← Back to Insights
Briefing · Transformation History

The Tool Curve Is Not the Value Curve.

A verified read across seven technology eras: what actually repeats, what has genuinely accelerated, and where the value has always been won.

Conditional sign-off verdict

It is reasonable to treat today's AI usage surge as real and still refuse urgency-driven scaling: across two centuries of technology, tool adoption keeps accelerating while the organizational redesign that produces value has never been shown to speed up. Sign off on scale when an initiative has a named owner funding workflow redesign, stated stall criteria, and a production path with its own budget; adoption numbers alone do not qualify.

The two curves, drawn

adoption · widespread individual use measured productivity payoff the gap · redesign not yet done 0 10 20 30 40 ELECTRIFICATION · AVAILABLE 1882 the gap · work not yet redesigned 1 in 10 US households · ~26 yrs factory productivity accelerates · ~40 yrs PERSONAL COMPUTING · AVAILABLE 1981 1 in 5 US individuals · yr 3 aggregate productivity break · late 1990s GENERATIVE AI · AVAILABLE 2022 payoff not yet measured ~45% of US adults · yr 2 years after the technology became available years after the technology became available 0 20 40 ELECTRIFICATION · AVAILABLE 1882 1 in 10 US households · ~26 yrs factory productivity accelerates · ~40 yrs PERSONAL COMPUTING · AVAILABLE 1981 1 in 5 US individuals · yr 3 aggregate productivity break · late 1990s GENERATIVE AI · AVAILABLE 2022 ~45% of US adults · yr 2 payoff not yet measured

Three technology eras · years after each became available · August 2026 · milestones differ by what each era's source measured; the comparison is conservative (the oldest era gets the lowest bar) · anchors verified against primary sources 2026-08-26; open an era below for its evidence and limits.

Electrification1 in 10 US households · ~26 yrs · factory productivity accelerates · ~40 yrs

The first central power station opened in 1882. US households reached 1 in 10 electrified around 1908 (Our World in Data household series). Measured factory productivity did not accelerate until the early 1920s: four decades, closed only when plants were rebuilt around unit drive (a motor on every machine) and a generation of capital turned over.

Where it stops: David's paper is an invited AEA piece, not refereed, and his own caveat travels with it: computers are not dynamos. The four-decade figure describes US manufacturing, not every sector.

Sources: David (1990), AER Papers & Proceedings · Our World in Data, US household adoption

Personal computing1 in 5 US individuals · yr 3 · aggregate productivity break · late 1990s

Computer use reached 20% of US individuals by year three (Bick, Blandin and Deming's comparison series measures individuals' use; household ownership was near 8% then and did not cross 20% until about 1991). Solow's 1987 paradox line described the wait; the aggregate productivity break arrived in the second half of the 1990s, with information technology accounting for about two-thirds of the speed-up (Oliner and Sichel).

Where it stops: The late-1990s break is an economy-wide reading with a live counter-analysis (Gordon 2000 located the revival narrowly in durables manufacturing), so the ~16-year marker is directional, anchored to the published break date rather than a single measured lag.

Sources: Oliner & Sichel (2000), JEP · Bick, Blandin & Deming (2026), Management Science

Generative AI~45% of US adults · yr 2 · payoff not yet measured

About 45% of US adults aged 18-64 had used generative AI by late 2024, two years after ChatGPT's launch: the fastest overall start yet measured, reaching about 55% by August 2025. Whether organizations convert that usage into measured productivity faster than in past eras is unstudied; no published evidence exists in either direction.

Where it stops: The open right edge of this row is the honest state of the evidence, not a prediction. Any claim that the organizational payoff is arriving faster than electrification's or computing's would outrun every source this figure rests on.

Sources: Bick, Blandin & Deming (2026), Management Science

The sign-off test

Name the curve

For every metric, vendor claim, or headline, ask whether it measures tool usage or redesigned work. Only the second has reliably predicted value.

Fund the redesign

The pilot and the governed production system are different products with different budgets, sponsors, and timelines. Fund the second explicitly.

State stall criteria

Name in advance what would show the initiative stalling rather than "in resistance": no workflow redesigned, pilot renewals without production budget, sponsor loss.

Gate, then cadence

Where deployment is regulated, keep pre-deployment approval as a gate under the applicable regime, then add a re-certification rhythm owned by counsel and compliance.

Owner, briefing, proof

Owner

A sponsor who owns the workflow being redesigned and the budget for its production form, not only the pilot that demonstrates it.

Briefing

Which workflow changes, the baseline it moves, the expected dip while redesign lands, and the stall criteria that would stop the initiative.

Proof

Measured process change: workflows redesigned, decisions rerouted, approvals restructured, reported beside usage so adoption cannot masquerade as absorption.

What leaders should take from it

1
In every era measured here, value followed reorganization, not installation.

Electrified factories showed no productivity jump for about four decades; gains arrived when plants were rebuilt around the new capability. The computer age repeated the shape, and the economics of general-purpose technologies explains why: the payoff rides on complementary investment the technology itself does not supply.

2
Tool adoption genuinely accelerates, era over era.

Across 15 technologies and 166 countries, technologies invented ten years later were adopted about 4.3 years faster. Generative AI is the fastest overall start yet measured: about 45% of US working-age adults used it within two years.

3
Organizational absorption shows no sign of the same acceleration.

Most organizations report regular AI use somewhere, and most have not begun scaling it. Whether firms restructure faster this cycle than in the electrification or IT eras is unstudied, so the confident version of "AI transforms business faster than any technology" has no evidence behind it in either direction.

4
The gap between those curves is where programs stall.

"Fastest-adopted technology in history" and "most programs show no measurable return yet" are the same fact seen from two curves. When adoption speeds up and absorption does not, the gap widens, and more organizations stand in it at once.

5
The transformation story is a vocabulary, not a law.

The familiar arc (trigger, resistance, pioneers, tipping point, new normal) describes transformations that completed; the ones that stalled left the sample. It stays useful for locating a program, provided stall criteria are named in advance so the story can be wrong.

Where the evidence stops

Three familiar claims run ahead of the evidence. That AI transforms organizations faster than past technologies: the organizational comparison is unstudied. That 95% of AI programs fail: the figure comes from preliminary research whose denominator includes organizations that never ran a pilot. That a specific window of years separates winners from losers: no reviewed source supports a window. The historical pattern itself is drawn from transformations that completed, which is a real selection limit.

The findings in full

8/10
The productivity paradox is the best-documented regularity across eras.

Factory electrification's four-decade lag closed through plant redesign and capital turnover. The Productivity J-curve formalizes the mechanism: intangible complementary investment precedes measurable gains. A 2026 field experiment adds causal modern evidence: startups prompted to reorganize production around AI generated 1.9x the revenue of peers using it only to speed tasks.

8/10
Adoption lags have compressed for two centuries, and generative AI extends the trend.

Mean adoption lag across technologies since 1820 runs about 45 years, shrinking roughly 4.3 years per invention decade. Generative AI reached about 45% of US adults 18-64 by late 2024 and about 55% by August 2025; use at work spread about as fast as the PC did.

7/10
No published evidence shows the organizational curve accelerating.

The pre-generative-AI baseline had under 6% of US firms using AI in production, concentrated in the largest firms. On the current wave, nearly two-thirds of surveyed organizations have not begun scaling, and at least 30% of generative-AI projects were projected to be abandoned after proof of concept. The restructuring-speed comparison across eras remains unstudied.

7/10
Adoption is uneven, and a meaningful share of it is unsanctioned.

Use concentrates by firm size and geography. Browser-telemetry research reports security tooling blind to most AI logins in its customer base, and a large 2026 survey finds about one in five organizations claiming mature governance for AI agents. Measured adoption overstates governed capability.

6/10
The five-phase arc survives as a diagnostic vocabulary, not a prediction.

Historians of technology show the genre over-weights innovation stories, and the arc's sample is completion-biased. On the current era the usual fifth phase is not yet reachable: the legal ground moved in both directions within a year, with obligations deferred in the EU, a state act repealed and reenacted before it applied, and a US federal consent order set aside.

First moves for a sponsor's team

01
Split every AI number your program reports into the two curves.

Usage counts, seat activations, and query volumes on one line; workflows redesigned, decisions rerouted, and approval chains restructured on the other. The second line is the value forecast.

02
Reset pilot expectations with the dip in view.

Tell the steering committee that measured productivity may fall while redesign lands, and set the review horizon accordingly. Programs killed in the dip pay the cost of the paradox without collecting its payoff.

03
Write stall criteria into every initiative charter.

Two quarters without a redesigned workflow, pilot renewal without production budget, or sponsor turnover without succession are stall signals, not resistance to wait out. Naming them in advance keeps the transformation story honest.

04
Fund the production system as its own decision.

Put the two-budgets choice in front of the sponsor who owns the workflow: the demo proved possibility; the governed, integrated, supported production form is a separate product with separate cost. Deciding it explicitly beats discovering it mid-rollout.

05
For regulated deployments, run approval as a gate plus a cadence.

Keep the pre-deployment approval your regime requires as a hard gate, then have counsel and compliance set a re-certification rhythm with named authority to withdraw a prior approval, because the rules are currently moving in both directions.

Where to start

Start with one workflow where usage is already high and redesign has not happened: that gap is the cheapest place to prove the two-curve model on your own ground. Measure the baseline, redesign the flow, and report process change beside usage. If the gap turns out to be wide across the portfolio, widen to a readiness look before scaling anything.

Claim ledger

29/29
Checked
citations traced to primary or official sources on August 26, 2026; one aggregate unreachable and held out
0
Fabricated
no invented source survived verification
17
Corrected
figures, samples, attributions, dates, and legal wording tightened against primaries
4
Demoted
preliminary, superseded, or telemetry-scoped claims held out of stronger findings
CorrectedDavid, 1990 (AER Papers & Proceedings): electrification's productivity lag was four decades, closed by unit-drive redesign and capital turnover; invited paper, not refereed.repec.org
CorrectedSolow, 1987: the "computer age everywhere but in the productivity statistics" line is verbatim, from a New York Times book review about a post-1973 slowdown, not a study.gwern.net
ConfirmedOliner & Sichel, 2000 (JEP): IT accounts for about two-thirds of the late-1990s US productivity speed-up; Gordon's counter-analysis noted.aeaweb.org
CorrectedBrynjolfsson, Rock & Syverson (AEJ: Macro 2021): Productivity J-curve; adjusted TFP levels 11.3% (2004) and 15.9% (2017) above official; the authors state intangibles do not explain the post-2004 slowdown.nber.org
CorrectedComin & Hobijn, 2010 (AER): 15 technologies, 166 countries, 1820-2003; mean adoption lag ~45 years; ~4.3 years faster per invention decade. An earlier 2004 paper's sample had been fused into this citation.aeaweb.org
ConfirmedComin & Mestieri, 2018 (AEJ: Macro): adoption lags converged across countries while intensity of use diverged.aeaweb.org
CorrectedBick, Blandin & Deming (Management Science 2026): ~45% of US adults 18-64 used generative AI by late 2024 (author-corrected from 39.4%); overall start faster than PC or internet; work adoption as fast as the PC; ~55% by Aug 2025.doi.org
CorrectedMcElheran et al. (JEMS 2024): US Census-based; under 6% of firms used AI in production, ~18% employment-weighted; reference year 2017, a pre-generative-AI baseline; startup adoption concentrated in a few cities.doi.org
DemotedMIT NANDA, 2025: "95% getting zero return" is verbatim but preliminary and not peer-reviewed; method is 52 interviews, 153 leader surveys, 300+ public initiatives; critique: the denominator includes non-pilots.report pdf
ConfirmedGartner, Jul 2024: prediction that at least 30% of generative-AI projects will be abandoned after proof of concept by end-2025; analyst forecast, not measurement.gartner.com
ConfirmedGartner, Jun 2025: prediction that over 40% of agentic-AI projects will be canceled by end-2027; supporting poll is self-selected webinar attendees.gartner.com
CorrectedMcKinsey State of AI, Nov 2025: 1,993 respondents, 105 nations; 88% report regular AI use somewhere; nearly two-thirds not yet scaling; 51% of AI users report at least one negative consequence; self-reported survey.mckinsey.com
DemotedMETR, 2025: RCT, 16 experienced developers, 246 tasks; 19% slower with AI while believing ~20% faster; METR flags the result as out of date, and its 2026 continuation is self-reported with different findings.metr.org
CorrectedKim, Kim & Koning, 2026: field experiment, 515 startups; reorganization treatment produced 1.9x revenue; working paper under review; previously misattributed to a consultancy via an article that does not contain the claim.ssrn.com
ConfirmedAcemoglu, 2024 (NBER w32487; Economic Policy 2025): upper bound of no more than 0.66% TFP over ten years, revised below 0.53%; a ceiling, not a central estimate.nber.org
ConfirmedGoldman Sachs Top of Mind 129, Jun 2024: "Gen AI: Too Much Spend, Too Little Benefit?"; a printed debate in which Covello argues the skeptical case and the firm's economists argue the opposite.goldmansachs.com
CorrectedHyperscaler 2026 capex: ~$700B combined for the four largest, up more than 60% from 2025, per primary reporting; a higher widely-shared triple traces to a different compilation, and guidance has since been revised.cnbc.com
CorrectedNVIDIA Q1 FY2027 (May 2026, 8-K): record Data Center revenue $75.2B, up 92%; the draft had cited the wrong fiscal year and a guidance figure that appears in no NVIDIA document.sec.gov
CorrectedOpenAI figures, 2026: $852B post-money valuation confirmed; revenue run-rate above $40B annualized by Aug 2026; the widely-quoted 2026 loss is a reported internal projection; the $1.4T figure is stated multi-year commitments, largely contingent.cnbc.com
DemotedCircular-financing aggregate: the "$800B+" total is paywalled and unverifiable; components split: Microsoft's 8-K confirms a $250B Azure purchase commitment; the Nvidia investment is a progressive letter of intent; the Oracle figure is reported, not disclosed.sec.gov
CorrectedRegulation (EU) 2026/1744: in force Jul 27, 2026; defers the AI Act's high-risk obligations to Dec 2, 2027 (Annex III) and Aug 2, 2028 (Annex I); the dates are unconditional, not tied to standards availability.europa.eu
CorrectedCJEU C-634/21 (Dec 2023): automated credit scoring is an Art. 22 GDPR decision where the receiving party draws strongly on the score; a conditional holding, not a flat rule.europa.eu
ConfirmedCJEU C-203/22 (Feb 2025): trade-secret claims route disclosure to the supervisory authority or court for balancing; the data subject receives the procedure and principles applied, not the algorithm.europa.eu
CorrectedColorado, 2024-2026: SB24-205 delayed, then enforcement paused by an agreed court stay (not an adjudicated injunction), then repealed and reenacted by SB26-189 with duties from Jan 1, 2027.colorado.gov
CorrectedFTC, 2024-2025: Operation AI Comply sweep; Workado order over AI content-detector accuracy claims; the Rytr final order reopened and set aside in Dec 2025 citing the administration's AI policy.ftc.gov
ConfirmedDittmar (QJE 2011): the printing press reached 205 European cities within fifty years; early-adopting cities grew about 60% faster over the following century.cepr.org
CorrectedEdgerton, 2007: The Shock of the Old critiques innovation-centrism (Concorde as celebrated failure, corrugated iron as underweighted use); the selection-bias point stands separately from his argument.oup.com
CorrectedUS household adoption speeds (Our World in Data): landline telephones took ~86 years to reach 80% of households; tablets went 3% to 51% in six years; comparisons re-sourced from a popular chart to the underlying dataset.ourworldindata.org
DemotedShadow-use figures: LayerX telemetry (no visibility into 89% of AI logins, self-selected customer base) and Deloitte's 2026 survey (21% mature agent governance) are separate methods; a prior citation attributed both to an article containing neither.deloitte.com
Contested signals: held out of the verdict

The evidence does not establish that the AI investment cycle is sustainable or circular, that individual productivity gains are large or small in general, or that macro forecasts converge: credible estimates span an order of magnitude. None of these entered the verdict above.

What would change this conclusion

Related work

Verified research · 29 citations checked · 0 fabricated · 17 corrected · 3 demoted