# AI Productivity Gains: What the Studies Show

Author: Marin Niehues, Niehues Consulting · Published: 2025-09-01 · Updated: 2026-11-16 · AI Strategy & Business Case
Source: https://www.niehuesconsulting.com/blog/ai-productivity-gains-studies/

Generative AI raises productivity by 14 percent in customer service and by up to 56 percent on a single programming task in field studies and experiments, depending on task and experience. The range is wide because the studies measure different things. A randomised study also shows that AI can slow experienced specialists down on complex tasks. For executive management, what matters is how much of the measured gain arrives as a result in the company's own operations. This article places the key studies in context and shows how to translate their findings into a robust plan.

**Studies measure gains at the level of individual tasks. Whether they become a result for the company is decided by delivery, with process and ownership as well as enablement.**

## What the most cited studies measure

The evidence comes from a manageable number of papers that compare employees with and without AI support under comparable conditions. Four of them shape the debate.

Brynjolfsson, Li and Raymond studied the work of more than 5,000 customer service agents at a software company, first as a working paper in 2023 and published after peer review in the Quarterly Journal of Economics in 2025. The field study uses the staggered rollout of an AI assistant in live operations. Issues resolved per hour rose by 14 percent on average. For novice and less skilled workers the increase was 34 percent. For the most experienced agents the effect was small.

Peng and colleagues, researchers at GitHub and Microsoft, compared developers with and without GitHub Copilot in a 2023 experiment with 95 participants. The group with the assistant completed a single defined task, building an HTTP server in JavaScript, 55.8 percent faster.

In a 2023 paper in Science, Noy and Zhang had 453 college-educated professionals complete occupation-specific writing tasks. With ChatGPT, time taken fell by around 40 percent. Quality as rated by evaluators rose by around 18 percent.

In a 2023 Harvard Business School working paper, Dell'Acqua and colleagues tested 758 consultants at an international strategy consultancy. On tasks inside the AI's capability frontier, participants completed 12.2 percent more tasks. They worked 25.1 percent faster, and quality was more than 40 percent higher. On a task outside that frontier, the rate of correct solutions with AI was 19 percentage points lower than without.

**The findings at a glance**
14 percent | more issues resolved per hour in customer service; field study in live operations with over 5,000 agents (Brynjolfsson, Li, Raymond)
34 percent | increase for novice workers in the same field study
55.8 percent | faster completion of a single programming task; experiment with 95 participants (Peng et al.)
40 percent | less time on writing tasks, with 18 percent higher quality; experiment with 453 participants (Noy, Zhang)
19 percentage points | lower accuracy on tasks outside the capability frontier; experiment with 758 participants (Dell'Acqua et al.)
19 percent | longer completion time for experienced developers in their own codebases; randomised study with 16 developers and 246 tasks (METR)

**Exhibit 1: Depending on the task, generative AI ranges from a clear gain to a measurable loss**

| Label | Value |
| --- | --- |
| Programming task, experiment | 55.8% faster |
| Writing tasks, experiment | 40% less time |
| Customer service, field study | 14% more issues resolved |
| Tasks outside the capability frontier | 19 percentage points less accurate |
| Experienced developers, own codebases | 19% more time |

Source: Brynjolfsson, Li, Raymond, Quarterly Journal of Economics, 2025; Peng, Kalliamvakou, Cihon, Demirer, 2023; Noy, Zhang, Science, 2023; Dell'Acqua et al., Harvard Business School, 2023; Becker, Rush et al., METR, 2025
Note: measures differ by study; negative values mean a worse result with AI
Graphic: [SVG](https://www.niehuesconsulting.com/blog/ai-productivity-gains-studies/grafik-1.svg) · [CSV](https://www.niehuesconsulting.com/blog/ai-productivity-gains-studies/grafik-1.csv) · [JSON](https://www.niehuesconsulting.com/blog/ai-productivity-gains-studies/grafik-1.json) · Niehues Consulting, [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/deed.en). Data: Brynjolfsson, Li, Raymond, Quarterly Journal of Economics, 2025; the source's terms apply.

## When does AI reduce productivity?

AI reduces productivity when it is used for tasks outside its strengths, or when experienced specialists in complex, familiar environments spend more time checking and correcting than they save. Two studies show this.

In a 2025 randomised study, the research organization METR had experienced open-source developers work on real tasks in their own large codebases, with each task randomly assigned to allow or disallow AI tools. With AI, they took 19 percent longer. Before the study they had expected a speed-up, and even afterwards they rated their productivity with AI as higher. The gain depends on task and context; the tool alone does not deliver it.

The study by Dell'Acqua and colleagues shows the same mechanism in knowledge work. On a task outside the AI's capability frontier, consultants using AI more often delivered wrong results because they followed convincingly worded answers. Two rules follow for planning: tasks are sorted by their position relative to this frontier before rollout, and perceived time savings do not replace measurement. How to assess the effect in software development is described in the article [AI in software development](/blog/ai-in-software-development-productivity/).

## How many hours per week does AI save?

Employees who use generative AI report saving around 5 percent of their working time on average, a little over 2 hours in a 40-hour week. That is the finding of a representative survey by the Federal Reserve Bank of St. Louis. Across all workers, including non-users, the saving was 1.4 percent of working hours.

A study by Humlum and Vestergaard covering around 25,000 workers in eleven occupations in Denmark arrives at a similar order of magnitude of around 3 percent. At the same time it found no measurable effects on earnings or recorded hours. The time saved flowed into other activities, part of it into new tasks created by the AI itself. That is the empirical core of the question why gains are missing from the financial figures.

**Exhibit 2: In everyday work generative AI saves a few percent of working time, 1.4 percent across all workers**

| Label | Value |
| --- | --- |
| Users of generative AI, self-reported | about 5% |
| Eleven occupations in Denmark | about 3% |
| All workers including non-users | 1.4% |

Source: Bick, Blandin, Deming, Federal Reserve Bank of St. Louis, 2025; Humlum, Vestergaard, NBER Working Paper, 2025
Note: Share of working time saved by generative AI
Graphic: [SVG](https://www.niehuesconsulting.com/blog/ai-productivity-gains-studies/grafik-2.svg) · [CSV](https://www.niehuesconsulting.com/blog/ai-productivity-gains-studies/grafik-2.csv) · [JSON](https://www.niehuesconsulting.com/blog/ai-productivity-gains-studies/grafik-2.json) · Niehues Consulting, [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/deed.en). Data: Bick, Blandin, Deming, Federal Reserve Bank of St. Louis, 2025; the source's terms apply.

## How much productivity does AI bring to the economy as a whole?

For the economy as a whole, economists expect a noticeable but limited effect: Daron Acemoglu estimates the gain in total factor productivity from AI over the next ten years at no more than around 0.66 percent in total. Estimates from consultancies and banks are several times higher because they assume broader automation and faster adoption.

In Germany, adoption limits the effect. According to the Federal Statistical Office, 20 percent of companies with 10 or more employees used AI in 2024, rising to 48 percent among companies with 250 or more. The economy-wide gain grows with this share and with the depth of use in processes.

## Why study results do not transfer directly

Four characteristics of the studies limit how far they transfer to a business.

The first concerns the level of measurement. Three of the four papers measure individual tasks under study conditions; the customer service study measures in live operations, but only one metric of one role, not an entire process. A 40 percent time saving on a subtask that accounts for a fifth of a role's working time reduces the role's time requirement by 8 percent. The process changes only when upstream and downstream steps are adapted.

The second concerns task selection. The researchers chose tasks that can be clearly defined and assessed. In day-to-day work, tasks are less sharply defined, and quality criteria often remain implicit.

The third concerns variation by experience. In customer service, novices benefited far more than experienced staff. For planning, the composition of the workforce therefore determines the gain.

The fourth concerns the capability frontier of the technology. The same technology that speeds up one task lowers accuracy on another.

## A calculation framework between study and result

The findings yield a framework that divisional heads can use to estimate the gain for their area. It combines four factors multiplicatively.

**The framework in four factors**
- Time share: What share of working time goes to the task that AI supports?
- Efficiency gain: By how much does AI shorten or improve the task, measured against your own baseline?
- Usage rate: What share of the target group uses the tool regularly in the intended process?
- Utilization rate: What share of the freed-up time flows into additional output or higher quality?

The product of the four factors gives the expected gain.

The first two factors can be measured in a prototype. The last two emerge only in operation and in practice decide the order of magnitude. A tool with a 30 percent efficiency gain that only half the target group uses, and whose freed-up time is half utilized, delivers a quarter of the measured efficiency gain.

For measurement in your own operations, a before-and-after comparison with a control group is recommended, as the studies themselves use. The approach, including the ROI calculation, is described in the article on [measuring the ROI of AI projects](/blog/measuring-ai-roi/). Suitable metrics are lead time per case and quality rate, supplemented by the usage rate.

## Where the gain lands in operations: the delivery perspective

The studies show that the technology works. Whether the gain reaches the company is decided by delivery, where four recurring breaking points appear.

The first breaking point is the process. If AI sits next to the existing process instead of being built into it, extra work arises from checking and transferring between systems. The time gained on the subtask is lost along the way.

The second is ownership. Without named owners for usage and impact, the usage rate is left to chance.

The third is enablement. Employees get different amounts out of the same tool. Enablement aims to ensure that experienced staff also use the tool effectively and recognize when to leave it aside.

The fourth is utilization. Freed-up time becomes value only once it is decided what it will be used for: more throughput, shorter delivery times or a capacity adjustment. Without this decision, the time spreads across other activities.

That many initiatives never reach these points is described in the article [Why AI pilots do not reach operations](/blog/why-ai-pilots-fail-to-reach-operations/). How executive management steers an AI portfolio as a whole is covered in the article [AI management: what executive leadership must steer](/blog/ai-management-what-executives-must-steer/).

## Niehues Consulting as a partner from study to result

[Niehues Consulting](/) supports companies from assessment through to live operation and works from a delivery perspective. Marin Niehues brings experience from projects ranging from VC-funded scale-ups to DAX 40 groups, in 20+ countries on six continents.

In AI management, business analysis takes shape together with a business case that fills the four factors with your own data. Selection and sequencing of applications follow, as does steering the AI portfolio with metrics. AI prototypes prove the benefit in the real process, and the transfer into operation takes place with internal or external teams. The approach follows the Niehues Delivery Model with the phases Understand, Build, Roll out. Three of its review questions matter most for productivity gains: What is the baseline, and how do we measure success? Is the business case confirmed with real values? Which metric shows after 90 days that the solution is landing?

The entry point is a free 30-minute initial consultation in which the four factors are assessed for your most important use case. [Book an appointment](/#kontakt).

**What decision-makers should take away**
The evidence shows real gains that vary widely by task and experience and turn negative on unsuitable tasks. Four factors count for planning: time share and efficiency gain as well as usage and utilization rates. The last two emerge in operation. Measure in your own process with a control group, and decide before rollout what freed-up time will be used for.

**Sources**
- Brynjolfsson, Li, Raymond: Generative AI at Work, Quarterly Journal of Economics, 2025 (NBER Working Paper 2023)
- Peng, Kalliamvakou, Cihon, Demirer: The Impact of AI on Developer Productivity, arXiv, 2023
- Noy, Zhang: Experimental evidence on the productivity effects of generative artificial intelligence, Science, 2023
- Dell'Acqua et al.: Navigating the Jagged Technological Frontier, Harvard Business School Working Paper, 2023
- Becker, Rush et al. (METR): Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 2025
- Bick, Blandin, Deming: The Impact of Generative AI on Work Productivity, Federal Reserve Bank of St. Louis, 2025
- Humlum, Vestergaard: Large Language Models, Small Labor Market Effects, NBER Working Paper, 2025
- Acemoglu: The Simple Macroeconomics of AI, Economic Policy, 2025
- Statistisches Bundesamt (Destatis): Jedes fünfte Unternehmen nutzt künstliche Intelligenz, press release no. 444, 2024

## Frequently asked questions

### How much does generative AI increase productivity?

Study results range from 14 percent more resolved issues in customer service, through around 40 percent less time on writing tasks, to 55.8 percent faster completion of a programming task. These figures apply to individual tasks. For a company, the gain follows only from time share and usage as well as utilization.

### Can AI decrease productivity?

Yes, a randomised study found 19 percent longer completion times with AI for experienced developers working in their own codebases. On tasks outside the AI's capability frontier, the rate of correct results also falls. Check with a measurement for each task whether AI speeds it up.

### How many hours a week does AI save?

Users of generative AI report saving around 5 percent of their working time on average, a little over 2 hours in a 40-hour week. Across all employees the saving is considerably lower because many do not use the tool. For your company, what counts is the time measured in your own process.

### How do you measure AI productivity gains in your own company?

The most reliable method is to compare a group with AI support against a group without, measured by lead time and quality on the same task. The baseline is captured before launch, the usage rate in parallel. Self-reports alone overstate the effect.

### Why do productivity gains not show up in the financial figures?

Because freed-up time becomes a result only when the company redeploys it. If a team saves hours each week, this affects costs or revenue only when the time is used for additional output or a capacity adjustment. A study from Denmark has so far found no measurable effects on earnings or working hours.

### Do experienced or less experienced employees benefit more from AI?

Less experienced employees benefit considerably more in the studies, with 34 percent in customer service against 14 percent on average. AI makes experiential knowledge more widely available and shortens onboarding. Experienced staff gain mainly on the routine parts of their work.
