A few years ago I was in the audience at a business presentation when the speaker said something that I have never quite been able to put down. As you move up an organisation, he said, the time over which your decisions have to hold without correction gets longer, and that, more than anything else, is what separates the work of a front line team member to the work of a CEO. The team member is judged on how the shift went. The CEO is judged on whether the company exists in 2035, or if it does, was it the right company to have built.
Simply put, how long can someone independently work before they need to be supported by their manager? And yes, even CEOs have managers.
The idea has a name. It belongs to Elliott Jaques, a Canadian psychoanalyst who spent decades working out why managers and the people they managed kept stepping on each other. Jaques called it the time span of discretion, and he organised it into seven levels (he called them strata). Front-line team members fix problems inside a shift. A unit manager carries one to two quarters of consequence. A general manager carries a couple of years. A CEO 3-5. And a genuine board chair is making decisions whose consequences will not be visible for ten years or more. Jaques used to have much longer time horizons than this, but I feel they have been greatly compressed in recent history due to the rate of change and speed of business.
Where does AI sit on this, and will you be reporting to it?
The sharpness of the boundaries between levels has been argued over for decades, like I said, I think they are shorter. What has held up is the general observation. The thing that changes as you climb up the org chart is not difficulty, it is the timeframe over which your judgment has to remain sound without anyone correcting it. That is a different cognitive task at each layer, not simply a harder version of the one below.
This would be an interesting anecdotal piece of management theory and nothing more, except that an organisation called METR, which evaluates frontier AI systems, has independently arrived at the same idea, but for AI. They call it the time horizon, and they measure it the same way Jaques did, by asking how long an AI agent can work autonomously before its judgment is tested. The frontier in early 2026 sits, at fifty per cent reliability, on tasks that take human professionals roughly two to four hours. That horizon was doubling every seven months for most of the period 2019 to 2025, and the doubling rate has tightened to roughly four months over the last year.
In Jaques's vocabulary, that places the frontier of AI capability today inside the lower half of level 1. A capable intern who needs checking just before lunch to make sure they are not deleting 20 years of historical data.
There are good reasons to be careful with the findings from METR. METR's tasks are overwhelmingly software work, which is structured and verifiable in ways that most executive judgment is not. AI failure modes are not human failure modes. And the doubling trend leans heavily on exponential compute growth that may not continue at the same pace. With those caveats, if the trajectory holds, the frontier crosses into one-day tasks in roughly two years, one-week in three, one-month in four, and one-year in around six. That last one is level III, or mid-level management.
This to me is nuts. By now I feel we mostly know the technology is impressive, but for the first time I am starting to think about where on the actual org chart AI currently sits and where it is moving. Today the frontier models sit below the bottom of the formal management structure. In two years at this rate, it sits at the layer of first-line management. In four, at the layer of unit management. In six, comfortably inside the work of a general manager.
Here is the part that I think is worth sitting with for a moment. Today, the relationship between you and the AI in your stack is unambiguous. You direct it. It executes within the small horizon it can hold. You are above it on the model, by a comfortable margin, and the question of who works for whom does not really come up.
That comfort has a shelf life, and the shelf life is shorter than most career planning currently assumes. As the AI frontier moves up the levels, it does not just take work off your plate, it takes layers off the org chart from the bottom upwards. The persistent intelligence with the longer working memory and the cleaner audit trail starts to occupy the layer you used to manage from. And once an AI agent is genuinely operating at level II or III, the people sitting at level I or II are no longer above it on the model. They are below it. The line of reporting.
None of this is fixed. Jaques's framework, in the version I find most useful, treats time-horizon thinking as a muscle. It is trainable, and it is profoundly under-trained in most senior executives, because the operating environment of modern corporate life rewards reactivity and quarterly cadence in ways that atrophy the longer-horizon thinking.
So what can you do to stop AI from becoming your boss? The way out is the way it has always been. Move up the levels, and fast. Train the muscle. Make decisions that look wrong for eight months and right in the third year, and stay in the role long enough to be the person who was right.
Sources and further reading: Elliott Jaques, Requisite Organization (1996) and Executive Leadership (1991, with Stephen Clement); METR, "Measuring AI Ability to Complete Long Tasks" (2025) and "Time Horizon 1.1" (2026).