Welcome to the era of AI sprawl

Tokenmaxxing became the latest buzz in AI this spring. As summer begins, the emptying trend is already continuing. Amazon has discontinued its AI leaderboards after some employees made useless AI work to manipulate rankings. Palantir CEO Alex Karp compared tokenmaxxing to porn addiction, and Duolingo reversed its decision to evaluate its use of AI in employee performance reviews. Meta and AT&T have reportedly started to curb their use of AI as costs soar.

The pressure to use AI for the sake of using AI is creating AI sprawl. Employees are adopting new agents and vibecode solutions with a myriad of AI tools that are difficult for businesses to keep up with. This means they often burn through expensive AI budgets creating duplicate work, fail to pass on the best tips and tricks to their colleagues, and waste time “bot-sitting” or giving the AI the context and edits it needs to make the output usable. A new study by Glean’s Work AI Institute of 6,000 digital workers in the US, UK, and Australia found that 77% of people using AI engage in multiple programs per week, a third use four or more tools, and 60% shuffle the same prompt across multiple tools if they feel the first output isn’t good enough. Employees using AI individually say they save an average of 11 hours each week, but only 13% of those surveyed said these savings “significantly improved” their company’s performance.

“There is so much pressure to signal innovation through mere AI awareness, knowledge, and appetite that it is leading us astray,” says Kate Niederhoffer, director of BetterUp Labs, a behavioral research center for coaching and workforce development firms. To make a big difference in your workforce, you need to answer big questions, she says. For example, “Why are we adopting these tools? What are we trying to accomplish here? And how can we communicate that in a really clear and compelling way to influence everyone to use these tools to achieve our goals?” But very few companies are answering the “big why” around AI.

Few companies are answering the “big why” for AI, says Kate Niederhoffer.

Elucidating that the use of AI is essential means more than removing leaderboards or adjusting reviews to focus on specific impacts on employees. The rhetoric around AI adoption—that if you don’t get the most out of it and master it, it will replace you, or someone better at using it will replace you—reinforces the urgency to pursue individualism. AI has the potential to enhance collaboration and decentralize some skills, such as coding and image editing. But much of the evidence so far suggests that instead of thriving during shortened work weeks, AI Maxers are becoming burnt out, losing trust in their colleagues, and working alone on isolated islands.

Protein maxing, looks maxing, Ozempic maxing, 9-9-6 maxing—in a post-pandemic era that celebrates profit at all costs, workers were primed to engage in token maxing. However, the lack of a consistent AI user manual also allowed tools to spread at will across organizations. Individual employees are not maximizing efficiency. Companies need to limit AI sprawl while protecting opportunities for people to innovate.

Technical updates and new workflows typically happen from the top down. A company decides to use Zoom over Microsoft Teams and Microsoft 365 over Gmail. Employees receive a login to a set of tools. But outside of some corporate subscriptions to OpenAI or Anthropic, employee use of AI often happens in the shadows. OpenAI took steps this year to integrate ChatGPT and Codex. People want to use apps that are made specifically for their role, whether it’s coding, marketing, or human resources. Two people working in sales might want to use AI in different ways, repeating prompts and tasks and spending tokens to create near-duplicate reports and materials when they once would have collaborated with colleagues to get the job done.

Lee Senderoff, chief transformation officer at Travelport, a retail platform for travel agents, said he has seen AI sprawl take hold as people seek to incorporate technology into their work. One worker spent 160 times more tokens in four days than the next most prolific AI user. When employees work in silos and are encouraged to use AI to do more, they may end up repeating the same tasks as their colleagues even if they experiment with AI alone. It’s not cheap. “You’re incurring hard costs, you’re spending more money on tokens you don’t need to spend, you’re incurring duplicative costs,” Senderoff said. “But there’s also a double soft cost. You’re just wasting effort. So who are the experts to write this?”

One worker burned 160 times more tokens than the next most prolific AI user.

When people work alone with AI, results can be diluted and the rewards of collaboration can be leveled in favor of quick solutions. Nobel Prize-winning researcher Herbert Simon witnessed this behavior decades before the advent of AI. Individuals will choose a solution that is good enough rather than considering all possible options. Simon described this as “satisfying”. “On an individual level, we do it all the time,” says Emily DeJew, a professor at Carnegie Mellon University’s Tepper School of Business. “The purpose of an organization is to bring these satisfied people together and coordinate them in a large and productive way to work toward a common goal.”

The firing of thousands of workers and the pivot to AI clashes with this theory. Meta, which laid off 8,000 employees last month, plans to increase spending on AI by 60% to 87% this year, following a “tumultuous year” in which it began cutting jobs to shift its focus. Mark Zuckerberg said that the fact that one person can now do tasks that once required an entire team threatens to erode the larger structures that make organizations and companies work.

Instead, AI in the workplace is heading toward the fate of past innovations: the tragedy of the commons, says Rebecca Hines, director of Glean’s Work AI Institute. The theory goes like this. Once individuals benefit from a shared resource, they use it until it is nearly depleted or, in the case of AI, use it to advance their own status or credibility, at the risk of downgrading the entire team or project. “Whenever there is a tool that increases personal productivity, we tend to reach for it first,” Hines says. “The problem is the neglect of coordination that occurs when we fail to consider the impact of our actions on the broader population.”

Malicious use of AI can reduce trust. BetterUp’s past research has shown that when people create AI-generated documents or PowerPoints without proper oversight, co-workers begin to lose trust in them. Over-reliance on AI can disrupt the communal aspects of work. People increasingly rely on chatbots to answer questions and rely on Gen AI to regurgitate tasks that may previously have required the expertise of colleagues.

But AI has also democratized innovation. Marketers can vibrate codes, agents can act like personal assistants, and startups can do more with less. The key is to transition the benefits individuals gain from AI and translate them into larger team-wide and company-wide workflows. “How do we get out of this unorganized, slightly crazy state of anarchy, and how do we start to be at least a little organized so we can make the most of it?” Senderoff says. She says her company is experimenting with centralizing AI workflows. If you know two people are working on the same thing, you can encourage them to collaborate and show them best use cases at the enterprise level. The larger the company, the more difficult centralization can be.

Senderoff acknowledges that everyone is still experimenting with how best to achieve this. But it’s becoming clear that there is no tokenmaxxing shortcut to getting there.

amanda huber I’m a senior correspondent for Business Insider, covering the technology industry. She writes about the biggest technology companies and trends.