Enterprise artificial intelligence is shifting from conversation to execution as automated systems consume far more model capacity than ordinary human users.

OpenRouter recorded about 7.3 trillion agentic tokens on a seven-day average by Aug. 10, roughly 14 times the level seen in early February. Human-driven usage reached about 1.4 trillion tokens over the same period, up 2.8 times from early February. That left AI agents consuming more than five times as many tokens as humans across traffic routed through the platform.

Tokens — the chunks of text that language models read and generate — are also the unit by which most providers meter and bill API usage, so the ratio is a direct measure of where inference compute is going. OpenRouter, a routing service that sends developer requests to models from multiple providers, distinguishes agentic activity using signals such as tool calls, conversation turns and timing patterns. The figures reflect OpenRouter traffic rather than the entire generative AI market, but they indicate how quickly automated workloads are scaling. Unlike a chatbot that answers a single prompt, AI agents can inspect information, call tools, test outputs, revise steps and repeat actions before completing work.

Enterprise AI shifts from chat to automated workflows

OpenAI’s Enterprise Signals report points to the same transition inside companies, especially among the heaviest users. Frontier firms, defined as the top 10% of enterprise AI users, generated 8.3 times more output tokens per active user than typical firms.

That gap was 2.6 times in January, showing a widening difference between ordinary enterprise adoption and the most intensive users. By June, Codex, OpenAI’s agent for software development, produced 64% of combined Codex and ChatGPT output tokens among enterprise customers.

Weekly active Codex users had increased 108 times in legal since February, compared with 41 times in sales and recruiting. The mix suggests enterprise use is expanding beyond chat into coding, document creation, research and other multi-step workflows.

OpenAI also reported that 21% of active users at frontier firms used Plugins weekly, compared with 9% at typical firms. That difference reinforces the shift toward systems that can use tools, create files and complete tasks rather than only answer questions.

Codex, Plugins and caching drive the shift beyond chat

Andreessen Horowitz, citing OpenRouter data, said more than 85% of agentic token usage now comes from cached prompts. Prompt caching allows systems to reuse previously processed context instead of recomputing identical input during repeated calls.

OpenAI says caching can reduce latency and input costs, making repeated-context workloads cheaper to operate. Because inference is generally billed by the token, reusing context instead of recomputing it directly affects the cost of the long, repetitive call sequences that agents run. However, those workloads still require infrastructure that can store and rapidly reuse large contexts alongside the GPUs handling inference.

That increases the importance of memory capacity and bandwidth as enterprises delegate more complex assignments to automated systems. The shift also creates a test for traditional workflow software.

A16z argued that rising agent adoption could pressure products built around fixed automation steps as users adopt systems that reason through tasks. Even so, web-traffic changes alone do not prove displacement, while Zapier, Make and n8n — platforms that connect business apps to automate multi-step tasks — are also adding AI features.

Overall, AI agents are generating more token demand as enterprise software performs longer chains of work on behalf of users. As that pattern expands, enterprise AI is becoming less defined by chat volume and more by the amount of work delegated to software, and the measures already in view — agents’ share of token traffic, the frontier-firm output gap and cached-prompt share — are the ones that will show how far it goes.