On Thursday afternoon, a table arrives at your door.
Your sister is coming on Friday. There will be six people for dinner. You remembered this on Tuesday, between two things you could not stop doing, and told your assistant to find a folding table that would arrive in time. You gave it a budget. It found an option, checked delivery, brought you the total, and asked you to approve the purchase. You did. The whole exchange took less time than a long voice message.
Now the table is here. It seats six. It was within the limit. Everything the assistant said about it was true.
You carry it inside and discover where the trouble begins. The table fits the room when open. When folded, it will not fit into the space where you need to store it. You had looked at the picture, seen the word folding, and pictured a corner of your apartment. The assistant had found a product. You had imagined a life in which the product would stay out of the way.
Dinner will be lovely. On Saturday, there will be a table in the hall.
This is a small failure. Nobody has stolen your identity or emptied your account. The seller has supplied what was ordered. The assistant followed the instructions you gave. You even approved it. Yet the errand has returned wearing a different outfit. It is now a return label, a box you threw away, and a trip you will need to make.
The attractive part of having an assistant is that it can take a piece of life off your mind. A capable one might have asked about folded dimensions, or found them in a previous conversation, or noticed that the apartment had little storage. It could have solved the problem better than you would have solved it yourself. That possibility is worth taking seriously.
So is the table in the hall.
Between those two outcomes lies the question that will follow us through this book. What must an assistant understand, and what power must you retain, for work done on your behalf to serve the life you are trying to live?
The work inside a small request
A surprising amount of work hides inside the phrase “Could you handle that?” Take the table. Someone has to turn a loose need into a set of requirements. That person has to search, decide which descriptions to trust, compare prices, check delivery, enter information, keep the receipt, and notice if the package fails to arrive. Most of this is too small to deserve a meeting. Together, it can consume an evening.
The visible purchase is only one part of the burden. You also have to remember to begin. An errand that takes ten minutes may occupy a small room in your head for three days. You can be doing something else while still carrying the knowledge that it remains undone. The relief when someone dependable takes it over is not adequately described by a stopwatch. This helps explain the appeal of software that can act. A system that gives you an answer may reduce one piece of the work. A system that carries a task through can remove several pieces, including the need to keep returning to it.
The difference is becoming part of the products people are offered. In September 2026, Meta introduced Muse as a personal agent that could work across connected services, take on tasks, and return for approval when needed. On September 29, OpenAI announced dots, always-on agents with their own cloud computers and access to apps the user chooses to connect. These are the companies’ descriptions of their products, including their proposed benefits and limits.[1][2]
The practical promise is easy to understand. You could go to sleep with something unresolved and wake to useful progress. The machine can keep working through the dull middle of a task, where human enthusiasm often becomes another open browser tab.
We should not sneer at that promise. A person with abundant time, good health, reliable internet and confidence with forms may call an administrative task trivial. Someone else may be locked out by exactly that task. If help can turn a confusing process into a completed application or a corrected bill, convenience becomes a form of access.
Even here, however, the description of success needs care. Completing a form incorrectly can be worse than leaving it unsubmitted. A bill can be lower because a useful service was removed. An appointment can be booked at a time that is technically free and practically impossible.
“Done” is a claim about a goal. Before accepting it, we need to know which goal the system was working toward.
Consider the table request again. The spoken task was to buy an object. The purpose was to host people without creating a lasting problem. A sensible assistant should be able to move between those levels. It should also recognize when it does not have enough information to do so.
There is no perfect instruction that contains a whole life. You will not describe the width of every doorway each time you ask for help. Nor should you have to turn an ordinary request into a contract that anticipates every possible misunderstanding. People hire help partly because they expect the helper to supply some judgment.
With AI, that expectation meets software whose behavior is shaped by more than your latest sentence. The company chooses how the service works. It decides which tools are available and which actions are forbidden. Websites decide what the service can reach. Other businesses supply information that may be incomplete or out of date. Your request enters this arrangement; it does not create the arrangement from scratch.
The assistant may speak to you as if there are only two of you in the conversation. There can be a great many parties inside the errand.
The moment an answer starts to move
Suppose you ask for a restaurant suggestion. If the assistant gives you a name, there is still a gap between the answer and the evening. You can check the menu, call the restaurant, ask a friend, or forget the whole idea. The answer can influence you, but you remain the person who takes the next step.
Now give the assistant permission to book. A suggestion acquires a time, a head count and perhaps a cancellation charge. If it can also arrange transport, invite guests and pay a deposit, the original choice begins to spread. Several systems now contain a record of what you intend to do.
That spread is the defining feature of delegation. A decision in one place starts to change the world in others.
The word agent can make this sound more mysterious than it is. Here it means software able to take steps toward a goal, often by using other tools or services. Some systems only suggest steps. Others can prepare them for review. Others can carry out certain actions under permissions already given. The name on the product is less useful than knowing which of these powers it actually has.
An agent may be impressive at a difficult part of a task and weak at an ordinary part. It might compare pages of technical specifications yet overlook which day you arrive. The complex work can make the result feel thoroughly checked. That feeling deserves examination. Difficulty in one dimension does not certify accuracy in another.
There is also a difference between the task the assistant is trying to do and the authority it has been given. “Find a suitable hotel” gives it a goal. “Book any room under this total, provided I can cancel until Wednesday” adds permission and a boundary. “Do whatever is best” supplies very little clarity while appearing to supply a great deal of freedom. The most dangerous ambiguity may be the one both parties find convenient. You want to stop thinking about the trip. The service wants to present itself as capable. A broad instruction lets the conversation move quickly. Only later does someone discover that “best” concealed a disagreement about cost, comfort or risk.
Yet asking a question about every small detail is no solution. An assistant that turns one errand into twenty decisions has merely reorganized your work. You have become the supervisor of a particularly needy intern, except that the intern is installed on your phone and can reach you during dinner.
Good delegation requires judgment about what to ask. A missing storage dimension could justify a question before buying furniture. The choice between two equivalent confirmation-email formats probably does not. The aim is to save the person’s attention while preserving it for choices that matter to that person.
This is harder than maximizing the number of completed tasks. A task counter can register an achievement while a human acquires a new inconvenience. It can also miss an achievement when the right result is to stop: the hotel is unsuitable, the claimed discount is unconvincing, or the purchase can wait. An assistant that sometimes says “There is no good reason to buy this today” may be doing excellent work. Whether its business has room to value that outcome is a question we will return to.
What it costs to keep watch
Imagine that the table had been perfect. It fits, folds neatly, arrives on time and costs less than the one you would have chosen. You are pleased enough to delegate the next errand. After several successes, you spend less time checking. This is not necessarily laziness. It is how the service becomes useful.
If every task must be repeated by the user, the assistant’s apparent productivity is misleading. It has generated output while leaving the cost of establishing its reliability with someone else. Think of a small piece of arithmetic. A task takes you thirty minutes. With assistance, giving instructions takes three minutes and reviewing the result takes four. You gain twenty-three minutes, assuming the result needs no repair. If checking the assistant requires twenty-five minutes, most of the gain disappears. If repair takes an hour, the transaction in time has gone the other way.
These are invented amounts, but the relationship matters. Speed of generation is only one contribution to the final result. Time spent supervising, correcting and recovering belongs in the account too.
The burden can also fall unevenly. A confident user knows which part of a result is worth inspecting. A newcomer may have to check everything, or may not know that anything needs checking. Telling both people to “use your judgment” does not give them equal protection.
This is one reason I am wary of any promise that offers effortless assistance and then places the whole cost of error on the person who accepted the offer. A warning has a role. It cannot do the work of a well-designed system.
Nor does trust have to arrive all at once. You can trust an assistant to organize reading material before trusting it to contact people. You can allow it to prepare a purchase while reserving the decision to pay. You can authorize a narrow recurring action without giving permission for every related action it later invents.
The difficult part is knowing whether those distinctions exist in the service itself. A polite exchange in which the assistant agrees to be careful is reassuring. A restriction that the system actually enforces gives the promise a different footing.
We will examine these controls in more detail later. For now, keep the distinction in view. You can tell a person not to spend more than a hundred dollars. You can also give them a payment method that cannot spend more than a hundred dollars. The instruction and the limit do different jobs.
Technical limits cannot settle the whole question. A purchase can stay within budget and still be a bad purchase. A message can go to an approved recipient and still say the wrong thing. Controls around the act protect certain boundaries; they do not supply all the judgment inside them.
This is why useful oversight must begin before the final click. The assistant needs a workable account of what matters. The user needs enough visibility to correct that account. The provider needs to make the limits real. None of these jobs can be completed by decorating the checkout page with the word control.
The uncomfortable case for letting go
It would be comforting to conclude that humans should simply retain every important decision. We could let machines do the paperwork and keep judgment for ourselves.
For any particular task, we need to ask what the human decision adds.
A 2026 study by Kehang Zhu and colleagues placed 243 participants in a three-person trading game with different forms of AI help. Participants preferred a mode that offered advice over one that could act as a delegate. Groups with access to the delegate had the highest average surplus, although its improvement over a historical human comparison group did not survive the study’s adjustment for multiple tests. The experiment was short and structured, with other limits including timing differences and participants’ lack of awareness of other players’ AI help.[3] The experiment leaves an awkward possibility worth testing further: people may prefer forms of help that do not produce the highest returns. It gives us a reason to investigate how control affects outcomes, while leaving the answer open beyond this setting.
Before deciding that the feeling is foolish, however, ask what the measurement leaves outside the frame. A person may care about how a bargain is reached, the relationship it creates, or the opportunity to learn. A result that looks efficient in a game may have costs in a lasting relationship. Those possibilities also need evidence; they are not a license to dismiss an inconvenient result.
We are left with a more demanding task. We must decide which kind of human involvement improves which kind of outcome. At times, involvement adds knowledge that the system lacks. At others, it interrupts work the system can perform better. Sometimes a person should review the objective at the beginning and the result at the end, rather than interfere with every step between them. Sometimes the stakes justify direct supervision throughout.
The right arrangement depends on what can go wrong, how soon the error becomes visible, and what it would take to put things right.
Compare two kinds of mistake. A shopping list includes an extra packet of rice. A private message is sent to the wrong person. You may be able to return the rice. The recipient cannot be made to unread the message. A low price does not make a disclosure reversible.
Now consider two kinds of learning. You want an assistant to reconcile a routine set of receipts. You also want to understand how to negotiate your first employment contract. In the first case, handing off the work may serve the purpose. In the second, the chance to build your own judgment may be part of the purpose.
An assistant should be able to help you become more capable, even when that means doing less for you today. The service that always reaches for maximum automation can miss this form of success.
None of this requires treating the machine as a child, a friend or an employee with feelings. It requires being clear about the relationship you are entering. You are allowing a service, built and operated by other people, to make certain things happen under your name or for your benefit. Its fluent voice should not make the arrangement harder to see.
Enough information to remain the author
When someone acts on your behalf, how much should you need to know about what happened?
Everything is an impossible answer. You do not need a record of every calculation, every page fragment or every discarded sentence. A mountain of detail can conceal a consequential choice as effectively as silence. Nothing is an unsatisfactory answer. “Trust me” becomes especially thin after a mistake.
Between those extremes is information that lets you remain the author of the task. You should be able to recognize the goal the assistant pursued. You should know about a material compromise it made. If the result rests on uncertain information, the uncertainty should reach you before the point at which it matters.
For the table, that might mean a short note: delivery is confirmed, the total is within budget, but storage size has not been checked. It would take a few seconds to read. Those seconds could save Saturday.
For a hotel, the important compromise might be a nonrefundable rate. For a subscription, it might be that cancellation removes stored work. For an email, it might be a new recipient the assistant inferred from a first name. The same generic approval message will not serve all of them.
A well-chosen question can preserve your freedom without consuming your evening. There are limits to what you alone can establish. You cannot inspect every commercial agreement behind a service. You may have no way to know which products were excluded before the search began. You cannot make a provider’s safety system reliable by asking it to be. These are responsibilities for the businesses that design the arrangement, and questions for the institutions that oversee them.
This book will follow those responsibilities beyond the conversation window. We will look at how an assistant is funded, how an option reaches you, and what happens when a website refuses to deal with the helper you chose. We will separate the permission to use information from the uses made of it. Later, we will follow a purchase through the records it leaves and the remedies available when it goes wrong.
The purpose is to make assistance worth having. A person should not need to become an expert in every system that helps them. That is an unreasonable price for an easier day. But they do need ways to decide what to trust, ways to set boundaries that hold, and somewhere useful to turn when a promise fails.
Go back to Thursday. The package arrives. You open it. The table fits its corner, as well as the room. There is nothing left to resolve before your sister comes.
You put the receipt away. Then you do something the assistant was never meant to choose for you.
Sources
[1] Meta, “Introducing Muse: The World’s First Personal AI Agent Built for Everyone,” September 8, 2026. Company announcement, used for the described task-taking design and approval structure, not independently established performance.
[2] OpenAI, “Introducing dots,” September 29, 2026. Company announcement; rollout and capabilities depend on the product’s actual availability and permissions.
[3] Kehang Zhu, Nithum Thain, Vivian Tsai, James Wexler and Crystal Qian, “Choose Your Agent: Tradeoffs in Adopting AI Advisors, Coaches, and Delegates in Multi-Party Negotiation,” arXiv:2602.12089v3, June 27, 2026. Preprint; see Table 1, section 4.1 and section 6. The group-level Delegate comparison had p=.033 before adjustment and p=.10 after Holm–Bonferroni correction. Its measure concerns the specified game, not general consumer welfare.