From Doom Loop to Abundance Path

Issue 01 sets the position the series marks to the tape from here on. Citrini Research's "2028 Global Intelligence Crisis" argues that AI agents substitute for white-collar work fast enough to break consumption and tip the economy into a doom loop. We take it seriously enough to test it, and the collapse it describes needs about five fragile assumptions to hold at once, while the brakes on it are already visible in the data through mid-2026.

Read the companion deck (PDF)→

THE CLAIM

Since this first issue we hold and mark to the tape a two-part view, constructive on where AI ends up and cautious about the path it takes to get there. The terminal technology is deflationary the way the steam engine, electrification, and the computer were deflationary, lowering the cost of a scarce input until output and real incomes rise. The route there runs through dislocation that is real, uneven, and worth planning for. Citrini Research's "2028 Global Intelligence Crisis" is the strongest bear case, and it argues that AI agents substitute for white-collar labor fast enough to break consumption and feed a self-reinforcing collapse. We do not dismiss it.

The doom loop is a conjunction of about five conditions that have to hold at once, and each brake on it is an ordinary one. Capability can jump in a single release, while diffusion still grinds through cost, complements, trust, and regulation. Collapse needs recursive adoption, extreme substitution, cheap compute, savings that never reallocate, and institutions that fail to adapt, all together, so dropping any one leaves a transition instead of a meltdown.

The data through mid-2026 reads as transition. Software-development job postings have rebounded about 11% year on year on Indeed's index after their 2023 to 2025 decline, on Citadel Securities' reading, where runaway substitution would show continued collapse (Exhibit 2). The occupations most exposed to AI grew employment 10.4% and real wages 3.9% from May 2019 to May 2025, against 3.8% and 2.6% for everyone else, on our reading of the federal wage survey (Exhibit 3). US business applications have run near 1.5 to 1.7 times the pre-2020 norm of about 3.4 million a year every year since 2021, the entrepreneurial pressure valve that opens when the cost of starting a firm falls (Exhibit 4).

For an operator the posture that follows is preparation. The doom narrative is already shaping decisions through hiring freezes, panic transformations, and frozen capital budgets, and the wrong posture is costly whichever way the path runs. We would build the plan that holds up in either case.

01 · THE BEAR CASE

The doom loop in its strongest form

Citrini Research's "2028 Global Intelligence Crisis" runs in four steps that feed back on each other, and the scenario deserves its best statement before we answer it, since dismissing a strawman would prove nothing. AI agents substitute for a large share of white-collar work; the resulting layoffs weaken consumption, which raises the pressure to automate further; cost-cutting then feeds on itself and spreads across sectors; and markets, pricing a structural demand shock where an ordinary downturn would clear, set off the deleveraging cascades that turn a contraction into a collapse. Each round of layoffs erodes the demand that justified the previous round of automation, so the mechanism compounds on itself, and it can be dangerous even if it develops slowly.

We steelman Citrini's memo without forecasting from it, because the history of general-purpose technology records painful reallocations that resolved well in aggregate and slowly for the people in the path, with employment rarely stable through the transition itself. The honest question is which assumptions have to hold for displacement to compound into the spiral the memo describes, given that AI plainly does displace particular tasks and every prior transition still dissipated into reallocation. The analytical work is in naming those assumptions, because a scenario that needs five things to be true at once is far less likely than a scenario that needs one.

02 · THE MECHANISM

Collapse is a conjunction of five fragile assumptions

The doom loop rests on five conditions, and each is contestable on today's evidence. First, recursive adoption, where capability gains translate almost immediately into deployment and diffusion friction stays low everywhere instead of concentrating in a few digital sectors. Second, extreme substitution, where most cognitive tasks become cheaper to do with machines than with people, across industries, in the same window. Third, compute that scales cheaply, where no binding energy, capital, or supply-chain constraint slows the rate at which substitution can be rolled out. Fourth, savings that do not reallocate, where the profits from automation are hoarded, no new firms or products or categories of demand emerge, and competition among capital owners stops. Fifth, institutions that fail to adapt, where policy and organizational change arrive too late everywhere with no offsetting response.

Exhibit 1 shows why the economics of each lever cuts against the conjunction. Diffusion is governed by the cost of complements and the build-out of trust and regulation, which is why adoption follows an S-curve and not a step (Exhibit 1). Substitution is bounded by the price of compute and the scarcity of the inputs that surround it. Reallocation is the base case of price theory, since a fall in the cost of a productive input sends the surplus to be spent or invested somewhere else, and the next round of demand and hiring forms wherever it lands. The doom loop lives in the corner of the distribution where every brake fails together, and through mid-2026 the brakes are mostly holding. The genuine novelty in the opposing case is breadth, the claim that this substitution is broad and simultaneous in a way earlier automation never was and that cognition is the general input, which we grant, though breadth still has to clear cost, complements, and diffusion, and on the evidence it has not yet.

03 · THE BRAKES

Forward labor signals are repricing, and the spiral is absent

Demand for the workers most directly in the path of runaway substitution is the clearest place to see it arrive, and that demand is not in free-fall. Software-development postings on Indeed's index fell hard from their 2022 zero-rate peak through early 2025 and then turned up, rebounding about 11% year on year on Citadel Securities' reading of the series (Exhibit 2). A rebound is consistent with repricing and retooling, with demand shifting toward higher-value, AI-assisted builders; runaway substitution would show continued decline. This does not prove that no disruption is under way, only that the specific claim of an already-present spiral fails to appear in the forward signal that would show it first. The level of the index is approximate between published anchors and the year-on-year figure is Citadel's reading of Indeed and not a series we compute, so we hold to the shape and not the decimals.

Occupational Employment and Wage Statistics data points the same way and lets us compute it ourselves. Taking the occupations Pew classifies as most exposed to AI and tracking them through the Occupational Employment and Wage Statistics, high-exposure groups grew employment 10.4% from May 2019 to May 2025 and raised real mean wages 3.9%, against 3.8% and 2.6% for all other occupations (Exhibit 3, Second Order calculations). The most exposed work grew faster and paid better, which is what an augmentation phase looks like: tools raise output per worker before they replace whole categories. One choice drives the result and we flag it, since excluding office and administrative support, the largest and most clerical of the exposed groups, leaves the remaining high-exposure occupations growing employment 25.2% while their real wages slip 2.3%, so the wage premium is concentrated in the professional groups. This is one window on published data, evidence against broad labor destruction in that window, and it does not prove that AI cannot displace.

US business applications, running near 5.6 million in 2025, are the third brake the doom loop leaves out, and they answer what becomes of the savings. A fall in the cost of cognition lowers the cost of starting and running a firm with it, and the historical response to a cheaper input is entry. US business applications have run between about 1.5 and 1.7 times the pre-2020 norm of roughly 3.4 million a year in every year since 2021, holding near 5.6 million in 2025 (Exhibit 4, Census Business Formation Statistics, our annual totals). This pressure valve is the one the collapse scenario assumes away, since disruption converts into new categories of demand and hiring as the surplus from automation is reinvested. The same logic turns inward for our own firm, because cheaper cognition lets a services business spin up new offers faster and at lower cost than before, which makes entrepreneurship an internal capability as well as a macro statistic.

SECOND-ORDER EFFECT · ONE SHOCK, TWO ENDINGS

The first-order reading of cheap cognition is the doom loop, where AI commoditizes thinking, incomes fall faster than prices, output rises while demand collapses, and credit stress amplifies the shock into Ghost GDP, a statistical economy producing more while the people in it can buy less. The mechanism is coherent, and we do not dispute that the first move, the repricing of routine coordination and reporting, is real and already under way.

The second-order reading starts from the same commoditization, and the difference is whether the surplus reallocates. Cheaper services raise household purchasing power even where wages are flat, businesses that redeploy the savings into expansion and new markets create new categories of demand and work, and every prior general-purpose technology, steam, electricity, and computing, raised output and real incomes through that channel. Repricing can be bullish because the surplus shifts from intermediaries and overhead toward consumers and investment, and that is the abundance path the series is named for.

WHAT WE'D DO · THE OPERATOR'S PLAYBOOK

Six moves make an operator win in either world, and each applies the brakes above inside a firm. First, be the governance layer by productizing AI-enabled reporting, the weekly and quarterly business reviews and the KPI narratives, with human sign-off and safe-by-default patterns for permissions, personal data, and an audit trail, because buyers keep paying for governance and sign-off even when the underlying model is a commodity.

Second, bank the saved hours into analysis, redeploying the time automation frees into root-cause work, scenarios, and stakeholder alignment. Third, launch agent-assisted offers, standing up intake and triage, roll-ups, and exception detection as services, which is the entrepreneurship brake applied inside the firm.

Fourth, pick two workflows now and automate them relentlessly, since the reporting pipeline and intake and triage carry the clearest return, and experiment liberally elsewhere while committing on those two. Fifth, write the AI delivery standard once, the security, quality assurance, citation and audit-trail, and human-sign-off rules, so every engagement ships to the same standard. Sixth, measure what AI does not supply, domain trust and operating rhythm, tracked through cycle-time-to-insight, rework rates, and executive confidence.

WATCHLIST · WHAT WOULD CHANGE OUR MIND

Four signals would move the view, and one falsifier would break it. AI-at-work intensity, a sharp upward inflection in daily use and in the share of tasks run autonomously against the wide-but-shallow usage of today, would move the diffusion brake, so we watch intensity and not whether people have tried the tools. White-collar unemployment that rises and stays elevated for 12 to 18 months or more, against a series of one-off cuts, would be the labor brake failing, and the duration is the signal.

The character of deflation is the third signal: service prices falling on weak demand mean something different from prices falling on productivity gains, even though both read the same in the index. Reallocation is the fourth, where business formation that accelerates and small-team output that rises, with savings absorbed into capital spending and research, is the abundance signal and its absence is the warning. The claim that the brakes hold and the path is a transition would break on a sustained, broad-based decline in high-exposure employment alongside a sharp upward inflection in AI-at-work intensity, with business formation rolling over. We are watching all four, and we will mark the call against the tape as the intensity data develops.

SCORECARD · CONSENSUS VS THE TAPE

Three consensus sentences look different against the tape through mid-2026. The claim that AI agents will gut white-collar employment within a couple of years does not appear in the forward signal yet, where software postings have rebounded about 11% year on year and the most-exposed occupations grew employment and real wages through May 2025, and we mark it against the tape as the intensity data develops. The claim that the macro choice is between boom and collapse is a false binary, since the base case is a transition of repricing plus reallocation that disrupts without a meltdown, and collapse is the corner where five brakes fail together. The claim that cheaper cognition is straightforwardly deflationary and therefore bearish is half right, because it is deflationary, and whether that is bearish turns on whether the surplus reallocates, which the entry data says it is doing so far.

NEXT ISSUE · THE WRONG DENOMINATOR

Cheap compute is one of the assumptions the doom loop needs, and the first brake to arrive turned out to be the price of cognition itself. Next week the series takes apart the measure everyone now quotes, since tokens measure a lab's revenue rather than a buyer's value, and the wrong denominator makes the AI economy look both cheaper and more fragile than it is.

SELECTED SOURCES

Citrini Research, "2028 Global Intelligence Crisis" (scenario memo), steelmanned and not endorsed. Citadel Securities on the software-postings rebound, reading Indeed Hiring Lab's US job-postings index (raw.githubusercontent.com/hiring-lab/job_postings_tracker). BLS Occupational Employment and Wage Statistics, national, May 2019 and May 2025; CPI-U deflator via FRED; AI-exposure groups per Pew Research Center, "Which U.S. workers are more exposed to AI on their jobs" (Jul 26, 2023); Second Order calculations. US Census Bureau, Business Formation Statistics, total business applications (series BABATOTALSAUS via FRED), Second Order annual totals. Bill Gurley, on the TBPN podcast, on curiosity as a compounding asset.

Estimates and interpolations are flagged on each exhibit. Exhibit 1 is an illustrative schematic and not a measured series. The software-postings level is approximate between published anchors and the +11% year-on-year figure is Citadel's reading of Indeed and not a Razor series. The exposure analysis is a descriptive cut of public wage data that makes no causal claim, and it is sensitive to the office-support group as flagged. Second Order is an independent research briefing, provided for discussion purposes only and not investment, legal, tax, or accounting advice. Copyright 2026.