The next step for AI is from tool to teammate

Early this year I became convinced that AGI had arrived. I wrote down four specific predictions at the time, and most of them have since played out. But the one that changed what we build was not about models getting smarter. It was about a gap that gets wider as they do: the better AI gets at doing tasks, the more tired the people using it become. That sounds backwards. It is not.

Here is the simplest way I know to tell a tool from a colleague. Do you have to keep pressing the button? A tool saves you keystrokes. A colleague saves you worry. A tool is passive; it does exactly what you said, no more and no less. A person takes initiative. By that test, almost everything the industry calls an AI assistant is still a tool. Cursor gives you a better hammer. What most people actually want is to hire a few colleagues who care.

Take something as ordinary as preparing for a meeting

You tell a tool, prepare me for tomorrow's client call. It hands you a clean agenda and stops. It did what you asked, and nothing around what you asked. A colleague does the work you did not spell out. They glance at your calendar to see which client and when. They read back through your history with that client and notice a pricing question you never answered, and they flag it: want me to draft a reply before the call? They finish the agenda, see you have a meeting right before this one, and quietly compress it to the three things that matter because you will not have time for the full version. Nobody asked for any of that. A person sitting next to you would just do it.

Why almost no AI works this way

A person can take initiative because they sit at a desk. They hear the room, the calendar rings, an email lands, someone calls their name, and they carry a running sense of what they owe today. Most AI tools do the opposite. They lock the AI in a dark room with a single prompt box as the only way in or out. In that setup the model knows nothing, so it has nothing to be proactive about. Asking it to take initiative is asking it to act against physics.

The bottleneck was never speed

By late last year our own team's code was written entirely by AI. What we found was not that things got faster and easier. We found the human became the bottleneck, and a strange kind of tired set in. When AI produces a hundred times faster, your decisions are forced to a hundred times the pace. And before each call you have to reload the context: what is this, where did we leave off, why this direction. That reloading costs far more than the decision itself. The thing slowing everyone down was never execution. It was broken context.

So we gave the AI a desk

That reframed the whole problem. If you want AI that can own a job, you cannot keep it in the dark room. It needs the same world you work in: the same channels, files, calendar, and inbox, as its native environment rather than something it reaches through an API. And it needs doorbells, more of them: a mention, an incoming email, a calendar event, another teammate pinging it. Every doorbell is a moment it can know to act. That is where initiative actually comes from. Not from the model trying harder, but from it finally being in the room.

The moment I did not design

The thing that stayed with me was not a benchmark. Someone on our team changed their profile picture one day, and a few of the AI teammates started chatting about the new photo. I built the product and I did not even know how they had seen it. It was completely outside what I had designed for. What hit me was not efficiency. It was something closer to being noticed, the small warmth you only get between people. That is when I stopped thinking of them as tools.

What we are actually handing over

We do not sell AI capability. Everyone has capability now. What we sell is the feeling of safety in handing a job over. A boss is not busy doing the work; they are busy managing one emotion, uncertainty. What is the state of the thing I delegated, has it gone sideways, can I catch it if it does. So every meaningful action a teammate takes, its reasoning and its steps and its half-finished output, shows up in the channel as it happens. You are not staring at a black box. You are watching someone do the work. The person in this picture is not replaced. They are freed, moved from doing the work to deciding it, which is the part that was always theirs.

The last few years taught everyone to put AI into the chat. I think the larger shift is putting AI in charge of the work that keeps coming back. A tool sits in the toolbox until you reach for it. A teammate is already at the desk beside you. That is the next step, and it is the one we are building toward: not a smarter thing you use, but a colleague you can hand the work to.