A VentureBeat Research study published July 24, 2026 covered five parallel surveys and 573 enterprise leaders. It describes gaps in the controls around deployed agents, with 81% of respondents recommending or deciding AI purchases at their companies.
Seventy-one percent of enterprises said a quarter or fewer of their deployed “agents” could complete multi-step work autonomously. Only 10% said autonomous agents were the majority of what they ran. That finding concerns the capabilities of deployed systems; it does not establish that 71% lacked governance or knowingly deployed without controls.
Start with identity and evaluation
VentureBeat examined five control layers: identity, evaluation, cost telemetry, the context layer, and orchestration. Each requires someone to decide what the workflow needs and verify that the control works.
Sixty-nine percent of companies let at least some agents share credentials. Organizations allowing credential sharing reported security incidents or near-misses at a 63.5% rate, compared with 40.9% where every agent had its own scoped identity. The comparison supports examining shared credentials; it does not show that identity alone explains the difference. Start by mapping production agents to their credentials and checking whether their access matches their jobs.
Two-thirds of enterprises either already allowed agents to push changes to production based on automated evaluation alone, with no human review, or were engineering toward that within 12 months. Only 5% fully trusted those evaluations. Half reported shipping an agent that passed internal evaluations and then caused a customer-facing failure in the past year.
For a deployment team, that is a reason to check what its evaluations cover. Include known failure cases and a review path for consequential changes. Passing an internal test leaves a need to monitor what happens in use.
Track costs and the context behind answers
More than eight in ten enterprises running their own GPUs reported utilization of 50% or less. Only 44% rigorously tracked what their AI compute cost and returned. A useful cost review needs a defined unit of work and enough telemetry to connect resource use with accepted output.
Fifty-seven percent of enterprises traced a confident, wrong agent answer in the past six months to missing or inconsistent business context. Stale definitions, missing documents, and conflicting metrics are examples to check in a workflow review. Assign an owner for the source material and a route for correcting it when reviewers find an error.
Review orchestration before changing vendors
Orchestration coordinates multi-step work. Here, 68% planned to adopt, add, or replace platforms within 12 months, and thirty-four percent planned to move within the quarter. Across all five layers, 57 to 68% planned to switch vendors or add new ones within 12 months; roughly a third, depending on the layer, planned to move within the quarter.
Those intentions describe planned purchasing activity. They leave open how much control each company already has and what a replacement will improve.
Before evaluating another platform, document the control gap it is supposed to address. Name an owner for identity, evaluation, cost visibility, business context, and orchestration. Test the relevant behavior in the current workflow, then make the same test part of the vendor trial. That gives the team a way to judge whether a purchase actually fixes the problem it identified.