OpenAI has been quietly repositioning ChatGPT away from being a conversational tool toward something it calls an operator - an agent that doesn’t just answer questions but takes actions. Books a reservation. Files a form. Clicks through a checkout. The infrastructure for this has been building since the launch of the Operator product in early 2025, and the feature set has expanded steadily since.
The technology part is largely solved, or close enough. The models can navigate web interfaces, interpret UI elements, and recover from minor errors mid-task with surprising reliability. What OpenAI hasn’t solved - and what rarely gets discussed in coverage focused on benchmark scores - is the trust gap between capable and trusted.
Most people do not want software acting on their behalf in any context where a mistake has real-world consequences. This isn’t irrational. A chatbot giving you a wrong answer about a French verb is recoverable. A chatbot submitting an insurance claim with incorrect details, or purchasing the wrong airline fare class, is not. The asymmetry of risk here is massive, and users feel it even when they can’t articulate it.
OpenAI’s current strategy seems to bet that repeated positive experiences will erode this hesitation over time. That’s plausible - it’s more or less what happened with online banking and mobile payments. But those transitions took over a decade and were driven by institutions with regulatory accountability behind them. OpenAI is a private company asking users to extend the kind of trust that took banks twenty years to earn, on a compressed timeline, in a product category with no established liability norms.
The enterprise side is a different story.
Corporate deployments of agentic AI are growing faster than consumer adoption, precisely because companies can constrain the action space. An agent that only touches internal ticketing systems, only routes to pre-approved vendors, only escalates above a dollar threshold - that’s a manageable risk profile. Enterprises have legal teams and audit trails. Individual users have neither.
The interesting question isn’t whether agentic AI will eventually become mainstream consumer infrastructure. It probably will. The question is what the liability framework looks like when it does - and whether OpenAI, or any of its competitors, is actually prepared to operate inside one. Right now there’s no clear answer, and the companies building these tools seem content to leave that problem for later.