Agentic AI promises unlimited productivity. Delegate your most tedious tasks to bots and let them work while you sleep. So why aren’t more people using agents?
It’s not because people aren’t using AI. Pew Research found that 60% of U.S. adults were already using LLMs in 2026—one of the fastest rates of technology adoption in history.
Agents are different. ChatGPT asks almost nothing of you beyond opening a browser and typing. Agents ask users to delegate: connect accounts, grant permissions, define workflows, supervise results, and trust software to act on their behalf.
That turns out to be a much higher bar.
And it points to a possibility hiding in plain sight: maybe most people were never supposed to “use AI agents” at all. Maybe they’re just supposed to use products that happen to have agents underneath them.
Why?
In Model Behavior, Wired’s AI newsletter, author Maxwell Zeff explores the same question, referencing a viral X post from Browser Company CEO Josh Miller in which Miller wrote, “Theoretically, the tech is ready for AI agents to totally transform how we work and live our lives… but alas the general public dgaf.”
Miller suggested agents would be more popular if there were more of them that functioned as standalone products that just did something people wanted, needed, or appreciated.
Miller, admittedly, runs a company with a prime example. The Browser Company’s Dia browser sends users a morning brief consisting of their to-do list and other personalized tidbits. Dia uses an AI agent, but users don’t know that and, more importantly, don’t need to understand how it works.
Though Zeff noted Miller’s skin in the game, he agreed that there are few agentic products designed for the average consumer.
The average consumer, in this case, is someone who doesn’t care how something gets done, only that it does. They’re not thinking, “Wow, I can set up an AI agent to plan my day.” They’re more into, “It’s handy that my calendar reminds me it’s 30 minutes until my meeting,” or, “I like this payments app because it tells me when invoices are overdue.”
Their needs are often already met by email, spreadsheets, calendars, and other familiar tools. Or they’re perfectly happy chatting back and forth with an LLM on simpler tasks, never wondering whether they should be delegating instead.
Paras Chopra, who runs AI research lab Lossfunk, writes that even if it does occur to someone to delegate a task, they must also “justify the additional cognitive cost of delegation.”
That may be the simplest explanation for the adoption gap.
Delegation isn’t free. Before an agent saves you any work, you first have to figure out what to delegate, explain what you want, connect the right tools, grant permissions, monitor what happens, evaluate the result, and potentially fix whatever went wrong.
In other words, the value of an agent isn’t simply the amount of work it can perform. It’s the work avoided minus the cost of managing the delegation.
For plenty of tasks, that equation still doesn’t work.
Setting up an agentic workflow may take more time and effort than just doing the work yourself. And the smaller or less frequent the task, the harder that overhead becomes to justify.
Chopra ultimately arrives at a similar conclusion as Miller: AI agents become far more adoptable when the delegation disappears into the interface.
We’re already beginning to use products powered by agentic systems without having to write a prompt or think about what’s happening underneath. That may ultimately be what mainstream adoption looks like: not millions of people “building agents,” but software quietly becoming more autonomous.
That invisibility also makes adoption harder to measure. Agentic AI may already be shaving minutes off repetitive tasks without users ever identifying themselves as “agent users.” These minutes pile up and may result in time back, but it’s not necessarily the kind of ROI that transforms an organization overnight. Even larger companies have struggled to show a clear ROI on their AI investments.
Author Cory Doctorow — who frequently dissects and criticizes Big Tech’s lofty ambitions, coining such terms as “enshittification” and the “reverse centaur” — brings up another problem: the internet itself wasn’t built for agents.
Websites, internal systems, and databases frequently hold information behind interfaces agents can’t reliably understand or access. For agentic AI to work broadly, much more of the digital world would need to become machine-readable and interoperable.
Doctorow argues there are economic reasons that may never fully happen. Some businesses benefit from keeping information difficult to compare or access — dynamic pricing being one example — giving them little incentive to redesign their systems for someone else’s AI agent.
Then there’s the most obvious problem: handing over a task also means handing over some degree of control to technology known to make mistakes.
That matters a lot more when AI can act instead of simply answer. In 2025, Replit’s coding agent reportedly ignored an explicit code freeze and deleted a live production database containing records for more than 1,200 executives. More recently, an OpenClaw agent tasked with booking a gym class discovered a vulnerability in the reservation system and canceled another customer’s spot while trying to move its user up the waitlist.
Neither user asked the agent to do those things. That’s what makes agentic mistakes fundamentally different from a chatbot hallucinating an answer. The more autonomy we give an agent, the more we’re trusting it not just to understand what we want, but to decide how to get there.
And sometimes, apparently, getting you into a gym class means taking someone else out of one.
These kinds of stories can turn off potential adopters, but a more realistic concern may not be one spectacular failure at all. It may be agent drift: systems that continue producing plausible results while subtly changing how they operate over time.
Take this example from CIO, in which a lending AI designed to help review loan applications began skipping the income verification step 20%-30% of the time, despite delivering recommendations that looked sound.
That’s arguably harder to catch than an obvious failure. The agent still appears to be working. The output still looks reasonable. Something underneath has simply begun to degrade.
While larger organizations may have teams devoted to managing their agentic systems, individual users may be reluctant to take on that oversight — especially when the entire reason for delegating the task was to think about it less.
So if Not Now, When?
Just because the technology exists — and despite incredible promise — doesn’t mean it’s ready or designed for the average person. For now, that creates a chasm between AI super users willing to manage agents and ordinary people who just want software that works.
For agentic AI to build real momentum, it will need to be embedded in tools people already use or in new products they actually want to use. Those products must be accessible to non-technical people and reliably perform specific tasks or workflows in ways that create demonstrable value — obvious time back, noticeable revenue gains, less work.
They must be permissioned, meaning an agent operates under clearly defined controls rather than having an open-ended ability to “go rogue.” And if something does go wrong, there must be a clear chain of accountability.
Perhaps most importantly, what an agent offers must be significantly and measurably better than delegating the task to a human employee, using a regular app, or just doing it yourself.
Until then, the cognitive cost of delegation may outweigh the productivity it promises.
The irony is that mainstream adoption of agents may eventually arrive without mainstream adoption of “agents” at all.
People didn’t need to understand cloud computing, recommendation engines, or APIs to benefit from the products built on top of them. Agentic AI may follow the same path. The breakthrough won’t be convincing hundreds of millions of people to configure, prompt, and supervise their own agents. It will be building products useful and reliable enough that nobody thinks about the agent underneath.
