How Much Does AI Increase Productivity? The Data From Real Studies

The headline numbers

56%

Average productivity gain for software developers using GitHub Copilot — measured by task completion speed in Microsoft's randomised controlled trial [Microsoft/GitHub]

40%

Productivity increase for business writers using AI assistance in the MIT/Harvard study — with higher-quality output ratings from independent judges [Noy & Zhang]

34%

Increase in issues resolved per hour for customer support agents using AI copilot — from Brynjolfsson et al. Stanford randomised trial [Stanford]

The productivity numbers for AI are now backed by randomised controlled trials, not surveys. The results are real, consistent, and larger than most people expected — but they come with important caveats about who benefits most and under what conditions, which is exactly what the productivity illusion pulls apart.

The evidence

MIT + Harvard — Noy & Zhang (2023)

Business writing productivity

Randomised controlled trial with 453 professionals (marketers, consultants, HR managers). Half used ChatGPT for writing tasks. AI group completed tasks 37% faster and received 18% higher quality ratings from independent graders. Notably: the biggest gains went to lower-skilled writers, narrowing quality gaps between workers.

37%

faster completion

GitHub / Microsoft (2023)

Software development with Copilot

95 professional developers asked to complete HTTP server task in JavaScript. Those using GitHub Copilot completed it 56% faster. Follow-up study with 2,000+ developers showed 55% reported increased productivity and 74% said it helped them focus on more satisfying work by handling boilerplate code.

56%

speed improvement

Stanford / Brynjolfsson et al. (2023)

Customer support with AI copilot

5,179 customer service agents at a large tech company. AI copilot provided real-time response suggestions. Average issues resolved per hour increased 14%. New employees improved 34%. Critically: the lowest-performing workers gained the most — top performers showed minimal gains, suggesting AI primarily helps people learn faster.

14%

avg (34% for new staff)

BCG / Fabrizio Dell'Acqua et al. (2023)

Consulting tasks at Boston Consulting Group

758 BCG consultants across 18 countries. Those using GPT-4 for tasks within AI's capability boundary completed 12.2% more tasks, 25% faster, and produced 40% higher quality work. However: on tasks outside AI's capability (requires physical world knowledge), those using AI performed 19% worse — they over-trusted the tool.

40%

quality improvement

Gains by task type

Code generation

+56%

Business writing

+40%

Customer support

+34%

Data analysis

+30%

Document summarisation

+70%

Image creation

+80%

Legal research (review)

+20%

The nuances

Junior workers gain more

Consistently across studies: lower-skilled and newer workers see the largest productivity gains. AI compresses the expertise gap. Senior experts gain less because their bottleneck isn't knowledge access — it's complex judgement AI can't replace.

Quality sometimes drops

The BCG study found workers on tasks outside AI's capability performed worse when using AI — they over-trusted outputs. AI adoption without critical evaluation skills can harm output quality, particularly for complex analytical tasks.

Gains require learning to use AI well

First-week AI users don't see the same gains as experienced users. The MIT study estimated a 2-4 week learning curve before workers internalise effective prompting habits — the kind of patterns laid out in how to use AI at work. Productivity gains compound over time as usage becomes habitual.

Aggregate GDP effects are modest so far

Despite impressive individual-level gains, macroeconomic productivity data hasn't yet shown the AI uplift predicted. The IMF and Goldman Sachs have both noted this gap — suggesting adoption is still in early stages and broad economic effects will take years to appear. Capturing them at an organisational level is what an AI strategy is supposed to do.

The honest bottom line

For well-defined, knowledge-intensive tasks (writing, coding, support, analysis): AI delivers 20-70% productivity gains backed by rigorous research. The gains are real. They're not universal — they require the right tasks, trained users, and critical evaluation of outputs. "AI will make everyone equally productive" is too optimistic. "AI produces measurable, significant productivity gains for knowledge workers" is well-supported.

FAQ

Do the productivity gains hold up over time?

The evidence is limited but encouraging. The GitHub Copilot study tracked developers for several months and found sustained gains. The Stanford customer service study found gains were durable and even grew slightly as agents learned to use AI suggestions more selectively. Novelty effects appear to be a minor factor — the gains are real, not just excitement.

Are productivity gains the same for all industries?

No. Industries with high proportions of knowledge work and information-intensive tasks see the largest gains: software, consulting, legal, marketing, customer service, finance. Industries where work is primarily physical (construction, manufacturing) or requires in-person presence (healthcare, hospitality) see smaller direct productivity gains from current AI tools.

If AI increases my productivity, will my employer reduce my hours?

Historical technological productivity improvements have generally created new work rather than reducing hours. The more immediate effect companies are reporting is redeployment — workers freed from routine tasks taking on higher-value work. Whether this translates to fewer total employees is a longer-term trend still playing out. The 4-day work week movement points to interest in capturing productivity gains as time rather than output.

Sources

[Noy] Noy & Zhang — "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence" (MIT, 2023)

[GitHub] GitHub — "Research: Quantifying GitHub Copilot's impact in the enterprise" (2023)

[Brynjolfsson] Brynjolfsson, Li, Raymond — "Generative AI at Work" (Stanford, 2023)

[BCG] Dell'Acqua et al. — "Navigating the Jagged Technological Frontier" (Harvard/BCG, 2023)

How Much Does AI Increase Productivity? The Data From Real Studies

Sources

04 — Don't watch from the outside