Most AI productivity numbers you see come from controlled studies of one task. Those numbers are real. But they are often quoted as if they were company results, and the company-level research says something much quieter. This post sets the two side by side, with the exact population and conditions behind each figure.
We build AI integrations through our AI integration service, so weigh that. Every finding below was checked against the papers and official reports themselves in October 2026.
Short answer: In controlled studies, AI made specific tasks 12–56% faster or more productive, mostly for writing, coding and support work. Firm-level surveys find far less: in one survey of nearly 6,000 executives, 89% reported no impact on labour productivity over three years. The gap is closed by redesigning the work, not by buying the tool.
What the Studies Found, at a Glance
As of October 2026, from the papers and official reports:
| Study | Who and what was measured | Finding |
|---|---|---|
| Noy & Zhang (Science, 2023) | 453 college-educated professionals, writing tasks, ChatGPT | Time down 40%, quality up 18% |
| Peng et al. (2023) | Recruited developers writing an HTTP server in JavaScript, GitHub Copilot | Task completed 55.8% faster |
| Brynjolfsson et al. (NBER) | 5,179 customer support agents | 14% more issues resolved per hour; 34% for novices |
| Dell’Acqua et al. (2023) | 758 consultants, 18 tasks | 12.2% more tasks, 25.1% faster inside the frontier; 19 percentage points less likely correct outside it |
| METR (2025) | 16 experienced open-source developers, 246 tasks | AI increased completion time by 19% |
| St. Louis Fed (2025) | US workers, November 2024 survey | Users saved 5.4% of work hours; 1.4% across all workers |
| NBER “Firm Data on AI” (2026) | Nearly 6,000 executives, US, UK, Germany, Australia | 89% report no impact on labour productivity |
| Humlum & Vestergaard (NBER) | Danish workers and workplaces | Null effects on earnings and hours, ruling out effects above 2% |
| Workday (2026) | 3,200 full-time AI users at $100M+ revenue firms | Nearly 40% of time savings lost to rework |
| PwC (2026) | 1,217 senior executives, mainly large listed firms | 20% of firms capture 74% of AI’s economic value |
| Acemoglu (NBER) | Macroeconomic model | No more than 0.66% TFP gain over 10 years |
Sources: Noy & Zhang, Peng et al., Brynjolfsson et al., Dell’Acqua et al., METR, St. Louis Fed, NBER w34836, Humlum & Vestergaard, Workday, PwC, Acemoglu.
What Controlled Studies Say About Task-Level Gains
A controlled study picks one group of workers, one task and one tool, then compares people with and without the tool. It strips out the friction of a real company: handoffs, approvals, checking, and the work around the task. Within those limits, the gains are large.
Where AI Consistently Delivers
Writing is the clearest case. In a preregistered experiment, 453 college-educated professionals did occupation-specific writing tasks, and half were given ChatGPT. Their average time fell by 40% and output quality rose by 18%, and the gap between stronger and weaker writers narrowed (source).
Coding shows a similar pattern on a contained task. Developers asked to build an HTTP server in JavaScript as fast as possible finished 55.8% faster with GitHub Copilot than the control group (source).
Customer support is the best real-workplace evidence. A study of 5,179 support agents found that an AI assistant raised issues resolved per hour by 14% on average. Novice and low-skilled agents improved by 34%, while experienced, highly skilled agents saw minimal impact (source).
Consulting work gains too, within limits. In a field experiment with 758 Boston Consulting Group consultants, people using AI on 18 realistic tasks inside the AI’s capabilities completed 12.2% more tasks, 25.1% more quickly, with significantly better quality (source).
Where AI Adds Minimal Value
The same consulting study chose one complex managerial task that sat outside what the AI could do well. There, consultants using AI were 19 percentage points less likely to produce a correct answer (source). The tool did not just fail to help. It made people worse.
Expert work in familiar territory can also slow down. In a randomised trial, 16 experienced developers completed 246 tasks in mature open-source projects they had worked on for an average of 5 years. They expected AI to cut their time by 24%, and afterwards believed it had cut it by 20%. In fact, allowing AI increased completion time by 19% (source).
Two lessons follow. Skilled people gain least on work they already do well. And people are poor judges of their own AI speed-up, so self-reported time savings need checking.
What Firm-Level Research Says: The Productivity Paradox
Firm-level research asks a different question: did the company as a whole get more productive? Here the numbers shrink sharply.
NBER Survey: 89% of Firms Report No Productivity Impact
The largest recent survey is “Firm Data on AI”, an NBER working paper that surveyed nearly 6,000 senior executives at US, UK, German and Australian firms. It is not peer-reviewed. 69% of firms actively use AI, yet executives’ own use averages only 1.5 hours a week (source).
89% of firms report no impact on labour productivity, measured as sales per employee, over the last three years. By country, that was 91% in the US and Germany, 89% in the UK and 79% in Australia. Executives expect AI to raise productivity at their firms by an average of 1.4% over the next three years, and 60% of firms expect no impact at all (source).
The paper puts the past three-year boost at around 0.29% across all firms. That is an estimate the authors build by assigning set values to each answer band, not a measured gain.
Danish data points the same way. Linking adoption surveys to official records, Humlum and Vestergaard find precise null effects on earnings and recorded hours for both workers and workplaces, ruling out effects larger than 2% two years after ChatGPT launched. Instead, employers absorb AI by reorganising tasks, adding new work such as content generation, AI oversight and AI integration (source).
Worker surveys are a little more positive. In a November 2024 US survey, generative AI users saved an average of 5.4% of their work hours, about 2.2 hours for a 40-hour week. Across all workers, including non-users, that falls to 1.4% of total hours, which the authors translate into a 1.1% rise in aggregate productivity (source).
Goldman Sachs: No Economy-Wide Relationship Yet
Goldman’s newer 2026 analysis is published for clients and is not public, so we do not quote it. Its public position, from May 2024, is plain. Economist Joseph Briggs said “We haven’t seen much of an impact on productivity growth so far,” because adoption was still limited, with about 5% of companies reporting regular production use of generative AI. Their forecasts did not assume any AI boost before 2027 (source).
The long-run macro view is also modest. Acemoglu’s model estimates no more than a 0.66% increase in total factor productivity over 10 years, and possibly less than 0.53% once harder tasks are included, because the early evidence comes from easy-to-learn tasks (source).
PwC 2026: 20% of Firms Capture 74% of AI Value
PwC’s 2026 AI Performance Study interviewed 1,217 senior executives, mainly at large, publicly listed companies across 25 sectors. It found that 74% of AI’s economic value is captured by 20% of organisations, while most businesses remain stuck in pilot mode. Performance was measured as revenue and efficiency gains attributed to AI, adjusted against industry medians (source).
The Variable That Separates Winners From Everyone Else
Workflow Redesign, Not Tool Selection, Drives Results
PwC’s leaders are twice as likely to redesign workflows around AI rather than simply add AI tools. They are about two to three times more likely to use AI for growth and business-model change, and the single strongest factor in their results was capturing growth from industry convergence (source). The Danish finding that firms absorb AI through task reorganisation fits the same story (source).
Note who was studied: large companies. The direction is useful for a smaller business. The exact ratios may not carry over.
The Rework Problem: 40% of AI Time Savings Are Erased
Workday surveyed 3,200 full-time employees at organisations with $100M+ in annual revenue, all active AI users. 85% said AI saves them one to seven hours a week. But nearly 40% of those savings are lost to rework: correcting errors, rewriting content and verifying outputs. Only 14% consistently get clear, positive net outcomes (source).
The same report found that in 89% of organisations, fewer than half of roles have been updated to reflect AI capabilities (source). Workday sells workplace software, so read it as vendor research. Still, the point matches the METR result: saved time on paper is not saved time in practice.
What This Means for SMBs Making AI Decisions in 2026
Which Functions Show Documented Gains
None of the studies above was run on small businesses specifically. What they do document is which kinds of tasks gain, and under what conditions:
- Drafting and writing: 40% less time and 18% higher quality on professional writing tasks (source).
- Contained coding tasks: 55.8% faster on a defined build task (source), but slower for experts on mature code they know well (source).
- Customer support: 14% more issues resolved per hour, mostly from newer staff (source).
- Analysis inside the tool’s range: more tasks done, faster and better, with a real accuracy drop outside that range (source).
If a function in your business looks like one of these, the evidence is on your side. If it does not, treat any number you are quoted as a guess.
How to Evaluate an AI Vendor’s Productivity Claims
Ask four questions before you accept a productivity figure:
- Was it measured on a task or on the business? A task speed-up is not a firm-level gain. Ask which one the number describes.
- Who was in the sample, and how big was it? A study of 16 developers and a survey of 3,200 employees answer different questions. So does a sample of large listed companies.
- Is the saving net of checking and rework? If review time is left out, the figure is too high.
- Was it measured or self-reported? People overestimated their own AI speed-up in the METR trial. Ask for before-and-after measures, and for clients you can speak to.
Related Posts
- What AI ROI Actually Looks Like for Mid-Market Businesses: turning time saved into money saved
- AI Automation Output Verification: Design It Before You Ship It: keeping rework from eating the gain
- Human review checkpoints in AI automation: where a person approves the output
- When AI Is Not the Answer: operations where it underperforms
Frequently Asked Questions
What is the average productivity gain from using AI tools?
There is no single average. Task studies range from 14% more support issues resolved per hour to 55.8% faster on a coding task, and one expert trial found a 19% slowdown. Across all US workers, a 2024 survey put the saving at 1.4% of total hours (source).
Why do most companies report no productivity improvement from AI?
Use is still light, and firms add tools without changing the work around them. In the NBER survey, executives use AI about 1.5 hours a week and 89% of firms report no productivity impact (source). Rework also eats part of the saving.
Which roles see the biggest productivity gains from AI?
Newer and less experienced staff on well-defined tasks. Novice support agents gained 34%, against minimal impact for experienced ones (source). Writing experiments also found the gap between workers narrowed.
How long does it take to see ROI from AI adoption for a small business?
No study above measures payback time for small businesses, so any fixed timeline is a guess. Start with one function that matches the evidence, measure it before and after, and include review time in the result.
Is AI productivity data reliable or mostly vendor-produced?
Both exist. Peer-reviewed experiments and NBER papers are the strongest evidence. Vendor and consulting surveys, such as Workday’s and PwC’s, are useful but self-reported and drawn from large companies. Check who funded the study and who was in the sample.
Can AI make work slower?
Yes. Experienced developers working on code they knew well took 19% longer with AI, while believing it had sped them up (source). Consultants on a task outside the AI’s range were 19 percentage points less likely to get it right (source).