An assistant can be so useful that leaving it becomes work.
It knows how you like a document arranged. It remembers the details you would otherwise repeat. It can find an old decision and connect it with the task in front of you. Over time, the service becomes a convenient place to begin.
That convenience has value. It also raises a question that is easy to postpone while everything is going well: what would happen if you needed to move?
Some of the answer concerns files. Some concerns permissions, ongoing tasks and knowledge that has accumulated inside the relationship. A folder of exported conversations may preserve a history without making another service able to use it. A new assistant may require time to learn what the old one understood. An account may connect differently elsewhere. We should neither pretend these costs disappear nor accept every avoidable difficulty as the natural price of progress. A person can appreciate a service and still need an independent way to reach their records, finish an urgent task and change providers. That independence makes the relationship stronger. Trust that depends on having no practical alternative is a poor foundation for something called personal assistance.
Keep one usable way back
You do not need a duplicate system for every task. Maintaining two assistants, two complete archives and two versions of every workflow could create more work than it prevents.
A more modest aim is to preserve the things that would matter if the service became unavailable at an awkward moment.
For a trip, that includes access to the booking record and the information needed to contact the seller directly. For an important document, it includes a usable copy in a place you control or can reach without the assistant. For an ongoing commitment, it includes knowing what remains active and how to stop or continue it.
The backup should match the consequence. Losing a casual recommendation is different from losing the only accessible record of a time-sensitive arrangement. The person should not have to prepare for every imaginable failure. They should be able to identify the few dependencies that would become expensive or disruptive if the service vanished.
This is an ordinary form of prudence. Its purpose is to let you use the assistant with less anxiety, rather than to turn every act of delegation into a rehearsal for disaster.
The provider can make the work easier. It can keep important records in usable formats, explain what an export includes and make ongoing permissions visible. It can support a clear exit without requiring the person to reconstruct every past instruction from memory. Those features deserve a place in how we judge the service. A product’s value is not exhausted by what it can do on the day you join. It also includes how well it supports a change of mind, a change of circumstances or a decision to leave.
A useful test is whether the service makes the next step clear when it is no longer the service you want to use.
A brief that can survive the errand
The previous chapters have followed many small distinctions. We should now put them to work without turning them into a private legal code.
A delegation brief can fit on one page. It should describe the result you want and the boundaries that matter. The assistant should help clarify it, and the provider’s actual controls must support it. Writing a limit does not create a technical capability the service lacks.
Here is a reusable form. Most ordinary tasks need only a few lines under each heading.
Purpose
What should be different when the task is finished? Include the reason when it changes the choice.
“Find a room for two nights” leaves more open than “Help me attend the Saturday event rested, within walking distance, with lodging and local transport under $350.”
What may proceed
Name the work the assistant can do without returning at every step. This could be reading specified sources, comparing options, preparing a draft or arranging files within a defined place.
Separate preparation from commitments. If you want the assistant to prepare a purchase but wait before paying, say that plainly.
What must be preserved
Identify the requirements that should not be traded away silently. These may concern location, accessibility, privacy, quality or a deadline.
If the requirements cannot all be met, the assistant should bring back the conflict. It should not quietly decide which one you probably did not mean.
Money and other commitments
State the total limit and what it includes. Specify whether the authority is for one transaction or an ongoing arrangement.
Include important nonfinancial boundaries too. A message, disclosure or account change can matter even when it costs nothing.
Information and recipients
Say which sources the assistant may use for the task and which information may be shared with which recipients.
A short instruction can be precise: use my calendar to find a free time, but do not disclose the reasons attached to other appointments. For sensitive information, identify the intended recipient rather than relying on a broad phrase such as “whoever needs it.”
Reasons to return
Name the changes that need your judgment. A different recipient, a higher total, an unavailable requirement or a hard-to-reverse consequence may be enough.
Avoid asking to review every trivial step. Concentrate the interruptions on decisions you would regret having made for you.
Evidence of completion
Specify the useful record: a confirmed booking, a delivered file, an accepted cancellation, or a source-supported answer with important gaps identified.
A request sent to another service is not necessarily the finished result. Ask the assistant to distinguish what it attempted from what it verified.
The ending
Say when the work should stop. For a recurring task, define the period or review point and how you can withdraw permission.
If you become unavailable, existing authority can remain in force within its bounds. Your silence should not become new authority for a step the assistant had asked you to decide.
The brief is a conversation aid. It should help the person and the service agree on the task before effort is spent, and help them recognize a material change later. It cannot anticipate everything. Its value lies in making the important parts clear enough to preserve.
For a low-stakes errand, two sentences may be sufficient. For an expensive or sensitive task, the extra detail is part of deciding whether to delegate at all.
Judge the help by what it gives back
Once the assistant has completed a few tasks, how should you assess it? A list of successful actions is useful, but it leaves out some of the relationship. Think about the effort required to get those actions right. Did you have to repeat instructions? Did you find yourself checking facts that the service presented as settled? Were its questions well chosen? When something went wrong, did it help restore the situation?
The comparison should be with a reasonable alternative, not with a fantasy of perfect human performance. You can forget a deadline or choose a poor hotel yourself. An assistant can improve the result even while remaining imperfect. The relevant question is whether it helps under the conditions that matter to you. Keep the assessment specific. A service may be excellent at organizing research and unreliable at completing transactions. It may draft well while needing close review of names and recipients. The response can be to delegate the useful part and retain the consequential step, rather than accept or reject the entire product as a single object.
This creates a more practical form of trust. You know where the service has earned room to act and where it has not.
Occasional checks can help maintain that judgment. For a recurring task, inspect a result at a sensible interval or after something important changes. A new tool, a new type of recipient or a change in the task can make an old pattern less reliable. The check should have a purpose; constant surveillance would consume much of the benefit.
There should also be room to notice when the service is doing too much. An assistant can generate a growing supply of suggestions, plans and projects. Some may be useful. Others may turn your life into a queue that you never chose.
A finished task is not automatically a reason to invent another one. The user’s goals should determine what deserves attention. If the assistant has saved an hour, it need not fill the hour with activities it can measure. Rest, conversation and unplanned time do not become wasted merely because they leave no completion record.
This is part of what it means for a service to work for you. It should help with the purposes you recognize, and make a suggestion without treating your acceptance as the only successful outcome. A good assistant should be able to hear “That is enough.”
The choices a settings page cannot provide
A personal brief can do useful work. It cannot settle every condition of delegation. You cannot create a reliable export feature by requesting one. You cannot make two services compatible by explaining how much you would like to move between them. You cannot independently establish every commercial influence on a search by looking at the final answer. These problems require action by providers and, in some cases, common rules or public institutions.
The 2026 NIST AI Agent Standards Initiative describes work on voluntary standards, interoperable protocols, identity, authentication and security evaluation.[1] It represents an effort to address shared foundations. It does not establish that every consumer assistant can already exchange data safely, or that any particular product is certified. The direction is important because independence depends partly on cooperation among systems. If services use incompatible ways to express authority, identify an agent or transfer useful records, the consumer may have to rebuild the same arrangements repeatedly. Better shared methods could reduce that burden.
But compatibility alone is not a complete answer. A record may include information about other people. A transferred permission may be inappropriate in a new setting. A receiving service may use data differently. Moving work safely requires attention to what is transferred, for which purpose and under whose authority. An exit route should preserve meaningful choice without becoming a route for uncontrolled disclosure.
There are tradeoffs here that should be discussed openly. Stronger identity checks can help a website recognize authorized activity, while requiring an agent to reveal more than the task needs can create new concerns. A common standard can make entry easier for new providers, while a costly or overly narrow certification process could make entry harder. The details determine whether an institution widens people’s options. We should judge these arrangements by their effects rather than their names.
What would a more useful market for assistance look like? Providers would explain the scope of their searches and material commercial relationships. People would have access to understandable records of consequential actions. Complaints would reach someone able to examine the evidence. Revocation would work as described. Moving important work would be practical enough that remaining with a provider was a choice.
These are proposed standards for a fair relationship. They do not require every service to be identical. A retailer’s assistant can differ from a general personal assistant. A specialist tool can be valuable within a narrow task. What matters is that the boundary is clear and that the product’s claims fit the role it actually plays.
Competition is most useful to the consumer when alternatives can be used, not merely when their names can be listed.
The public interest also includes people who never chose the assistant. They may receive its messages, appear in the information it processes or bear the effects of its actions. A service cannot define responsible behavior solely by the wishes of the account holder. The boundaries we have examined protect the conditions in which many people can use assistance without making one person’s convenience another person’s imposed burden.
That is a reason to build better forms of delegation, not a reason to retreat from the possibility of help.
The controls you keep
In Virtue of the Gods, I wrote: “By control, one refers to the noun, not the verb.”[2] The sentence distinguishes the means of influence from the effort to personally direct everything. It returns here with a practical demand. What controls does the person actually have, and will those controls remain usable when they are needed?
A person can keep meaningful authority while letting much of the work proceed. They can set the purpose, reserve a difficult choice, inspect a consequential result and end a task. They can also decide that a particular form of delegation asks too much of them or gives too little back. The provider should make those decisions practical. A control that exists only in promotional language gives the user little to hold. A real one changes what can happen.
This has been the question beneath the title. The assistant has a company behind it, and that company has its own powers, duties and incentives. Other businesses also influence what the assistant can see and do. The user’s task passes through that arrangement. Calling the software personal cannot make the arrangement vanish.
It can, however, make the standard clearer. A service that asks to know a person and act on their behalf should preserve that person’s ability to direct, question and leave the relationship. It should make important interests visible. It should carry useful work without claiming authority over the whole life around it.
We do not need to solve every argument about the future of AI before asking for these things. They are available questions now, at the level of an ordinary task: what are you trying to do for me; what can you reach; what have I permitted; what happened; and what can I do if this no longer works? The answers will change as products change. The questions can remain useful.
Go back to the table from the beginning. The assistant found one that seats six and fits the place where you store it. It preserved the budget, checked the delivery and kept a usable receipt. It did the work well enough that the work could leave your attention.
On Friday, your sister arrives. Dinner begins.
Six plates.
A chair pulled close.
The hallway is clear.
Someone begins a story,
and you are there
for all of it.
Sources
[1] NIST, “AI Agent Standards Initiative,” created February 17, 2026; updated August 14, 2026. The stated work concerns voluntary standards, protocols, identity and security research. It is not evidence of universal working portability or a consumer-product certification.
[2] Kevin L. Michel, Virtue of the Gods (2020), “Control.” The quoted complete sentence was checked against the author’s supplied text. Its use here concerns practical means of influence; the earlier work’s metaphysical assertions are not offered as evidence about AI.
The closing verse is original to this book.