Enterprises Deployed AI Agents Without the Safety Nets. Now They're Paying to Retrofit.
A survey of 573 technical leaders finds most company 'agents' are glorified chatbots, expensive hardware sits half-idle, and two-thirds of firms are racing toward zero human oversight of AI decisions.
Key points
- 86% of enterprises running their own AI chips report those chips operate at 50% capacity or less, based on a June 2025 survey of 573 technical leaders.
- 54% of companies experienced an AI agent security incident or near-miss in the past 12 months.
- 71% of enterprises say a quarter or fewer of their deployed "agents" can actually complete a multi-step task without a human driving each step.
- 34% already let AI agents push changes directly to production systems with no human review, and another 33% are engineering toward that same setup.
- 69% of companies allow multiple agents to share a single login credential, and those companies suffered security incidents at nearly twice the rate of companies that don't.
Enterprises knew their AI safety controls were not ready. They deployed anyway.
That is the headline finding from a VentureBeat Research survey of 573 technical leaders at companies with 100 or more employees, published this month. The results paint a picture of organisations that sprinted to deploy AI agents (software that can carry out multi-step tasks on its own, rather than just answering a single question) and are now scrambling to bolt on the guardrails they skipped.
The bill is already arriving. Fifty-four percent of companies had an AI security incident, or a near-miss caught before real damage, in the past 12 months.
What does this mean for ordinary workers and customers?
It means the AI making decisions at the companies you work for or buy from may be operating with fewer checks than most people assume. Half of surveyed enterprises shipped an AI agent that passed their own internal tests, then caused a customer-facing failure in the real world. A quarter watched that happen more than once.
The hardware story is equally striking. Eighty-six percent of enterprises running their own GPUs (the specialised chips that do the heavy number-crunching AI needs) say those chips run at 50% capacity or less. Companies spent heavily to build AI infrastructure, and most of it sits underused. Yet 45% of these same companies plan to evaluate a specialist AI cloud provider in the next 12 months, and roughly one in three is actively considering chips that are not made by Nvidia.
The agent label itself turns out to be largely marketing. Seventy-one percent of enterprises say a quarter or fewer of their so-called agents can complete a multi-step task without a human steering each step. Most are single-question chatbots wearing an "agent" badge. Analysts at Gartner have a word for this: "agentwashing".
Security is the sharpest near-term risk. Sixty-nine percent of companies let multiple agents share one login credential during operation. Organisations that did this suffered security incidents at a 63.5% rate, compared with 40.9% at companies where every agent has its own separate, limited-access identity.
Wrong answers are also a growing liability. Fifty-seven percent of enterprises traced a confident, incorrect agent answer back to their own missing or outdated business data, like a stale definition or an absent document.
The honest takeaway here is simple: measure what you have before you buy more. Check your existing chip utilisation. Give every agent its own login. And before you remove a human from any approval step, test whether your automated checks actually catch real-world failures, not just the ones in your internal test suite.
