Tech companies have gotten distracted by a misleading measure of AI success: how many tokens their employees are burning through. According to Scott Wu, CEO of Cognition Labs (the company behind Devin, the widely-used AI coding agent), this obsession with "tokenmaxxing" has led companies to chase the wrong metric and waste significant resources.

What Is Tokenmaxxing and Why Did It Backfire?

Over the past year, major tech companies including Meta and Amazon created internal leaderboards and incentive programs to encourage employees to use AI tools more frequently. The idea was simple: measure token consumption (the computational units that power AI interactions) and reward high usage. But the strategy backfired spectacularly. Employees began using AI tools just to climb leaderboards, deploying bots to complete useless tasks rather than solving real problems. Amazon eventually told staff to stop, with one senior vice president reportedly saying, "Please don't use AI just for the sake of using AI".

The financial consequences have been severe. Uber burned through its entire 2026 AI budget in just four months and capped employee token spending at $1,500 per month. Despite token prices dropping 90% since 2023, overall AI spending has actually increased as companies feel emboldened to consume more of the cheaper resource.

"It is directionally correct, but I think there are definitely some places where people have gotten carried away. People are like, 'We rank our engineers by how many tokens they're spending.' Well, let's try and rank people by how much output they're actually producing," said Scott Wu.

Scott Wu, CEO at Cognition Labs

The Real Problem: Measuring the Wrong Thing?

Wu's argument reflects a deeper issue that Boston Consulting Group (BCG) identified in its 2026 Global AI at Work report. The study surveyed nearly 12,000 frontline employees and found a striking paradox: 42% of workers reported that AI tools saved them eight hours per week (roughly one full workday), yet 66% received little to no guidance on how to use that time productively, and half weren't spending the saved time on strategic projects.

In other words, companies are spending heavily on AI without understanding what employees should actually do with the productivity gains. The problem isn't the technology; it's leadership's failure to communicate clear goals. David Martin, global leader of BCG's People and Organization practice, explained that senior leaders struggle to articulate a coherent vision for AI adoption.

"Senior leaders are really struggling to articulate what the vision and strategy is on AI. Consequently, it increases employee fear. It makes it harder for them to even understand what objectives they're pushing for, and it trickles through to adoption, usage, and the like," said David Martin.

David Martin, Global Leader of People and Organization Practice at Boston Consulting Group

How Should Companies Actually Measure AI Success?

Wu advocates for a return to basics: measure AI by its return on investment, not by consumption. He points to Cognition's own approach, which focuses on increasing engineering capacity and shipping more output. For Devin's enterprise customers like Goldman Sachs, Mercedes-Benz, and Rivian, the value comes from tangible results: faster research and development, more code shipped, or more problems solved.

Both Wu and Martin recommend treating AI like any other workplace tool. Rather than giving AI access to every employee regardless of role, companies should ask targeted questions: What is the business case? Who actually needs this? Are we delivering measurable results? And critically, are we holding people accountable to those targets ?

• Define Clear ROI: Identify specific returns on investment such as revenue growth, efficiency gains, or cost savings rather than focusing on token consumption metrics.
• Align AI Access With Role: Provide AI tools selectively based on job function and business need, not universally to all employees.
• Communicate Strategic Vision: Senior leaders must articulate clear goals and objectives for AI adoption so employees understand how to use productivity gains effectively.
• Measure Output, Not Input: Track what employees actually produce with AI assistance, not how much AI they consume.
• Establish Accountability: Hold teams responsible for meeting targets tied to AI adoption, just as you would for any other business initiative.

The Deeper Engineering Challenge Behind AI Agents

The tokenmaxxing problem also reflects a broader misunderstanding of how AI agents like Devin actually work. A developer who spent months building an AI agent discovered that the underlying issue wasn't the AI model itself, but the "plumbing" around it. The agent had five critical systems that were never actually connected: permission toolsets that were computed but never granted, a spend meter that didn't track the agent's own token usage, a citation verification system that was never invoked, and a verification step that logged findings but didn't enforce them.

This engineer's experience reveals why simply throwing more tokens at a problem doesn't work. The math is unforgiving: if an AI agent is 95% reliable per step, chaining 10 steps without verification between them drops overall reliability to roughly 59%. Public benchmarks confirm this; frontier AI agents succeed on fewer than 25% of retail tasks when run once, and the best agents complete only 24% of simulated office tasks.

The solution isn't better models or more tokens. It's better engineering: verification inside the loop, explicit exit conditions, bounded resource limits, and clear checkpoints. Wu's message aligns with this reality: companies need to focus on whether their AI investments are actually producing results, not on how much computational fuel they're burning.