Tokens are AI compute, and you pay for every one, including the ones your agent wastes. Published tests found three common leaks. Each has a simple fix with big measured savings. Don't load every tool's instructions into each request. Let the agent look up tools when it needs them. Anthropic measured 150,000 tokens cut to 2,000. Don't feed it raw HTML pages. Give it the text as Markdown. Cloudflare measured one page going from 16,180 tokens to 3,150. Don't resend the agent's whole history every time. Trim the stale parts. Xiao et al. found input tokens fell 40% to 60%. These are separate tests, so don't add them up. Sources and methods:

Business token usage: the full research

Business AI usage is splitting in two. Token volume is moving to cheaper models. Spending is not moving with it.