Nobody Is Using AI Agents Yet. What That Means for Your Budget

Every software vendor you buy from is about to add an “agent” tier and raise your price. Before you approve that line item, look at how many actual humans use agents today. The number is small enough to change how you budget for the next twelve months.

“The tech industry is realizing it needs to build agents based on what regular consumers want, not just what its AI models can do.”

— Maxwell Zeff, WIRED

Our take

Start with the arithmetic, because it is the part that gets waved past in most coverage. OpenAI’s Codex and ChatGPT Work agents sit somewhere around 10 million weekly users. Anthropic’s Claude Code and Cowork are in the same neighborhood. Meanwhile ChatGPT and Gemini each carry roughly a billion monthly actives. Agents are running at something like one percent of the audience that plain chat reaches.

That gap did not come from bad marketing. Josh Miller, CEO of The Browser Company, kicked off this whole conversation with a blunt post: he had not heard a single person outside the tech industry mention an agent they actually use. He blames the category itself, arguing that “AI agents” is an industry invention rather than something a normal person ever asked for.

He is half right, and the half he misses is the one that matters to anyone deploying this stuff. The blocker is not the name. It is reversibility.

Look at what people do use. Miller’s own success story is a personalized morning briefing: your calendar, your tasks, your reading. Notice what that is. It reads, it summarizes, it never touches money, and if it gets something wrong you lose eight seconds. Now look at the demos that keep failing to land: book the flight, order the groceries, negotiate the cancellation. Those spend money and produce consequences that a human has to unwind by phone.

Consumers are not confused about agents. They are correctly refusing to delegate irreversible actions to software with no audit trail and no one to sue. Adoption is going to track the cost of being wrong, not the intelligence of the model. Read-only agents will go mainstream fast. Agents holding a credit card will take years, and the delay will be legal and financial, not technical.

There is a second-order effect here that the source article does not touch, and it decides where your money goes. If consumer agents stay read-only for a while, then the panic about “AI agents will do all the shopping and your website will never be seen” is premature. But the adjacent panic is not. Answer engines and AI summaries are already intercepting the informational query, and that is a live, measurable problem right now. Those are two different threats on two different timelines, and a lot of agencies are selling them as one bundle. Budget for the summary problem this year. Watch the agent problem.

The flip side is the internal one. Agents are hitting real adoption in exactly the place you would expect: work, where a human reviews the output before it ships and the blast radius is one draft. That is the shape that works for a small business too. Not “an agent runs my ads.” Rather, an agent drafts thirty ad variants and a person kills twenty-seven of them.

One more thing worth noticing. Miller points out that every lab leader keeps citing the movie Her as the north star, which means the frontier labs are all building roughly the same product. When the biggest players converge, differentiation moves to the interface and the workflow, and those are things a small team can actually win. The model is a commodity. The thing wrapped around it is not.

What this means for your business

• Do not pay an “agent tier” premium on faith. Pick three recurring tasks, time how long they take a human today, run the agent against them for two weeks, and compare. If the agent does not beat the human on total time including your verification pass, you are paying for a demo.
• Deploy agents where mistakes are cheap and reversible. Drafting, summarizing, first-pass research, bulk variant generation, transcript cleanup. Keep them away from anything that sends money, sends email to a customer, or publishes without review.
• Split your AI-search budget from your AI-agent budget. Losing clicks to AI summaries is happening now and deserves real spend on schema, brand queries, and content that answers rather than teases. Consumer purchasing agents are not yet buying from you. Do not let a vendor sell you one problem using the other one’s urgency.
• Build the review step before you build the automation. Whoever signs off on agent output needs to be named, and the checkpoint needs to exist in the workflow. Teams that skip this find out the hard way, usually in a customer-facing channel.
• If you sell software or a service, stop leading with “agentic.” Your buyer’s customers are not asking for it. Lead with the outcome and the hours saved, and let the agent be an implementation detail.

The technology is real. The consumer product is not here yet, and the gap between those two facts is where a lot of budgets are going to get wasted this year.

Read the full original reporting at WIRED: Why Normal People Aren’t Using AI Agents.

---