Companies routinely spend over $10M on agentic AI with limited understanding

According to EY’s latest U.S. AI Pulse Survey, corporate budgets are pouring into agentic AI projects across industries, but most decision-makers still don’t know what they’re buying.

The term is getting tossed around everywhere, slapped onto anything that remotely smells like generative AI, and executives are greenlighting tens of millions with no real grasp of what these systems actually do. That’s creating a disconnect between cash and capability, and it’s not small.

One out of every five senior leaders surveyed said their company already dropped over $10 million into AI, and nearly one-third plan to do the same next year.

Dan Diasio, EY’s global AI leader, didn’t seem surprised. “Agentic AI has a buzz about it that many in the market want to capitalize on,” Dan said. “We’ve seen an incredible rebranding of anything related to generative AI presented as ‘agentic AI.’”

The problem? Most of what companies are calling agentic still works like an assistant. You type something in, it spits something out. It might recommend a next step or automate some admin work, but it’s not doing anything independently. Dan said real agents know when a task needs doing, understand the context, and handle every step without being told.

Leaders spend more, adopt less

Despite the surge in spending, implementation is crawling. Only 14% of surveyed leaders said their company had fully rolled out agentic AI. Everyone else is stuck in pilot purgatory. Dan said the gap is because companies aren’t ready for the demands.

“This includes having organized, high-quality knowledge to guide these systems and a clear understanding of how companies navigate the massive change between the current and future states.” Translation: no foundation, no rollout.

Even with returns from earlier AI tools, most firms are hesitant to move forward. Dan said it’s the mix of technical weakness and change resistance that’s slowing things down. “While this combination creates a climate of uncertainty, it provides a clear roadmap for organizations,” he said. That roadmap? Fix the internal mess first. Otherwise, it’s just more expensive pilots going nowhere.

Deepankar Mathur, associate director at Searce, said the whole idea of full-scale adoption is kind of useless now. “It’s like trying to hit a constantly moving target,” Deepankar said.

The way agentic AI evolves, there’s no single launch moment. Instead, it’s about constant upgrades. Identify what needs automating, decide what matters most, use the best tools available, and then improve them again, immediately. “This cycle of improvement isn’t a temporary project; it’s an ‘always-on’ operational imperative,” he said.

Executives define roles, secure systems, and unlock teams

Dan said the way to avoid fear and confusion is to treat the AI-human mix like a real partnership. Spell out who does what. “This means crafting a strategy that outlines what tasks AI will handle and what roles humans will play,” he said. That removes doubt and gives employees room to work with the tools instead of against them.

But that only works if the AI has something to work with. “Jobs are performed through know-how and experience, which is information that may exist only in workers’ heads,” Dan said. That kind of knowledge doesn’t sit in a database. It has to be captured and turned into structured material. Agentic systems need that to make smart decisions. No input, no output.

And then there’s cybersecurity. Dan said more agents in production means more vulnerabilities. “We’re starting to get more news of the cyber implications in many agents,” he said. That means companies need to build AI-focused cyber plans from day one. Set rules around data use, privacy, ethics, and when a human has to step in. “By proactively addressing these governance questions, leaders can build a trustworthy and transparent system,” he said.

Deepankar also pushed for giving teams direct access to AI tools. He said being an engineer isn’t required anymore to build something useful. “The barrier to AI implementation has significantly lowered,” he said. But relying on steering committees or centralized AI boards just slows things down. “True progress requires leaders to actively champion and enable this widespread adoption.”

He said the most forward-leaning companies are setting up internal AI centers of excellence. These aren’t giant departments, just tight teams of experts who embed into different business units, train them, and get them building their own agentic workflows. “The most successful enterprises are building small, elite teams of ‘AI blackbelt’ specialists,” he said.

The smartest crypto minds already read our newsletter. Want in? Join them.