Explainer

The Productivity Question

AI tools are making individual workers dramatically more productive. But that productivity isn’t showing up in the macro data yet, and it hasn’t translated into job losses. Understanding this gap is central to every prediction on this site.

The story so far: 14–56% individual productivity gains in controlled studies, early signs of macro impact in BLS data, near-zero measured job loss, and economists split on what comes next. This page connects the dots.

01

From productivity to jobs

Higher productivity doesn’t map one-to-one to job loss. History shows three distinct paths that firms take, often simultaneously. The ratio between them determines whether AI creates a net employment boom or a contraction.

What Happens When Workers Get More Productive?

AI-driven productivity gains lead firms down three paths — often simultaneously. Click each to see the research.

AI makes workers more productive

Research

Employment for 22–25 year olds in AI-exposed occupations declined ~16% since late 2022, while experienced workers remained stable or grew.

Brynjolfsson, Chandar & Chen (2025) — Stanford Digital Economy Lab

Corroborated by ADP Research (2025) payroll data showing hiring slowdowns concentrated in entry-level positions within AI-exposed industries.

Example

IBM replaced several hundred back-office HR positions with AI agents, consolidating routine processing that previously required large junior teams.

Research

Customer support agents using AI resolved 14–15% more issues per hour, with gains up to 35% for the least experienced workers. Microsoft/Accenture field experiments showed ~26% increase in completed pull requests for developers.

Brynjolfsson, Li & Raymond (2025) — Quarterly Journal of Economics

Why gains are uneven: Gans (2026) — formalizes “jagged intelligence” — AI boosts productivity on some tasks while degrading it on similar-looking ones, making local calibration critical.

Example

A 50-person support team handles the volume that previously required 70 people, with faster resolution times and higher customer satisfaction scores.

Research

This is Jevons Paradox applied to labor: efficiency doesn't shrink demand, it grows it. When AI dramatically reduces the cost of work, previously unviable projects become viable. The net effect depends on two conditions:

A: Demand elasticity

Does cheaper output create more demand?

B: High task exposure

Enough AI-augmented tasks to change the cost equation?

Agentic coding example: A project needs 50 engineers but the ROI doesn't justify it — so the company hires 0. AI agents make it a 10-engineer problem. Now the company hires 10. Net: +10 engineers, not -40.

Historical precedent: ATMs didn't eliminate bank tellers — they made branches cheaper to operate, so banks opened more branches, and total teller employment rose for decades (Bessen, 2015). Spreadsheets didn't eliminate accountants — they made analysis affordable, expanding the market for accounting services.

Note: expansion is a net effect, not frictionless. Some workers will be displaced at individual firms even as the sector grows overall.

Explore the full demand elasticity framework

Example

A marketing agency uses AI to cut production costs 40%, then expands from 3 service lines to 7, hiring specialists for the new offerings. A mid-sized company starts a software project that was previously too expensive, hiring 10 AI-augmented engineers for work that would have required 50 before.

The Productivity J-Curve

Most firms do all three at once. And it takes time — measured productivity often dips before it rises as firms invest in AI tools, reorganize workflows, and retrain workers before reaping the gains. The same pattern played out with electricity and computers.

The automation tools tracked above are the leading indicator. The BLS data tells you which path each industry is taking.

02

The core tension

Three facts coexist uncomfortably. First: at the task level, AI tools are producing some of the largest productivity gains ever measured in economics experiments. Second: at the economy level, aggregate productivity is only slightly above trend. Third: despite widespread predictions of AI-driven job loss, macro employment data shows essentially none.

These three facts aren’t contradictory. They’re the predictable signature of a technology still in its adoption phase. The research below helps explain why.

~21%

Median task-level gain

Median of 18 micro studies

+2.2%

Macro productivity vs. trend

BLS data, Q4 2025 vs. CBO forecast

~0%

Measured job loss

Yale, Goldman, Dallas Fed

03

What the research shows

The gap between micro and macro is the central puzzle. Individual workers are clearly more productive with AI, but the economy-wide numbers are just starting to move. Here are the studies, sorted by effect size.

What the research actually shows

18 task & firm studies on productivity and 6 macro studies on economy-wide effects. The micro evidence leans heavily positive; early signs of macro gains are now appearing in BLS data. Click any row for details and source.

Task & Firm Productivity

RCTs, field experiments, and firm-level surveys measuring individual, team, and business productivity changes with AI tools

14 positive1 negative3 nullMedian: +21%
0%Productivity change →

Macro Productivity

Economy-wide studies measuring aggregate productivity gains: TFP, labor productivity, and time savings

3 small positive3 negligible

Key gap narrowing: Micro studies show 14–56% individual productivity gains. BLS data now shows labor productivity 2.2% above pre-pandemic forecast, but how much is AI vs. other factors remains an open question.

24 studies · Studies range from peer-reviewed (Science, QJE) to working papers · Click any row for source

Why the micro-macro gap?

1
Adoption is still earlyCensus BTOS data shows only ~7% of US firms actively use AI in production. Even if individual gains are large, the aggregate effect is diluted by the 93% not yet using it.
2
Complementary investment takes timeFirms need to reorganize workflows, retrain workers, and redesign processes before AI tools deliver their full value. This is the J-Curve effect.
3
Measurement lags realityNational statistics treat intangible investments (workflow redesign, training) as expenses, not capital. So measured productivity actually looks worse during the investment phase.
4
Task gains don't sum linearlyA worker who is 50% faster at writing code may not be 50% more productive overall. Bottlenecks shift to meetings, reviews, and coordination.
04

What economists expect

If the micro gains are real, how much will they eventually move the macro needle? This question determines everything downstream: displacement forecasts, wage projections, policy responses.

The labor market effects above depend fundamentally on how much AI actually moves the productivity needle. Here's what leading economists currently think.

What economists expect from AI-driven productivity

Probability estimates for US productivity growth 2025–35, from Jason Furman's February 2026 exercise on X. 12 economists, 5 scenarios. Click any bar for source details.

Individual economistCrowd average (n=23)12 economists · 5 scenarios · Feb 2026
Stagnation< 1%
2%
1%
2%
10%
20%
1%
18%
5%
15%
15%
5%
12%

% probability

Status Quo1–2%
35%
20%
49%
20%
35%
1%
65%
35%
35%
40%
50%
35%

% probability

Solid Breakout2–3%
50%
60%
43%
30%
30%
5%
12%
42%
35%
30%
39%
27%

% probability

Boom3–4%
10%
15%
6%
30%
12%
35%
3%
15%
10%
10%
5%
17%

% probability

Phase Change> 4%
3%
4%
0%
10%
3%
60%
2%
3%
5%
5%
1%
9%

% probability

Based on 12 economists · X exercise by @jasonfurman · February 2026 · Click any bar for source

05

When productivity creates demand

There’s a critical counterforce to displacement that most forecasts underweight: when AI makes work dramatically cheaper, it can unlock demand that didn’t previously exist.

Explore demand elasticity and Jevons Paradox
06

The measurement problem

There’s a deeper issue: even when AI does boost productivity, our statistics may systematically undercount it. This isn’t speculation — it’s a pattern that has played out with every major technology adoption.

The Productivity J-Curve

When a major new technology arrives, measured productivity often stagnates or declines for years — not because the technology doesn’t work, but because the massive complementary investments it requires are invisible to official statistics. Electricity took ~40 years. Computers took ~25. AI’s timeline is the open question.

Read the full J-Curve explainerBrynjolfsson, Rock & Syverson (2021)
Historical patternElectricity: ~40 yrsComputers: ~25 yrsAI: ???
07

What this means

The productivity story is the engine behind every other prediction on this site. If AI productivity gains stay confined to narrow tasks and early adopters, displacement will be modest and gradual. If they compound economy-wide — which is what most economists expect within the next decade — the effects on jobs, wages, and industry structure will be much larger.

The current evidence suggests we are still in the early innings. Micro gains are real and large. Macro gains are just beginning to appear. Displacement is near zero. The question is not whether AI improves productivity, but how fast that improvement diffuses, and whether firms use it to reduce headcount, amplify output, or expand into new work.

For workers

Task-level productivity gains are reshaping job descriptions faster than they're eliminating jobs. Adaptability matters more than any specific skill.

For firms

Simply deploying AI tools without redesigning workflows captures almost none of the value. The complementary investment is where the real returns come from.

For policymakers

The J-Curve means the window for proactive policy is now, during the investment phase, before the productivity surge makes labor market transitions sudden.

For forecasters

Anyone predicting mass displacement from current productivity data is premature. Anyone dismissing AI's labor impact because displacement hasn't happened yet is equally wrong.

Productivity is the bridge between AI capabilities and labor market outcomes. Until we understand how that bridge works — how gains diffuse, how firms respond, how measurement catches up — every displacement forecast is incomplete.

Share this view