Amazon Employees Are "Tokenmaxxing" — When AI Adoption Becomes Theater

MeshClaw, Amazon's internal AI agent tool inspired by OpenClaw, was supposed to help employees automate repetitive tasks. Instead, some employees are using it to inflate their AI usage metrics.

Earlier this week, the Financial Times reported that Amazon employees have been "tokenmaxxing" — running up massive AI token counts by feeding long documents into AI tools without actually using the output. The goal? Improve their standing on internal AI usage leaderboards.

This isn't just a funny new slang word. It's a symptom of a deeper problem: AI adoption theater.

What's Actually Happening

Amazon deployed an internal AI agent tool called MeshClaw (inspired by the open-source OpenClaw project) that can triage emails, deploy code, and interact with Slack. The company's metrics showed thousands of employees using it daily.

But here's what some employees discovered: the easiest way to look productive with AI is to maximize token volume, not outcomes. Feed a 50-page document into the AI, let it process, and your usage stats skyrocket. Whether you actually use the summary is a different question.

Both Amazon and Meta employees have reportedly engaged in this behavior, prompting managers to be explicitly discouraged from using token counts as performance metrics.

Why This Matters Beyond Amazon

The "tokenmaxxing" phenomenon is the natural consequence of treating AI adoption as a metric to maximize rather than a capability to integrate.

It's the same pattern we've seen before with:

• Developers committing code at midnight to look busy
• Workers keeping their Slack status green while doing nothing
• Meeting culture where "being in the room" counts as productivity

When you measure the input (tokens consumed, prompts submitted) instead of the output (tasks completed, problems solved), people will optimize for the metric.

The Three Layers of AI Theater

Layer 1: Employee-Level Gaming

Employees generate volume without value. This is what "tokenmaxxing" is — the AI equivalent of sending 50 emails about nothing to look busy.

Real fix: Measure outcomes, not usage. Did the task get done faster? Was the quality higher? Did it free up time for meaningful work?

Layer 2: Manager-Level Pressure

Amazon reportedly posted team-wide AI usage statistics. When your manager can see your token count compared to your peers, the rational response is to increase your count — not to use AI more effectively.

Real fix: Stop leaderboards. Stop public AI usage stats. Replace them with case studies of actual time savings.

Layer 3: Company-Level Mandates

The top-down push to "use AI" — any AI, right now — creates perverse incentives. This is what MitchellH called "AI psychosis": companies applying AI as a solution without understanding the problem.

Real fix: Start with a problem, not a tool. "What's the biggest bottleneck in our workflow?" not "How can we increase AI token consumption by 200%?"

What Genuine AI Productivity Looks Like

After yesterday's article about why AI won't make your processes faster, the pattern is clear: AI helps when it's applied to a real bottleneck by someone who understands the problem.

Real examples of AI productivity that don't require tokenmaxxing:

• Claude or ChatGPT for clarifying vague requirements before writing any code
• GitHub Copilot for generating boilerplate once you know exactly what to build
• Perplexity for research synthesis where you need to verify against multiple sources
• AI tools for documentation that actually gets written — because documentation was the bottleneck, not "lack of AI usage"

The difference in every case: the bottleneck was identified first, and the AI tool was chosen to address it. Not the other way around.

The Security Angle

Multiple Amazon employees also raised concerns about granting AI agents permission to act on their behalf. One employee told the FT: "The default security posture terrifies me. I'm not about to let it go off and just do its own thing."

This is a legitimate concern that's often glossed over in the "move fast with AI" narrative. An agent with access to email, code deployment, and Slack can make mistakes at massive scale. The companies that win with AI won't be the ones deploying the most agents — they'll be the ones deploying them safely.

The Bottom Line

"Tokenmaxxing" is a warning sign for every company pushing AI adoption from the top down.

If your employees are gaming the metrics, it means:

1. They don't see real value in the AI tools you've given them
2. They're being measured on the wrong things
3. The culture is oriented toward looking productive rather than being productive

The fix isn't more AI. It's understanding why AI isn't helping — and fixing that first.

---

Want to discover AI tools that actually help you work smarter? Browse 觅·Mee's curated collection of 230+ AI tools, organized by category with real use cases.