 specifically, i.e. the mix shifts from productivity-led to demand-led as the diffusion matures. PwC further disaggregates by geography and sector: China stands to see the largest proportional boost (up to 26% of GDP), North America the second largest (14%), and retail, financial services and healthcare the sectors with the greatest combined productivity-and-product-enhancement potential.

Why does ENSI Foresight Division treat this as the low-probability tail rather than the central case, despite the credibility of the institutions behind it? Three reasons, each traceable to a different angle of this library, and each a condition the Boom scenario requires to hold simultaneously — which is precisely what makes their joint probability low even when each is individually plausible.

• It requires diffusion speed that the historical record argues against. Paul David’s electricity study and Nicholas Crafts’ steam study — the two general-purpose technologies with a full-century data record — took 40 and roughly 100 years respectively to show up in aggregate productivity, because the complementary reorganisation of firms, not the invention itself, was the binding constraint. The Federal Reserve Board’s 2025 working paper “Generative AI at the Crossroads: Light Bulb, Dynamo, or Microscope?” tests generative AI explicitly against these historical diffusion analogues and finds the honest answer is not yet knowable — the paper’s title is itself an admission that AI could be any of the three, with very different growth implications.

• It requires broad gain-sharing that the market-structure evidence directly contradicts. Jan De Loecker and Jan Eeckhout’s NBER working paper “The Rise of Market Power and the Macroeconomic Implications” documents that average US markups rose from roughly 18% above marginal cost in 1980 to 67% today, concentrated in an increase in high-markup firms rather than a broad-based shift — and they link this rise causally to falling labour share, falling low-skill wages, and a slowdown in aggregate output. Autor, Dorn, Katz, Patterson and Van Reenen’s companion NBER paper, “The Fall of the Labor Share and the Rise of Superstar Firms,” finds the same winner-take-most dynamic in industry concentration data. If AI’s productivity gains flow disproportionately to the handful of firms that already control the compute, the cloud infrastructure and the frontier models — exactly the concentration the UK Competition and Markets Authority’s “AI Foundation Models” technical report and the FTC’s 6(b) study of cloud-AI partnerships were opened to investigate — then the Boom scenario’s assumption of broadly shared productivity gains breaks down at the first link in the chain.

• It requires no binding compute or energy constraint, which Angle 04’s evidence says is not a safe assumption. The IEA’s “Energy and AI” special report models data-centre electricity demand as a potential bottleneck on how fast AI compute can scale at all — grid capacity, transformer lead times and permitting are not software problems that can be diffused at ChatGPT’s adoption speed. The BIS’s “The AI Investment Race” (Phurichai Rungcharoenkitkul, July 2026) goes further and models the entire AI buildout as a winner-take-most contest in which competing firms rationally over-commit resources: calibrated to balance-sheet and deal data, the paper finds over-investment running at roughly 1.5 times the efficient level, rising to around 3 times where demand is less elastic — financed substantially through debt and circular equity ties between hyperscalers, model labs and chip suppliers, with a network-cascade risk if any one node disappoints on revenue.

None of this means the Boom scenario is impossible — Goldman and PwC’s methodologies are serious and their authors are not naïve about diffusion lags. It means the scenario requires fast diffusion, broad sharing and unconstrained capital deepening to all hold at once, against a historical base rate, a market-structure trend and a physical-infrastructure constraint that each independently argue the opposite. Compounding three low-conditional-probability requirements is why ENSI Foresight Division prices this scenario at roughly 15% — high enough to take seriously in capital allocation and infrastructure planning, low enough that a state should not build its ten-year fiscal plan on it arriving.

Scenario 2 — The Base Case: Modest, Uneven Growth (~45–50% likelihood)

This is the mainstream institutional consensus, and it deserves to be argued for on its own terms rather than treated as the residual “everyone else’s number.” Its intellectual anchor is Daron Acemoglu’s “The Simple Macroeconomics of AI,” prepared for Economic Policy and circulated as NBER Working Paper 32487. Acemoglu builds a task-based model in which AI’s macroeconomic effect is bounded by a version of Hulten’s theorem: aggregate productivity gains are given by the fraction of tasks AI actually touches multiplied by the average task-level cost saving. Using the best available exposure and productivity-improvement estimates, he finds the resulting TFP effect is “nontrivial but modest — no more than a 0.66% increase in total factor productivity over 10 years.” He then argues this may still be an overestimate, because current evidence is drawn disproportionately from easy-to-learn tasks, while much of AI’s future effect will have to come from hard-to-learn tasks with context-dependent judgment and no objective performance metric to train against — on that adjustment, his predicted 10-year TFP gain falls to under 0.53%. This is not a paper written to be contrarian for its own sake: Acemoglu explicitly engages Goldman’s 7% and McKinsey’s $17.1–25.6 trillion range in his introduction and argues the gap is explained by his more conservative estimate of which tasks are genuinely automatable versus merely exposed.

The OECD’s “Miracle or Myth? Assessing the Macroeconomic Productivity Gains from Artificial Intelligence” (OECD Artificial Intelligence Papers No. 29, November 2024) arrives independently at a strikingly similar order of magnitude through an entirely different method — a novel micro-to-macro multi-sector general-equilibrium model with input-output linkages, rather than Acemoglu’s task-exposure algebra. Its headline finding: annual aggregate TFP growth attributable to AI of 0.25–0.6 percentage points, equivalent to 0.4–0.9 percentage points of labour productivity growth over a ten-year horizon. That two independent methodologies — one a stylised task-based bound, one a full general-equilibrium simulation with sectoral input-output linkages — converge on the same order of magnitude, an order of magnitude roughly a fifth to a tenth the size of Goldman’s or PwC’s headline figures, is the single strongest piece of evidence in this library for treating the Base Case as the modal outcome rather than a competitor to the Boom scenario.

Ten to fifteen supporting reasons this is where ENSI Foresight Division places the largest probability mass:

• It is where two independent, methodologically distinct institutional estimates converge, as above — Acemoglu’s task-exposure bound and the OECD’s general-equilibrium simulation, arrived at without coordination, land in the same 0.3–0.9 percentage-point-of-productivity-growth range.

• It matches the observed pattern of real deployment, not just theory. The BIS’s “Artificial Intelligence and Growth in Advanced and Emerging Economies: Short-Run Impact,” a 56-economy, 16-industry cross-country study, finds a real but modest growth effect concentrated in advanced economies, consistent with a diffusion process still in its early innings rather than a step-change already realized.

• It is consistent with firm-level field evidence showing real but bounded gains. Erik Brynjolfsson, Danielle Li and Lindsey Raymond’s “Generative AI at Work” studies 5,179 customer-support agents given access to a generative AI assistant and finds a 14% average productivity gain — economically significant, globally scalable in principle, but a 14% task-level gain is a different order of magnitude from a Goldman-style productivity boom, and the paper finds the gain is concentrated among novice and low-skilled workers (34% improvement) with minimal effect on already-experienced staff — a distributional pattern, not a uniform lift.

• Software-development field experiments tell the same story of real, bounded, unevenly distributed gains. Microsoft Research’s three-firm RCT across 4,867 developers finds a 26% increase in task completion; the earlier GitHub Copilot field experiment finds developers completed a standardised task 55.8% faster. These are large individual-task effects that nonetheless net out, at the Acemoglu-style aggregate level, to a modest TFP contribution once weighted by the share of the economy actually composed of tasks this exposed.

• The historical diffusion base rate argues for “modest and slow” over “large and fast.” The Productivity J-Curve literature (NBER Working Paper 25148) formalises why GPT adoption should be expected to show up first as a dip in measured productivity — as firms spend on complementary intangible investment that national accounts do not capture as capital — before any acceleration appears, which is exactly consistent with a base case that looks unremarkable for years before it looks real.

• It is a growth effect real enough to matter for a finance ministry, wrong enough to disappoint an equity analyst pricing a Goldman-sized re-rating — which is itself diagnostic. Sustained modest TFP acceleration compounds: even Acemoglu’s more conservative 0.53% figure, sustained and extended past the initial ten-year window as diffusion continues per the historical base rate above, is not nothing over a multi-decade horizon.

• It is unevenly distributed by construction, not by exception — the IMF’s “The Global Impact of AI: Mind the Gap” (WP/25/76) finds the estimated growth impact in advanced economies could be more than double that in low-income economies once sectoral exposure, technological preparedness and data/technology access are fed into a multi-region dynamic general-equilibrium model, with AI-driven productivity gains concentrated in the non-tradable sector large enough to disrupt the traditional exchange-rate-adjustment mechanism (an inverse Balassa-Samuelson effect, in the paper’s own framing).

• It coexists with — rather than reverses — the concentration trend already documented in Angle 06. Nothing about a modest aggregate TFP gain requires that gain to be evenly shared; De Loecker and Eeckhout’s markup evidence and the superstar-firm literature describe a economy-wide trend already three decades in train, and the Base Case scenario simply assumes AI does not interrupt it, which is the more conservative and more defensible assumption than assuming AI reverses it.

• It is what “Miracle or Myth?” itself frames as the honest middle — the OECD paper’s own Figure 1, comparing predicted macro-level productivity gains across the published studies, is explicitly built to show how much the estimates vary depending on methodology, and situates its own general-equilibrium estimate as the more disciplined, assumption-transparent number against which the use-case-based (McKinsey) and task-exposure (Goldman) estimates should be read as upper bounds rather than central forecasts.

• It requires the fewest simultaneous strong assumptions. Unlike the Boom scenario, the Base Case does not require fast diffusion, broad gain-sharing and unconstrained capital deepening all at once — it only requires diffusion to proceed at something like the historical GPT pace, under the market structure we already observe, which is the lowest-assumption, highest-prior scenario of the four.

The Base Case is not a comfortable answer for anyone selling AI transformation at Goldman-scale multiples, nor for anyone hoping AI will single-handedly resolve a decade of weak productivity growth. It is, on the weight of two independently-derived institutional estimates and a wide base of firm-level field evidence, the most probable single outcome — which is exactly why the reframe in the closing section of this report matters: a Base Case aggregate can still conceal a Scenario 4 dynamic underneath it, region by region and sector by sector.

Scenario 3 — Stagnation / The Productivity Paradox Redux (~15–20% likelihood)

The stagnation case is not merely “the Base Case, but slower.” It is a structurally distinct claim: that diffusion stalls hard enough, or the investment boom decouples far enough from realized value, that AI’s net contribution to measured growth over the coming decade is close to zero — with a meaningful tail risk of an outright investment bust dragging growth briefly negative.

Erik Brynjolfsson, Daniel Rock and Chad Syverson’s NBER working paper “Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics” is the intellectual anchor here, and it is worth reading closely because it is not a pessimistic paper about AI’s ultimate potential — it is a paper about why potential and measured statistics can diverge for a long time. The authors open bluntly: “measured productivity growth has declined by half over the past decade, and real income has stagnated since the late 1990s for a majority of Americans” even as AI systems match or surpass human performance in more and more domains. They offer four explanations for the paradox — false hopes (the technology is simply less transformative than believed), mismeasurement (statistics fail to capture real gains), redistribution (private gains that are zero-sum at the aggregate level, e.g. one firm’s AI-driven market-share gain is another’s loss), and implementation lag — and conclude, after weighing the evidence, that implementation lags have likely been the biggest contributor to the paradox: the most impressive AI capabilities have not yet diffused widely, and like every prior general-purpose technology, their full effect will not be realized until waves of complementary innovation — organisational redesign, new skills, new business processes — catch up, a process the paper models as a form of unmeasured intangible capital investment.

The companion Productivity J-Curve paper formalises the mechanism precisely: measured productivity should be expected to dip before it rises, because firms are spending real resources on GPT-complementary intangible capital that standard national accounts do not capitalise, so the investment shows up as a cost with no offsetting asset on the books — a mechanically depressing effect on measured TFP that has nothing to do with AI’s true productive potential and everything to do with accounting convention. This is the same phenomenon the OECD’s own Annex A.2 addresses under the heading “Baumol’s growth disease” — a decomposition of how factor reallocation and relative-price changes can drag on aggregate productivity growth even while sector-level productivity is genuinely rising.

Ten supporting points for treating this as a real, not merely academic, near-term risk:

• The theoretical ceiling on automation-driven growth is a Baumol constraint, not a technology constraint. Philippe Aghion, Benjamin F. Jones and Charles I. Jones’s NBER paper “Artificial Intelligence and Economic Growth” models AI as the latest wave of a 200-year automation process and finds that growth may ultimately be constrained not by what AI is good at but by what remains essential and yet hard to improve — the irreplaceable-task bottleneck that is the AI-era restatement of Baumol’s cost disease. If a large share of aggregate value-added sits in tasks AI cannot yet touch (skilled trades, elder care, complex judgment, physical world interaction), automating everything else at zero cost still caps aggregate growth at a rate set by the stubborn residual.

• The AI capex boom is already showing the structural signature of a contest-theory overbuild, not an efficient one. The BIS’s “AI Investment Race” model — over-investment calibrated at 1.5–3x the efficient level, financed through debt and circular equity ties between hyperscalers, chipmakers and model labs, with cascading network exposure — is not a hypothetical; it is a description of financing structures already observed in balance-sheet and deal data as of the paper’s July 2026 publication.

• NBER’s own investment-flow analysis is used to bound, not inflate, plausible cumulative GDP effects. “What Investment Data Implies about the AI Transition” (NBER Working Paper 35290) explicitly uses observed AI infrastructure investment to discipline what cumulative GDP effect is consistent with the capital actually being deployed — a methodological choice that, by construction, produces more conservative implied growth than a use-case-based forecast like McKinsey’s or PwC’s.

• Energy is a hard, not soft, constraint on how fast the capex can even be converted into usable compute. The IEA’s “Energy and AI” special report and RAND’s “AI’s Power Requirements Under Exponential Growth” both model grid capacity — not chip supply — as a binding near-term bottleneck: transformers, transmission permitting and generation buildout all move on multi-year timelines that do not compress just because model capability is advancing on a compute-doubling cycle Epoch AI measures at roughly every six months in its “Rising Costs of Training Frontier AI Models.”

• The historical base rate says stalling for a decade or more inside a multi-decade diffusion curve is the norm, not the exception. Paul David’s electricity study documents exactly this pattern — a long trough of disappointing productivity numbers between invention and full economic realisation — and Bresnahan and Trajtenberg’s original NBER theoretical treatment of general-purpose technologies, “Engines of Growth?”, explains why: GPTs require complementary innovation in every using-sector before their productivity potential is unlocked, and that complementary innovation is itself constrained by organisational and human-capital adjustment costs that do not move at software speed.

• Mismeasurement can cut either way, and the false-hopes channel cannot be dismissed. Brynjolfsson, Rock and Syverson’s own taxonomy keeps “false hopes” — that generative AI’s real economic contribution is simply smaller than current enthusiasm implies, once hype-driven capital allocation is stripped out — as one of the four live explanations, and do not claim to have falsified it, only to have found implementation lag the larger of the four in their reading of the evidence.

• The redistribution channel means some of the measured stagnation could be real even as reported corporate AI adoption rises. If AI’s gains are substantially redistributive — one firm’s win is a rival’s loss, netting to roughly zero at the aggregate level — then rising firm-level AI adoption metrics (the kind reported in surveys like the WEF’s “Future of Jobs Report 2025”) are compatible with flat aggregate productivity, exactly the paradox Brynjolfsson, Rock and Syverson set out to explain.

• A financial-stability tail risk is explicit in the library, not merely implied. The BIS investment-race paper’s network analysis shows that stress in one AI-buildout firm could cascade to others through chains of financial exposure — meaning the downside of this scenario is not simply “slower growth than hoped” but includes a discrete probability of an investment-bust event that subtracts from measured growth for a period, echoing the dot-com capex cycle but with debt and circular financing structures the BIS paper flags as a distinct amplifying mechanism.

• The OECD’s own scenario modelling treats slow-diffusion paths as a first-order sensitivity, not an edge case. “Miracle or Myth?” explicitly models the sensitivity of its central 0.25–0.6 percentage-point estimate to alternative assumptions about sectoral gains, demand response and reallocation frictions (its Figure 11), meaning the paper’s own central estimate already sits inside a distribution whose lower tail overlaps meaningfully with a stagnation outcome.

• The gap between AI capability benchmarks and AI economic diffusion is now a measured, tracked quantity, and it is widening, not narrowing — Stanford HAI’s “2026 AI Index Report” is the annual instrument the library uses to track compute, investment and model-performance trends, and the persistence of a capability-diffusion gap year over year is itself evidence against the fast-diffusion assumption the Boom scenario requires and in favour of the multi-year lag this scenario describes.

ENSI Foresight Division prices Stagnation at roughly 15–20% — meaningfully more likely than the Boom scenario, because it requires only one thing to go wrong (diffusion friction, or an investment overbuild correcting) rather than three things to go right, but still a minority outcome because the weight of the firm-level field evidence in Angle 15 (the call-centre, developer and robot-adoption studies) shows AI is already delivering some measurable productivity gain at the point of deployment — the stagnation case therefore requires that gain to fail to aggregate up, not that it fails to exist at the micro level, which is a narrower and less probable claim than pure technological disappointment would be.

Scenario 4 — Demand-Driven Shrinkage (~10–15% likelihood)

This is the scenario a policymaker is least likely to have internalised, because it inverts the intuition built by every supply-side headline number in this report. Goldman’s 7%, PwC’s $15.7 trillion, Acemoglu’s 0.66%, the OECD’s 0.25–0.6 percentage points — every one of these is, at root, a statement about potential output: what the economy is capable of producing once AI’s productivity effect is fully realized. None of them is a full macroeconomic model that closes with a demand side and asks whether anyone has the income to buy what the supply side now can produce. Scenario 4 is what happens when you ask that second question and the answer is no.

The mechanism, stated precisely, is this: automation raises output-per-worker (labour productivity rises; potential GDP rises) but simultaneously reduces the wage income of displaced or wage-suppressed workers. If the Acemoglu-Restrepo reinstatement effect — new tasks created for displaced labour — and government redistribution together fail to replace that lost labour income fast enough, aggregate demand falls. Because realized GDP in the short-to-medium run is demand-constrained, not supply-constrained, realized GDP can fall even as output-per-worker (and therefore measured “productivity”) rises. This is not a contradiction; it is what happens when a supply-side productivity shock is fed into a demand-side model rather than assumed away. The same underlying technological event that reads as unambiguous “growth” in Goldman’s task-exposure framework or Acemoglu’s Hulten’s-theorem bound can read as contraction the moment the model is closed with a Keynesian or New Keynesian demand block instead of assumed at full employment and full income pass-through.

The BIS’s “The Impact of Artificial Intelligence on Output and Inflation” (Aldasoro, Doerr, Gambacorta and Rees, BIS Working Paper 1179) makes the pivot condition explicit and precise. Modelling AI as a permanent, sector-differentiated productivity shock inside a macroeconomic multi-sector model, the authors find that the entire character of the response — expansionary or contractionary, inflationary or disinflationary — hinges on whether households and firms anticipate the future productivity gain. In their own words: “if they do not anticipate higher future productivity, AI adoption is initially disinflationary” and only gradually becomes moderately inflationary through general-equilibrium demand effects; “in contrast, when households and firms anticipate higher future productivity, inflation rises immediately” as spending is pulled forward against expected future income. This is precisely the fork between Scenario 2 (anticipated, gain-sharing, demand recovers) and Scenario 4 (unanticipated or undelivered, demand lags, output can contract before it expands) — and which fork an economy lands on is an empirical, observable question about expectations formation and income distribution, not a technological one.

Anton Korinek and Joseph Stiglitz’s NBER working paper “Artificial Intelligence and Its Implications for Income Distribution and Unemployment” builds the taxonomy this scenario needs. They identify the two main channels through which AI-driven inequality operates — the surplus captured by innovators, and the redistribution that occurs through factor-price changes (i.e., wages falling as labour’s bargaining position weakens relative to capital) — and they formalise two distinct channels of technological unemployment: an efficiency-wage channel, where firms find they can pay less and still retain adequate labour once AI substitutes for worker leverage, and a transitional-phenomenon channel, where unemployment is a temporary but potentially long-lasting feature of the adjustment path rather than a new steady state. Crucially, Korinek and Stiglitz show that even in the theoretically favourable case where AI could deliver a Pareto improvement — where redistribution could in principle make everyone better off — that outcome requires non-distortionary taxation and active redistribution to actually occur. The Pareto-improvement result is conditional on policy action, not automatic; absent it, their model produces exactly the winners-and-losers dynamic that defines this scenario.

This is not a purely theoretical construction bolted onto abstract macro models. It has a real, close empirical precedent, and the library’s demand-side and market-structure angles converge on why it is plausible rather than merely logically possible:

• Secular stagnation was already underway before AI, which lowers the bar for a demand-side shock to bite. Gauti Eggertsson, Neil Mehrotra and Lawrence Summers’ “Secular Stagnation in the Open Economy” formalises how a persistent demand shortfall can trap an economy in a low-growth equilibrium; Łukasz Rachel and Lawrence Summers’ “On Secular Stagnation in the Industrialized World” documents that neutral real interest rates across the industrialized bloc had already fallen by at least 300 basis points over the preceding generation on standard estimates — and by “as much as 700 basis points since the 1970s” on their private-sector-neutral-rate measure — reflecting chronic weakness in demand relative to desired saving, largely independent of AI. An economy already leaning toward demand insufficiency is an economy with less slack to absorb a labour-income shock without realized output actually falling.

• The mechanism by which gains fail to reach households as wages is not hypothetical — it is the same market-power evidence documented in Scenario 1’s rebuttal. De Loecker and Eeckhout’s finding that average US markups have risen from roughly 18% to 67% above marginal cost since 1980, alongside their explicit finding that this coincides with a falling labour share, falling low-skill wages, and slower aggregate output growth, is the empirical channel through which “AI raises productivity but wages don’t rise proportionally” stops being a theoretical possibility and becomes a continuation of an already-observed thirty-year trend.

• The scale of exposure is large enough for the demand channel to matter in aggregate, not just at the margin. The ILO’s “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” and Goldman’s own two-thirds-of-jobs exposure estimate both describe an exposed labour share large enough that even a partial failure of the reinstatement effect would move aggregate labour income by a materially larger amount than any plausible near-term productivity offset.

• Reallocation, empirically, takes far longer than models with instantaneous labour-market clearing assume — the China Shock is the closest real-world analogue in scale and mechanism. Autor, Dorn and Hanson’s “The China Shock” is explicit: “adjustment in local labor markets is remarkably slow, with wages and labor-force participation rates remaining depressed and unemployment rates remaining elevated for at least a full decade” after the shock begins, and “offsetting employment gains in other industries... have yet to materialize” even a decade-plus after the initial exposure. If an AI-driven demand/employment shock displaces workers at even a fraction of the geographic concentration the China Shock exhibited, the Autor-Dorn-Hanson evidence says the reinstatement effect Acemoglu’s Base Case model relies on cannot be assumed to arrive within any policy-relevant timeframe.

• The Autor-Salomons finding that own-industry productivity gains reduce own-industry employment is the microfoundation for why aggregate demand, not just aggregate supply, needs to be modelled explicitly. Their ECB conference paper “Does Productivity Growth Threaten Employment?” finds industry-level productivity gains do reduce employment within the gaining industry, even where economy-wide effects vary — exactly the granular mechanism Scenario 4 needs to be more than a stylised macro story.

• The direction of the demand response is not fixed by the technology — it is fixed by policy and expectations, which is itself the actionable finding. The BIS output-and-inflation paper’s central result — that the entire sign of the near-term response depends on anticipation — means Scenario 4 is not a prediction that shrinkage will happen; it is a precise statement of the condition under which it happens (unanticipated or undelivered gains, weak redistribution, concentrated market structure), which is exactly the kind of condition a government can monitor and intervene against.

• The IMF’s own productivity-headwinds work shows how weak demand and balance-sheet damage produce lasting hysteresis, not a one-off dip. “Gone with the Headwinds: Global Productivity” documents how demand weakness and balance-sheet damage can scar potential output itself over time — meaning a Scenario-4-style demand shortfall is not necessarily self-correcting even after the initial adjustment period, if it damages investment and human capital accumulation along the way.

• This scenario does not require universal wage stagnation — it requires the reinstatement-and-redistribution race to lose to the displacement-and-concentration race in enough of the economy, for long enough, to show up in realized output, which given the evidence above (thirty years of rising markups, a decade-plus of trade-shock evidence on reallocation speed, and an explicit BIS finding that the outcome hinges on anticipated income) is a materially more plausible bar to clear than “AI causes a literal, economy-wide, permanent contraction.”

ENSI Foresight Division prices outright, sustained national demand-driven shrinkage at roughly 10–15% — a real, non-trivial probability, but a minority one, because governments retain fiscal and monetary tools (the BIS output-and-inflation paper’s own point about anticipation being manageable through policy communication and demand management) that make a full national contraction an avoidable rather than an inevitable outcome. But — and this is the hinge on which the rest of this report turns — pricing the national aggregate version of this scenario at 10–15% radically understates how often the underlying mechanism fires at sub-national scale, which is exactly the gap the Bifurcation reframe below is built to close.

One paragraph on where the agentic layer fits — and why it belongs to Report 3

Everything above treats AI as a tool that humans deploy inside existing firms, tasks and labour markets. A qualitatively different mechanism is emerging alongside it: AI agents that transact, negotiate and compete as autonomous economic actors rather than simply assisting a human worker. Gillian Hadfield and Andrew Koh’s “An Economy of AI Agents,” prepared for the NBER Handbook on the Economics of Transformative AI, surveys how agents capable of planning and executing complex tasks over long time horizons with little human oversight could reorganise markets and institutions in ways the task-based automation literature was not built to anticipate. The University of Cambridge’s “When AI Agents Compete for Jobs” offers an early empirical glimpse of the dynamics this could produce: in a simulated AI labour market, agents equipped with metacognition, competitive awareness and long-horizon strategic planning achieve roughly 1.5 times the market share of agents using standard prompting approaches, and the simulation exhibits price-deflation and market-concentration dynamics as capable agents out-compete less capable ones at machine speed. Whichever of the four scenarios above turns out to dominate, an agentic layer would not create a fifth outcome so much as accelerate the arrival of whichever one is already winning — compressing the Boom scenario’s diffusion timeline, deepening the Stagnation scenario’s overbuild, or intensifying the Demand-Shrinkage scenario’s concentration mechanics. This is deliberately not developed further here: it is the entire subject of Report 3 in this series, and flagging the hand-off rather than pre-empting it is the correct discipline for Report 1.

The Bifurcation — the reframe, stated directly

Here is the argument this report has been building toward: the sharpest, most underpriced risk in the AI-growth debate is not which of the four scenarios above the global economy lands in. It is that the question is being asked at the wrong level of aggregation.

Every headline forecast surveyed in this report — Goldman’s 7%, PwC’s $15.7 trillion, Acemoglu’s 0.66%, the OECD’s 0.25–0.6 percentage points, the IMF’s advanced-versus-low-income growth differential — is a national or global aggregate, built from a supply-side, task-exposure or general-equilibrium model that does not fully endogenize the demand-side feedback loop described in Scenario 4. This is not a criticism of the methodology on its own terms: Acemoglu’s paper is explicitly a supply-side bound; the OECD’s is a productivity-focused general-equilibrium exercise; Goldman’s is a task-exposure extrapolation. Each is doing what it says it does. The problem is what happens when a policymaker reads the aggregate number as if it were a description of what will happen everywhere inside the country, uniformly, when the models were never built to make that claim in the first place.

Consider the arithmetic directly. The IMF’s “Mind the Gap” finds AI’s estimated growth effect could be more than double in advanced versus low-income economies — that is already a statement that the “0.4pp of global TFP” aggregate is a blend of very different sub-experiences. Push the same disaggregation down one more level, from country to region and sector within a country, and the same logic applies with more force, not less, because sub-national labour markets are less mobile and less fiscally insulated than national economies are from each other. The BIS’s cross-country study explicitly finds AI’s growth effect concentrated in advanced economies at the country level; nothing in the model architecture prevents an equivalent concentration within an advanced economy, between its AI-exposed financial and professional-services metros and its AI-exposed-but-poorly-reinstated manufacturing and back-office regions. A national GDP figure that averages a compute-cluster metro seeing a Goldman-style boom with three deindustrialising service corridors seeing a Korinek-Stiglitz-style demand contraction can report as an entirely unremarkable Base Case number — Acemoglu’s 0.5%, the OECD’s 0.4 percentage points — while concealing exactly the outcome this report’s readers most need to see coming.

The China Shock is the closest real-world precedent in the library for how long-lasting and geographically concentrated this kind of shock can be, and it is worth stating why it is the right analogue rather than a loose metaphor. Trade exposure to China, like AI exposure, was not evenly distributed — it hit specific industries (furniture, textiles, electronics assembly) concentrated in specific commuting zones. Autor, Dorn and Hanson’s evidence that those commuting zones saw depressed wages, depressed labour-force participation and elevated unemployment for at least a full decade, with offsetting employment gains in other industries that “have yet to materialize” even years later, is precisely the pattern a geographically- and sectorally-concentrated AI exposure shock would be expected to produce — and precisely the pattern that a national GDP or national employment aggregate is constructed to average away. The US economy as a whole absorbed the China Shock without a recession attributable to trade alone; the counties and workers directly exposed experienced something that looked, to them, indistinguishable from Scenario 4, for a decade, inside an aggregate reading Scenario 2.

This is why ENSI Foresight Division treats the Bifurcation not as a fifth item in the probability table but as a modifier that sits on top of all four scenarios simultaneously. Whichever aggregate outcome wins — even the Boom scenario — the distributional mechanics documented in the market-structure and labour-reallocation angles of this library (rising markups, superstar-firm concentration, decade-plus reallocation lags, more-than-double advanced-versus-low-income growth differentials) do not disappear inside a good aggregate number. They are, if anything, easier to miss inside a good number, because a rising national GDP figure is exactly the kind of signal that makes a finance ministry stop asking where the gains are landing.

What to watch: leading indicators for policymakers

A state cannot wait for the ten-year TFP number to resolve before deciding how to prepare. The following indicators, each drawn from a mechanism documented above, are observable well before the headline growth print confirms which scenario is materializing — and, read together, they are built specifically to catch the Bifurcation the aggregate number would miss.

• Track AI-exposed employment and wage data at the regional and occupational level, not only the national level. The ILO’s occupational-exposure index and the OECD’s sectoral-exposure mapping (used to build the “Miracle or Myth?” general-equilibrium model) already exist as instruments; the discipline is to run them quarterly against regional labour-force data the way the Autor-Dorn-Hanson China Shock methodology tracked commuting-zone-level trade exposure, rather than waiting for a national unemployment print to move.

• Watch the gap between AI capital expenditure and realized enterprise value creation. The BIS investment-race paper’s over-investment ratio (1.5–3x efficient level) and the NBER’s investment-flow-implied GDP bound are both designed to be recomputed as new balance-sheet and deal data arrive; a widening gap between announced AI capex and delivered productivity gain is the earliest observable signal of the Stagnation scenario’s bust tail, and it is visible years before an aggregate recession would be.

• Monitor markup and labour-share trends by sector, specifically in AI-adjacent industries. If cloud, compute and frontier-model markups continue the De Loecker-Eeckhout trajectory (18% to 67% above marginal cost since 1980, concentrated in a rising share of high-markup firms) rather than compressing as AI capability commoditises, that is a direct, sector-level signal that Scenario 4’s rent-capture mechanism — not broad gain-sharing — is the one operating.

• Track household and firm expectations of future income directly, not just realized income. The BIS output-and-inflation paper’s central finding — that the entire sign of the demand response depends on anticipation — means survey-based expectations data (of the kind central banks already collect for inflation expectations) can be repurposed as an early-warning instrument for whether an economy is on the Scenario 2 or Scenario 4 path, months or years before realized GDP data would reveal it.

• Watch reallocation speed against the China Shock’s decade-long benchmark, not against an assumed frictionless model. If displaced-worker reemployment rates, wage recovery timelines and labour-force-participation trends in AI-exposed regions are tracking faster than the China Shock’s documented decade-plus adjustment period, that is meaningful evidence the reinstatement effect is working in something closer to real time — a genuinely different, more optimistic signal than if they are tracking the same or slower.

• Watch neutral real interest rates and secular-stagnation indicators as a macro backdrop, not a separate story. Rachel and Summers’ finding of a 300–700 basis point decline in neutral real rates over the preceding half-century describes the demand-side terrain AI’s productivity shock is landing on; a further decline, rather than stabilisation or reversal, would indicate the economy has even less capacity to absorb a labour-income shock without realized output actually falling — raising, not lowering, the probability mass on Scenario 4.

• Track energy and grid-capacity data as a hard constraint on diffusion speed, independent of model-capability announcements. The IEA and RAND data on data-centre power demand and grid buildout timelines are a genuinely exogenous check on how fast the Boom or Base Case scenarios can physically proceed, regardless of how fast frontier-model capability itself advances — a useful discipline against extrapolating growth forecasts purely from compute-scaling trends.

• Watch for the emergence of agentic, autonomous AI transaction volume as an accelerant signal, not a fifth scenario. Early indicators — the kind Anthropic’s own Economic Index Report already tracks in real Claude usage data on task delegation and autonomy — are worth monitoring as a leading sign of how fast whichever scenario is already winning will arrive, which is the hand-off this report makes explicitly to Report 3.

The bifurcation, restated

The honest answer to “will AI grow or shrink the economy” is: probably grow it, modestly, unevenly, over a longer timeline than the enthusiasts assume and a shorter one than the historical base rate would predict for a typical general-purpose technology — while concealing, inside that same modest positive aggregate, regional and sectoral episodes that will look and feel like outright shrinkage to the people living through them, for years, exactly as the China Shock did. The four scenarios in this report are not a menu from which the world will choose one. The Base Case is where ENSI Foresight Division places the largest single share of probability, at roughly 45–50%, anchored to the convergence of Acemoglu’s task-based bound and the OECD’s independent general-equilibrium estimate. But the number that should worry a policymaker most is not the probability attached to any of the four rows in the summary table above — it is the near-certainty that whichever row wins, the Bifurcation modifier is already operating underneath it, and that the standard forecasting toolkit used by Goldman, McKinsey, the OECD and the IMF is not built to show a finance minister where. A state that only watches its national GDP print will find out it was living through Scenario 4 at the regional level roughly a decade after the fact — which is exactly how long it took to fully recognise the China Shock for what it was. The leading indicators above exist so that this report’s readers do not have to wait that long again.