Companies spent the first half of 2026 telling employees to use more AI. Now they’re dealing with the wreckage of that instruction.
Amazon scrapped its internal AI token leaderboard after workers figured out how to game it with busywork. Duolingo dropped its plan to factor AI usage into performance reviews. Palantir’s Alex Karp compared the whole tokenmaxxing trend to a porn addiction. Meta and AT&T have reportedly started reining in AI spend as costs spiral. The era of “use AI or fall behind” is over — but what replaced it isn’t efficiency. It’s chaos.
A New Word for an Old Problem
The phenomenon has a name: AI sprawl. Employees across organizations are independently spinning up agents, vibe-coding solutions, and subscribing to tools their IT teams have never approved and their managers can’t see. A Glean Work AI Institute survey of 6,000 digital workers across the US, UK, and Australia found that 77% of AI users juggle multiple tools weekly, a third use four or more, and 60% routinely shuffle the same prompt across platforms when the first answer doesn’t land.
Workers individually report saving 11 hours a week through AI. Only 13% say those savings have meaningfully improved company performance. The gap between those two numbers is where AI sprawl lives.
The 160x Problem Nobody Budgeted For
Lee Senderov, chief transformation officer at travel retail platform Travelport, watched one employee burn through 160 times more tokens than the next heaviest user over four days. The output wasn’t 160 times more valuable. It was mostly duplicate work, generated in isolation, disconnected from what colleagues were building across the hall.
This is the hidden cost structure of AI sprawl — and it doesn’t show up cleanly in any budget line. There’s the direct token spend, the duplicated effort, and then what Senderov calls the “soft costs”: eroded expertise, institutional knowledge that never transfers, and the quiet question of who actually owns quality when everyone’s outsourcing judgment to a chatbot. If you want to understand why enterprise AI bills keep climbing even as per-token prices fall, sprawl is a significant part of the answer — cheaper tokens expand usage faster than they reduce spend.
“Who’s the expert that should be writing this?” she asked. That question used to answer itself.
Satisficing at Scale
Carnegie Mellon’s Emily DeJeu points to Herbert Simon’s concept of satisficing — the human tendency to grab the first good-enough solution rather than exhaust every option. AI doesn’t fix this. It accelerates it. Workers satisfice faster, generate output faster, and skip the collaborative friction that once forced better decisions.
That friction was doing structural work. When someone had to ask a coworker for help, knowledge moved. Trust accumulated. BetterUp’s research found that when employees began circulating AI-generated output without adequate review, teammates trusted them less — not more. The economics of runaway AI compute tell one version of this story; the erosion of workplace trust tells a quieter, slower one.
Rebecca Hinds at Glean’s Work AI Institute frames what’s happening as a tragedy of the commons. AI promises individual productivity gains, so individuals optimize for those gains — and the collective absorbs the cost. “The problem is this coordination neglect that happens when we don’t consider the impact of our actions on the broader collective,” she says.
The Governance Gap Is Actually a Design Flaw
Traditional software rollouts move top-down. A company picks a tool, provisions accounts, and everyone operates inside the same system. AI adoption hasn’t worked that way. Most organizations have landed on a patchwork of individual subscriptions, department-level experiments, and shadow tools that IT discovers only when an audit surfaces a vendor renewal nobody can trace to a decision.
Kate Niederhoffer at BetterUp Labs puts the root cause plainly: companies adopted AI without answering the “big why.” Why this tool? For what outcome? Without those answers, AI adoption becomes performance — employees signaling fluency rather than solving real problems. Zuckerberg’s claim that one AI-equipped person can do the work of a whole team is technically defensible in isolation. As an organizational design principle, it’s producing the opposite of its promise: more output, less collective intelligence, and a workplace where everyone is productive, and nobody knows what anyone else is doing.
Senderov says Travelport is trying to centralize AI workflows — surfacing who’s working on overlapping problems, building shared pipelines, promoting best-use cases across the enterprise. She admits everyone is still figuring this out.
The tokenmaxxing era ended not with a correction but with a mess. Cleaning it up will require companies to do something they avoided all year: slow down long enough to ask what they actually needed AI to do — and whether the answer changes when it’s the whole organization asking, not just one employee with a prompt box open.
Related: What Are AI Tokens? How Models Read Text, Images & Code
