The meter was running while he ate
Branden Jenkins, CEO of Maxio — a private-equity-backed software company headquartered in Atlanta on a path toward $100 million in annual revenue — was at dinner when he glanced at his AI usage dashboard and saw the damage: $1,000 in tokens consumed over a single weekend coding session, charged automatically in $1,000 increments to a card set on auto-renew.
'A thousand is not that much, I would say, but for one weekend, it's pretty annoying,' Jenkins told Fortune. His token wallet had been configured to auto-refill every time it ran dry — silently, without a second approval required.
The mechanism behind the blowout is well documented. Gartner has estimated that agentic AI models can require between 5x and 30x more tokens per task than a standard chatbot exchange. A WitnessAI survey found that 68% of U.S. companies say at least some of their AI initiatives ran over budget in the past year, with a third reporting overruns happen 'mostly or always.' Jenkins's weekend tab is a rounding error against those benchmarks: Uber reportedly burned through its entire 2026 AI coding budget within four months, and Amazon reportedly spent $500 million on AI in a single month after rolling out access without usage caps.
Drift, not ambition, is the cost driver
Jenkins traced much of his own waste to two culprits: model selection and what he called runaway conversational drift — an AI system pulling a user down paths they never intended. 'A lot of times it's the agent's own mistakes that's burning your money,' he said. 'You kind of find yourself just chatting, and things getting away from you.'
His fix came not from his engineering team but from optimization tips circulating on TikTok. He now routes tasks by complexity — lighter models like Claude's Haiku for basic math, mid-tier models for routine coding, and the highest-reasoning models reserved for genuine strategic planning. He also adopted orchestration layers: third-party tools distributed as free GitHub repositories designed to compress AI output. One tool he called 'Caveman mode' forces the assistant to reply in short, blunt sentences, which Jenkins estimated cuts token use by 70%. Another, 'grunt mode,' compresses replies to a word or two. He named 'Superpowers' and 'Ponytail' as part of the same underground ecosystem of cost-saving hacks.
The problem, Jenkins acknowledged, is that none of it is accessible to a typical employee. 'These are nerdy things,' he said. 'Do we need sales leaders and service leaders and marketers finding this stuff?'
The human tab is harder to audit
Jenkins, a self-described technical CEO who writes code from his phone using Claude even while away from his desk, sits near the top of his company's internal AI spending leaderboard — an unusual position for a chief executive. He said he has no spending governors, unlike much of his staff, who must request approval when they hit limits. That asymmetry has surfaced a deeper organizational problem: employee 'insecurity' about being outpaced by the technology — and by him.
George Sivulka, CEO of Hebbia, captured the broader dynamic when he wrote that using agents means 'you just hired a million bad employees.' Jenkins's $1,000 dinner bill is a useful parable precisely because it is small enough to be relatable and large enough to be real. 'There's no refund button. There's no dispute button in Claude,' he said, adding that maybe there should be.
The numbers come first, and here they are unambiguous: unmanaged agentic AI is a cost center masquerading as a productivity tool. Free enterprise rewards discipline in capital allocation — whether the capital is equity, debt, or a token wallet set to auto-refill. Companies that treat AI access as a limitless utility will find the bill arrives whether or not anyone was watching the meter. The ones that build governance around model selection, task routing, and spending visibility will convert the technology into margin. The rest will be explaining the invoice at dinner.
