Your top 5% of AI users are driving most of the bill

man using MacBook

Your AI spend isn’t spread evenly across your team. A small group, probably five people or fewer, accounts for the bulk of it. That concentration is the most useful fact you can pull from your usage dashboard, and your per-seat line item won’t show it to you.

When you find those heavy users, there are two explanations. They’re either your best operators (the people who figured out how to hand real work to AI and run with it) or they’re running leaky workflows: an agent looping on the same task, a 200-page PDF re-uploaded into every single chat, or the largest model doing nothing more than grammar checks. Usually it’s both. Power users do genuinely useful work and waste a surprising share of their budget doing it.

Five fixes that stretch a plan without slowing anyone down

  • Start a new chat for each new task. Anthropic lists conversation length as one of the primary things that consumes usage limits. Every reply works from the full thread history. That three-hour chat is expensive on purpose.
  • Put documents you reuse into a project. In Claude, project content is cached and counts less against your limits on reuse. Upload the style guide once instead of pasting it into every new conversation.
  • Ask everything at once. Batch related questions into a single message and edit your prompt instead of sending a correction. Every round trip has a cost.
  • Match the model and tools to the job. Model choice, the Research feature, and web search all affect how quickly you hit limits. Save the largest model and deep research mode for work that actually needs them.
  • Queue API work that can wait. Anthropic’s Message Batches API runs jobs at half the standard price when you don’t need a real-time response.

The point isn’t to spend less on your top users. They should probably get more budget, not less. They should just spend it on real work rather than re-reading the same long thread.

What to do this week

Pull the usage report from your plan’s admin console or API usage page. Find the five people at the top. Buy them lunch and ask two questions: what are you actually using this for, and what would break if we capped it? The answers will tell you whether you have a leverage problem or a waste problem.

Stay on top of AI & Automation with BizStack Newsletter
BizStack  —  Entrepreneur’s Business Stack
Logo