As you might know by now, for the past year I have been building and using a personal AI assistant. It reads my mail, finds documents, watches open commitments, prepares replies and helps keep several businesses and a busy personal life moving.
That description usually leads to the same question: what do you actually allow it to do?
The honest answer is more interesting than either “everything” or “nothing.” I do not think of AI as an employee with a fixed job description. I think of it as working inside three different boundaries.
Work it may prepare
The first is preparation. Here I give it considerable freedom.
It can search through mail, collect the documents needed for a permit application, compare versions of an agreement, summarise a long thread, identify what is still missing and draft the next message. It can watch for a reply after I send something. It can assemble a morning brief from my calendar, inbox and open obligations. Of course, it can also look things up on the internet, find best practices and use reputable sources to add information and insight when necessary.
AI should compress preparation, not replace judgment.
Processing large amounts of data, tapping into many sources, working quickly and relentlessly: this is where AI excels. None of these tasks requires the assistant to make the final decision. They require patience, context and the willingness to do twenty small steps before presenting one useful answer.
Case studyA recent example started with a simple request: find the personal and company documents needed for an application. The assistant searched several mailboxes and drives, separated final documents from questionable versions, created a shared dossier, drafted one message to the accountant for the missing records and another to the permit adviser with the documents already available.
I still reviewed and sent the messages. But instead of spending an evening locating files and reconstructing the request, I spent a few minutes making decisions.
Work it may execute
The second boundary is execution with approval.
My assistant can prepare an email, a calendar change or a website update. It can show me exactly what will happen. But sending, publishing or changing an external system requires a clear approval from me.
That happened while preparing this article. I supplied a new photograph for my website. The assistant cropped it, replaced the old portrait and built a preview. Only after I had checked the result did it publish the change.
I can and will write more about this later, because there is a lot to say about harnesses and the specific architectural measures I use to counter prompt injection and make approval flows genuinely reliable. A hint: do not just give a language model rules it can break, ignore or forget.
What is still mine
The third boundary contains decisions I do not delegate.
The assistant may prepare a payment, but it does not decide that money should move. It may summarise a legal position, but it does not determine which risk I should accept. It may compare candidates, investments or business options, but the decision stays with me. The same is true when a situation involves personal relationships, sensitive family matters or consequences that cannot easily be reversed.
This is not because the system never performs well. It is because performance is not the same as responsibility.
Responsibility is another topic worth exploring in more depth. It touches the ethical question and the genuinely irreplaceable role of the human in the loop, something we cannot outsource to AI no matter how smart it becomes. But that is for another article. I want to keep this one practical.
The four questions you should use to decide when something is a task for AI
The boundary is not fixed by the type of task. I use four questions to decide where a task belongs:
- How costly would an error be? A mistake in a file search is inconvenient. A mistake in a payment or legal position can be consequential.
- How much of my time does the task consume? The more searching, sorting and chasing involved, the more useful delegation becomes.
- How much does my personal judgment add? Some work mostly needs persistence. Other work depends on taste, context, experience or responsibility that is specifically mine.
- How much do I enjoy doing it? Efficiency is not the only objective. If I enjoy the work, I may want AI beside me rather than in my place.
Finding and organising documents scores low on the cost of a reversible error, high on time saved, low on the value of my personal skill and very low on enjoyment. I delegate it broadly.
Publishing a website change is different. The work is reversible, but personal taste matters. The assistant prepares the change and I approve the result.
A legal or financial decision scores high on both the cost of error and the value of judgment. Here the assistant gathers evidence, identifies gaps and lays out options. I make the decision.
Writing sits somewhere else again. It takes time, but I enjoy shaping an argument and the voice is mine. AI is useful as an editor, researcher and sparring partner. I would not want it to remove me from the part I care about.
This fourth question is easy to overlook. A perfectly automated life is not necessarily a better one. Some friction is waste. Some friction is the work itself.
The distinction also matters because the assistant gets things wrong. It has selected the wrong version of a document, misunderstood a date and occasionally produced a confident answer from incomplete information. When that happens, the useful response is not to stop using AI. It is to change the process so the same kind of mistake becomes harder next time.
That has made working with the assistant feel less like using a clever chatbot and more like building a working relationship. I learn what context it needs. It accumulates corrections and operating rules. The boundary moves slowly, based on evidence rather than enthusiasm.
The popular conversation about AI is obsessed with autonomy. I think that misses the point. I do not need a machine that independently runs my life. I need one that removes the searching, sorting, chasing and reconstructing that sits in front of good decisions.
The goal is not to take me out of the loop. It is to make the loop smaller.
That is what I actually let my AI assistant do: prepare broadly, act within clear boundaries and leave responsibility, and the fun, where they belong.