AI Agent Governance Gap: 86% of Enterprises Run GPUs at Half Capacity

What happened

A large-scale survey of enterprise technology leaders has found that most organisations have deployed AI agents before putting adequate governance and control structures in place — and did so knowingly. VentureBeat Research's June 2025 study, drawn from 573 technical leaders at companies with 100 or more employees across five parallel surveys of the so-called "agentic stack," reveals a sector now scrambling to retrofit the guardrails it skipped on the way in.

The research maps five distinct control layers where enterprises are actively building or rebuilding: agent identity (defining which agent can act under whose authority); output evaluation (assessing whether agent work meets quality standards); cost telemetry (tracking what each agent actually costs to run); the context layer (the business data and definitions agents draw on); and orchestration controls (governing how agents coordinate with one another). Across all five, roughly six in ten enterprises plan to switch or add vendors within the next twelve months, with approximately one third intending to act within the current quarter.

The GPU utilisation finding — that 86% of enterprises report their graphics processing units running at half capacity or below — cuts against the prevailing Wall Street narrative of runaway AI infrastructure demand, suggesting that the bottleneck is not compute power but the organisational and governance scaffolding required to deploy it effectively.

Why it matters

For customer experience and service-design practitioners, this is a story about the gap between capability and control — a gap that customers ultimately feel. AI agents are increasingly the front line of service delivery: they handle queries, make recommendations, escalate issues and, in some cases, take consequential actions on a customer's behalf. When the identity, evaluation and orchestration layers governing those agents are underdeveloped, the result is inconsistent, unpredictable and sometimes harmful customer interactions. The "move fast, retrofit later" posture documented here is not merely a technical risk; it is a trust risk.

Behavioural economics offers a useful lens: customers extend trust based on perceived competence and consistency. An agent that behaves erratically — because its cost telemetry is absent, its output is unevaluated, or its permissions are poorly scoped — erodes the psychological safety that underpins loyalty. The mass vendor-switching now planned across all five control layers signals that enterprises are beginning to internalise this, even if the motivation is operational efficiency rather than explicit customer empathy.

By the numbers

• 573 technical leaders surveyed by VentureBeat Research in June 2025, all at companies with 100 or more employees.
• 86% of enterprises report GPU utilisation at half capacity or below, undermining the case that compute is the primary AI constraint.
• ~60% of enterprises plan to switch or add vendors across each of the five agentic control layers within the next twelve months.
• ~1 in 3 enterprises, depending on the layer, plan to make those vendor changes within the current quarter.
• 5 distinct control layers identified: agent identity, output evaluation, cost telemetry, context, and orchestration.

The Renascence take

Most commentary on this research will focus on the infrastructure story — underutilised GPUs, vendor churn, budget cycles. The more consequential finding for anyone who designs or manages customer-facing services is that enterprises have been running agents without knowing whether those agents are doing good work, at what cost, or with what authority. That is not a technology gap; it is a service-design failure dressed in technical language.

The instinct to deploy first and govern later is a classic present-bias trap — the benefits of speed are immediate and visible, while the costs of poor agent behaviour accumulate slowly and are often attributed to something else entirely. What customer-obsessed operators should do right now is treat agent evaluation and identity controls not as an IT backlog item but as a core component of their service promise. If you cannot answer "what did this agent do, why, and was it any good?" you cannot make a credible commitment to your customers — and no amount of compute will fix that.