AI Agents Have 10M Weekly Users. Chatbots Have 1 Billion. Here’s the Gap.

OpenAI's Codex and ChatGPT Work agents report about 10 million weekly users combined, while ChatGPT and Google Gemini each sit at roughly 1 billion monthly active users. That gap is what prompted Josh Miller, CEO of The Browser Company, to post a now-viral observation this week: the general public does not care about AI agents. Miller argues the problem is not the technology but the framing. The industry ships impressive demos rather than products people actually want, and it is stuck in a groupthink loop shaped by the 2013 film "Her."

What happened

Data point	Figure
OpenAI Codex + ChatGPT Work weekly users	~10 million
ChatGPT and Gemini monthly active users (avg.)	~1 billion each
The Browser Company acquisition price (Atlassian)	$610 million
Miller’s first company (Branch) acquired by	Facebook, 2014

Josh Miller, CEO of The Browser Company, posted on X this week that it is “kinda crazy that nobody is really using AI agents.” He says he wrote it half asleep during a family trip to Europe, but he stands by the point. The post spread quickly among Silicon Valley insiders, many of whom privately share the concern.

According to OpenAI, its two agent products (Codex, a coding agent, and ChatGPT Work) collectively reach about 10 million people per week. People close to Anthropic told Wired that Claude Code and Cowork are seeing similar numbers. Against a baseline of roughly 1 billion monthly users for mainstream AI chatbots, agents represent a tiny slice of actual usage.

Why it matters

The labs have spent billions training models capable of multi-step, autonomous tasks. Agents are the main commercial vehicle for recovering that investment. If agents stay a niche tool used mostly by developers and productivity enthusiasts, the business case for those training costs gets harder to defend.

Most people using generative AI today stick to simple tasks: asking questions, summarizing text, chatting. That behavior does not require an agent at all. The industry needs to move users up the value chain, but it has not yet found the product that does it.

What is actually stopping adoption?

Miller’s core argument is that “AI agent” is not a consumer category, it is an internal technical label the industry invented. He put it plainly: “No one wants AI agents, because AI agents aren’t a thing.” What people want is a product that makes them calm, focused, and productive. The fact that an agent architecture powers it under the hood is irrelevant to the user.

He points to the most popular feature in The Browser Company’s AI-powered browser, Dia: a personalized morning briefing. When users open the laptop they see a greeting, a to-do list pulled from their calendar and email, and a daily piece of art. Technically it runs on an agent. Users never need to know that.

Miller also flags a groupthink problem. When The Browser Company was exploring acquisition talks last summer, he met with leaders at most of the major AI labs. He says nearly every lab except one described their product vision by referencing the 2013 Spike Jonze film “Her.” That kind of shared sci-fi imagination narrows the range of products anyone tries to build.

The current trend of letting users build personal automation tools (pitch decks, dataset organisation, workflow scripts) has a ceiling. Only a small population wants to build their own productivity software. That is not a path to mainstream adoption. For more on how AI integration can be designed around what users actually do rather than what the technology can theoretically do, the framing matters as much as the capability.

Our take

Miller is right, and the numbers back him up. A 10-to-1 ratio between agent weekly users and chatbot monthly users is not a slow start. It is a signal that the product category has not been defined yet in terms consumers recognize.

The Dia morning briefing example is the most useful thing in this whole conversation. It is a concrete case where the agent architecture disappears and the user experience stands on its own. That is good product thinking, and it is rare right now.

For business owners evaluating whether to invest in agent-based tools: the raw capability is genuinely there. What is missing is the wrapper. If a vendor is selling you “an AI agent,” ask what specific problem it solves and whether a non-technical person on your team can use it without training. If the answer is no, it is still a demo. We have covered similar dynamics in the security risks that emerge when agents are given real-world capabilities, which is another reason consumer-facing products need to be designed carefully, not just shipped fast.

The labs will get here eventually. The ChatGPT moment for agents has not happened yet, but it probably looks less like “autonomous AI browsing the web for you” and more like Dia’s morning briefing: quiet, useful, invisible infrastructure.

What to do about it

1. Test agent products on a real task your team does weekly, not a demo task. See if non-technical staff can complete it without help.
2. Ignore the label “agent” when evaluating tools. Focus on the outcome: does it save measurable time on a specific job?
3. If you are building a product that uses agents, bury the terminology. Name the feature after what it does for the user, not how it works.
4. Watch adoption numbers, not launch announcements. OpenAI’s 10 million figure for agents is the honest benchmark to beat.

The killer consumer agent product does not exist yet. That is an opportunity as much as a warning.

Source: WIRED · AI