A survey by VentureBeat Research of 573 technical leaders reveals an uncomfortable truth for the enterprise world: most AI agents currently in production are not autonomous agents at all, but simple chatbots responding to one prompt at a time. According to the study, 71% of companies say that a quarter or fewer of their agents can complete multi-step tasks independently. Only 10% can claim a majority of true autonomous agents.
86% of enterprise GPUs run at half capacity or less
The report highlights significant waste in hardware infrastructure: 86% of enterprises that run their own GPUs report utilization of 50% or less. This comes as Wall Street debates whether the AI buildout is overbuilt. Based on direct measurements from buying companies, the research suggests that the most expensive hardware in data centers operates at less than half its capacity. Only 44% of organizations rigorously track the costs and returns of their AI compute, while the rest rely on estimates.
Sponsored Protocol
66% of companies allow or are engineering toward automated agent deployment without human review
An alarming finding concerns trust in automated evaluations. 34% of enterprises already allow an AI agent to push code or system changes to production based solely on automated test results, with no human oversight. Another 33% are actively working to enable this practice within the next 12 months. Yet only 5% fully trust the automated evaluations driving these decisions. The distrust is earned: half of companies shipped an agent that passed internal tests and then caused a customer-facing failure in the past year.
Sponsored Protocol
Credential sharing increases security risks
69% of companies allow credential sharing among agents during runtime, using the same API key or service account for multiple agents. These organizations experienced a security incident or near-miss rate of 63.5%, compared to 40.9% for those that assign each agent its own scoped identity. The study recommends giving every agent its own scoped identity, starting with those that touch production systems.
57% traced incorrect answers to missing or inconsistent business context
More than half of companies traced at least one confident but incorrect agent answer in the past six months to missing or inconsistent business context: wrong metrics, stale definitions, absent documents. To address this, 25% have already implemented a governed semantic layer, a unified business definition that every AI system reads, while 34% are still building one and 41% have not started. The study urges companies to govern their business definitions before scaling the agents that depend on them.
Sponsored Protocol
Enterprises are now racing to catch up, with about six in ten planning to switch or add vendors in each of five control layers (identity, evaluation, cost telemetry, context, orchestration) within the next 12 months. As reported in a related article, the US has eased export restrictions on AI chips, a factor that may influence hardware availability. For further details, see the original article on VentureBeat.
