If current penetration is roughly 15–20% meaningful augmentation, the next four years follow a fairly predictable arc:
- 2026: invisible margin lift, hiring freezes, quiet attrition
- 2027: earnings visibly show AI leverage in leading sectors
- 2027–2028: role compression becomes obvious; restructuring discontinuities drive earnings beats
- 2028+: pricing compression and industry restructuring
The shape of that curve is broadly consensus now. What’s contested is the timing, the threshold, and the second-order effects, and most of the published models are getting all three subtly wrong in the same direction.
I’ve argued previously that the structural endpoint of this transition is an L-shaped economy rather than a K… a drop followed by managed stagnation for the majority while a smaller tier accelerates away.
This piece is the timing companion to that structural argument: why the financial market signal arrives later than the labor market signal, and what actually determines whether we land in a productivity boom or in the managed-stagnation regime.
Where We Actually Are
My estimate is that roughly 20% of company activities are currently being AI-augmented in any meaningful sense. At the bleeding edge, some organizations may be pushing 50–60%, though many such claims likely overstate true operational substitution versus superficial assistance.
The aggregate adoption numbers are genuinely large. McKinsey’s latest survey reports that 88% of organizations now use AI in at least one business function, with 72% using generative AI specifically; up from 33% a year earlier.1 But the same data shows the gap between adoption and scaled value capture remains wide: only 23% of respondents say their organization is scaling an agentic AI system in even one business function.
That gap is the whole story.
The aggregate number obscures sharp bimodality: contact centers, content operations, and parts of software engineering are probably past their internal tipping points already. Legal, accounting, and routine analytics sit mid-curve. Healthcare delivery, skilled trades, and regulated industries lag by years. The restructuring shock hits in sector waves, not as a synchronized inflection, which means the tradable signal will appear in sector rotations well before it surfaces in aggregate margin data.
“AI is no longer a futuristic concept, and adoption numbers are genuinely staggering. But value remains rare.” — McKinsey, State of AI in 2025
The Mechanics Are Step-Shaped, Not Linear
Even modest adoption matters, but the relationship between workflow automation and margin expansion isn’t smooth. Early adoption captures high-volume, low-judgment work where per-unit savings are real but team-size thresholds aren’t crossed; you get a few percent of efficiency gain without ever actually reducing headcount. The large margin gains come at discrete points where a layer of management can be eliminated entirely, or a function collapsed, or an outsourcing contract terminated rather than renewed.
Roughly, every additional 5% of meaningful workflow automation translates into a 3–5% operating margin improvement on average, but with high variance clustered around those org-restructuring discontinuities. Coupled with headcount reductions, process redesign, and reduced vendor spend, that impact could plausibly double over a 24-month window in companies that hit the discontinuities early.
This matters for how earnings get modeled. Consensus models using smooth extrapolation will systematically underestimate the variance, and the largest earnings beats of 2027 will come from companies that hit org-restructuring discontinuities the quarter before they report. The discontinuity is invisible in guidance until it lands.
The earnings discontinuity is invisible in guidance until the quarter it lands.
The Real Tipping Point Isn’t a Number
The market is currently fixated on aggregate penetration thresholds, the idea that somewhere around 28–34% average enterprise adoption, AI stops being experimental optimization and becomes visible operating infrastructure. That framing is directionally useful but wrong in its mechanism.
The actual trigger isn’t a penetration percentage; it’s visible peer outperformance.
Goldman Sachs has begun framing this transition as a phase shift: a move from “Phase 2 of the AI trade (infrastructure) to Phase 3 and 4 (execution and productivity)”, with the firm projecting AI-driven productivity gains lifting S&P 500 earnings per share by 0.4% in 2026 and 1.5% in 20272. Those are conservative numbers, and they’re conservative on purpose, the firm explicitly notes that a widening “productivity gap” between top-tier companies and laggards could drive consolidation and M&A activity.
That productivity gap is the trigger mechanism.
Once 2027 earnings clearly attribute margin expansion to AI in any leading sector, the boardroom calculus shifts immediately. Every CFO faces a peer-comparison problem that’s fiduciary rather than strategic, and the response curve is nonlinear.
The mechanism that compresses the timeline is what I'd call the Leaker effect… the first firm in any given sector to successfully execute a large-scale agent-driven restructuring doesn't just outperform on margin; it triggers boardroom panic across every direct competitor.
The moment one credible peer demonstrates that the playbook works, every other CFO in the sector has their board asking why they aren't doing the same, and the question stops being "should we" and becomes "how fast." That's how an adoption curve that should take three years collapses into one. The leak compresses the window from quarters into months, and laggards get tagged immediately rather than after a few cycles of comparative underperformance.
The penetration curve plausibly compresses from 20%→40% in 18 months rather than 36, pulling all downstream timelines forward by roughly half a year.
The Labor Market Damage Has Already Started
The negative second-order effects — role compression, wage pressure, reduced hiring, organizational flattening — are already visible in leading indicators, just not yet in aggregate earnings reports.
A Stanford study found employment among early-career workers in AI-exposed occupations has dropped 16% since the launch of ChatGPT in 20223. Big tech firms saw entry-level hiring fall to just 7% of new hires in 2024… a 25% drop from 2023 and more than 50% below pre-pandemic levels. Revelio Labs data shows entry-level postings down 35% since January 2023.
“AI is doing what interns and new grads used to do. Now you can hire one experienced worker, equip them with AI tooling, and they can produce the work of multiple people.” — SignalFire research4
Anthropic’s own CEO has been blunter than most: Dario Amodei has warned that AI could eliminate 50% of entry-level white-collar jobs within five years, potentially pushing unemployment to 10–20%5. Whether that ceiling is right or not, the directional pattern is now visible in BLS data: white-collar unemployment has risen each year since 2023 while blue-collar unemployment has stayed flat or declined.6
The earnings reports will lag this by a year or more.
By the time analysts model margin expansion from AI, the labor market damage is already underway, and the lag itself is shortening while the underlying acceleration steepens, which narrows the window for any meaningful policy response.
The Demand-Side Recycle Is the Whole Game
The piece most macro models under-weight is what happens on the demand side. Supply-side margin expansion via cost reduction is well-modeled; every sell-side desk has a version of it. But pricing compression in 2028+ recycles consumer surplus, returning some of that captured value to the still-employed via real wage gains.
The L-shape thesis I’ve developed elsewhere depends on the displaced not recapturing meaningfully, that the “treadmill at the top is what makes the flatline at the bottom permanent,” because intelligence as a production layer compounds with the capital, data, and networks already in place to absorb it. Empirically the cohort-level data supports that. But at the macro level, the sequencing of pricing compression versus role compression determines which regime we actually land in.
If pricing compression lags role compression by 18–24 months, you get a demand air-pocket. If they arrive synchronously, you get something closer to aggregate-neutral but distributionally catastrophic.
A synchronous outcome looks something like the late-1990s productivity surge: prices fall, real wages rise for those still employed, aggregate demand holds up even as labor share of income shifts. A lagged outcome looks like a deflationary bust; costs collapse first, prices stay sticky, and the demand air-pocket pulls everything down before the consumer surplus can be returned. The L turns into something uglier.
This is where the policy response window matters.
The Federal Reserve doesn’t have tools well-suited to managing a productivity-driven deflationary shock layered on top of a labor-displacement shock. Fiscal policy could in principle address the distributional consequences, but the political feasibility window closes once the earnings beats start arriving and the equity market is celebrating margin expansion.
What This Means
The investable thesis is straightforward enough… long sector winners in Phase 3, short the laggards Goldman has already named, “zombie” balance sheets and high-labor-intensity sectors with low automation potential7.
The defining characteristic isn't financial structure per se; it's the intersection of high labor intensity and low automation potential. Companies whose cost base is heavily human and whose workflows resist current-generation AI tooling — certain traditional retail formats, legacy transportation sub-sectors, parts of regulated healthcare delivery, low-margin business services with high compliance friction — will watch their margin gap to peers widen quarter after quarter without an obvious path to close it.
That's the short. The macro thesis is harder, because it depends on a sequencing question that hasn’t been decided yet.
What’s clear is that the inflection arrives faster than the consensus timeline suggests, that the labor market signal is already running ahead of the earnings signal, and that the question of which macroeconomic regime we land in — productivity boom or managed stagnation — isn’t decided in earnings reports at all.
It’s decided in the gap between when costs fall and when prices follow. That gap is being set right now, in the architectural decisions companies are making about how to absorb the productivity gains and whether to pass them through.
The earnings story will be loud. The labor story will be louder. The pricing story will be the one that actually matters, and it’ll be the last one anyone is watching.
The Window You’re Still Inside
The same clock applies personally. Every adjustment that runs on the policy timeline — career repositioning, business model restructuring, capital reallocation, geographic optionality, the deliberate acquisition of leverage rather than wages — closes at the same moment the earnings beats arrive and the equity market starts celebrating margin expansion.
The beats are the signal that the window has already closed for the people who needed it most.
Anyone reading a piece like this is, by definition, still inside that window. The question isn’t whether the adjustment is worth making. It’s whether it gets made while there’s still optionality on the table, or after, when every move remaining is a reactive one, and reactive moves under deflationary labor conditions are the moves that don’t compound.
The 18-month window between now and the earnings inflection isn’t a forecast.
It’s a runway.