AI agent performance jumped on benchmarks, but only about 2 in 10 respondents have scaled agents organization-wide

Key findings

Terms used in this report

• AI agents, in McKinsey's 2025 report, are "systems based on foundation models capable of acting in the real world, planning and executing multiple steps in a workflow." Unlike a chatbot that only answers questions, an agent calls tools and keeps working through a task.
• The scaling phase is one of the adoption stages in McKinsey's survey. The report describes it as "expanding the deployment and adoption of the technology within at least one business function." It comes after experimenting and piloting. In this report, organization-wide scaling refers to responses that McKinsey's 2026 report describes as "reaching the scaling phase across their organization." It is a stricter bar than scaling within a single function.
• EBIT is earnings before interest and taxes, a measure of profit from operations. McKinsey separately counts respondents who say AI use contributed any amount to EBIT.
• A share of respondents is not a share of companies. McKinsey's survey asks individuals about their own organization, and to correct for different response rates by country, it weights each response by the respondent's country's share of global GDP. This report says "respondents," as the source does.
• OSWorld is a benchmark that gives agents 369 tasks on Ubuntu, Windows and macOS, covering desktop and web apps, file operations and workflows across several apps. The human baseline of 72.35% is the share of tasks solved by computer science students.
• Telemetry is usage data that software records automatically while it runs. Instead of asking people, it aggregates those records.

Capability: on benchmarks, agents are close to the human baseline

The AI Index 2026 describes 2025 as the year agents "advanced from answering questions to completing tasks." The evidence comes from benchmarks that have agents work in real environments and then check the resulting state. Every value below appears in the AI Index 2026 text or figures, and the original data source for each benchmark is listed alongside.

Two qualifiers are easy to lose when these numbers are repeated.

• "SWE-bench went from 60% to nearly 100%" is not a resolve rate. The AI Index summary says SWE-bench Verified performance rose from about 60% to close to 100% in one year. That value comes from an index the AI Index built to compare different benchmarks on one scale: it converts the best model's score each year into a percentage of the human baseline. The actual resolve rate in the body of the same report is about 76.8% as of February 2026, with the top models clustered between 70% and 76%.
• The 12% in "from 12% to 66%" is not a single year's value. The AI Index body says the strongest models "have historically reached only 1%–12% success" on OSWorld, and the summary shortens this to "from roughly 12% to 66.3%." The same summary adds that agents still fail roughly one in three attempts on structured benchmarks.

Benchmark scores are measured on fixed task sets. The AI Index itself notes that these benchmarks "may not fully capture the breadth of capabilities required for real-world AI deployment." That is why they need to be read alongside the enterprise surveys in the next section.

Enterprise adoption: in the surveys, scaling is still limited to some companies and some functions

Most of the enterprise adoption figures in the AI Index 2026 are taken from McKinsey's 2025 survey. The AI Index itself says these figures are self-reported survey data and should be read as showing direction rather than a complete picture. McKinsey published its 2026 survey results on 2026-08-25, so this report puts the 2025 and 2026 results side by side.

Three things stand out in the table.

• "88% AI adoption" and "single-digit agent scaling" are different questions in the same survey. The 88% asks whether the organization regularly uses AI in any one function. Agent scaling is a much narrower question. Summarizing the 2025 survey, the AI Index writes that "across most business functions, a majority of respondents reported no agent use at all" and that "scaled use was in the single digits for nearly all functions." Even in IT and knowledge management, the functions with the most activity, about two-thirds or more of respondents reported no agent use.
• Scaling is concentrated in large companies and the technology sector. In the 2026 survey, the share of respondents from organizations with more than $1 billion in revenue who report scaling agents rose from 27% to 40%, while smaller organizations stayed at 22%. In the 2025 survey, the technology sector had higher rates of scaled agent use than other industries in software engineering (24%), IT (22%) and service operations (21%) (AI Index Figure 4.3.8).
• Financial impact has not followed wider use. The share saying AI contributed to EBIT went from 39% to 37%, roughly the same. "AI high performers," who attribute 5% or more of EBIT to AI and say they have seen "significant" value from it, were about 6% of respondents in both years. Meanwhile, 80% of respondents said AI had improved their own productivity.

The surveys also asked separately about barriers to scaling AI agents. In the 2025 responsible AI survey run by the AI Index and McKinsey (which excludes responses from China), 62% named security and risk concerns as the main obstacle to fully scaling agentic AI, followed by technical limitations and regulatory uncertainty at 38% each. This barrier is consistent with the June 2026 note's emphasis on "permissions, audit logs and execution control." But these surveys did not measure whether such controls actually speed up scaling.

The word "adoption" covers three different measurements

In the AI Index 2026 summary, 88%, 53% and 28.3% all appear under the word "adoption." Tracing them back to their sources shows that each counts something different.

• The 88% is about organizations, not people. If one employee in one function uses AI regularly, the organization counts as an adopter.
• The 53% is a U.S. figure. The AI Index summary calls it "53% population adoption" without naming a country, but a footnote in the body says the sample is U.S. adults aged 18–64.
• In the U.S. alone, survey and telemetry methods produce results that differ by nearly twofold. Survey-based figures are 53% to 61.8%; Microsoft telemetry gives 28.3% to 33.0%. The age range, period and method all differ, so these sources cannot settle which is right. Microsoft notes that its metric "reflects usage, not capability or impact."
• South Korea showed large gains on Microsoft's metric. It went from 25.9% in the first half of 2025 to 30.7% in the second half, 37.1% in Q1 2026 and 40.6% in Q2 2026, and Microsoft says South Korea posted the largest gain of any major economy for three consecutive reporting periods. This is a measure of individual generative AI use; none of the sources in this report give a separate agent adoption rate for Korean companies.

Revisiting the claims in the June 2026 note

The original note covered a "Top 10 trends" list, business opportunities and a recommended architecture all at once, but did not give evidence for each claim. We separated what we could and could not confirm this time.

Checklist

What is not settled yet

• This report is a 2026-10-06 rewrite, to our current standard, of a note first published on 2026-06-16. The original note's Top 10 ranking and recommended architecture were removed because their basis could not be verified. We made no new measurements; every figure is a value stated in the sources listed below.
• Benchmark values are as of when the AI Index 2026 compiled them. The AI Index says domain performance results reflect early 2026 unless otherwise noted, so scores from later models are not included. We did not recheck the current rankings on each leaderboard.
• Enterprise figures come from self-reported surveys. Several people from one organization can respond, and responses are weighted by each country's share of global GDP. The 2025 (1,993) and 2026 (1,719) samples differ. McKinsey says that looking only at the 552 respondents who answered in both years gave the same results on workforce changes, but it does not report the same check for other questions.
• We could not reconcile one figure in McKinsey's text. The 2025 report says "approximately one-third" of respondents report scaling AI organization-wide, while the 2026 report gives the prior-year value as 38%. We could not confirm from the article text whether these refer to the same question.
• We read McKinsey's reports through Internet Archive copies. We could not access mckinsey.com directly, so we read a copy saved on 2026-09-21 and could not check detailed exhibit values that are not in the article text. For per-function figures we used only the AI Index's citations and McKinsey's own sentences.
• Microsoft's usage metric is built by a vendor from its own telemetry. The adjustment method is described in Microsoft's technical paper; we did not verify it. South Korea's rank and gains are comparisons within this metric.
• This report has no agent adoption data specifically for Korean companies. McKinsey does not publish country-level figures in its article text.

References

All sources were retrieved and checked on 2026-10-06.

The 2026-06-16 initial note was rewritten on 2026-10-06 to our current standard. Sources the original note listed, such as Gartner, NVIDIA and Hugging Face, were not used for any claim in this report and have been removed from the list. If the AI Index or McKinsey revises its figures, we will update this report and note the date.