Why AI’s payoff hides in productivity statistics
AI is a general purpose technology that demands heavy, hard-to-measure complementary investments—from process redesign to human capital—that initially depress measured productivity before unleashing outsized gains. This measurement gap creates a predictable productivity j-curve: early underestimation, later overestimation, then normalization as intangibles mature. Strategically, leaders should budget for sizable intangible complements, retool metrics beyond GDP and TFP snapshots, and prepare for a lag between AI spend and visible performance.
Points clés
- General purpose technologies like AI require large complementary intangible investments—business process redesign, co-invention, and human capital—that are poorly captured by traditional accounting.
- These hidden investments produce a productivity j-curve: early underestimation of TFP and output, followed by overestimation as intangible assets begin generating measurable returns.
- AI and machine learning meet GPT criteria, implying substantial implementation lags and mismeasurement, as documented by Brynjolfsson, Rock, and Syverson.
- US GDP reached $19.5 trillion in 2017; growth averaged 2.17% (2010–2017), down from 2.72% (2000–2007).
- Explaining the entire post-2010 growth slowdown would require roughly $107 billion per year in unmeasured intangible investment.
- AI startup funding rose from $500 million in 2010 to $4.2 billion in 2016, including a 40% jump between 2013 and 2016.
- Corporate AI investment in 2016 totaled $26–$39 billion—about 300% growth since 2013—largely concentrated in IT; industrial robot shipments doubled since 2004 and quadrupled in consumer electronics.
- For AI to account for the 0.55% “lost” 2017 GDP, correlated intangibles must be about 2.7–4.1 times tangible investments; prior work valued computer capital at up to $11 in market value per $1 of measured spend.
- Software’s j-curve is least mature: mismeasurement understated productivity by an annualized 0.86% by end-2016, and by about 1.6% at the late-1990s peak.
- Hardware-related intangibles are smaller: adjusted TFP was 4.4% above measured by end-2016, with recent hardware slowdowns briefly overstating productivity.
À retenir
So, if your AI program isn’t instantly pumping up the productivity charts, congratulations—you’re probably doing it right. Budget 2–4x your tangible spend for the invisible stuff (process redesign, data plumbing, training), track leading indicators (deployment velocity, model uptime, adoption), and give software capabilities top billing. And when the numbers look flat, resist the panic button; the payoff curve is shaped like a “J,” not a firework—less “boom now,” more “patience, then compounding.”
Sources
Quiz sur le document: 10 questions
