Technology · Economy · Work
Can AI actually deliver the productivity gains it's promising — or is it hype?
At the task level
Real and measurable
14–80% gains on specific tasks, documented by research
At the macro/economy level
Not showing up yet
89% of companies report no productivity impact; GDP data flat
Verdict based on the perspective of 11 sources — publications, academic research, and central bank analysis.
Last updated Mar 19, 2026 · Atemporal — review quarterly
⚡ The Central Paradox
AI is measurably helping individual workers complete specific tasks faster — sometimes dramatically so. Yet at the company level, almost no one can explain the upside in their filings. And at the economy level, AI simply does not show up in productivity statistics, GDP, employment, inflation, or earnings data outside of tech. Individual gains are real. Aggregate gains are, so far, missing.
Background
Since ChatGPT launched in late 2022, artificial intelligence has been celebrated as a once-in-a-generation productivity tool — yet the economic data has been slow to show the gains that advocates promised. The gap between AI's visible impact on individual tasks and its absence from macroeconomic productivity statistics is the central paradox this page examines.
What the numbers actually say
14–80%
Task-level productivity gains documented by controlled studies (MIT, Anthropic, Stanford)
89%
of companies (6,000 CEOs/CFOs surveyed) report no impact on productivity or employment in the last 3 years
0.01pp
AI's estimated contribution to total factor productivity growth in 2025 — essentially zero
95%
of enterprise AI pilots fail to generate meaningful ROI, per MIT NANDA study (though methodology is debated)
1.9%
US productivity growth in 2025 — just below the long-run average of 2%, far short of internet-era boom
0.5%
Total productivity gain Daron Acemoglu (Nobel laureate) projects over the entire next decade
📚 Historical Parallel
We've been here before: Solow's Productivity Paradox
In 1987, Robert Solow observed: "You can see the computer age everywhere but in the productivity statistics." Despite the PC revolution, productivity growth actually slowed during the 1970s and '80s. It only showed up in the data in the late '90s — after 30+ years of electrification-era-style adoption, redesign, and workforce retraining. The San Francisco Fed notes it took nearly 100 years for electricity to fully boost productivity after Faraday's discovery. The question isn't whether AI will deliver — it's when, and the honest answer is: probably not yet.
The case for real gains
✅ Evidence it's working
- •MIT (2023): AI raises skilled worker performance by ~40% on structured tasks like writing, coding, analysis
- •Anthropic (2025): Claude speeds up typical 90-minute workplace tasks by ~80%
- •Customer service, coding, legal research: documented time savings of 30–55% in multiple controlled studies
- •BIS study (EU, 2019–2024): AI adoption raises labour productivity by 4% on average across firms
- •US GDP grew 2.2% in 2025 while employment grew just 0.1% — suggesting each worker produced more
- •GitHub Copilot: developers complete tasks 55% faster; code quality improves for junior engineers
❌ Why it's not showing up
- •Individual gains are being consumed by coordination costs, quality checks, and rework loops before reaching the bottom line (Asana, 2025)
- •Workers save time — but take it as leisure, not output. Employers don't capture the gain (St. Louis Fed, 2025)
- •Most AI usage is light: executives average only 1.5 hours/week. 25% of firms don't use it at all (NBER, 2026)
- •90% of US GDP growth in H1 2025 came from data centre investment, not AI-driven productivity (Harvard/Furman)
- •"No signs of AI in profit margins or earnings expectations outside tech." — Apollo chief economist Torsten Slok
- •Worker confidence in AI utility fell 18% in 2025 even as usage rose 13% (ManpowerGroup, 2026)
The case it's just a matter of time
- •Wharton projects AI will contribute 2.7% to US productivity by 2028 as adoption scales — still below optimists' predictions, but meaningful
- •History supports the "J-curve" theory: new technologies reduce productivity first (learning costs, integration), then raise it — the internet boom arrived 30 years after the PC (Brynjolfsson et al., 2026)
- •Agentic AI (like OpenClaw) may be the real catalyst: agents that act autonomously could be the shift that makes productivity measurable at scale — but we're at the very start
What the FT found in S&P 500 filings
An FT analysis of hundreds of corporate filings and earnings call transcripts found that 374 of 500 S&P 500 companies mentioned AI in earnings calls — nearly all framing it positively. But almost none could explain how it was actually improving their business.
Customer service and data companies: clear use cases
Paycom (payroll) called AI "an important differentiator for attracting and retaining clients." These sectors have the most structured, repetitive tasks — where AI works best.
Most companies: innovative uses, no growth correlation
Huntington Ingalls applying AI "for battlefield decisions." Zoetis speeding up horse medical tests. Dover tracking hail-damaged vehicles. Creative — but Dover's stock barely matched the S&P 500 Equal Weighted index since ChatGPT launched.
The honest companies: legal risk, cyber, failure
Many non-tech companies in S&P 500 filings express concern over cybersecurity risks, legal liability, and the potential for AI to fail in practice — a starkly different tone from the optimistic earnings calls.
The new variable: AI agents
The Economist and NYT both flag that the move from chatbots to agents may change the productivity equation — for better and worse.
The upside: tasks that used to require humans can now run autonomously
A San Francisco founder asked an AI agent to arrange a Davos speaking slot while he slept. It succeeded — by texting people, negotiating, and finding connections across the web. Tasks that previously required hours of human effort.
The downside: agents can cause expensive mistakes
The same founder woke up to find his agent had agreed to a 24,000 Swiss franc corporate sponsorship — $31,000 he couldn't pay. Agents operating with too much autonomy can create new costs that dwarf the productivity gains.
Young coders: productivity anxiety, not productivity gains
"When I don't have agents running, I feel this angst," said one 28-year-old founder. Silicon Valley's most AI-immersed workers are managing AI systems like employees — adding a new cognitive overhead that may not net out as "more productive."
Voices on the debate
"AI is everywhere except in the incoming macroeconomic data. You don't see AI in the employment data, productivity data, or inflation data."
Torsten Slok, Apollo chief economist — Fortune, Feb 2026
"That's better than zero. But it's just disappointing relative to the promises that people in the industry and in tech journalism are making."
Daron Acemoglu, MIT / Nobel laureate, on his 0.5% productivity projection — Fortune, Feb 2026
"Artificial intelligence is advancing at startling speed. Its effect on output, not so much."
"The productivity gains are real at the individual level, but they're being consumed by coordination costs, quality taxes, and rework loops before they reach the bottom line."
"Something Big is Happening" [viral essay title] — AI optimists argue we are at an inflection point where task-level gains will begin showing up in macro data within 2–3 years.
Cited in The Economist, Feb 2026
How sources frame the debate
The Economist
Centre-right · UK financial press
Not yet
"Solow, the Sequel" — the most rigorous mainstream analysis. Acknowledges output/employment gap but concludes AI gains are not yet visible and the Bessent/Warsh productivity optimism lacks evidence.
Financial Times
Centre-right · UK financial press
Unclear
Documented that S&P 500 companies talk about AI constantly but can't explain the business impact. Sober and data-driven: real use cases exist, but no correlation to growth.
NYT
Centre-left · US
Mixed
Covers both sides: genuine individual task gains (Cade Metz) and the anxiety/overhead of managing AI agents. Does not take a clear macro verdict but the evidence presented is mixed-to-skeptical.
Fortune
Centre · US business press
Skeptical
Highlighted the NBER CEO study (89% see no impact) and Acemoglu's "disappointing" projection. Strong on the Solow parallel framing.
SF Federal Reserve
Central bank research
Patient
Most balanced. Uses electricity analogy: transformations take time. "Limited evidence of a significant AI effect" in macro-studies so far — but expects gains to come, just slowly.
MIT / Wharton / NBER
Academic research
Task yes
Task-level gains are solid and well-documented. Macro gains are near-zero today, projected to grow but modestly. No consensus on when the inflection happens.
The bottom line
The honest answer is: yes at the desk, not yet at the economy. AI demonstrably helps individual workers with specific tasks. Those gains are not yet visible in GDP, employment, or earnings data at scale. The most likely explanation isn't that AI is a fraud — it's that we're at the beginning of a transformation that historically takes decades to show up in statistics.
The uncomfortable question isn't whether AI will deliver. It's whether the current $250 billion/year in AI investment (2024) is priced on a timeline that matches the historical precedent — which suggests the payoff is real, but years away.