Enterprise agent cost controls remain post-hoc for 21% in Pulse survey

Enterprise agent cost controls remain incomplete for a meaningful share of a self-selected group of AI-active organizations: 21% said they track agent spending only through logs reviewed after the fact, with no real-time means to halt a runaway execution loop, according to a July 2026 VentureBeat Pulse Research survey. The finding sits alongside reported spending priorities led by agent monitoring and debugging at 31% and security and permissions enforcement at 30%.

The survey covered 107 organizations with at least 100 employees. It is a single, cross-sectional wave, not a probability sample, and it skewed toward large companies and technology respondents. More than half of participants worked at companies with 10,000 or more employees, while technology and software made up 53% of the sample. The results are directional for that cohort, rather than a market-wide estimate.

The practical distinction is narrow but consequential. The survey records whether respondents reported post-hoc spend logging and whether they had a real-time stop mechanism. It does not establish why those controls were absent, which products were responsible, or whether monitoring and permission controls work better than spend controls.

What does the survey show about real-time agent spend controls?

It shows that observing agent activity and enforcing policies are prominent stated investment areas, but that one in five respondents reported no real-time mechanism to interrupt a runaway execution before costs accrue. The survey does not measure cost attribution by team or workload, budget enforcement, dollar exposure, or the effectiveness of any specific vendor's tools.

For context, IBM defines agent orchestration as coordinating multiple specialized AI agents in one system toward shared objectives. In practice, respondents to the Pulse survey reported that deployment is still at an early stage: 47% said between 26% and 50% of their so-called agents were genuinely orchestrated; 37% put the share at a quarter or below; and 16% said more than half were orchestrated.

Several platforms are the norm in this survey cohort

Most respondents reported using more than one orchestration platform. Eighty-five percent used two or more, 64% used three or more, and the average was 3.1 platforms per organization. Those are overlapping footprints, not exclusive vendor market shares.

Microsoft AI Foundry/Copilot Studio appeared in 70% of respondents' stacks, OpenAI's Agents SDK in 68%, and Anthropic's Claude Platform in 47%. When asked to identify one primary platform, 46 respondents selected more than one answer and could not be counted. Among the remaining 61 unambiguous answers, Microsoft led with 41%, followed by Anthropic at 28%, LangChain/LangGraph at 10%, and OpenAI at 7%.

Flexibility across models and tools was the most frequently reported buying factor, cited by 29% of respondents, ahead of security and permissions at 17%, production reliability and execution control at 15% each, and alignment with a preferred base model at 10%.

Looking toward the end of 2026, 53% expected a hybrid control plane combining provider-native and external orchestration. Security and permissioning limitations were the most-cited concern about provider-resident control, at 37%, ahead of vendor lock-in at 23% and limited visibility at 22%. Operators evaluating those concerns may consider them as part of an enterprise security program, alongside the separate question of whether an agent can be stopped in real time when its spending runs out of bounds.

This story draws on original reporting from VentureBeat.