Every Technology Arrives Twice

Thesis

New technologies arrive twice. First as capability: machines can do something that was impossible yesterday, and the early adopters demonstrate it. Then, often decades later, as value at scale: organizations have rebuilt their buildings, workflows and habits to the point where capability turns into measurable impact. Electricity needed roughly four decades for that second journey, the computer a good two.

Generative AI stands in the middle of its first arrival, and the numbers currently causing disillusionment simply describe the distance to the second: roughly six percent of companies draw measurable value at scale, while the capability curve keeps climbing steeply and practitioners’ trust is even falling. Measured against the hype, that is a disappointment. Measured against how general purpose technologies have historically arrived, AI is on schedule – if anything, ahead of it. The bottleneck of the second arrival has never been the technology: it sits in structures, workflows, and in the trust of the people who are supposed to rely on the new thing every day.

Two Curves Coming Apart

The usual story measures a single curve: what models can do. METR measures how long the tasks are that a model reliably completes on its own, and finds that span doubling roughly every four to seven months. That’s the curve forecasts, newsletters and conference stages talk about.

Value, however, appears on a second curve: the uptake of that capability into actual work. And it runs differently. A study by MIT (Project NANDA, State of AI in Business 2025) found that about 95 percent of the enterprise generative-AI pilots it examined showed no measurable effect on the bottom line – against 30 to 40 billion dollars of spend. More precisely: roughly five percent of initiatives measurably accelerate revenue, the rest stall. The figure is often shortened to “95 percent failed”; the precise claim is about the absent P&L impact, not a measured failure.

How long the road from capability to impact actually is becomes visible when the stations are laid out side by side – from trying a tool, to daily use, to trust, to measurable value:

A Forecast Correcting Itself

This becomes clearest at the most honest place the field has to offer: among the forecasters themselves. The scenario AI 2027 (AI Futures Project, April 2025) traces an acceleration to superintelligence by the end of 2027 – the most aggressive serious takeoff forecast. In February 2026 the authors graded their own 2025 predictions and arrived at about 65 percent of the predicted pace. Raw capability roughly kept up, with the METR time-horizon following their curve almost one to one. What didn’t keep up they stated themselves – marketed AI agents struggled to find widespread usage. In late 2025 they moved their median for superintelligence out of 2027 and into the early 2030s.

The point sits in the mechanism of that correction: the scenario’s models got the capability curve nearly right – what they missed is the curve on which that capability passes into actual work and value creation.

Arvind Narayanan and Sayash Kapoor distilled the same observation into a frame in 2025 (AI as Normal Technology, Knight First Amendment Institute): impact comes from slow diffusion through regulation, organization and habit, not from the capability leap – on the timescale of decades, as they write explicitly. Their image for it is a trickle, not a tsunami.

Four Decades for Electricity, Two for the Computer

That decades lie between capability and broad impact is no AI peculiarity but the recurring pattern of general purpose technologies – and for electricity and the computer it is unusually well measured.

The first commercial central power station went online in 1882 (Edison’s Pearl Street Station in New York). Seventeen years later, in 1899, only 4.8 percent of mechanical drive power in US factories came from electric motors; the 50-percent mark fell around 1919, almost four decades after availability (Warren Devine, From Shafts to Wires, Journal of Economic History 1983). The economic historian Paul David showed why: factories had been built around steam engines, central shafts and leather belts. Hanging an electric motor into that architecture achieved little – its value only emerged once factories were rethought, with individual motors on individual machines, single-story halls, new workflows. The measurable productivity surge of US industry came in the early 1920s, “four decades after the first central power station opened for business” (Paul David, The Dynamo and the Computer, American Economic Review 1990).

The computer repeated the pattern in compressed form. The IBM PC appeared in 1981; six years later Nobel laureate Robert Solow quipped that the computer age was visible everywhere “but in the productivity statistics” (New York Times Book Review, July 1987). The statistics proved him right – until the mid-1990s: between 1995 and 2004, US productivity growth doubled from roughly 1.4 to roughly 3.1 percent per year, and about two-thirds of that acceleration traces back to information technology (Oliner & Sichel, Journal of Economic Perspectives 2000). From the PC to the surge: 14 to 23 years. The cloud, as the more recent case, needed about 14 years from AWS (2006) to the point where companies first spent more on cloud than on their own data centers – and even in 2024 only about half of enterprise data lived there.

The mechanism behind this rhythm has a name: economists call electricity, the computer and AI general purpose technologies, whose value only emerges through co-invention in the fields of application – new workflows, new roles, new buildings (Bresnahan & Trajtenberg 1995). Erik Brynjolfsson and colleagues formalized this as the productivity J-curve: early on, investment flows into invisible complements such as process redesign and training that appear on no balance sheet, and only afterwards does the value show up in the numbers. How large that invisible layer is shows in their estimate that one dollar of installed computer capital goes along with a multiple of accompanying organizational assets in market valuations (Brynjolfsson, Hitt & Yang, Brookings Papers 2002).

Where the Six Percent Sit in This Rhythm

Against this background, the present reads differently. McKinsey counts roughly six percent “high performers” in AI’s third year – companies with substantial bottom-line impact – alongside 88 percent usage (State of AI 2025). Electricity in its third year stood at practically zero; even in its seventeenth it hadn’t reached five percent of factory drive power. Taking history as the yardstick, AI is not behind schedule but ahead of it.

The number stays uncomfortable nonetheless, in two directions. First, it barely moves: McKinsey lowered the bar for “high performer” over the years from 20 to 10 to 5 percent of bottom-line impact, and the share stayed in single digits throughout (about 8 percent in 2020 through 2022, about 5 percent in 2024, about 6 percent in 2025) – on a constant bar, the series would read flat to declining. Second, something moves underneath: BCG counts 35 instead of 22 percent of companies in the scaling phase over the same period. The top isn’t growing, but the middle is closing in – exactly what one expects shortly before the steep section of a diffusion curve.

Two readings follow for the projection. If AI tracks the rhythm of the computer and the cloud, broad value realization lands in the early to mid-2030s; Gartner meanwhile puts the productivity plateau at two to five years out and expects GenAI in production at over 95 percent of enterprises by 2028 – usage, mind you, not value. Arguing for compression: software diffuses without rebuilding factories, and individual adoption is breaking every historical record, with 39 percent of the US working-age population using generative AI just two years after ChatGPT and around 55 percent after three (Bick, Blandin & Deming, NBER). Arguing for the old rhythm: the real complementary investment has stayed the same – workflows, accountability, trust – and intensity remains thin, with only one to five percent of work hours assisted by generative AI so far.

The Diffusion Curve for Europe and Germany

For Europe, this projection can now be quantified officially. The Eurostat series on AI use in enterprises (10+ employees, at least one AI technology) shows the typical early S-curve: 7.7 percent (2021), 8.1 (2023), 13.5 (2024), 20.0 percent (2025) – most recently a doubling pace of under two years. Germany sits above the EU average at 26 percent (2025), and the diffusion gradient by size is stark: 55 percent of large EU enterprises use AI, but only 17 percent of small ones. Ask German companies openly about “using AI”, however, and the ifo business survey already reports 54.5 percent (May 2026) – self-reported use has crossed the majority threshold while the official technology checklist is nowhere near it. That, too, is the gap between trying and reliable integration, just at the company level.

The political expectation stands in open contradiction to this. The EU’s Digital Decade programme sets the target that by 2030 at least 75 percent of enterprises use cloud, big data or AI. Yet in its 2025 State of the Digital Decade report, the Commission itself projects only 35.9 percent AI adoption by 2030 – and the 75-percent mark reached only around 2042. That deserves a second look: 2042 is year 20 of generative AI. The Commission’s own projection lands almost exactly on the rhythm the computer needed from availability to breadth.

The Bottleneck Is Organizational, Not Technical

When adoption lags, it’s rarely about the interface. The Boston Consulting Group sums up the success condition of AI initiatives in a rule of thumb: ten percent algorithms, twenty percent technology and data, seventy percent people and process. Most organizations invest in the reverse order. McKinsey confirms the pattern from the other side: bottom-line impact correlates most strongly with redesigning workflows, and that is exactly what stays rare – only one percent of executives call their own rollout mature.

This explains why individuals and organizations move at different speeds. A designer or a developer in the early-adopter layer integrates a tool in days. An organization doesn’t follow the capability curve, because structure, accountability and habit don’t follow the capability curve – Paul David’s factories that kept their steam-era architecture are the hundred-year-old exhibit for this.

Trust Falls While Usage Rises

Beneath the adoption gap sits a finding that looks contradictory at first: usage is rising, trust is falling. Stack Overflow’s 2025 developer survey shows 84 percent who use or intend to use AI tools – yet trust in the accuracy of the output fell within a year from around 40 to 29 percent, while 46 percent actively distrust it. Two-thirds report spending more time fixing “almost-right” AI output. Among designers the pattern is the same: in the Figma 2025 AI Report only 32 percent say they can rely on the output – the lowest score measured, alongside a high perception of efficiency.

The finding reaches beyond the tool level. The global study by KPMG and the University of Melbourne (2025) finds only 46 percent of people willing to trust AI, and 70 percent expecting regulation. Edelman’s Trust Barometer (Flash Poll, November 2025) calls trust the missing ingredient in the AI boom: adoption depends on trust, not on the technology. Between trying a tool for the first time and relying on it daily lies a question of trust, then, and no question of convenience.

How Trust Gets Into the Tools

If trust paces the second arrival, the lever sits in the diffusion layer, not in more capability. Four approaches are emerging, each evidenced:

• Demonstrated value over promises. Edelman measures the strongest jump in trust – around 40 points – where AI helps people understand something. Trust comes from lived, verifiable results, not from marketing. The practical consequence: make small, verified wins visible instead of asserting capability.
• Provenance and traceability. Stack Overflow’s own diagnosis of the developer trust gap ends at a simple rule: cite your sources. Whoever can trace where an output came from can trust it for a reason – or disagree for a reason.
• Calibrated reliability. Research on calibrated trust shows the goal isn’t maximal but appropriate trust: the reliability of the output has to match what is claimed about it. Both over- and under-trust degrade decisions. Human oversight at the right points is calibration here, not distrust.
• Governance as a signal. The NIST AI Risk Management Framework, the certifiable ISO/IEC 42001 standard and the EU AI Act are becoming a trust signal in procurement. They say less about a system’s capability than about the auditability of its use – which is exactly what organizations ask before they go to scale.

What This Means in Practice

History gives a more precise instruction than the hype does. The value of electrification did not emerge in the power plants but in the rethought factories; the value of IT not in the computers but in the rebuilt workflows. Economists call this co-invention, and it is the reason the second arrival cannot be waited out – it gets built. The work that creates value now sits on the second curve: integration into real workflows, traceability of results, a build-up of trust that has to outpace the skepticism. Whoever designs that layer shortens the historical rhythm. Whoever waits for the next model repeats the factory that kept its steam architecture and wondered why the electric motor achieved nothing.

Assessment

This assessment comes from product and design practice – design leadership, design ops, AI in everyday product work, with a focus on SMEs and mid-sized organizations in the German-speaking region. What this perspective sees: the operational shift in teams, the trust behavior of practitioners, the economics of small and mid-sized engagements. What it does not see: the inside of capability research, where the real precision of capability forecasts is decided; the multi-year adoption cycles of global corporations; and the full depth of economic history – an economic historian would test the electricity and IT analogies more strictly against their limits, an economist would weight the diffusion mechanics differently. The historical numbers themselves are taken from the primary sources; their transfer to AI remains an analogy with assumptions.

Critical Assessment

What Holds Up

• The historical series are verified against primary sources: Devine’s electrification table (4.8 → 25 → 53 → 78 percent), David’s “four decades”, the 1995–2004 productivity doubling from the official BLS series
• The self-downgrade of AI 2027 is documented: the authors’ own grading (February 2026) and the shift of the superintelligence median into the early 2030s
• The scaling gap is convergently evidenced across four independent 2025 sources (MIT, McKinsey, BCG, Gartner); the decline in trust is measured among practitioners and the public alike
• The record speed of individual adoption is solidly measured (Bick, Blandin & Deming: faster than the PC and the internet in comparable windows)

What Needs Context

• The McKinsey series is, strictly speaking, not comparable across years due to definition changes (20 → 10 → 5 percent of bottom-line impact) – it shows no climb, but no precise trend can be read from it either
• The analogy has limits: Devine’s percentages measure drive capacity, McKinsey’s percentages measure survey self-reports – a shared rhythm does not mean a shared measurement
• Software diffuses without rebuilding factories; the cloud needed 14 rather than 40 years. A substantial compression of the rhythm is plausible, and the projection is a range, not a date
• Trust is one bottleneck among several: cost, latency, maturing capability and plain inertia pace adoption too
• Practitioners in the early-adopter layer overestimate the majority’s pace because they mistake their own surroundings for the curve – this assessment itself is written from inside that layer

Discussion Questions

01 Shortening the rhythm: Electricity needed forty years, the computer twenty, the cloud fourteen – which properties of generative AI justify assuming further compression, and which argue against it?

02 Co-invention as a task: If value is co-invented in the fields of application – which roles and disciplines occupy that layer: design, engineering, a new function?

03 Speed of trust: Can trust be built faster than an organization can change its structures – or is building trust itself the structural change?

04 The measuring-bar question: If McKinsey has to lower the threshold for “high performer” to keep the number stable – what would be an honest, durable measure for the second arrival?

05 Public sector: Public administration carries especially heavy structures and especially high trust requirements – does that imply an even longer rhythm, or does the sector skip ahead later with proven patterns?

Sources

• Warren D. Devine Jr. – From Shafts to Wires: Historical Perspective on Electrification (Journal of Economic History, 1983; full text of the 1982 ORAU study)
• Paul A. David – The Dynamo and the Computer (American Economic Review, 1990)
• Robert Solow – “We’d better watch out” (New York Times Book Review, July 12, 1987; citation record)
• Oliner & Sichel – The Resurgence of Growth in the Late 1990s (Journal of Economic Perspectives, 2000)
• Brynjolfsson, Rock & Syverson – The Productivity J-Curve (AEJ: Macroeconomics, 2021)
• Brynjolfsson, Hitt & Yang – Intangible Assets: Computers and Organizational Capital (Brookings Papers, 2002)
• Bresnahan & Trajtenberg – General Purpose Technologies: Engines of Growth? (Journal of Econometrics, 1995)
• Bick, Blandin & Deming – The Rapid Adoption of Generative AI (NBER Working Paper 32966)
• Bick, Blandin & Deming – The State of Generative AI Adoption in 2025 (St. Louis Fed, November 2025)
• AI 2027 – A Scenario (AI Futures Project)
• AI Futures Project – Grading AI 2027’s 2025 Predictions (February 2026)
• MIT Project NANDA – State of AI in Business 2025 (Fortune coverage)
• McKinsey – The State of AI 2025
• McKinsey – The State of AI in 2022 and a Half Decade in Review (high-performer series)
• BCG – The Widening AI Value Gap (2025)
• Gartner – Press release on GenAI project abandonment (July 2024)
• Synergy Research / TechCrunch – Cloud spending passes data centers (2021)
• Eurostat – 20% of EU enterprises use AI technologies (December 2025)
• European Commission – State of the Digital Decade 2025 (COM(2025) 290 final)
• Decision (EU) 2022/2481 – Path to the Digital Decade (75% target)
• ifo Institute – More than half of companies use artificial intelligence (June 2026)
• Stack Overflow – 2025 Developer Survey (AI)
• Figma – 2025 AI Report
• KPMG & University of Melbourne – Trust, Attitudes and Use of AI (2025)
• Edelman – Trust Barometer Flash Poll: Trust and AI (November 2025)
• Narayanan & Kapoor – AI as Normal Technology (Knight First Amendment Institute, 2025)
• METR – Measuring AI Ability to Complete Long Tasks
• NIST – AI Risk Management Framework
• AI 2027: A Scenario – earlier field note for context

Glossary

General Purpose Technology A technology applicable across nearly all industries that keeps improving – like steam power, electricity, the computer. Its value only emerges through adaptive invention in the fields of application, which is why its impact trails its availability by years to decades.

Solow Paradox Named after the economist Robert Solow, who observed in 1987 that the computer age was visible everywhere except in the productivity statistics. The apparent contradiction resolved in the mid-1990s, when the preceding years of IT investment became measurable.

Productivity J-Curve Model by Brynjolfsson, Rock and Syverson: with new general purpose technologies, investment first flows into invisible complements (process redesign, training, new structures) that initially depress measured productivity and later lift it – the trajectory resembles a J.

Co-invention Term from the research on general purpose technologies: the users invent the usage along with the technology – new workflows, roles and structures, without which the technology does not unfold its value. In electrification, the co-invention was the rethought factory.

High Performer (McKinsey) Companies reporting substantial bottom-line impact from AI in McKinsey’s State-of-AI surveys. The threshold was lowered over the years (from 20 to 10 to 5 percent of EBIT impact), while the share stayed in single digits throughout.

EBIT Earnings Before Interest and Taxes – operating profit before interest and taxes. In the cited studies, the reference measure for whether AI contributes measurably to business results.

Digital Decade The EU’s digital-policy target programme to 2030 (Decision (EU) 2022/2481). One of its targets: at least 75 percent of enterprises use cloud, big data or AI. The annual “State of the Digital Decade” report measures progress against these targets.

GenAI Divide Term from the 2025 MIT study for the gap between companies that draw measurable value from generative AI (few) and the majority whose pilots stay without P&L impact.

Diffusion (of innovations) The process by which a novelty spreads through a population or organization – typically slower and more uneven than the technical maturity of the novelty itself. The transition from the early-adopter layer to the majority is considered the hardest point.

Early Adopter The early user layer in diffusion research: people and organizations who take up a novelty before the majority trusts it. Their pace says little about the pace of the majority.

Takeoff Forecast A prediction that AI capability becomes self-accelerating – for instance by AI taking over AI research – and reaches superhuman levels within a short time. AI 2027 is the best-known fully elaborated forecast of this kind.

Calibrated Trust Research term for appropriate rather than maximal trust: the reliability you assign to a system should match its actual reliability. Both over- and under-trust lead to worse decisions.

NIST AI RMF The AI Risk Management Framework of the US National Institute of Standards and Technology – a reference framework defining what makes an AI system trustworthy (reliability, safety, transparency, explainability, privacy, fairness).

ISO/IEC 42001 The first certifiable international standard for AI management systems (2023). Organizations can have their governed use of AI externally certified – increasingly a trust signal in procurement.

EU AI Act The European Union’s AI regulation – the first comprehensive legal framework for AI systems, with risk-based obligations for providers and operators. Enforcement phases in through 2026/2027.

METR Research organization that measures the ability of AI models to complete long, autonomous tasks. Its “time-horizon” metric serves as one of the more robust measurements of the raw capability curve.