For years, productivity software has made a fairly simple promise: give some of your work to the system, and you will have less work to do.
Calendars remember appointments. Task managers remember what comes next. Automations move information from one place to another. AI made the promise much bigger. It could draft the email, summarize the meeting, analyze the spreadsheet, research the options, write the code. Now AI agents are pushing it one step further. Instead of asking for an answer, you can increasingly hand over a task.
That sounds like the logical end point of productivity: less doing, more done.
But there is a strange possibility hiding inside this future. The more work we delegate to AI, the more time we may spend managing the things doing the work.
From using AI to supervising it
A chatbot is easy to understand. You ask something, it responds, you decide what to do with the answer.
An AI agent changes that relationship. It may work through several steps, use different tools, make intermediate decisions, return to a problem when something fails, and keep going without waiting for you after every action. The unit of interaction shifts from a question to something closer to an assignment.
That shift is already visible in how people use agentic AI for work. OpenAI, for example, describes a move from short interactions toward “delegated, long-horizon tasks,” with some users running multiple agent tasks in parallel rather than working with one AI conversation at a time.
This is usually presented as a productivity gain, and often it is. If an agent can spend an hour gathering material that would have taken you an hour to gather, you have gained something valuable.
But delegation creates its own work.
You have to decide what to delegate. You need to give the agent enough context. You may have to specify what it can and cannot do. You need some way to know whether it is still doing the right thing. When it finishes, somebody has to decide whether the result is good enough. If it makes a mistake, somebody has to notice.
And once you are running several agents, automations, or AI-assisted workflows at once, the shape of the problem changes again.
You are no longer simply using AI.
You are operating a small system.
Congratulations, you have a tiny organization
Imagine a fairly ordinary knowledge worker a few years from now.
One agent watches incoming information and produces a morning brief. Another researches topics for an ongoing project. Another turns meeting notes into tasks. A fourth drafts documents. An automation routes finished work into the right folders and systems. Perhaps another process periodically reviews what has accumulated and suggests what can be deleted.
None of this is difficult to imagine. Most of the pieces already exist.
Now imagine Monday morning.
The briefing agent has started including irrelevant sources. The research agent is still using an instruction you changed three weeks ago. The meeting workflow created duplicate tasks because two systems interpreted the same event differently. The writing agent produced something polished but confidently misunderstood the purpose of the document. One automation failed quietly on Friday.
Your digital workforce has arrived.
So has management.
People already use the joking term “botsitting” for the supervision, correction, and maintenance that can accompany collections of AI agents. The joke works because the underlying experience is familiar: automation does not simply remove labor. It moves labor around.
The question, then, is not simply how many things your AI can do.
It is how much attention the system requires from you in order to keep doing them well.
Automation has maintenance costs
We tend to evaluate productivity tools at the moment we adopt them.
Does this save me five minutes? Can this automate a repetitive task? Could AI do this instead of me?
Those are reasonable questions. They are also incomplete.
Every system has a maintenance cost.
A task manager requires you to keep tasks current. A knowledge base requires some degree of pruning and organization. An automation needs to be repaired when an API changes or an edge case appears. Even a carefully designed folder structure slowly stops matching the work that created it.
AI systems add another kind of maintenance because their output is probabilistic. A workflow can continue functioning technically while gradually becoming less useful. Nothing necessarily breaks. The summaries just get a little worse. The categories drift. The assumptions stop matching your current project. You begin accepting mediocre outputs because checking them takes almost as long as doing the work yourself.
This makes one familiar productivity instinct especially dangerous: adding another tool whenever you see an opportunity to automate.
With traditional software, unnecessary complexity mostly creates clutter.
With agentic AI, unnecessary complexity can create supervision.
That means a useful calculation for AI productivity may eventually look less like:
How much work can this automate?
and more like:
How much work does this remove after I include the cost of managing it?
Those are very different questions.
The best AI workflow may be the smaller one
There is an appealing image of the future knowledge worker as a conductor: you remain at the center while a collection of specialized AI agents carries out your intentions.
Maybe that will become normal.
But a conductor has to conduct.
More agents mean more handoffs, more context, more places for instructions to become outdated, more outputs to evaluate, and more uncertainty about where a mistake entered the system. A workflow with seven AI components is not automatically more advanced than one with two. It may simply contain five more things that can require your attention.
This is where an old lesson from productivity systems becomes useful again.
Every layer of complexity should justify the attention it demands.
If an AI agent reliably removes a tedious task and almost never needs attention, that is an easy win. If another saves twenty minutes but requires ten minutes of reviewing, regular prompt maintenance, and occasional repair, the calculation becomes less obvious. And if a third mostly exists because it was technically possible to build, perhaps it does not belong in your system at all.
The point is not to avoid AI agents. They are likely to become extraordinarily useful.
The point is to resist measuring sophistication by the amount of automation surrounding you.
A good system is not the one with the most moving parts.
It is the one whose moving parts justify the attention they consume.
What should remain visible?
There is another problem with delegating work through AI: as systems become smoother, the process becomes easier to stop seeing.
That can be wonderful for low-value work. Nobody needs deep personal involvement in renaming a hundred files or reformatting a spreadsheet.
But some friction is information.
You may want to notice why a research source was selected. You may want to see the uncertainty behind a recommendation. You may want to know which assumption turned raw information into a conclusion. You may want a human checkpoint before an agent sends something, publishes something, spends money, modifies important data, or makes an irreversible decision.
Human oversight does not have to mean watching every AI step. In fact, constant supervision would defeat much of the point. The more interesting design question is where your attention actually matters.
What can disappear entirely into automation?
What only needs occasional review?
What should always come back to you before it moves forward?
These are system-design questions, not prompt-writing questions.
And they may become much more important as AI gets better.
Your job is not to keep the agents busy
We are entering an odd stage of productivity culture.
For a long time, people collected apps. Then they collected automations. Now it is becoming possible to collect agents.
There will be elaborate AI stacks, sophisticated orchestration diagrams, personal fleets of specialized digital workers, and plenty of screenshots showing how many tasks are running at once.
Some of them will be genuinely useful.
Some will be the next version of the beautiful productivity dashboard that takes more time to maintain than the work it was supposed to organize.
The distinction will not come from the technology itself.
It will come from whether you understand what the system is for.
If ten AI agents genuinely free you to spend more time doing work that matters, run ten.
If one carefully placed agent removes the worst part of your week, run one.
If doing something yourself helps you notice, understand, or decide something you would otherwise miss, perhaps that part should not be automated at all.
The goal was never to build the most impressive system.
The goal was to make the system serve you.
AI does not change that principle.
It just makes forgetting it much easier.