Construction adopts late. Now there is a number.
The folk claim says construction is the last industry to pick up any technology. This measures it: for ~20 technologies and seven industries, the first fiscal year each technology sustainably entered 10-K disclosure, FY1996–2025, and how many years construction trailed the leading industry.
When each technology entered each industry's disclosure
A dot per industry at its first sustained entry (5% of filings, 3 firms, held two consecutive years). Construction is the large blue dot. A lag marked “≥” is a lower bound: the leader was already above threshold when the panel starts in 1996. Hover any dot for detail.
The lag, technology by technology
Reading the matrix
The lag is real but structured. Technologies dated after 2005, where the panel is uncensored, put construction 4–7 years behind the leader: cybersecurity 4, IoT 4, cloud 6, big data 6, blockchain 7.
The internet is the exception that dates the claim. In FY1996 construction was already among the leading industries on internet language. When a technology arrives with a consumer-scale wave, construction moves with everyone else: the same pattern the companion AI study finds for AI after ChatGPT.
Empty rows are findings. RFID, wearables, VR/AR, 3D printing and autonomous equipment each cleared the threshold in at least one contrast industry and never in construction.
Risk-first entry is a signature, not a rule. Of the technology-industry entries dated 2006 or later, only a minority arrive risk-first, overwhelmingly cybersecurity, plus blockchain. Most technologies enter as capability language.
AI's lag is a lower-bound story told elsewhere. The AI term family failed the precision check on the pre-2015 eras (document part numbering and initials masquerade as “A.I”), so its adoption numbers await passage-level coding; the companion AI study carries the AI contrast on the modern panel.
Thirty years of filings, one click from the original
One square per company-year, FY1996–2025. Colour is how many of the confirmed technology families that annual report mentions. Hover a square to see which; click it to open that filing on sec.gov, scrolled to its first technology sentence where a verified anchor exists. Each industry splits into its SIC subgroups, including construction's upstream machinery suppliers, the Caterpillar and Deere band, or search across all eight.
Every technology, every industry, thirty years
Share of each industry's operating filings mentioning the technology, FY1996–2025. Construction is the heavy blue line.
The lag, tested
Four formal tests of the paper's claims. The headline twist: the folk claim does not survive: construction sits mid-pack on entry order, while the statistically late industries are pharma, retail, utilities and, of all bands, construction's own machinery suppliers. What is real is the gap to the frontier. Every estimate is an association about corporate language.
Can a keyword be trusted?
Every adoption number on this site comes from a term family that passed a precision check: three open-weight language models from three labs (Alibaba, Microsoft, Mistral AI) read random hit contexts from the filings and judged whether the matched term really denotes the technology. A family stays keyword-measured only if at least 80% of its hits are true positives; a family that fails is measured by passage-level coding instead.
How it was measured
Data
Every 10-K variant (10-K, 10-K405, 10-KSB, 10-KT and amendments) filed by US registrants in seven SIC-defined industries, fiscal years 1996–2025: 41,880 filings, frozen under a SHA-256 manifest. FY1996 is the panel floor because EDGAR's electronic-filing mandate phased in through 1996; earlier years exist only on paper or in licensed archives, which the series' public-data-only rule excludes.
Term families
~25 boundary-guarded term families, tiered: clean families are measured by keyword counts (subject to the precision check); noisy families (lean, modular/prefab, automation, TQM, six sigma) never carry adoption numbers from raw counts and await passage-level LLM coding. Bare abbreviations (BIM, ERP, RFID, GPS, IoT, AI) are matched case-sensitively so initials and tickers never count.
Dating an entry
An industry adopts a technology's language in the first fiscal year at least 5% of its operating filings and at least 3 distinct firms mention it, sustained for two consecutive years. Sensitivity variants (2%/2 firms, 10%, no-sustain) are computed alongside. Leaders already above threshold in FY1996 are left-censored, and their lags are reported as lower bounds.
The screen
Filings under 5,000 words are excluded (shells and blank-check registrants). The series' financial screen needs XBRL data that does not exist before 2009, so it is a robustness check on the 2014+ overlap rather than a panel-wide filter.
What this is not
Disclosure entry is when a technology reached the industry's conversation with investors, not when the first machine arrived on a site. Every result is a fact about corporate language.