In its latest edition of Charts of the Week, a16z compiled a clear picture of how AI agents are altering usage patterns, drawing on data from OpenAI, OpenRouter, and Similarweb.
The starting point is divergence. Overall AI demand is climbing, but unevenly: token output at a typical company has only roughly doubled, whereas the top 10% of enterprise users have surged more than 17x since April 2025. The gap between the two cohorts now stands at ~8x across all sectors, and widens to nearly 12x in the tech sector, where the top decile generates 32.5x more tokens.
Power users are not burning more tokens simply by chatting more. They have migrated to more advanced tools like MCP, plug-ins, and skills. This is evidenced by the top decile’s plug-in adoption rate being ~2x higher, and skill usage 6x higher, than that of a typical company. Measured by on-the-job Codex usage, the standout sector is surprisingly legal, where adoption has surged 108x since February 2026, though how much of that jump stems from Codex’s broader rollout remains unclear.
The most striking finding lies in how agents consume tokens. According to OpenRouter, agents consume five times more tokens than humans, and agent usage has surged roughly 14-fold since February alone. The core difference comes down to cached tokens: more than 85% of the tokens agents burn come from cached prompts—meaning context preserved across steps rather than reloaded from scratch. While humans interact through turn-by-turn Q&A, agents are designed for iterative, goal-oriented loops; the initial prompt carries a token-heavy pre-fill (policies, coding guidelines, task context), after which the system only performs incremental reads and writes to the cache.
This architecture produces two contrasting effects. Cached tokens are far cheaper than upfront inputs, making agent operational costs increasingly economical. Yet maintaining a cache also means tying up hardware memory, which is one reason high-bandwidth memory for AI remains in short supply: millions of parallel agents require memory to hold context without having to reprocess everything from scratch on every cycle.
Even more notably, agents may be siphoning users away from legacy automation tools. Similarweb data shows that apart from Gumloop, traffic across automation tool websites has seen consecutive declines. Three pre-LLM incumbents—N8N, Zapier, and Make—still lead the category in traffic, but all three have suffered double-digit percentage drops over the past 12 weeks. Meanwhile, Gumloop, an agent-building platform founded in 2023 and designed AI-native from the outset, was the sole platform to post growth.
The author concludes that it is too early to write off Zapier, given that they also have AI at their disposal. Yet agents have only just arrived, and the competitive landscape is already shifting significantly.