Enterprises plan AI vendor churn as agent governance lags
Companies are preparing to replace or add AI control vendors within the next year, creating a significant market shift as they race to secure autonomous systems they deployed before proper safeguards existed.
Between 57% and 68% of enterprises intend to switch or add vendors for AI agent controls within 12 months, with roughly a third planning to do so this quarter. This churn spans five critical layers: identity, evaluation, cost tracking, data context, and orchestration. The spending surge reflects a belated scramble to retrofit governance onto systems that companies knowingly launched ahead of proper safety measures.
The urgency stems from a stark disconnect between the autonomous agent narrative and operational reality. A June survey of 573 enterprise technology buyers found that 71% of companies report a quarter or fewer of their so-called "agents" can actually complete multi-step tasks independently. Only 10% run a majority of true autonomous agents, suggesting most current deployments are single-prompt chatbots that do not yet require the rigorous controls true autonomy demands.
Security and reliability risks mount
Governance shortcuts are already impacting operational stability. Half of the enterprises surveyed reported a customer-facing failure in the past year caused by an agent that had passed internal evaluations. Despite this, two-thirds of companies either already allow automated agent code pushes to production without human review or plan to within a year. This is occurring even though just 5% of organizations fully trust the evaluation frameworks making those deployment decisions.
Credential mismanagement is compounding the security exposure. Sixty-nine percent of companies let at least some agents share API keys or service accounts. Among those allowing credential sharing, 47 out of 74 experienced a security incident or near-miss—a 63.5% failure rate. That compares to a 40.9% incident rate at companies where every agent operates under its own scoped identity.
Capital expenditure on AI infrastructure is also facing scrutiny. Over 80% of enterprises running their own GPUs report utilization rates of 50% or less, yet only 44% rigorously track what their AI compute actually costs and returns. For investors, the immediate financial opportunity lies not in selling more hardware, but in the software that optimizes utilization and per-workload costs of existing chips.
Flawed internal data presents another drag on ROI. Fifty-seven percent of enterprises traced confident but incorrect agent responses in the past six months directly to missing or inconsistent business definitions. Because no single control layer has an entrenched incumbent, the default tools shipping with major AI platforms are highly vulnerable to displacement. Orchestration platforms face the highest switching intent, with 68% of companies planning vendor changes within a year, creating a wide-open window for specialist challengers.
