Lately I spend remarkably little time writing code myself. I ask an AI to implement something, inspect what comes back, and have it fix whatever I don't like. Meanwhile, I wander off to check on another task. The time my hands spend on the keyboard has shrunk, while the time I spend deciding what to ask for and tracking how everything is going keeps quietly growing.
Before I knew it, I had become, more or less, middle management for a team of AIs. My reports never complain and need no overtime pay. So far, so good. But the moment I thought that, an unwelcome question followed: won't AI be doing this managing, too, before long?
AI takes the dispatcher's seat
Handing work to several AIs, sorting out order and dependencies, and pulling the results together is called orchestration. Like a conductor leading an orchestra, one AI keeps the others in line.
Picture a coordinating AI that tells a research AI to gather information. It passes the material to an implementing AI, and once the implementation comes back, has another AI review it. Problems get sent back for fixes, and only where human judgment is needed does it come to a person, a little sheepishly, to ask. With a setup like that, people no longer have to make the rounds instructing each AI one by one.
Some of what I do now, "that one's done, let me check it," "next, I'll hand off this," could surely be delegated as well. Of course, deciding what counts as success and verifying the results would remain. Even so, "move to the side that uses AI and the managing stays with humans" looks, frankly, like wishful thinking.
What's more, there's no reason the AI's charges should be limited to other AIs. If it can grasp the state of human work too, it wouldn't be strange for it to take on the team's project management.
What "how's it going?" is really for
At work, exchanges meant to check on status come up all the time. "How's it going?" "Stuck on anything?" "When do you think it'll be done?" Hardly anyone enjoys hearing them. The person asked stops to explain; the person asking tidies up the answer and passes it on. Someone builds slides for the weekly meeting, and a boss uses them to report further up. A grand game of telephone.
Part of why this game is needed is that information about the work is scattered. The code is on GitHub, the discussions are in Slack, the specs are in documents, the schedule is on a calendar. And to top it off, some of it lives only in someone's head.
Gathering all that and reshaping it into a picture of the work as a whole is, in effect, one of a manager's jobs. If AI could collect the necessary information and sort out how pieces of work relate, people would surely spend less time going around asking.
To talk about that, though, we should first separate the AI that supports an individual's work from the AI that looks across the whole organization.
The AI at hand, and the AI in the clouds
On supported Macs, Apple Intelligence's on-device model is already running. AI being built into the OS has, in other words, already begun.
What I imagine beyond that is an AI sitting right next to a person's work. Which document is open, which code is being changed, what the person is trying to do. Within what it's been permitted to see, it grasps that context and lends a hand. If I didn't have to hand over files and explain the background from scratch every time, the effort it takes to ask an AI for help would drop sharply.
That said, my Mac knowing my situation is a different thing from understanding the whole company's. Seeing across a team takes a system that handles shared code, documents, conversations, and schedules. That part could plausibly be handled by an AI in the cloud.
Say it connects A's "waiting on an answer about the spec" with C's "that answer needs confirmation from the customer." Suddenly the report "work is blocked" turns into "once C checks with the customer, it moves." No more tragedies where both people dutifully wait until someone notices in a meeting room.
Not that local and cloud split the roles neatly. A device may call a cloud model, and a company may run models in its own shared environment. What matters is less where the AI runs than what information it can see and what it has been entrusted with. If the right information flows between the AI at hand and the AI in the clouds, individual work and the team's progress get tied together.