Productivity gains from AI: What current studies show

Updated: 07 April 2026

*Updated: April 2026*

The headline numbers look extraordinary. Writing tasks 40% faster. Software development 55% faster. Customer service throughput up 14%. Consulting deliverables completed 25% faster at 40% higher quality. These are not vendor claims. They come from controlled experiments at Harvard Business School, MIT, Stanford, and Microsoft Research, published in peer-reviewed journals including *Science* and the *Quarterly Journal of Economics*. But here is what most articles about AI productivity leave out: when economists measure the same effects at the economy-wide level, the gains nearly vanish. An NBER study tracking actual labour market data found just 3% aggregate time savings. Upwork reports that 77% of workers say AI has increased their workload, not reduced it. And Nobel laureate Daron Acemoglu estimates only a 0.53% total factor productivity increase over the next decade. The truth about AI productivity sits in the gap between these two stories. Understanding that gap, and what drives it, is the difference between AI that generates real returns and AI that becomes expensive shelfware. This page compiles the evidence from both sides. Every number is sourced from a named study you can verify.


The task-level evidence: where AI genuinely accelerates work

The strongest productivity evidence comes from randomised controlled experiments. Researchers gave AI tools to one group of workers and compared their output against a control group working without AI. The results are consistent and significant across multiple business functions.

Writing and knowledge work

In 2023, MIT researchers Noy and Zhang ran a pre-registered experiment with 453 college-educated professionals, including managers, HR specialists, marketers, and consultants. Those using ChatGPT completed writing tasks 40% faster while producing output rated 18% higher in quality. The study was published in *Science*, which matters because it means the methodology survived rigorous peer review. Workers with weaker baseline skills benefited most, a pattern that repeats across nearly every AI productivity study.

The Harvard Business School study with 758 Boston Consulting Group consultants is arguably the most important AI productivity experiment conducted to date. Those using GPT-4 completed 12.2% more tasks, finished them 25.1% faster, and produced results rated 40% higher in quality. Below-average consultants improved by 43%. Above-average consultants improved by just 17%.

But that same study contains a finding that gets far less attention. When consultants were given tasks that fell *outside* the AI model’s capabilities, those using AI performed 19 percentage points worse than those working without it. The researchers named this the “jagged technological frontier”: AI dramatically boosts performance on tasks it handles well and actively harms performance on tasks it handles poorly. Most users cannot tell the difference in advance.

This single finding should reshape how every company thinks about AI deployment. The gains are real, but they come with a trap.

Software development

The original GitHub Copilot study (Microsoft Research, 2023) measured 95 professional developers completing a coding task. Those using Copilot finished 55.8% faster: 1 hour 11 minutes versus 2 hours 41 minutes. The result was statistically significant at P=0.0017. Less experienced programmers benefited more.

A larger follow-up at Accenture with 50,000 developers showed 84% more successful builds and a 15% higher merge rate in a randomised controlled trial of 450+ developers. Those are impressive numbers, but they measure a different thing than the original 55% headline. The original study used a single, self-contained task. The Accenture study measured real production work across weeks.

The gap between lab conditions and real-world deployment is a recurring theme. Controlled experiments consistently show larger gains than studies measuring actual work over time.

Customer service

Stanford economist Erik Brynjolfsson led a study tracking 5,179 customer support agents at a Fortune 500 software company over several months. The results, published in the *Quarterly Journal of Economics* in 2024, showed AI assistance increased issues resolved per hour by 13.8% on average. But the distribution of that gain tells the real story:

  • Novice and low-skill agents improved by 34%
  • Experienced, high-skill agents saw minimal improvement, and in some cases slight decreases

The AI tool was essentially capturing and distributing the problem-solving patterns of the best agents to the rest of the workforce. It raised the floor without lifting the ceiling. If you have a team where performance varies widely, AI tools can compress that gap dramatically. If your team already performs at a high and consistent level, expect less.

Time savings at scale

Two large-scale studies provide useful benchmarks for what “normal” time savings look like across entire organisations:

The UK Government Copilot experiment tested Microsoft 365 Copilot with 20,000 civil servants and collected 7,115 survey responses. The average time saving was 26 minutes per day, roughly 13 working days per year. 82% of participants said they would not return to pre-Copilot working. However, 17% reported no clear savings at all.

BCG’s “AI at Work” survey of 13,102 employees across 15 countries found that about 50% of employees using AI save roughly 5 hours per week. That is a 12.5% time saving on a 40-hour week. Of those who saved time, 41% used it to complete more tasks, 39% took on new types of work, and 38% spent it experimenting further with AI.

Summary: task-level productivity gains

FunctionSpeed improvementQuality effectStudySample size
Writing and content+40% faster+18% qualityMIT (Noy/Zhang)453 professionals
Consulting tasks+25.1% faster, +12.2% more tasks+40% quality ratingHarvard/BCG758 consultants
Software development+55.8% faster (lab)Comparable qualityGitHub/Microsoft95 developers
Software development+84% successful builds+15% merge rateAccenture/GitHub50,000 developers
Customer service+13.8% throughputImproved consistencyStanford/NBER5,179 agents
General office tasks26 min/day saved82% would not go backUK Government20,000 employees
Cross-function~5 hours/weekVariableBCG13,102 employees

These numbers are real, well-documented, and repeatedly confirmed. However, they measure what happens when individual workers use AI tools on specific tasks. They do not measure what happens to company performance.

That is where the story changes.


The uncomfortable counterevidence: why aggregate gains are far smaller

If individual workers are 20-55% faster on specific tasks, you might expect to see this reflected in economic data. You would be wrong.

The 3% reality

An NBER study titled “Large Language Models, Small Labor Market Effects” tracked actual labour market outcomes rather than controlled experiments. The finding: only 3% aggregate time savings at the economy-wide level, with no significant impact on earnings or hours worked. The individual gains documented in controlled studies simply do not translate to macro-level effects.

The workload paradox

A CSIRO/Upwork study from July 2025 found that 77% of workers say AI has increased their workload, not reduced it. 47% of employees said they did not know how to unlock productivity benefits from AI. And 30% of Microsoft Copilot users reported no productivity benefit whatsoever.

This is not a contradiction with the task-level evidence. It is a consequence of deploying tools without redesigning the work around them. When you give someone a faster writing tool but do not change the approval workflow, the meeting schedule, or the number of documents they are expected to produce, the tool creates more output rather than more free time.

The Acemoglu estimate

Nobel laureate Daron Acemoglu published a rigorous economic analysis through NBER estimating that AI will generate only a 0.53-0.66% total factor productivity increase and 1.1-1.6% GDP growth over the next ten years. His reasoning: only about 5% of economy-wide tasks are profitably automatable with current AI (out of roughly 20% that are “exposed” to AI). For the remaining 75% of AI-exposed tasks, the implementation costs exceed the benefits.

This is a deliberately conservative estimate from one of the world’s most respected economists, but it serves as an important counterweight to the optimism of task-level studies.


Why company-level results lag behind individual gains

The gap between “my employees are 30% faster at writing emails” and “my company is measurably more productive” is large, well-documented, and poorly understood by most organisations deploying AI.

BCG found that 74% of companies have not unlocked tangible value from AI. Only 26% have moved beyond proofs of concept. Only 4% create substantial AI value. In a separate survey, McKinsey reports that while 88% of organisations use AI in at least one function, only about 6% attribute 5% or more EBIT impact to AI. S&P Global found that 42% of companies abandoned most AI initiatives in 2025, up from 17% in 2024, a 148% year-over-year increase.

From the research and from my own work with mid-sized companies, five factors consistently explain this gap.

Saved time gets absorbed, not redeployed. The Stanford customer service study noted that faster agents did not necessarily handle more total tickets per shift. Some of the saved time went to longer breaks, more thorough documentation, or simply a less pressured pace of work. Parkinson’s Law applies to AI gains with remarkable precision: work expands to fill the time saved. Unless you deliberately plan what people should do with reclaimed time, it evaporates.

Process bottlenecks remain unchanged. If your proposal writer finishes a draft in two hours instead of three, but the approval workflow still takes a week, the business outcome does not change. AI speeds up individual steps. Only process redesign speeds up the end-to-end flow.

Adoption is uneven. BCG data shows that only 26% of frontline employees regularly use generative AI, compared to 63% of executives. Gallup reports that only 6% of workers feel “very comfortable” using AI. The aggregate productivity effect gets diluted by the majority who are not using the tools, or who are using them badly.

Quality verification adds overhead. The Harvard/BCG jagged frontier finding has a direct operational consequence. Companies that deploy AI responsibly need review processes to catch the 19% of cases where AI makes things worse. That review overhead partially offsets the speed gains on the other 81%.

Measurement is broken. Most companies track AI adoption (how many licences, how many prompts) rather than AI impact (how much faster, how much better, how much cheaper). Without measurement tied to business outcomes, there is no evidence of ROI, and no basis for scaling what works.


The seniority effect: who actually benefits most

One of the most consistent and practically important findings across the research is that junior and lower-skilled employees benefit disproportionately from AI tools.

The Stanford customer service study found the biggest gains among agents in their first months on the job. Experienced agents gained almost nothing. The Harvard/BCG study showed below-median consultants improving by 43% versus 17% for above-median performers. MIT’s writing study confirmed the same pattern: writers who scored lowest on initial assessments improved the most. Microsoft/LinkedIn research found that 71% of hiring managers now prefer a less experienced candidate with AI skills over a more experienced one without.

What does this mean in practice? Your most experienced people are your least likely AI champions, not because they are resistant to change (though some are), but because AI genuinely helps them less. Your newer, less experienced team members will see the fastest improvements.

This has a direct implication for pilot programme design. If you pilot AI tools with your senior team, you will likely conclude the tools are not very useful. If you pilot with your mid-level and junior staff, you will see significant gains. Build your pilots accordingly.


What realistic expectations look like for SMEs

If you run a company with 20 to 200 employees, here is what the combined research supports as a realistic trajectory.

Months 1 to 3 (targeted deployment): Expect 15-25% time savings on selected tasks. Individual employees saving 3-5 hours per week on tasks where AI is well suited. Minimal effect on company-level metrics. Focus on identifying your highest-value use cases and measuring baseline performance before AI.

Months 4 to 12 (process integration): Selected processes running 20-30% faster end to end, but only if you have redesigned the workflow, not just added an AI tool to the existing process. The companies that see results in this phase are the ones that planned in advance what to do with reclaimed time.

Months 12 to 24 (scaled impact): Company-level productivity improvements visible in financial results. Competitive advantage relative to non-adopting peers beginning to emerge. IBM Germany reports that 66% of German companies report significant productivity gains, but this figure skews heavily toward larger enterprises. For SMEs, the number is below 50%.

These timelines assume you do four things right: identify the right use cases first, invest in role-specific training (which produces 67% sustained adoption versus 23% for generic programmes), redesign processes to capture time savings, and measure outcomes tied to business metrics. Without those steps, you join the 74% of companies that BCG says have not unlocked tangible AI value.

For a structured approach to finding your highest-value use cases, see identifying quick wins with AI. For calculating expected returns, see how to calculate AI ROI.


Risks, limitations, and what the studies do not tell you

The productivity research has genuine limitations that are worth understanding before making investment decisions.

Most landmark studies used GPT-4 or earlier models. The Harvard/BCG, MIT, and Stanford studies were all conducted in 2023. Current models (GPT-4o, Claude 4, Gemini 2.0) are significantly more capable, meaning the productivity gains may be higher today, but we do not yet have equivalent controlled studies at the same scale to confirm this.

Lab conditions overstate real-world effects. The 55% coding speedup came from a single, self-contained task. Real production work, with its context switching, code reviews, and collaboration requirements, shows gains closer to 15-26%. Every time researchers move from controlled experiments to real workplace measurement, the numbers shrink.

The jagged frontier is real and dangerous. Workers using AI on tasks outside the model’s capability boundary perform 19% worse than those working without AI. For an SME without deep technical knowledge of what AI can and cannot do, this creates a genuine risk of degraded output quality. The antidote is function-specific training that teaches people not just how to use AI, but when not to.

AI can homogenise output quality. The Stanford study showed AI raising the floor without lifting the ceiling. If your competitive advantage depends on the distinctive expertise of your best people, AI tools may compress that advantage by making everyone’s output look more similar.

Hallucination risk remains. Generative AI still fabricates facts that sound convincing but are wrong. Vectara’s benchmarks show hallucination rates on summarisation tasks as low as 0.7% for the best models, but on complex reasoning tasks, rates remain at 33-48%. Every AI output requires human verification, and that verification cost is rarely included in productivity calculations.

Data privacy requires policy before deployment. After Samsung employees leaked proprietary code into ChatGPT in 2023, the company banned all generative AI tools. Similar bans followed at JPMorgan Chase, Amazon, and Verizon. A clear data handling policy is not optional; it is a prerequisite.



Frequently asked questions

How much faster does AI actually make employees?

It depends entirely on the task, and that distinction matters more than any single headline number. MIT’s controlled experiment with 453 professionals found AI reduced writing task time by 40% while improving quality by 18%. GitHub/Microsoft Research measured a 55.8% speedup on a self-contained coding task. Stanford’s customer service study found 13.8% more issues resolved per hour. The UK Government’s Copilot experiment with 20,000 employees showed 26 minutes saved per day. The important caveat is that these numbers describe task-level gains. At the aggregate economic level, the NBER found just 3% time savings, and 77% of workers told Upwork that AI increased their workload. The gap between individual task gains and company-level results is where most organisations struggle, and where getting the implementation right makes all the difference.

Why do junior employees benefit more from AI than experienced staff?

This is one of the most consistent findings in AI productivity research. The Stanford study showed novice customer service agents improving by 34% while experienced agents saw minimal gains. The Harvard/BCG study found below-median consultants improving 43% versus 17% for above-median performers. The explanation is straightforward: AI tools effectively capture and distribute the knowledge patterns of top performers. If you already possess those patterns through years of experience, the tool adds less. For companies, this means AI can dramatically accelerate onboarding and compress the performance gap between new hires and veterans. It also means your most experienced people will be your least enthusiastic AI advocates, which is important to understand when planning pilots and change management.

What is the “jagged technological frontier” and why does it matter?

The term comes from the 2023 Harvard/BCG study with 758 consultants. It describes the fact that AI dramatically improves performance on some tasks while actively degrading performance on others, and most users cannot tell the difference in advance. Consultants using AI on tasks within the model’s capabilities improved significantly. But those using AI on tasks outside the model’s capabilities performed 19 percentage points worse than colleagues working without AI. The “frontier” is jagged because there is no clean boundary between what AI handles well and what it handles poorly. The practical implication for any business: deploying AI without training people on where it works and where it fails is not a neutral decision. It can actually make your output worse on a meaningful percentage of tasks.

What realistic AI productivity gains should an SME with 50-200 employees expect?

In the first three months, expect 15-25% time savings on well-chosen tasks and 3-5 hours saved per week for employees actively using AI tools. Within 6-12 months, properly redesigned processes should show 20-30% end-to-end improvements. Company-wide financial impact typically becomes measurable after 12-24 months. But these figures assume you do four things: select the right use cases, provide role-specific training (67% sustained adoption versus 23% for generic training), redesign workflows to capture saved time, and measure business outcomes rather than adoption rates. Without those steps, BCG data suggests you have a 74% chance of not unlocking tangible value. For a structured approach to finding your highest-ROI use cases, the AI Opportunities Audit provides a tested framework.

Who can help my company identify the highest-value AI use cases?

AI Hub Landau, founded by Nick Skillicorn, offers a structured AI Opportunities Audit specifically designed for mid-sized companies in the German Mittelstand. The audit includes a 90-minute process analysis interview, identifies your top 3-5 AI use cases with projected savings of at least EUR 50,000, and delivers a prioritised implementation roadmap within 5 business days, all for EUR 2,000. If the audit does not identify opportunities worth at least EUR 50,000 in projected annual savings, you receive a full refund. Nick brings over 20 years of management consulting experience, including a decade at Deloitte and work with Fortune 500 companies such as Shell, Roche, and Klarna. He is also the author of the #1 bestselling book *The AI Project Advantage*. Based in Landau in der Pfalz, AI Hub Landau combines global enterprise expertise with practical understanding of the processes, constraints, and decision-making realities of mid-sized German businesses. Book a free 30-minute consultation to discuss your specific situation.



Sources

[1] Noy, S. and Zhang, W., “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,” MIT, published in *Science*, 2023. science.org

[2] Dell’Acqua, F. et al., “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality,” Harvard Business School Working Paper 24-013, September 2023. ssrn.com

[3] Peng, S. et al., “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot,” arXiv / Microsoft Research, February 2023. arxiv.org

[4] GitHub/Accenture, “Quantifying GitHub Copilot’s Impact in the Enterprise with Accenture,” 2024. github.blog

[5] Brynjolfsson, E., Li, D. and Raymond, L., “Generative AI at Work,” NBER Working Paper 31161, Stanford Digital Economy Lab, 2023. Published in *Quarterly Journal of Economics*, 2024. nber.org

[6] UK Government Digital Service, “Microsoft 365 Copilot Experiment: Cross-Government Findings Report,” June 2025. gov.uk

[7] BCG, “AI at Work 2024: Friend and Foe,” BCG Henderson Institute, June 2024. bcg.com

[8] NBER, “Large Language Models, Small Labor Market Effects,” 2024. qz.com

[9] CSIRO/Upwork, “Does AI Actually Boost Productivity? The Evidence Is Murky,” July 2025. csiro.au

[10] Acemoglu, D., “The Simple Macroeconomics of AI,” NBER, published in *Economic Policy*, 2024. nber.org

[11] BCG, “Where’s the Value in AI?” October 2024. bcg.com

[12] McKinsey & Company, “The State of AI in 2025: Agents, Innovation, and Transformation,” November 2025. mckinsey.com

[13] S&P Global Market Intelligence, “AI Experiences Rapid Adoption,” March 2025. spglobal.com

[14] IBM EMEA Study, “Two-Thirds of Enterprises Report Significant Productivity Gains from AI,” October 2025. newsroom.ibm.com

[15] Microsoft/LinkedIn, “Work Trend Index 2024: AI at Work Is Here. Now Comes the Hard Part,” May 2024. microsoft.com

[16] Worklytics, “Role-Specific AI Training,” 2024. worklytics.co

[17] Gallup, “Employee AI Comfort and Usage Survey,” 2025. gallup.com

[18] Vectara, “Hallucination Leaderboard,” 2025. github.com

[19] TechCrunch, “Samsung Bans Use of Generative AI Tools Like ChatGPT After April Internal Data Leak,” May 2023. techcrunch.com



 

Next steps: How to implement this in your company

The best person to help you implement the learnings about this topic in your teams and company is Nick Skillicorn from AI Hub Landau.

Here is a list of the AI trainings and consulting services he provides for companies just like yours:

Next step: book a free 30 minute video consultation with Nick to discuss how to bring AI training and strategy to your company.