Instead of measuring AI by usage, look to metrics that measure what gets accomplished and improved. 

Companies moving quickly on AI are facing a new leadership problem: “tokenmaxxing.” This is the act of judging AI adoption by volume, tokens, prompts, and outputs, rather than by what the work accomplishes. Managers want to know whether employees are adopting the tools. So they look for visible signals: tokens used, prompts entered, outputs created. These metrics are easy to collect.  

But AI use on its own is not a productivity metric. The logic behind tracking and rewarding token consumption and output volume is that more AI usage signals more efficiency. However, not only do these metrics not necessarily translate directly to ROI, but they may in fact signal wasted time and money. One company reportedly spent $500 million in a single month on AI token use earlier this year.  

By focusing on tokenmaxxing and similar usage metrics, businesses are rewarding activity for its own sake, while simultaneously harming employees’ intrinsic motivation, which comes from solving the right problem well. Instead, organizations should ask, “Does this AI application make it easier for people to bring their best judgment to their work?”  

AI Signals Don’t Translate Directly to ROI  

AI usage data can be useful. It can show where tools are being tried, where adoption is uneven, and where more training may be needed. But usage data is an input, not an outcome. 

Instead, focusing on where AI is improving the work means looking at decision quality, speed to resolution, customer impact, reduction in rework, and whether teams are spending more time on the parts of the job that require human judgment. In some roles, higher AI usage may be a good sign. In others, it may suggest people are spending too much time managing the tool. 

That is the risk of “botsitting.” Employees can end up supervising outputs, correcting errors, rewriting generic drafts, and feeding systems more prompts because of the focus on activity over impact.  

The organizations that get AI right are the ones that treat it as a work design challenge. What should AI handle as its capabilities improve? What should people still own? Where does automation create new capacity, and where do risks outweigh the benefits? 

Producing the first draft fast is becoming table stakes. The differentiator is having the judgment and the motivation to question the output and to refine it and push past the first result when it isn’t right.  

Focus on Feedback  

Good feedback helps people improve. It is transparent, connected to expectations, and tied to development and business outcomes. Control is different. It is opaque, one-directional, and focused on compliance. The distinction matters because intrinsic motivation runs on autonomy, the sense that a person, not a tracker, owns how the work gets done.  

If employees believe AI tools are being used to monitor how often they prompt, how fast they produce, or whether they appear busy, trust will suffer. Once that happens, adoption becomes 

performative. People comply with the metrics, while withholding the judgment that makes AI useful.  

If the work product is weak, the answer is not always more monitoring. Leaders should look upstream. Are expectations clear? Have teams been trained well? Do managers know what good AI-enabled work looks like? Are people being rewarded for better outcomes or just more output? Those are harder questions than pulling a usage report, but they are a lot more useful.  

Measure What Matters 

AI can make work better. It can help employees move faster, learn faster, and spend less time on repetitive tasks. But the technology will not fix unclear goals, weak management, or poorly designed work. 

HR has an important role to play here, not as a function that slows AI down but as a function that helps companies adopt it well. That means making sure measurement systems reinforce judgment, learning, accountability, and customer impact.  

The goal should be great work, not maximum AI usage. Employees are asked to build new AI skills and prove their own unique value as workers at the same time, and right now those two goals compete instead of reinforcing each other. It shouldn’t work that way. 

Use the data. Learn from it. But do not confuse tokenmaxxing with progress. The real measure is whether AI helps people do work that is faster, smarter, and more valuable than before.