You wake up. Your AI has already answered your emails, drafted your daily brief, checked the pipeline, flagged five leads ready to move phases, and proposed a new way to deliver your service.
Sounds fantastic. It's also overwhelming, and some of it is subtly wrong. The AI did exactly what you told it, but what you told it needs adjusting for the operational context only you carry in your head. So now there's a pile of work. For you. You spend the day sorting the good from the plausible-but-wrong, and by Friday you've quietly ditched the whole thing and gone back to the slow way. No AI. No pre-drafted anything. Just the same old slog.
Task completion is a vanity metric
That's the hard part of getting AI right, and why we're not yet seeing it show up in the economy. Every demo measures the same wrong thing: forty tasks overnight, a hundred emails drafted, a stack of reports generated. Volume.
But the number that matters isn't how much the machine produced. It's how much of it you accepted, shipped to a client, sent to a customer, put into the work, divided by the hours of your attention it burned to get there.
Reward agents for volume and you get volume: plausible drafts, half-right analyses, confident proposals with a buried error. Every one of those lands in your inbox for review. Optimize for task count and you don't build a workforce. You build a review backlog, with yourself as the slowest station on the line.
So the goal isn't automation. It's a ratio: accepted work produced per hour of human attention.
The question the ratio forces
Here's what took me too long to see. A ratio like that has to be owned by someone. If no one owns it, it defaults to the worst case, where every agent maximizes its own output and dumps the sorting on you. You become the owner by default. You become the drowning owner from the top of this email.
So the ratio doesn't first produce a fleet of AI workers. It produces exactly one hire. An AI Chief of Staff.
Your newest employee. The AI Chief of Staff
Which is fitting, because that's what the role has always been. A great Chief of Staff's entire job is protecting the principal's attention and deciding what actually reaches their desk. So the first AI you commission shouldn't do the work at all. It should sit between you and everything the work produces.
Filter, then memory, then controller
The CoS earns its place in three stages.
It starts as a filter. Day one, its only job is triage: pass through what clears your bar, hold back the rest. You review less, not more.
It becomes a memory. Every time you reject something, saying "this was wrong because you didn't know how we actually operate," that correction has to land somewhere instead of evaporating. It lands in the CoS. The operational context you used to re-explain every morning gets captured once.
It becomes the controller. Once it holds your bar, your context, and your accept/reject history, you stop launching new agents into your inbox. You launch them through the CoS, the one thing that already knows which patterns produce accepted work and which produce review debt.
That's the quiet shift. You're not managing a pile of agents. You're managing one, and it manages the rest. Every agent you commission after that starts at the center of everything the CoS has already learned.
One thing to be clear about: the CoS controls the fleet, not you. It concentrates the review burden so you don't have to hold it. You're still the one who sets the bar. That never moves.
What to watch
Your accept rate over time. If a new agent's output needs heavy editing after two weeks, the problem isn't the model. It's that its context isn't flowing back into the CoS.
Where corrections go. If the same mistake needs correcting twice, you don't have a memory yet. You have a very fast intern.
Your own calendar. The real signal isn't tasks completed. It's whether your mornings are getting quieter.
How to build it
So you might be asking yourself, "Great, but how do I actually build one without an engineer?"
That's next week. I'll walk through two versions: a Starter version you can set up today in Cowork or Codex with a single Markdown file. That works well for a single person and will be all most people need.
But I’ll talk you through an advanced version where the agent builds you a real database, no code required. The file remembers. The database measures. More soon.
Alan