The Displacement Cascade: What Happens If AI Empties the Office

A note on dates: this framework was developed in February 2026, and the labor, credit, and confidence data in it are deliberately preserved with their February 2026 as-of dates rather than refreshed, because the point of publishing it now is to show the framework as it stood and to test it against what has happened since. Developments since February appear in their own dated section, and everything in that section comes from our own previously published notes.

The Question the Framework Asks

The catalyst for the February work was a widely covered prediction by Andrew Yang that AI would eliminate millions of white-collar jobs by late 2027, alongside a McKinsey Global Institute update from November 2025 estimating that 57% of current work hours are potentially automatable with existing technology, per McKinsey's published work. We did not adopt the aggressive timeline as fact. We treated it as a scenario worth mapping in full, because in our reading the market in early 2026 was pricing only the most obvious consequences (AI infrastructure up, staffing and office landlords down) while the deeper consequences of a genuine displacement wave had barely been examined. Estimates commonly place the U.S. white-collar workforce near 70 million people; the framework asks what happens, in order, if a meaningful fraction of them lose their income within a compressed window.

The February 2026 Evidence Base

The framework rested on conditions that were already observable in February, not on projections. All figures below are as of the framework's writing in late February 2026, per the sources indicated.

Two things were true about that table in February and remain true about it now. First, none of it proves AI causation; a cooling labor market and a stretched consumer have many possible causes, and we return to that in the bear case. Second, whatever the cause, the household balance sheet described above has very little cushion. Our view then and now: a consumer entering a displacement shock at a 3.6% savings rate with delinquencies already at levels the New York Fed data last showed around the financial crisis does not absorb the shock; it transmits it. That transmission path is what the framework maps.

The Framework: Six Orders of Effects

We call the framework the Displacement Cascade. Its core claim is that mass white-collar displacement, if it occurs, propagates through the economy in a predictable order, and that the later orders are where the least-examined consequences sit. In February 2026, in our reading, the first order was consensus, the second was partially priced, and orders three through six were largely absent from mainstream research discussion.

First and Second Orders: The Consensus Layer

The obvious trades were already crowded when we wrote the framework, and we said so internally. Every workflow handed to an AI agent consumes compute, networking, and platform capacity, which is why companies like NVIDIA (which market-share estimates give a dominant position in AI accelerators), Microsoft, and Broadcom sit at the foundation of any displacement scenario. On the other side, staffing firms place exactly the categories most exposed to automation: by February 2026, Robert Half had fallen roughly 65% from its 52-week high and ManpowerGroup roughly 55%, per market data at the time, which told us the market was already pricing the first order. Office landlords faced the same logic compounded: remote work had already cut reported utilization 30-40%, and displacement removes the workers entirely. And at the consumer end, discount and value retailers (Walmart, Costco, Dollar General) are the standard beneficiaries when households trade down. None of this was differentiated by February 2026. The framework exists for what comes after.

Third Order, Part One: The Per-Seat Software Question

Business software is overwhelmingly priced per user per month. AI agents do not need dashboards; they work through APIs. If an enterprise replaces a large share of its sales, marketing, or support staff with agents, the per-seat licenses attached to those employees are cancelled in the same budget cycle. The framework's claim, stated as opinion: headcount-linked software pricing is structurally exposed to displacement, and the value migrates toward software priced on consumption, data volume, or outcomes, which earns more as AI activity rises regardless of how many humans remain. Companies like Palantir and Snowflake price on that consumption side; per-seat marketing and CRM platforms sit on the exposed side, and by February 2026 the market had begun to agree, with HubSpot down roughly 71% from its 2021 high, per market data at the time.

One piece of intellectual history we publish deliberately: our own February work moved Salesforce from the beneficiary side of this framework (where our earlier draft had placed it as an automation enabler) to the exposed side, on per-seat logic. What happened next is discussed in the dated section below, because it is, in our view, the most direct public test any leg of this framework has received, and it did not resolve the way the simple version of the argument predicted.

Third Order, Part Two: The Healthcare Payer Shift

The market treats healthcare as recession-proof. Under a displacement scenario, we think that assumption inverts, because U.S. healthcare margins are subsidized by employer-sponsored insurance. Commercial plans pay hospitals a large premium over government rates, commonly estimated near 2.5 times Medicare in published hospital price studies. The displaced white-collar worker (median income around $82,000 per year, per BLS-derived estimates) is precisely the highest-margin covered life. If millions of those workers fall from commercial coverage onto Medicaid or subsidized exchanges, margin compresses across insurers with commercial-heavy books and hospital operators with commercial-heavy payer mix, while Medicaid-focused managed care organizations absorb the migrating lives. The February evidence that the mix was already straining: Elevance had guided its 2026 earnings down 16% with benefit expense ratios reported near 90%, per its disclosures. This is the leg of the framework we believe remains least priced, and also the leg most exposed to policy intervention, as the bear case discusses.

Fourth Order: The Insolvency Infrastructure

A household sector with a 3.6% savings rate and $1.28 trillion of card debt, per the February data above, does not default gracefully at scale. It defaults through machinery: debt purchasers who buy distressed receivables, auction platforms that process repossessed vehicles, and self-storage operators who warehouse the possessions of downsizing families. In a displacement scenario, that machinery sees volume it has not seen since the financial crisis. This is an uncomfortable category to write about, and we think that discomfort is exactly why it goes unexamined. We would rather name it plainly: businesses like Encore Capital, Copart, and Public Storage are the processors of the cascade's fourth order, and their volumes are a direct real-time gauge of whether the cascade is happening.

Fifth Order: The Escapism Economy

The standard macro assumption is that unemployed consumers simply stop spending. The behavioral evidence from prior dislocations suggests something darker and more specific: people stripped of professional identity and given 60 or more empty hours a week migrate toward low-cost, high-duration escapism. A subscription game or a betting app offers hundreds of hours of engagement and a visible ladder to climb for the price of a single restaurant meal. Sports betting and iGaming platforms such as DraftKings sit on the most accessible version of that vector, and live entertainment sits on its premium mirror image: Live Nation reported 159 million annual attendees with operating income up 52% year over year in its fiscal 2025 results, per company disclosures, monetizing an experience that cannot be automated. We want to be explicit that this section is a description of predictable behavior under income shock, not an endorsement of it, and the public-health costs of a gambling-heavy version of this order would be real.

Sixth Order: The Physical-World Reallocation

The last order is the reversal of a 40-year flow of human capital into credentials. Our opinion, stated bluntly: digital reskilling is a value trap in a world where AI performs digital work, and the displaced 45-year-old manager with severance is likelier to buy a plumbing or HVAC business from a retiring owner than to learn prompt engineering. The market had already rendered a partial verdict on the reskilling narrative by February 2026: 2U had gone through bankruptcy and Chegg had lost more than 95% of its peak value, per market data at the time. The winners of this order are unglamorous: equipment suppliers to the trades (United Rentals reported EBITDA margins near 46%, per company filings), franchise systems, and the small-business acquisition ecosystem. This order arrives last, moves slowest, and, we suspect, lasts longest.

What Has Happened Since February

This note is published roughly six months after the framework was written. Everything in this section is drawn from our own published notes of August 30 and 31, 2026, with the original dates and sources preserved, and no new outside claims are introduced here.

The seat-compression leg went through a full public cycle. In the first quarter of 2026, the fear at the center of our third order was priced violently: the IGV software index fell 24%, its worst quarter since the fourth quarter of 2008 per index data, with roughly $2 trillion of software market value erased in about thirty days, per press aggregations. The market, in other words, briefly agreed with the strongest version of the per-seat argument. It then partially reversed, because the disclosed numbers cut the other way: Salesforce reported Agentforce annual recurring revenue of $800 million, up 169% year over year, in its February 25, 2026 results; Palantir reported first-quarter revenue up 85% year over year on May 4, 2026; and Snowflake's May 27, 2026 report of accelerating consumption revenue was followed by what press reports described as the stock's largest single-session gain on record. A Deutsche Bank survey reported through press coverage, which we flag as secondary, could not identify a software company expecting a negative 2026 revenue effect from AI.

What we take from that. Both halves matter. The agent revenue inflection is, in our reading, direct evidence that enterprises are buying displacement: agent products are being purchased at scale precisely to do work people previously did, which supports the cascade's premise. At the same time, the company our February framework moved to the exposed side, Salesforce, turned out to be among the fastest builders of disclosed agent revenue, which argues against the simplest version of the per-seat bear argument and for a more careful claim: the risk concentrates in undifferentiated per-seat tools, while platforms that convert agents into their own revenue line can cross the transition. We publish the original call and the complication together, because the complication is the finding.

The scale of AI activity kept compounding. As our August 30 note recorded, Google disclosed at I/O 2026 that it processes 3.2 quadrillion tokens per month, roughly seven times the prior year, and combined hyperscaler capital spending was running near $725 billion for 2026 with street estimates above $1 trillion for 2027, per company guidance and sell-side aggregations. None of that proves job displacement, but it is the physical signature of automated work being done in rapidly compounding volume, and the cascade requires exactly that input.

What has not yet shown up. We have published no evidence since February of the fourth, fifth, or sixth orders accelerating, and we will not manufacture any here. The framework itself anticipated this: the internal version flagged a 6-18 month lag between initial job losses and the second-order and later effects. Six months in, that lag window is still open, which means the later orders remain a forecast, not an observation, and we hold them accordingly.

The Bear Case, Steelmanned

The strongest objections to this framework, stated as forcefully as we can state them:

The causation gap. The February evidence base shows a stressed labor market and a stretched consumer; it does not show AI causing either. A JPMorgan estimate cited in our framework put only about 55,000 of roughly 1.2 million announced 2025 job cuts as explicitly AI-linked. Rate policy, post-pandemic normalization, and ordinary cyclical cooling could explain most of the February table, in which case this framework is an elaborate narrative draped over a garden-variety slowdown.

The timeline is the whole ballgame. The aggressive displacement window the framework maps (drawn from Yang's public prediction) has not been validated by attributed data. If displacement stretches over five to ten years instead of two, the economy absorbs it the way it has absorbed every prior automation wave, the later orders never concentrate enough to be visible, and the cascade dissolves into ordinary churn.

Job creation could outrun destruction. The World Economic Forum has projected 170 million new jobs created by 2030 against 92 million displaced, per its published work. Every prior automation panic, from the spreadsheet to the ATM, ended with more employment in the affected sectors, not less. The burden of proof sits with anyone claiming this time differs, and we acknowledge that we carry that burden.

Policy blunts the cascade. Basic-income pilots were already launching as we wrote (South Korea in early 2025, and Ireland moving its Basic Income for the Arts scheme toward permanence, per press reports at the time). Extended unemployment benefits, retraining subsidies, or expanded health-coverage subsidies would directly interrupt the first, third, and fourth orders. The healthcare payer leg in particular can be neutralized by a single act of Congress.

The six-month evidence cuts against the cleanest version. The one leg of this framework that received a real public test, per-seat software compression, resolved with more nuance than the framework's original form predicted, as we documented above. It is fair to ask whether the other five orders would survive contact with reality any better.

Second-order timing risk. Even if the thesis is right, the framework's own internal work flagged a 6-18 month lag before the later orders manifest. A thesis that is right eventually and wrong for a year is, for most practical purposes, wrong for a year.

What Would Prove Us Wrong

We publish break conditions on every framework we maintain. For the Displacement Cascade, the conditions are qualitative and observable:

1. The unemployment crossover reversing. If white-collar unemployment falls back below blue-collar and stays there across several quarters of BLS data, the labor-market signature the framework rests on was cyclical noise, and the premise fails.

2. AI attribution staying trivial. If, as attribution methods improve, explicitly AI-linked job cuts remain a low single-digit share of total announced cuts through 2027, the displacement wave is not arriving on any timeline this framework contemplates.

3. Household credit stabilizing. If serious delinquency rates in the New York Fed data plateau or decline from their early-2026 levels while savings rates recover, the transmission mechanism the cascade depends on is not loaded, whatever happens to jobs.

4. Seat counts holding at the platforms. If disclosed retention and seat metrics at the large per-seat software platforms show no compression through fiscal 2027, the third order's software leg was wrong in both its original and its refined form.

5. Payer mix holding. If commercial covered lives and insurer benefit-expense ratios remain stable through 2027, the healthcare shift is not happening, whether because displacement stalled or because policy absorbed it. Either way that leg breaks.

6. New-role absorption showing up. If job creation in AI-adjacent categories visibly absorbs displaced workers at scale, per BLS category data, the WEF's projection was right and ours was wrong, and we will say so in those words.

As of this writing, none of the six is in evidence, but neither is decisive confirmation of the cascade's later orders. That symmetry is the honest state of the thesis, and we revisit it as the data arrives.

Our View, Stated as Opinion

Three Douglas Research believes the market is examining the first order of AI labor displacement and largely ignoring orders three through six, and that the later orders, if displacement materializes on anything close to an aggressive timeline, contain the least-priced consequences: a healthcare payer shock, a repricing of headcount-linked software, an insolvency volume cycle, an escapism consumption shift, and a capital reallocation toward the physical trades. Those are opinions, formed in February 2026, published here with their original as-of dates, and falsifiable by the break conditions above. The six months since have strengthened the premise (agent adoption and AI work volume are compounding in disclosed numbers) while complicating the cleanest expression of it (the per-seat software leg resolved with more nuance than our original form predicted, and we own that in the text above). The framework's weakest link remains causation and timing, and we have printed the bear case at full strength for that reason. Our company-level work sits behind the membership wall.