Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsiTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Estimate AI productivity gains by comparing the same work with and without the tool, measuring both speed and quality, and stating exactly whose work and which outcome changed. A writing experiment found faster, higher-quality work on its tested assignments; a customer-support rollout found more issues resolved per hour. Neither result is a universal forecast for every worker, company, or the economy.
What counts as an AI productivity gain?
Productivity is output relative to the inputs used to produce it. For an AI-assisted task, a useful starting measure is completed work per labor hour—but output must be defined for the job. A support agent might resolve issues; a writer might deliver a usable draft; a team might complete a project milestone.
Do not treat speed, output, and quality as interchangeable. A shorter task may produce the same amount of work, more work, or work that needs more correction. Likewise, higher-quality work is not necessarily a larger quantity of output. Name the measure whenever you report a gain.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- Time: minutes or hours spent on a defined task.
- Quantity: completed items, tasks, or issues within a defined period.
- Quality: an assessed score, error rate, rework requirement, or recipient outcome.
- Broader outcomes: such as work outside regular hours, customer satisfaction, earnings, or employment.
A gain in one measure does not establish a gain in the others. In particular, time saved is not automatically converted into more output: it may instead reduce overtime, support higher-value work, or remain unused capacity.
#1 Best Overall
How do you measure AI productivity gains?
- Set the unit of analysis. Decide whether the estimate concerns a task, workflow, team, firm, or economy. A result for one task should not be relabeled as a firm-wide effect.
- Define the work and comparison. Specify the task, expected deliverable, quality standard, and comparison condition. State which AI tool or model was available, how it was integrated, what training or access rules applied, and what the comparison group used instead.
- Measure speed and quality together. Record time and output quality for the same work. Where relevant, also track output quantity, errors, rework, human effort, and outcomes for customers or other recipients.
- Choose a measurement window. State when measurement began and how long it lasted. A short task experiment, a phased workplace rollout, and a six-month field experiment answer different questions.
- Separate access from use. Record who was offered access and who actually used the tool. If a study reports outcomes among users as well as outcomes for everyone offered access, keep those estimates distinct.
- Check who benefits. Look for differences by task, skill, or experience. An average can conceal large gains for one group and little change for another.
- Follow the saved time. Measure whether it became additional output, reduced work outside regular hours, supported different tasks, or remained slack. Do not assume a conversion that was not measured.
- Limit the conclusion to the design. Random assignment can support a causal estimate in the tested setting. Applying that estimate to different jobs, firms, or the broader economy requires additional evidence.
How much does generative AI improve productivity?
There is no single rate that applies across work. The studies below measure different tasks and outcomes, use different designs, and observe different populations. Treat the figures as setting-specific evidence, not as directly comparable entries in a league table.
| Study and setting | Design and population | Reported result | What the measure means |
|---|---|---|---|
| Noy and Zhang, professional writing (2023) | Preregistered online experiment; 453 college-educated professionals assigned incentivized, occupation-specific writing tasks, with half randomly given ChatGPT access. Result summarized by Stanford SCALE. | Average time taken fell 40%; output quality rose 18%. | Task time and assessed writing quality in that experiment—not expected savings across all knowledge work. |
| Brynjolfsson, Li, and Raymond, customer support (2023) | Workplace study of a staggered introduction of a generative AI assistant, using data from 5,179 customer-support agents. | The NBER Working Paper reports a 14% average increase in issues resolved per hour, including a 34% improvement for novice and low-skilled workers; effects were minimal for experienced and highly skilled workers. The NBER Digest reports 13.8% more issues resolved per hour, about 9% less time per chat, and no significant change in customer satisfaction. | Resolution rate and chat time in a workplace rollout. The two NBER summaries describe the same study, not independent replications. |
| Dillon, Jaffe, Immorlica, and Stanton, work patterns across firms (2025) | Field experiment across 66 firms and 7,137 knowledge workers; workers were randomly selected for access to a generative AI tool in familiar work applications during a six-month experiment. | In the second half, the 80% of treated workers who used the tool spent two fewer hours on email each week and reduced work outside regular hours. Researchers did not detect shifts in task quantity or composition from individual-level AI provision. | Email time and work patterns among users, alongside a separate finding about tasks. The reported time change is not evidence of increased total output. |
| Dell’Acqua and co-authors, product innovation (2025 paper; published online in 2026) | Preregistered field experiment with 776 professionals working on real product innovation challenges at Procter & Gamble. | Individuals using AI matched the performance of teams without AI on the experiment’s challenges. | Performance on those specific challenges and measures; it does not establish that AI replaces teams in other kinds of work. |
| Cruces, Fernández Meijide, Galiani, Gálvez, and Lombardi, workplace-style problem solving (2026 working paper) | Randomized online experiment with 1,174 adults aged 25–45 completing an incentivized business problem-solving task with or without generative AI, followed by a module without AI. | AI raised task performance for all participants and narrowed the education-based performance gap. In the follow-up, participants did not perform worse than controls without AI, lower-education participants retained part of their gain, and performance was higher when intensive AI use was paired with sustained task effort. | Bounded experimental performance, including a follow-up without AI—not a long-term workplace productivity estimate. |
The OECD’s 2025 summary reports experimental productivity gains ranging from 5% to over 25% across selected customer-support, software-development, and consulting studies. That is a cross-study range across different tasks and measures, not a pooled estimate or general forecast. Its 2025 review also emphasizes that results depend on task and user experience, identifies human–AI collaboration as important, and notes gaps in evidence on long-term business effects and workers’ understanding of AI limitations.
Why do estimates differ so much?
Two percentages are not comparable just because they are both described as productivity gains. Before comparing them, check what changed, for whom, and how it was measured.
Recommended Free Tools
- Task and industry: drafting text, resolving customer issues, and handling email have different outputs and constraints.
- Population and baseline experience: skill and familiarity can affect how much a worker benefits. In the support-agent study, novice and low-skilled workers improved much more than experienced and highly skilled workers.
- Access and actual use: offering a tool is not the same as using it. A study of access estimates a different effect from an analysis limited to users.
- Study design: a randomized online task experiment, a staggered workplace rollout, and a multi-firm field experiment provide different kinds of evidence and support different conclusions.
- Outcome definition: time per task, output quality, issues resolved per hour, and time spent on email are distinct measures.
- Duration and persistence: a result during AI access does not tell you whether it continues after the tool is removed or over a longer period.
- Quality checks and downstream results: a faster first draft may still need review or rework. Where the study measures customer satisfaction or other recipient outcomes, those findings should be reported separately from speed.
When does task-level time saved become a wider productivity gain?
Only when the saved capacity changes what people or organizations produce—or improves another valued outcome—does a task-level time saving translate into a broader benefit. The connection depends on workflow, demand, staffing, quality control, and how the organization uses the time.
The six-month experiment across 66 firms illustrates the distinction: users spent less time on email in the latter half of the study, but researchers did not detect changes in the quantity or composition of workers’ tasks following individual-level AI access. The finding does not show that no other workplace could turn saved time into additional output; it shows why time saved and output gained need to be measured separately.
Evidence on wider effects remains unsettled. The ILO’s June 2026 review drew on experiments, firm-level data, platform studies, and worker and firm surveys in Australia, Denmark, Germany, Korea, Kuwait, the United Kingdom, and the United States. It concluded that gains are real but uneven and often unverified; in the evidence it reviewed, worker-reported time savings of a few percent of working hours had not yet translated into higher measured output, earnings, or employment. This is the ILO’s synthesis of the evidence reviewed, not a claim about every firm. The OECD likewise identifies long-term business effects as a research gap.
Rank #4
- Stay present in every scenario: Every conversation is covered, in person, on calls, and online. 4 MEMS + 1 VPU microphones with AI beamforming capture every voice across the room. Smart Dual-Mode Recording switches automatically between phone calls and in-person. The free Plaud Desktop captures online meetings without a bot
- Walk out of every meeting with notes ready to act on: Plaud Intelligence transcribes in 112 languages with speaker labels and turns each recording into action items, decisions, and follow-ups, structured and ready to use. Choose from 10,000+ customizable templates tailored to your role and industry
- AI summary ready before you reach your desk: Auto Transfer moves each recording to the Plaud app automatically, and AutoFlow transcribes and summarizes so your notes are ready before you are back at your desk. Upgrade anytime to Pro (1,200 min/mo) or Unlimited
- Access your AI workspace anywhere: One connected workspace across Plaud Desktop, Plaud Web, and the Plaud mobile app, so your conversations and finished work follow you everywhere
- Your conversations stay private and yours: Compliant with ISO 27001, ISO 27701, SOC 2, HIPAA, GDPR, and EN 18031, with zero data used to train AI models. Trusted by 2.5M+ professionals, including legal, medical, and business professionals handling sensitive information
How to report an estimate clearly
A useful estimate should let a reader understand both the result and its boundary. Include the following in a report or internal evaluation:
- the task or workflow and the unit being evaluated;
- the AI tool, access conditions, and comparison condition;
- who took part and relevant baseline skill or experience;
- the study design, measurement window, and whether the estimate covers everyone offered access or only users;
- the exact outcome and its unit, plus the quality or error checks used;
- variation across worker groups, if measured;
- what happened to saved time, and whether effects persisted after access ended.
For example, “AI increased productivity by 20%” is incomplete without a defined task and measure. A clearer claim would name the task, population, comparison, outcome, period, and quality conditions—and avoid extending the finding beyond the setting observed.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
