AI Agents Might Be Burning Your Budget, Not Saving It

What if the tool you hired to cut costs is the one running up the bill?

That uncomfortable question hung over two separate AI events in Silicon Valley this week, where engineers and executives from Google, Amazon, Microsoft, and Meta sat down to talk honestly about what it actually takes to build and run AI agents at scale. The verdict was more cautionary than the boardroom pitch decks suggest.

The "Never-Sleeping Intern" Has a Spending Problem

The sell is seductive: AI agents that work around the clock, handling tasks that once required human attention. Nvidia CEO Jensen Huang called it "definitely the next ChatGPT" just last month. Since OpenAI released OpenClaw—a framework that lets developers build and manage fleets of AI assistants using various models—the industry has been racing to deploy these digital workers across enterprise workflows.

But at the Generative AI and Agentic AI Summit in San Jose on Wednesday, Kevin McGrath, CEO of AI startup Meibel, named what he sees as the industry's most pressing problem: the assumption that every task needs to go through a large language model.

"Just give all of your tokens and all of your money to an AI bot that will just waste millions and millions of tokens," McGrath said—describing not a hypothetical disaster, but a pattern he's actively watching unfold.

Google software engineer Deep Shah put a sharper point on it. The first challenge when deploying multi-agent systems at scale, he said, is inference cost—the fee incurred every time an AI agent calls an LLM to process a task. A poorly designed monitoring system for those agents doesn't just fail to save money; it can actively destroy it.

Complexity Is the Real Boss

Sponsored

Advertise with Us

[email protected]

Sponsored

Ravi Bulusu, CEO of startup Synchtron, framed the problem differently: it's not just cost, it's chaos. Every enterprise organizes its data differently, runs different tech stacks, and builds software in its own idiosyncratic way. AI agents touch all of those dimensions simultaneously.

"No single dimension is solved in isolation," Bulusu said. "The interdependencies are what make this hard—in fact, chaotic even."

This isn't a small caveat. It's the central tension of the current AI agent moment: the tools are being sold as solutions to enterprise complexity, but they require enterprises to first solve that complexity before the tools can work properly. It's a bit like selling a high-performance engine to someone who hasn't yet built the car.

The conversation continued Thursday at a separate event in Mountain View, where ThinkingAI—a Shanghai-based company that recently pivoted from mobile game analytics to AI agent management—made a pointed critique of OpenClaw itself. Co-founder Chris Han said the popular framework is "too complicated and too prone to security flaws" for serious enterprise use.

"OpenClaw is a good tool for personal things, but definitely cannot reach the enterprise level," Han said. "In terms of the enterprise level, you have to figure out a lot of things—your memory, how to manage your agents, teams, communications."

ThinkingAI has partnered with MiniMax, a Chinese AI lab that went public in Hong Kong in January and is considered one of China's leading open-source AI developers. Han noted that ThinkingAI's platform supports models from OpenAI and Google as well—a hedge against potential U.S. restrictions on Chinese AI models. Asked whether a U.S. ban on Chinese open-weight models would hurt the company, Han was sardonic: "If that happens, maybe we are successful."

The Gap Between the Pitch and the Pipeline

None of this means AI agents don't work. They do—in specific, well-scoped contexts with careful cost management. But the gap between C-suite enthusiasm and engineering reality is wide enough to matter for anyone making investment or deployment decisions right now.

For investors, the signal here is nuanced. The companies building AI agent infrastructure—monitoring tools, cost optimization layers, orchestration platforms—may have a more durable business than the headline-grabbing agent applications themselves. If every enterprise deploying agents eventually hits the wall of inference costs and operational complexity, the picks-and-shovels play becomes more attractive.

For enterprise buyers, the question is whether the ROI math actually closes. AI agents are frequently justified on labor cost savings, but those projections rarely account for the engineering overhead required to build, monitor, and continuously tune the systems. A poorly governed agent fleet can generate cloud bills that dwarf the salaries it was supposed to replace.