Uber's CTO says the company burned through its entire 2026 AI budget in four months, driven by Claude Code and Cursor usage.
Meta had a viral internal leaderboard ranking employees by AI token usage.
Microsoft is suggesting AI agents may need their own software licenses.
Across enterprises, AI agents are arriving faster than budgets, permissions, architecture, or governance can handle, and even the biggest companies in the world acknowledge they haven't figured it out yet.
Several stories this week painted the same picture from different angles: enterprises are applying AI agents to their operations without the budgets, permissions, architecture, or governance to support them. Uber CTO Praveen Neppalli Naga told The Information that the company burned through its full-year 2026 AI budget in just four months, driven primarily by usage of Claude Code and Cursor. "I'm back to the drawing board because the budget I thought I would need is blown away already," he said.
The Uber story frames a larger debate over "tokenmaxxing," a Silicon Valley term for maximizing AI consumption inside a company as a proxy for AI adoption. Meta recently had an internal dashboard ranking employees by token usage, with titles such as "Token Legend" that went viral. The dashboard was taken down. Writer CEO May Habib told the Wall Street Journal that driving token use is "existential" for her company. One Writer employee used nearly 11 billion tokens in March, costing north of $50,000. HubSpot CEO Yamini Rangan pushed back publicly: "Outcome maxxing >> token maxxing."
Meanwhile, OpenAI launched Codex for (almost) everything, giving its coding agent background computer use across any Mac app plus more than 90 new plugins. Microsoft executive Rajesh Jha told Business Insider that AI agents may eventually need their own software licenses, since one human supervising 50 agents breaks SaaS seat economics. Box CEO Aaron Levie posted a widely shared thread arguing that enterprises will have to keep "dramatically upgrading" their AI architecture, with patterns like RAG and GraphRAG already obsolete. And Amazon is reportedly dealing with AI sprawl: duplicate internal tools, disconnected data, and growing operational risk.
SmarterX founder and CEO Paul Roetzer broke down what all of this means for business leaders on Episode 210 of The Artificial Intelligence Show.
4 months - Time it took to burn Uber's 2026 AI budget, driven by Claude Code and Cursor
11 billion tokens - Used by a single Writer employee in one month, costing $50,000+
90+ plugins - Added to OpenAI's Codex in its latest update, extending agent reach across common business apps
Under 1% - Enterprises Roetzer believes are prepared to integrate agentic technology responsibly right now
Nobody has this solved. This includes the companies most associated with AI leadership. "I think it's presented as though people have a grasp on what is happening, but every day that goes by, you realize just how early we are in the integration of agents into workflows and businesses," Roetzer says. The companies pushing hardest are the ones most likely to find the potholes first.
The capabilities are compounding faster than anyone can plan for. Coding agents including Codex and Claude Code are turning into general-purpose agents that non-technical knowledge workers can run. New abilities land every week. "It's moving so fast. Almost to where it's almost impossible to just keep track of what is going on," Roetzer said. That's what broke Uber's budget. When engineers run parallel agents on complex workflows, Claude Code gets expensive fast.
The risks rise when non-coders can develop code. Roetzer pointed to a reported data visibility issue at AI app-builder Lovable, where default project settings exposed source code, database credentials, AI chat histories, and customer data to other free accounts. He also flagged a reported breach at Vercel that started with a compromised employee account connected to an outside AI platform and cascaded from there. "Lovable is a really good company with tons of funding," he said. "And yet they made this apparently intentional choice. There's just so many growing pains we have to go through."
"If you asked me point blank, how many companies, let's say 250 employees or more, fully understand generative AI and have properly integrated it so they can really scale transformation?
It's single digits. Low single digits. And if you extend that to which ones are prepared to integrate agentic technology and scale it within their company in a responsible way, it's well under 1%."
— Paul Roetzer, founder and CEO of SmarterX, Episode 210 of The Artificial Intelligence Show
Seat economics and architecture assumptions are both wobbling. Microsoft's Jha is right that SaaS pricing doesn't work when one human supervises 10 or 50 agents. Levie is right that the architectural patterns enterprises standardized on two years ago are already getting replaced. Amazon's sprawl problem is what happens when those two pressures collide inside a real company with real data.
The headlines make it feel like legacy enterprises are behind. They are not as behind as the narrative suggests. Y Combinator startups and VC-backed AI-native companies are supposed to be out on the bleeding edge. That is their job. The legacy enterprise's job is something harder: integrating AI agent technology into real workflows, with real data, real employees, and real regulatory exposure, in a way that does not blow through budgets or leak customer information.
"Do not feel like if you're not racing into agents, you are way behind. You're not. It's okay. If you're a legacy enterprise trying to drive transformation in a responsible way, agent stuff like this is going to be sandboxed through a technical team for the next 12 to 24 months before you start seeing this really living in the wild."
— Paul Roetzer, founder and CEO of SmarterX, Episode 210 of The Artificial Intelligence Show
The companies that take the next year to genuinely understand what they are deploying, and how it reshapes change management, org structure, and tech stacks, will come out ahead of the ones tokenmaxxing their way to a blown budget and an unexplained breach.
What happens when non-technical agents have their Claude Code moment. Claude Code existed for more than a year before a cluster of improvements late last year turned it into the capability that is rearranging how engineering gets done. A similar breakthrough is coming for general-purpose agents aimed at non-technical knowledge workers. Whether that is Cowork, OpenClaw, or something else, most companies are not ready for what it will mean when marketers, analysts, and operations staff can run the same kind of parallel agent workflows that just maxed out Uber's AI budget.
How this shows up in outside relationships. Many enterprises are relying on consultants, agencies, and systems integrators to stand up their agent infrastructure. The Lovable and Vercel incidents are a warning about what happens when outside tools get plugged into core systems without leadership fully understanding what access they are granting. Leaders should know exactly what their partners are setting up, what identities those systems run as, and what data they can reach. Treat AI agents less like software and more like new hires with login credentials, because that is increasingly what they are.
OpenAI: Codex for (Almost) Everything → openai.com
WSJ: Why Some Companies Say AI Tokenmaxxing Is Key to Survival → wsj.com
The Information: Uber CTO Shows How Claude Code Can Blow Up AI Budgets → theinformation.com
Business Insider: Microsoft Exec Suggests AI Agents Will Need Software Licenses → businessinsider.com
Aaron Levie on Constantly Upgrading AI Architecture → x.com
Business Insider: Amazon's AI Boom Is Creating a Mess of Duplicate Tools and Data → businessinsider.com
Heard on The Artificial Intelligence Show, Episode 210
Paul Roetzer and Mike Kaput discuss why enterprises are struggling to manage AI agents and what business leaders should do about it. Listen →