What's the right way to measure whether your company is actually using AI? Reid Hoffman, LinkedIn co-founder and longtime venture investor, stepped into the "tokenmaxxing" debate this week with a measured position: token counts can signal adoption, but they are not a productivity metric.
A token is the basic unit AI language models process - roughly four characters, or about three-quarters of a word. When you send a prompt to ChatGPT or Claude and get a response, every word in both directions gets counted in tokens. "Tokenmaxxing" is the shorthand for treating this volume number as proof that AI is delivering value.
Hoffman's argument is that context matters. High token volume could mean employees are doing genuinely productive work - drafting, analyzing, coding. Or it could mean they're burning tokens on low-value prompts, poorly structured queries, or experimentation that never produces usable output. The same 10 million tokens could represent 500 productive tasks or 5,000 failed attempts.
Token Counts Are Also Gameable
Tie token usage to performance reviews or departmental KPIs and you'll get more tokens, not more productivity. Long prompts, unnecessarily wordy outputs, multi-turn conversations that could have been a single query - these inflate counts without improving outcomes. Any metric that can be gamed will be gamed.
Hoffman's broader point is that token data is most useful as a leading indicator, not a lagging one. Near-zero token usage definitively signals a team isn't using AI, regardless of what they say in surveys. Sudden spikes often correlate with a new tool rollout or a newly discovered use case. But tokens tell you input - they say nothing about what actually shipped.
What Should Companies Track Instead
Pairing token data with output metrics gives a more honest picture: how many tasks actually completed, how many support tickets resolved faster, how many code reviews processed per sprint. Tokens measure what went in. Those questions measure what came out.
The "tokenmaxxing" debate reflects a broader shift in how organizations think about AI adoption. The first question was "are people using it?" - token counts help answer that. The harder second question - "is it working?" - requires a different measurement entirely.