AI Agent Productivity Statistics 2026: Studies and ROI Data

Headline Productivity Numbers

The strongest evidence on AI productivity comes from experiments that compare workers with and without the tool. They agree that gains are real on well-defined tasks, larger for less experienced workers and uneven across tasks. Customer-support agents resolved 14% more issues per hour, software developers completed 26.08% more tasks, consultants finished tasks 25.1% more quickly and professionals took 40% less time on writing tasks. One trial points the other way: experienced open-source developers took 19% longer with early-2025 AI tools.

National data puts those task-level results in proportion. US workers who used generative AI said it saved them 5.4% of their work hours in late 2024, and across all workers, users and non-users together, the time workers reported saving came to 2.2% of work hours in Q2 2026. In Denmark, two years after ChatGPT’s launch, researchers could rule out any effect on earnings or recorded hours larger than 2%.

Time Saved by Function: Controlled Studies

No public dataset reports hours saved by department across many companies. What exists are studies of specific kinds of work, and they measure different things: tasks completed, time on a task, or time on email. Read each row as the result for that task and population, not as a department-wide rate.

Software development has the most trials, and the spread is wide. A 2023 GitHub Copilot experiment found a 55.8% faster completion of one programming task; Google’s internal trial found about 21% less time on a complex enterprise task, with a large confidence interval; the pooled Microsoft, Accenture and Fortune 100 trials found 26.08% more completed tasks. For office work, a field experiment across 66 firms found users of an AI tool built into their email and writing apps spent two fewer hours a week on email, with no detectable change in the quantity or composition of their tasks.

Occupation matters too. In the St. Louis Fed’s survey, workers in computer and mathematical occupations saved 2.5% of their work time against 0.4% in personal service occupations, and only about one user in five said AI saved them four hours or more in a week. For a method that turns these effects into a business case, see our guide to AI agent ROI measurement beyond task completion.

What Workers and Developers Report

Surveys ask people whether AI has helped, and most say it has: 80% of McKinsey’s 2026 respondents report better individual productivity, as do 65% of US employees in organisations implementing AI (Gallup) and more than 80% of the technology professionals in Google’s DORA research. These are perceptions, and the METR trial shows how far perception can drift from measurement: its developers believed AI had cut their completion time by 20% when it had raised it by 19%.

Self-reported time savings also vary with who is asked. OpenAI’s survey of its enterprise users gives 40 to 60 minutes a day, while in Gartner’s surveys 19% of employees reported no time saved at all (2026) and 7% said AI had cost them time (2025). Developer surveys disagree too: 52% of Stack Overflow’s 2025 respondents said AI tools or agents had a positive effect on their productivity, well below the share in Google’s DORA research.

Cost per Task: What Companies Disclose

Cost-per-task comparisons between an AI agent and a person are rarely published with their method, and none of the widely repeated per-ticket or per-review figures could be traced to a source. Company filings are the most reliable alternative, because they are made under securities law and state their basis. Klarna’s is the most detailed: its 2025 annual report on Form 20-F says its AI assistant handled 80% of customer service chats in 2025, does the work Klarna estimates at more than 700 full-time agents and delivered $39 million of cost savings in 2024.

The same filing shows what the savings look like in the accounts. Customer service and operations expenses fell from $240 million in 2023 to $203 million in 2024, then rose 2% in 2025 while transactions grew 25%. Klarna’s headcount fell from 4,352 to 2,831 over the same period, which it attributes to using AI to drive efficiency. These are one company’s results, measured on its own definitions, not a benchmark for other businesses.

For a first estimate of what an agent saves in your own operation, our AI agent ROI calculator for marketing operations sets out the inputs; the figures above are a reminder to count review time and running costs, not only the minutes an agent saves.

Pilot to Production and Time to Value

No public study measures time to first value for vendor agents against custom builds. What surveys do measure is how far pilots get. In Deloitte’s survey of 3,235 leaders, 25% had moved 40% or more of their AI pilots into production; IBM’s CEOs said 16% of AI initiatives had scaled enterprise-wide. Agent adoption is rising quickly among large companies: in KPMG’s Q3 2026 pulse of 314 leaders at US companies with $1 billion or more in revenue, 62% said their organisations were building, deploying or developing agents, and 25% were working on multi-agent systems, up from 6%.

Official statistics show a much lower base across the whole economy. The Census Bureau’s survey for November 2025 to January 2026 found 18% of US firms using AI in a business function, and most of those used it in three or fewer functions. Larger firms adopt first: 37% of firms with at least 250 employees used AI. For adoption detail by function and company size, see our AI agent adoption evidence for 2026.

ROI and Payback: What Executive Surveys Report

Payback is where the evidence is thinnest and most often invented. No public dataset gives median payback by department or use case. The best available figures are executives’ own answers, and they describe years rather than months: most of the 1,854 executives in Europe and the Middle East in Deloitte’s 2025 survey said a typical AI use case reaches satisfactory ROI in two to four years, against a typical payback of seven to 12 months expected for technology investments. Only 6% reported payback within a year.

Agentic AI is further behind. Among Deloitte’s respondents already using agentic AI, 10% said they were realising significant ROI from it, compared with 15% of generative AI users reporting significant, measurable ROI. Broader surveys point the same way: 56% of CEOs in PwC’s survey had seen neither higher revenue nor lower costs from AI, IBM’s CEOs said 25% of AI initiatives had delivered the expected ROI, and 37% of McKinsey’s 2026 respondents attribute any EBIT impact to AI.

A minority of companies are pulling ahead. BCG classified 5% of companies as “future-built” for AI in 2025 and 7.5% in 2026, with another 41% actively scaling; its future-built group reported 2.3 times the shareholder return of laggards. That is an association between AI maturity and performance, not proof that AI produced the difference. Gartner adds a caution for anyone counting savings in headcount: among large organisations deploying autonomous technologies, workforce reduction rates were nearly the same for those reporting higher ROI and those reporting modest or negative results.

Vendor-Reported Agent Metrics

Vendors now publish usage metrics for their agents, and the ones in company filings are the most dependable. Salesforce reports Agentforce annual recurring revenue of more than $1.5 billion at the end of July 2026 and 7.0 billion Agentic Work Units, its count of tasks executed by agents in production, delivered to date. Both definitions are Salesforce’s own, and it widened the ARR definition in that quarter to include Slackbot and Headless 360, so the series is not like for like. Microsoft reports more than 30 million paid Microsoft 365 Copilot seats.

Usage data shows what businesses delegate. Anthropic’s September 2025 Economic Index found 77% of business uses of its API follow automation patterns rather than augmentation. None of these figures measures productivity or ROI directly: revenue, seats and task counts show adoption and spend. Gartner’s estimate that only about 130 of the thousands of vendors marketing agentic AI are real is a reminder to check what a product labelled “agent” actually does.

Where the Productivity Story Breaks

The same research that shows gains also shows where they stop. In the BCG consultant experiment, AI users were 19% less likely to reach a correct answer on a task chosen to sit outside what the model could do well. METR’s developers were slower with AI and did not notice. METR’s late-2025 follow-up suggests that may have changed, but the institute calls its new data only very weak evidence because of selection effects: for example, 30% to 50% of developers chose not to submit some tasks they did not want to do without AI.

Agent reliability is the second limit. METR’s time-horizon measurements, which express task length as the time a human expert needs, show frontier models handling much longer tasks at 50% success than at 80%: 12 hours against 70 minutes for Claude Opus 4.6. In a simulated software company the best agent finished 30% of tasks on its own, and seven open-source multi-agent systems failed on 41% to 86.7% of tasks, each on a different benchmark. Trust follows: 46% of developers distrust the accuracy of AI tools. For why many pilots stall before production, see our analysis of the AI agent scaling gap.

The labour-market evidence is still early. The Census Bureau found AI-related employment decreases in only 2% of US firms, and Stanford’s Digital Economy Lab finds no evidence of widespread, economy-wide displacement, but employment of 22 to 25 year-olds in AI-exposed occupations is now 19% below where it would be had it kept pace with less-exposed peers. Saved time also has to be put to use: in Gartner’s survey of 114 HR leaders, only 7% of organisations gave employees guidelines on how to use the time AI frees up.

Outlook: Forecasts and Macro Estimates

Forecasts of AI’s economic effect differ by an order of magnitude, depending on how much of the work they assume AI can do and how fast companies adopt it. The Penn Wharton Budget Model projects AI raising the level of US productivity and GDP by 1.5% by 2035, with its largest annual boost to total factor productivity growth, 0.2 percentage points, in 2032. Daron Acemoglu’s task-based estimate is no more than 0.66% added to total factor productivity over ten years. Anthropic’s estimate is far higher at 1.8% a year of added labour productivity growth, but it extrapolates Claude’s own estimates of task-level time savings and assumes AI is adopted across the whole economy within ten years.

Research-firm forecasts for agents mostly concern adoption and failure rather than measured productivity. Gartner expects agentic AI to resolve 80% of common customer service issues by 2029, cutting operational costs by 30%, but also expects more than 40% of agentic AI projects to be canceled by the end of 2027 and 40% of enterprises to demote or decommission autonomous agents by 2027 because of governance gaps. Treat each as a forecast with a target year, and compare it with the measured figures above before planning around it. Our agentic AI statistics collection tracks adoption and spending forecasts in more detail.

Figures to Check Before You Quote Them

Several AI agent productivity figures circulate with a named source that did not publish them. Each one below is compared with what the named publisher, or the closest primary source, actually says.

Sources

Every figure on this page was read at the publisher’s own page, paper, data release or filing on October 21, 2026. Each publication is listed once below, grouped by publisher type.