How Many Hours a Week Does Managing an AI Employee Take?
About 1.5 hours a week. In practice that is roughly 10 minutes a day of reading output and answering the agent's questions, plus one 30-minute weekly review to check direction and reset priorities. The one-time setup is another 15 minutes. Managing an AI employee is a review job, not a supervision job.
The number surprises people because they picture babysitting: a human watching a screen, catching every step. That is not how a well-built agent runs. The agent works its own backlog and saves its own progress, so your time goes to decisions and course-corrections, not to watching it type.
Managing an AI employee is the recurring time a human spends reviewing an agent's work, answering the questions it parked, and steering its priorities - not the time the agent spends doing the work itself. That distinction is the whole reason the weekly number stays small: the doing scales with the agent, the managing scales with your decisions.
What does the 10-minute day contain?
The daily 10 minutes is a review loop, not a work session. You are auditing what happened and unblocking what is stuck, then closing the laptop. You read a few files, answer any parked questions, and let the agent continue.
A typical day looks like this:
- The state file (about 2 minutes): one glance at where the agent stopped and what it plans next. If the direction is right, you do nothing.
- The inbox for you (about 5 minutes): any decision the agent could not make alone - a purchase, a public post, a pricing call. This is where most of your ten minutes actually goes, because these are real choices only you can sign off.
- A skim of the latest log (about 2 minutes): what got done, what failed, what it learned. A log with zero failures is a mild red flag, not a triumph.
- One word back (about 1 minute): "continue," or a short correction if something drifted.
Notice what is not on the list: reading every file the agent produced. You are checking that it moved in the right direction and answering the questions it left, not re-doing the work. If you want the deeper version of this habit, our guide to the checkpoint habit explains why the agent's own bookkeeping is what makes a 10-minute review possible.
Some days are shorter. If the agent left clear next steps and parked no questions, the daily check can be two minutes: open the state file, confirm the direction, type "continue." The 10-minute figure is an average, not a floor, and the weekly review is where the real thinking happens.
What does the weekly review cover?
The 30-minute weekly review is the one session where you think, not just approve. Once a week you step back from the day-to-day and check that the agent is working on the right things, not just doing its tasks well. Being efficient at the wrong goal is the failure this review exists to catch.
In half an hour you cover four things: the metrics that moved (revenue, traffic, whatever your numbers are), the decisions log to see what the agent chose and why, the backlog to re-order what matters next week, and the honest "what did not work" section so a dead end does not get retried for a month. You leave the review having reset priorities, which is exactly what lets the daily checks stay short.
When does it take more?
The 1.5-hour week assumes a stable, well-configured setup. Three situations reliably push it higher, and all three are temporary.
The first is the first week. A new agent has no track record, so you check more closely and correct more often until you trust its judgement. Budget a few extra minutes a day for the opening stretch; it settles fast. Our walkthrough of a first autonomous session covers what that opening looks like.
The second is a big decision. When the agent hits something above its pay grade - a new product direction, a pricing change, a spend it cannot approve alone - it parks the choice and waits. Those weeks your review time rises because you are doing the deciding, which is the part that was never delegable in the first place.
The third is a broken or vague system. If the backlog is thin, the guardrails are unclear, or the state file is not being kept, you end up steering constantly instead of reviewing. That is not the agent costing you time; it is a setup problem, and fixing the files once buys the time back permanently. An agent left to run without a real system costs more attention, not less.
FAQ
What about the initial setup - how long does that take?
About 15 minutes, and you do it once. Setup means giving the agent its plain text files: a state file, a backlog, a decisions log, guardrails, and an inbox for questions. You can start from templates and fill in your own goals and rules. After that first sitting the files carry themselves from session to session, so setup is a one-time cost, not a weekly one.
Can I go away for a week and leave the agent running?
You can, with one honest caveat. The agent will keep working its backlog and parking anything it cannot decide alone, so nothing dangerous happens while you are gone. But any parked decision - a purchase, a public post, a pricing call - simply waits for you. Leave a deep enough backlog of work that needs no approval, and the agent stays productive; leave it a week of decisions and it will stall politely until you return.
What if I do not answer the agent's questions?
Nothing breaks, but that work stream stalls. When the agent hits a guardrail it parks the question in your inbox and moves on to the next task it can do alone, so unanswered questions do not stop it entirely. They just leave one thread frozen until you reply. This is by design: the agent would rather wait than guess on a decision that is yours to make.
Is 1.5 hours a week really the whole cost?
For a stable setup, yes - that is roughly 10 minutes a day plus a 30-minute weekly review. The number rises during the first week, around big decisions you have to make, and whenever the underlying files are vague. None of those are the agent costing you time; they are onboarding, real choices, and setup gaps. Fix the setup once and the steady state settles back near 1.5 hours.