AINEWS
About 10 million people used an AI agent, software that books, codes, or shops for you, last week. The chatbots at those same companies have around a billion users each.
Dia, a product from The Browser Company, can read your email, open a tab, and book a flight without supervision. That is an AI agent: software you hand a goal, then leave to do the multi-step work across apps on its own. OpenAI's agent products, Codex and ChatGPT Work, had about 10 million weekly users between them.
The chatbots those agents live inside have around a billion monthly users each. The two numbers describe the same industry, the same year, and a 100x gap. Why Normal People Aren't Using AI Agents
Miller. "Nobody is really using AI Agents," he wrote, citing the absence of the "college friends talking about it in group chat" feeling that marked the rise of Instagram or TikTok. Miller's X post
In an interview with Wired, Miller drew the contrast directly: "I just have not heard a single person outside of the tech community talk about an agent that they use." Wired
The 10 million weekly figure is OpenAI's count for Codex and ChatGPT Work combined. Anthropic's Claude Code and Cowork are at "similar levels of adoption," per people close to the company, which is directional rather than audited. ChatGPT and Gemini have each cleared around a billion monthly users on average. By the source's framing, the agent base is "basically a rounding error" against the chatbot base.
The missing piece is the job-to-be-done itself.
"Chat" had an obvious job: ask a question, get an answer. The first billion users arrived because the verb was already part of the language. Agents are being asked to invent their own verb. The consumer use cases that have crystallized so far are still narrow: writing and running code for developers, summarizing inboxes for knowledge workers, browser-side booking for early adopters like Dia. The broader industry worry, captured in the same Wired piece, is that there is no killer consumer product for agents despite the billions already poured into model training.
The lab-side bet is that the verb is coming, and that the consumer product will follow the infrastructure. Anthropic shipped a managed-agent product this month meant to lower the build cost for agent developers. Anthropic launches Claude managed agents OpenAI's pitch for Codex is the same direction from a different angle: give every knowledge worker a junior engineer.
Trade groups are buying the agentic-commerce thesis at full price. The US Chamber of Commerce and the platform commercetools have both framed 2026 as the year agentic shopping moves from demo to revenue. Wired
Two open design questions sit underneath the adoption gap. Trust: a chatbot that hallucinates is annoying, but an agent that hallucinates while moving money or canceling a flight is a different category of problem. Handoff: when a goal is too big or too ambiguous, the agent has to either ask or stall, and the design choice between the two determines whether a consumer ever feels in charge of the workflow.
The question worth holding is what an agent would have to do, in language a person would actually use, before ten million becomes a billion.
Anthropic's next managed-agent release is the next test. So is whichever consumer product, in 2026, finds the verb that "chat" already had.