The Timing Valley
The real story on jobs one year after “The Last Normal Year”.
A year ago we published a projection most people still found hard to believe. The consensus put meaningful AI-related job displacement three to five years out. The data, read through the causal chain, said the window would open in Q1 2026 and that the first visible effects would arrive faster and more quietly than the classic layoff story suggested. Our forecast was closer to six months.
That projection held. The window opened on schedule. The mechanisms turned out quieter and more structural than we mapped. And the most important thing the original model under-weighted is now the center of the story: the two arcs—reduction of traditional jobs and creation of durable new ones—are both real, but they run on different clocks. The gap between them is the Timing Valley.
This piece does three things. It accounts for what the model got right, what it missed or under-weighted, and where the numbers still mislead. It offers a fresh projection for the path ahead. And it names the practical levers that can still compress the valley before it deepens.
What held
The original timeline was roughly right. Visible displacement signals appeared in Q1 2026. Administrative and routine cognitive roles in early-adopting organizations moved first. Creative and analytical functions followed. The causal chain—capital intensity, procurement, executive language, hiring freezes, then payroll and participation effects—activated ahead of the base case.
The compression against the prior 3–5 year consensus was the focal point. Previous technology transitions took years from infrastructure to employment impact. AI compressed the cycle. That gap between consensus expectation and signal velocity remains the most useful finding from last August.
We also correctly identified that formal layoffs would not be the dominant mechanism. Attrition, non-backfill, and role reclassification would do more of the work. The data since then has confirmed it. A large share of displaced workers never appear in unemployment claims. Severance packages, savings, stigma, and optimism create a buffer. People exit the measured labor force rather than register as unemployed. The official rate therefore understates the structural shift.
What the model under-weighted
Three things emerged over time as we explored the numbers.
First, the bottom-up path. Workers have been automating their own workflows faster than most organizations formally redesign roles. The job is a bundle of discrete sequences. When enough of those sequences get absorbed by tools, the container thins even if no one issues a layoff notice. This evaporation process runs on a personal clock, not a corporate planning cycle, and it compounds.
Second, intermediate process capture. Work that looks like it has simply moved offshore often follows a different sequence: the process gets documented, exceptions get catalogued, and the standardized version later gets automated. Medical-records review is one clean example. The domestic job does not disappear in a single announcement; it migrates, stabilizes, and then vanishes from the cost structure. These pathways leave almost no trace in conventional labor statistics. We treat this as a plausible and observable mechanism that deserves systematic tracking, not as settled macro fact.
Third, the growing mismatch between what headline BLS series measure and the transition now underway. The unemployment rate counts people who are both without work and still actively looking. Once someone stops looking—whether because of severance runway, freelance income, caregiving, disability reclassification, or simple discouragement—they leave the labor force. Absolute numbers of people not in the labor force have reached record levels. Participation has declined even after accounting for demographics. Repeated downward revisions to earlier strong prints are consistent with a survey that systematically over-reads strength when the underlying mechanisms are quiet. The numbers are not lying. They are answering a narrower question than most readers assume.
None of this invalidates the original model. It enriches it. The direction and the approximate timing were correct. The texture of the transition is more subtle, more distributed, and harder to see in the aggregate data than a classic displacement wave.
What the numbers can and cannot see
A cleaner way to say it: each major series answers a different question.
The unemployment rate tracks active job-seekers inside the labor force. It is silent on people who have stopped looking.
The employment-population ratio and labor-force participation rate capture broader attachment. Absolute NILF levels and prime-age cohort cuts by education and occupation tell a sharper story than the headline rate.
JOLTS, payroll, and job-board data pick up flows—postings, hires, separations, hours—before stocks fully adjust.
Firm-level indicators (adoption language in filings, revenue or profit per employee trajectories, contractor use) often move earlier than the national aggregates.
Official statistics aren’t useless. But if you’re looking for the transition, it’s currently more visible in micro flows, firm panels, and task-level changes than in the lagging aggregates most commentary still treats as the full picture.
The Timing Valley
Both sides of the public argument are true. AI reduces demand for traditional cognitive roles. AI will also create new roles. Timing between the two is the key.
Reduction in jobs is running ahead. Production deployment of multi-agent systems in large enterprises has moved from pilot to scale with unusual speed. (Definitions are essential here: we mean systems that own multi-step workflows with measurable volume, error budgets, and fallback, not copilots or experiments.) Revenue and profit per employee at AI-native and heavy-adopting firms already sit multiples above traditional benchmarks in many reported cases; the divergence is suggestive and directionally important even after we adjust for sector mix and capital intensity. Entry-level pipelines in exposed occupations remain under pressure or are being “seniorized”—the remaining junior roles demand judgment and context that used to sit higher in the hierarchy. These are current structural signals, not future risks.
Creation of new work runs on a slower institutional clock. New durable roles in agent oversight, complex exception handling, energy and compute infrastructure, and the institutional forms that will organize post-wage work do not appear at the same velocity. Educational systems, corporate redesign cycles, and physical build-out all move more slowly than software deployment.
The result is a valley. Traditional containers compress first. New containers that can absorb the displaced labor at scale arrive later. The depth and length of that valley aren’t predetermined. They’re functions of how deliberately we build the bridges.
The path ahead
Anticipated outcome. Through late 2027 into mid-2028, the primary pattern in cognitive employment continues to be further compression of traditional job containers. This registers as elevated NILF pressure and softer participation among prime-age cognitive cohorts, continued weakness or seniorization in entry-level pipelines, and rising output-per-employee metrics at adopting organizations. Net-new role creation scales more slowly. The Timing Valley has opened. Its depth is still partly controllable.
Consensus gap. Most current commentary still treats the labor impact as muted, largely augmentation, or roughly balanced on the same clock. That view rests on inherited timelines and on aggregate statistics that lag the micro mechanisms. The gap between that consensus and the current signal set is itself the finding.
Primary causal chain.
Enabling conditions (capability, cost, willingness) are largely complete.
Production-scale multi-agent deployment across knowledge workflows is accelerating.
Workflow compression produces non-backfill or headcount reduction in affected functions.
Structural labor-market effects become clearer in the micro data.
Macro recognition and narrative lag follow.
Scaled net-new role creation arrives last and remains the bottleneck.
Parallel paths stay active: individual evaporation, intermediate process capture, and anticipatory corporate positioning.
Institutional clocks. Deployment and compression move on market and corporate clocks (quarters). Creation of durable new roles and the physical infrastructure that supports them move on slower educational, redesign, and construction clocks (years unless deliberately accelerated). The chain can only advance as fast as its slowest necessary link.
Three scenarios with public signposts.
Compression (base case). Agent production scales; non-backfill dominates; bridges start late but become visible. Signposts for 2027–28: entry postings in exposed occupations stay weak; prime-age exposed cohorts soften further; adopter revenue- or profit-per-employee gaps widen. Planning posture: prepare transitions now rather than wait for aggregate unemployment to confirm.
Extension. Complementarity and demand growth absorb more of the task savings than expected. Signposts: hiring mix changes without broad NILF pressure; hours and wages hold in exposed roles. Planning posture: invest in augmentation; keep the falsification triggers armed.
Constraint. Regulation, reliability limits, or power and capex bottlenecks slow deployment. Signposts: production-agent share stalls; incident and liability language rises in filings; build-out delays outpace labor impact. Planning posture: shift focus from labor shock to capacity, safety, and governance.
We’ll treat these as living scenarios. When the signposts hit or miss, the probabilities update. That’s the public rule.
What would falsify or delay the valley thesis. Sustained, broad acceleration in true entry-level cognitive hiring that’s not merely residual seniorized roles; clear, large-scale net-new role creation visible in both postings and payroll within the next several quarters; reversal of the revenue-per-employee divergence between heavy adopters and the rest of the economy; or a regulatory or institutional brake that meaningfully slows production deployment.
Competing explanations we’re watching. Demographic aging, immigration shifts, interest-rate effects, return-to-office dynamics, sector mix changes, contractor substitution, pure measurement revisions, and ordinary cycle effects. Any of these can move the aggregates. The valley claim rests on the residual pattern after these are considered: exposed cohorts and adopting firms showing compression while less-exposed groups do not.
Compressing the valley
Two practical levers stand out. Both are still early and need real experiments.
Apprenticeship on-ramps into shortage sectors. When a cognitive role compresses, the person who held it can move directly into structured training for work that still requires human presence, judgment under uncertainty, or physical execution. Career change is already common. What’s missing is a low-friction on-ramp that does not require years of formal education before income resumes.
Scale question: which shortage occupations can absorb displaced cognitive workers within 6–18 months at meaningful volume? Design risks include training for yesterday’s shortages, wage suppression, and weak employer commitment. First experiments should be employer-consortium pilots tied to real vacancies, wage floors, and completion-linked funding.
Private infrastructure programs focused on energy and compute. Companies that benefit from the AI transition have both the capital and the long-term interest in a deeper talent pipeline and in public legitimacy. A program that employs people to build generation, transmission, data centers, and integration systems while training them in the relevant trades creates demand for labor on the creation side of the valley at the same time it accelerates the physical layer. Think Works Progress Administration (WPA) types of work, but this time privately funded and intended to build the next economy.
Scale question: how many sustained jobs can these projects absorb per dollar and per year once permitting and supply-chain constraints are priced in? Design risks include public-relations value without net hiring and boom-bust exposure. First experiments should require participating firms to commit to apprenticeships, transparent hiring data, and measurable additionality.
Neither idea requires ideological consensus. Both are available to firms and coalitions of firms that decide the valley is a risk worth managing rather than a feature to be endured.
Limits
This piece can’t yet deliver a full firm-level panel linking adoption to payroll, hours, and contractor substitution. It can’t yet produce clean, sector-adjusted elasticities of headcount to task automation. The process-capture path remains a hypothesis with supporting cases rather than a quantified national flow. The probability ranges are planning ranges, not precise forecasts. We’ll update them as the signposts resolve.
What the piece can show is the shape of the asymmetry, the measurement gaps that hide it, and the levers that still remain open.
The element of surprise
A year ago the surprise was that the disruption would arrive sooner than most people expected. The surprise now is the shape of the transition itself. Reduction and creation are both underway. They’re not synchronized. The interval between them is the Timing Valley.
The original projection was useful because it showed a gap between the consensus timeline and the causal chain. The useful gap today is between the binary public argument and the actual asymmetry of the clocks. If we work from a clear sense of the asymmetry between job displacement and job creation, we can begin building the bridges while the valley is still relatively shallow. That will determine how deep and how long it becomes.
The data are already moving. How they move depend on how deliberately we respond.