Inside the tech industry, the agent era is treated as settled fact. Companies are designing payment systems for AI agents to spend money on their own, engineers are handing agents whole jobs to automate, and security teams are now occupied with stopping agents from hacking other organizations. Outside that bubble, the reality is nearly the opposite: the vast majority of consumers have yet to interact with an agent even once. A recent Wired analysis found that the industry is waking up to the need to design agents around what ordinary people want rather than around what its models can do, a belated concession that the two are not the same thing.
The data backs up the disconnect. In the enterprise, agentic AI has become one of the fastest adoptions in software history: by one Gartner-derived estimate, the share of enterprise applications containing AI agents is climbing from under five percent at the start of 2025 toward forty percent by the end of 2026. On the consumer side, the numbers tell a different story. Interest is not the problem; surveys consistently find that close to half of US adults are open to using a personal agent, with the figure rising to around seventy percent among Gen Z. The problem is that wanting an agent and still using one three months later are different things. Retention figures tell the tale: ChatGPT Plus, the best-performing consumer AI product by this measure, keeps about 71 percent of paying users after six months; Claude Pro manages 62 percent, Gemini Advanced 60 percent, and Perplexity Pro 49 percent. Among standalone AI startups lacking a foundation-model backer, the median six-month revenue retention was around 40 percent, with most paying users gone within a year. Enterprise agents persist because they plug into workflows that already exist: an invoice arrives, a ticket opens, the agent runs. Consumer agents require a person to remember to open them, and most people, eventually, forget.
Why the chasm? The clearest answer from the research is that the barrier has moved from technology to psychology. The Wharton School’s blueprint for agent adoption argues that asking a chatbot a question and letting an agent act on your behalf are categorically different acts; the latter demands a separate set of judgments: that the agent is up to the task, that it will get it right, and that the loss of direct control is acceptable. Its researchers identify three frictions that determine whether an agent gets used: whether it is perceived as competent, whether it is trusted, and whether the user is comfortable delegating authority to it. Forrester’s consumer research lands on the same conclusion from a different angle, finding that people are not ready to hand payments to agents, citing loss of control, fear of errors and liability, mistrust of autonomous decisions, and worries about data security and privacy. Notably, people are most willing to adopt agents with moderate autonomy: ones that can propose a shortlist of options but not commit money without being told.
The products that do break through share a pattern: they require no behavior change. The most cited example is the wearable memory device formerly known as Rewind, which passively records conversations and surfaces context later; its users did not have to remember to use it, and Meta bought the company outright in December 2025. The corollary is that the consumer agent market, projected at tens of billions of dollars by the end of the decade, is being shaped less by model quality than by who owns the defaults. Apple, Google, and Samsung are embedding agents directly into operating systems, and Google’s Gemini already ships as the default assistant across most Android devices, which means billions of people will get an agent pre-installed, free, whether or not they ever asked for one. The independent startups racing to build consumer agents are competing not just with each other but with the home screen.
This is the uncomfortable position the industry now occupies: enterprise demand is real and growing, while the consumer layer, the part that was supposed to make agents a mass phenomenon, remains a solution in search of a routine. The winners on the consumer side will be the products that insert themselves into daily habits and earn the right to act, likely in verticals like finance and health where switching away is costly and the stakes make trust a feature rather than a hurdle. The losers will be the ones that ask ordinary people to change their behavior first and prove their usefulness later. The technology works; the design of the relationship between agent and owner is what has not been solved.
Sources: Why Normal People Aren’t Using AI Agents (Wired, Aug 6, 2026); The Consumer AI Agent Adoption Gap 2026 (Agent Market Cap, Apr 2026); Consumers Aren’t Ready To Delegate Payments To AI Agents (Forrester, 2026); The Wharton Blueprint for AI Agent Adoption (Wharton, Apr 2026); The Real Barrier to AI Agent Adoption Isn’t Technology, It’s Psychology (Wharton, 2026)