OpenAI has been positioning its Operator product - and the broader push toward agentic AI - as the logical next step after chatbots. You stopped typing queries into Google; now you’ll stop clicking through websites yourself. The agent does it. Sounds convenient. It probably is, in narrow cases.
But the framing of this as a productivity revolution obscures what’s actually being built: a layer of habitual dependency that locks users and businesses into OpenAI’s infrastructure in a way that a chat interface never could.
ChatGPT is easy to leave. You close the tab. A competitor offers a better model, you switch. The friction is basically zero. That’s a problem for OpenAI’s long-term business, and agents are a structural answer to it.
When an AI agent has access to your email, your calendar, your browser sessions, your purchasing history, and your stored preferences - and when it has learned how you like things done over months of use - switching costs become real. Not contractual lock-in, but practical lock-in. The kind that makes enterprise software sticky for decades even when it’s mediocre. OpenAI is chasing that stickiness deliberately.
The Integration Trap
Agentic systems require deep integration to be useful. That’s not a flaw; it’s the point. An agent that can’t touch your tools is just a chatbot with ambition. But every integration is a tendril. Operator connecting to your Outlook, your Stripe account, your internal Notion workspace - each connection raises the cost of switching to Anthropic or Google or whoever comes next.
This is the same playbook Salesforce ran in the early 2000s. Get inside the workflow, not just adjacent to it.
The Reliability Problem Nobody’s Solved
There’s a more immediate issue that tends to get buried under the marketing: current LLM-based agents are not reliably good at multi-step tasks in live environments. They hallucinate intermediate steps. They misread UI states. They occasionally do the wrong thing confidently. For low-stakes tasks - drafting an email, booking a reservation - mistakes are annoying. For anything touching money, contracts, or customer data, they’re serious.
OpenAI knows this. The rollout cadence reflects it. But the public pitch consistently runs ahead of what the technology can actually deliver without supervision, and that gap tends to erode user trust in ways that are hard to walk back.
The agent future might still arrive more or less as described. But it will be shaped more by who controls the infrastructure than by which product is genuinely most useful.