UPLIFT

From the working paper — methodology

The sector is defined in two layers. The core production system covers firms whose output is the creation of built assets: residential and nonresidential building, heavy and civil engineering, specialty trade contractors, architecture and engineering services, construction materials manufacturing, building products distribution, and construction machinery and equipment. The extended layer covers the operation and transaction of the stock the core produces: lessors, transaction services, facilities services, an attributed finance/insurance slice, public planning and building regulation, and surveying/testing.

Sector accounting lives or dies on boundary judgments. Rather than bury them, the panel publishes each one in the register above, ranked by severity — the estimated impact of the judgment on sector totals.

From the working paper

"Construction" as reported in the national accounts understates the economic system that produces and operates the built environment. Following the sector-definition approach developed by Gerard de Valence, the panel assembles fourteen components covering 2005–2025: a core production system (contractors, designers, materials, distribution, equipment) and an extended layer (property operation and transaction, facilities services, attributed finance and public administration).

The extended layer is dominated by property operation — the annuity the core production system feeds. That relationship is the sector's economic signature: a cyclical production system attached to a stable operating annuity roughly its own size.

From the working paper — reading the cycle

The three work-done categories cycle on different clocks. Residential follows household credit and rates: it peaked first (2006), fell furthest (more than half), and was first to boom when rates hit zero in 2020–21. Nonresidential buildings lag the business cycle — the 2008 peak while housing was already collapsing is the classic pattern — and the 2023–25 surge is a policy artifact: CHIPS and IRA subsidies pulled factory construction to record shares of the mix. Infrastructure is the slowest, steadiest clock, funded by budgets rather than markets; ARRA (2009) and IIJA (2022 onward) read as plateaus where the private categories dipped. A strategy that ignores which clock its demand rides on mistakes policy tailwinds for skill.

From the working paper — the productivity puzzle

Nominal value added per core job rises steadily, but most of that rise is price, not output: deflated construction productivity has been roughly flat for decades while manufacturing multiplied. The canonical explanations are structural — every project is a prototype built by a temporary coalition of firms; output happens outdoors on unique sites; the industry is fragmented into hundreds of thousands of small contractors with thin margins and thinner R&D. The sector's employment, meanwhile, swings with the cycle far less than its nominal value does: firms hoard skilled labor through troughs because crews are the scarce asset. This is precisely why the AI uplift question in L7 matters — the sector has absorbed forty years of tools without moving the productivity needle, because the gains kept leaking out through the same contract structures analyzed next.

From the working paper — geography

Construction employment tracks population growth and capital formation, not existing stock — which is why Texas and Florida punch far above their share of the built stock while the Northeast punches below. Concentration compounds: migration drives housing demand, housing demand drives trade employment, and trade availability then becomes a site-selection factor for the factory and data-center wave. The state slice is the panel's thinnest layer (employment only, top ten states); the QCEW ingest upgrades it to full state coverage when run live.

From the executive report — the governance agenda

• Data rights before productivity. Field capture, model data, and cost history are the fuel for every uplift area, and contracts rarely say who owns them. Owners should claim project-data rights at procurement; contractors should treat historical cost data as the asset that survives repricing.
• Procure with the new cost base. The owner's share of the uplift arrives only if procurement actually reprices: updated should-cost models, AI-assisted estimating disclosures, term agreements rebid on the post-AI baseline.
• Watch the vendor toll. A durable toll on the sector's workflow accrues to whoever owns the platform layer. Resist single-vendor lock-in on data formats — portability clauses are cheap now and expensive later.
• Model risk in a life-safety sector. Design automation operates under professional liability and code compliance. Human-in-the-loop sign-off, versioned model records, and clear responsibility chains are the license to automate.
• Labor transition is a capacity question. The sector is capacity-constrained; AI that removes back-office and coordination cost frees skilled field labor rather than displacing it. The exposure is in estimating, drafting, and administration roles — plan that transition explicitly.
• Honest data, honest claims. This analysis recomputes from a quality-flagged database; every seeded cell is labeled and every boundary judgment published. Hold sector AI claims to the same standard.

From the working paper — data quality appendix

Every cell in the panel carries a quality flag (censused live below). seed = seeded structure calibrated to published magnitudes (2017 Economic Census anchors, stylized cycle paths); bea = BEA GDP-by-Industry / NIPA; qcew = BLS Quarterly Census of Employment and Wages; census = Census construction spending (VIP); cbp = County Business Patterns. Ingest scripts replace seeds in place, so the census always reflects the current database — and this page re-tells whatever the database currently holds.