[DigitalToday reporter Chi-gyu Hwang (황치규)] An analysis says AI agents are consuming more AI tokens than people, in what it called a somewhat ironic finding. It also found a sharp increase in AI use in the legal field.

Andreessen Horowitz, a Silicon Valley venture capital firm, recently shared these and other notable trends in finance and technology on its website.

According to recent data from OpenRouter, a platform that helps users access various models, bots use nearly five times more tokens than people. Bot token usage has increased about 14-fold since February.

Bots also use tokens differently from people. Bots use cached tokens to perform much more work than people do.

Cached tokens can be seen as tokens that let AI reuse stored processing results, rather than recalculating from scratch content it has already read and processed.

More than 85 percent of the tokens handled by AI agents were based on cached prompts. Most of the recent increase in token use came from cached tokens.

People tend to have conversations in a question-and-answer format, while agents are designed to try repeatedly and constantly toward a goal, it said.

Andreessen Horowitz said users first input large amounts of information needed for tasks, such as policies, work procedures and code guidelines. It said agents then proceed step by step, reading new information and producing results, while needed information continues to accumulate in cache for reuse.

Andreessen Horowitz said, "Cached tokens are much cheaper than tokens entered for the first time. This is also lowering the cost burden of operating AI agents. But maintaining a cache requires a lot of memory. This is also why demand for high bandwidth memory (HBM) is surging. Many AI agents need that much memory to keep remembering and using previous work. It is still early, but the rapid spread of AI agents is also pushing up memory demand further," it reported.

Andreessen Horowitz also noted that the spread of AI agents could reduce demand for existing automation and workflow tools. Citing Similarweb data, it pointed to a steady decline in site traffic for automation tool sites, except for Gumloop.

N8N, Zapier and Make were all launched before large language models (LLMs) emerged. They lead the overall automation market by site visits, but visits to these services have posted a cumulative double-digit decline over the past 12 weeks. By contrast, Gumloop, launched in 2023 as an AI-native agent builder, was the only one to show growth.

Andreessen Horowitz also pointed to a widening gap in AI usage between average users and power users.

Some companies are rapidly widening the gap with typical companies by increasing AI use explosively, and OpenAI's recent data also showed this trend. While output tokens for typical companies roughly doubled, usage by the top 10 percent of companies increased much more steeply.

The legal sector is also drawing attention as a key area leading AI growth after coding. Even in coding AI, growth in the legal field stands out.

Adoption of Codex by legal professionals has increased 108-fold since February 2026. Andreessen Horowitz said it is not certain whether this change is because Codex has spread widely, but it is clear that changes in legal work are far more pronounced than in other areas.

Moves by leading companies targeting legal AI also appear related. Anthropic introduced Claude for Legal in May and entered the legal AI market, and Google recently unveiled a law-focused Gemini Enterprise aimed at law firms and in-house legal teams.