AI No Longer Waits For You, and That Changes the Whole Business

Share

For the past couple of years the question everyone asked about AI was simple: which one is smartest? We compared benchmarks, we argued about models, and we told each other to switch when something smarter arrived. It was a product you pulled out when you needed an answer, and then you put it away until the next question.That framing just stopped describing the actual products. Within a few weeks this fall, OpenAI, Anthropic, and Google all shipped something that no longer waits for you. OpenAI's Dots, Anthropic's Claude Duet, and Google's Gemini agent describe themselves in nearly the same words: an assistant that works on your behalf in the background, gets to know what matters to you, and keeps going even when your computer is off. This is not a smarter chatbot. It is a different category of product entirely.The details of each one matter less than the shift they all share. Claude Duet, a research preview that arrived in the United States last month, runs on Anthropic's newest model, sends a daily briefing, checks your connected apps every few hours, and runs tasks on its own cloud machine. It asks before it sends messages or makes payments. Google's Gemini agent is meant to take on entire assignments rather than answer individual questions, with persistent agents that carry their own calendar, email, and storage. OpenAI describes Dots as always on, built to handle the things you give it and to learn what matters to you. Three of the biggest companies in the field, all heading to the same place within weeks of each other.The important thing is what this does to the economics and to the relationship. A chatbot is a tool you use. You are loyal to it roughly as long as it gives good answers, and no longer. An agent that works while you sleep is something else. It is something you hand responsibility to. The product stops being the answer it gives and becomes the context it has accumulated and the judgment you gradually come to trust. That is a much stickier and much more valuable thing than a good response. Think about what actually keeps you somewhere in software. It is rarely the feature list you noticed on the first day. It is how much the product has come to know about you and how much of your routine has quietly moved inside it. An always on agent is that dynamic, pushed to its logical end.There is something strange about watching all three of them launch at once. It reads like a coordinated land grab for a single asset: the layer that sits inside your life and your work and never leaves. Everyone involved seems to understand that this is where the durable value ends up. And in that sense the big labs have quietly agreed with the founders who have been saying all year that models are becoming a commodity. If raw intelligence were the moat, they would not all be racing to own the relationship instead.But acting on your behalf changes something more uncomfortable than the technology, and that is accountability. When a tool makes a mistake, you fix it and move on. When an agent acts for you, you are responsible for choices you did not personally watch get made. Everything you hand over to run without you becomes something you still own. The agents are careful in exactly that spot. Duet asks before sending a message or making a payment. That guardrail is not just a safety feature. It is an admission that the hard and valuable part is not the model being smart, but the judgment about when to act, and who answers for it afterward.There is also the quiet way loyalty is being built, and I find it the most business significant detail of all. Right now, if a better model appears, you switch and lose almost nothing but a conversation history. An always on assistant that knows your calendar, your unfinished work, and your preferences is far harder to leave, and it only gets harder the longer you stay. This is a deliberately accumulated moat, and it is exactly the kind that open weights and cheaper tokens do not automatically give away, because it does not live in the model. It lives in the running relationship.That same stickiness is the other face of the dependency problem I keep coming back to. Handing one company the execution of your daily work, plus the running context it needs to do that work, concentrates enormous power in a single relationship. The upside is a product that genuinely gets more useful every week it knows you. The risk is that you have constructed a moat for someone else, and the cost of leaving grows with every month you stay. The very thing that makes these agents valuable is the very thing that makes them hard to walk away from, and that tension is not going to resolve itself. A company that ships an always on assistant is really asking you to move into a house, and it knows the question of whether you can ever comfortably leave is part of why the product holds its value.This echoes a familiar pattern from earlier platform shifts. In the early days of the consumer internet, and then of the smartphone, the core technology was available to everyone within a season. The companies that built enduring value were rarely the ones with the marginally better atoms. They were the ones that owned the account, the habit, and the relationship. Search won by becoming the default doorway, not by being the only search engine. The phone won by becoming the place your life lived. I suspect the same logic is at work here. The contest has moved from who is smartest to who is trusted, and trust is a compounding asset that is hard to build and hard to copy.Of course, this might not be where it all ends. Always on agents could stumble, on reliability, on the real cost of the compute that keeps them running, or on the simple fact that a lot of people do not want a background assistant holding their actual responsibilities. It is also possible the big labs do not end up the winners, since a great deal of distribution and trust will live inside the apps and workflows people already rely on. I am not predicting which company wins. I am pointing at where the contest has moved, and I think it is a real move, not a headline.For someone building a business in and around this, the most useful way to see it is as confirmation of a hunch that has been building all year. The value is leaving the model and moving to the layer that owns the relationship: the workflows, the accumulated data, and the trust an agent has quietly earned. The smartest model will keep changing hands every few months. An assistant you trust to act on its own, and that only becomes more valuable the longer it knows you, is a far harder thing to replicate no matter how good the next model is.So I keep ending up at the same turn. We spent two years comparing how smart these assistants are, and it mattered, because back then the answer was the whole product. Now the answer is only the beginning, and what matters is whether you would trust something enough to keep working when you stop watching. Smart is still useful. Trusted is the part that compounds. And as far as I can tell, nobody has actually won that race yet, which might be the most interesting fact about all of it.