Enterprise AI agents in production: what the mid-2026 surveys actually show

Two surveys published in June give the clearest recent picture of how many organizations run AI agents in production, and how well they run them. The headline numbers agree: a little over half. The numbers underneath the headline are more instructive, because they show where the gap between "deployed" and "managed" sits.

Two surveys, two different populations

LangChain's State of Agent Engineering collected 1,340 responses between November 18 and December 2, 2025, and was published on June 12, 2026. It reports that 57.3% of respondents have agents running in production, up from 51% a year earlier, and a further 30.4% are building with plans to deploy ^[1].

KPMG's Q2 2026 AI Pulse, released June 24, surveyed 204 U.S. C-suite and business leaders at organizations with at least $1 billion in annual revenue, fielded April 28 to May 25. It puts agent deployment at 53%, "on par" with 55% the previous quarter ^[2].

These are not the same measurement. LangChain's sample is drawn from people building with agent tooling and was fielded about seven months before publication. KPMG's is a small panel of senior leaders at large companies, and "deployed" is the respondent's own definition. Our reading is that the LangChain figure likely overstates the share of the wider market, because builders are more likely to respond to a builder survey, but the report does not quantify that effect. The fair summary is that production agents are now common among organizations actively working on them, not that every other company has one.

The more interesting number: multi-agent doubled

KPMG reports that the share of organizations orchestrating multiple agents across workflows doubled from 9% to 18% in a quarter, even as overall deployment was flat ^[2]. Single-agent deployments are mostly bounded tasks. Orchestration means agents handing work to each other, which raises the stakes for tracing, permissions and cost, because a failure in one step propagates.

What is blocking production

LangChain asked what stands in the way. About a third (32%) named quality as the primary blocker, 20% named latency, and cost was cited less often than in earlier years. Among enterprises with 2,000 or more employees, security ranked second at 24.9% ^[1]. The most common use cases were customer service (26.5%), research and data analysis (24.4%) and internal workflow automation (18%) ^[1].

The KPMG data points the other way on cost. Thirty-five percent of leaders cite AI cost management and economic literacy as a barrier, only 26% have full, real-time visibility into what their AI systems cost to run, and only 36% have direct token or usage controls ^[2]. Both can be true: individual builders find cost secondary to quality, while finance-minded executives find it hard to see and govern. Leaders also plan to invest a weighted average of $202 million over the next 12 months, versus $207 million the quarter before ^[2].

KPMG also found employee resistance rising to 20% from 5% in the prior quarter, with trust and ethics concerns behind 53% of it and workload or complexity concerns behind 51% ^[2]. That is a single-quarter move in a small panel, so we would watch it rather than build a plan around it.

Observability is where production teams differ

The LangChain data shows a consistent pattern: organizations with agents in production monitor and test more.

Among teams with agents in production, 22.8% are not evaluating their agents, compared with 29.5% of all respondents ^[1]. In other words, nearly one in four teams running agents live reports no evaluation.

A dependency the surveys do not capture

Neither survey asks what happens when the underlying model is withdrawn or unavailable. An agent deployment counted as "in production" in a survey may still have a single point of failure that no one has tested.

What to take from this

If your agents are live or about to be, the useful questions are whether you can see what they cost per workflow, trace a bad outcome end to end, and switch models without a rebuild. Those are cheaper to put in place before a second or third agent depends on the first.