Adapted from the Hebrew guide "How to work with a personal AI assistant without giving up control" on AI NEWS IL (12 September 2026): https://ainewsil.co.il/articles/working-with-personal-ai-assistant
The labels chatbot, assistant and agent do not mark a uniform boundary between products. Some chat services keep context or memory between conversations; others do not, or let you turn the capability off. Assistants and agents may also act (coordinate, draft, track), depending on the product and the permissions you grant. The capabilities that make these tools useful are also the risk: accumulated memory is a store of personal data, and the ability to act is a permission that has to be managed.
The category is young and products change fast, so the framework below is built to stay correct when tools are swapped: defined tasks, minimal permissions, approval before irreversible action, and routine human review.
Two lists before you connect anything
The first list is not "what is the assistant good at" but "which tasks do I delegate". A task fits delegation if it repeats, its output can be checked in a minute, and a mistake is reversible. Good candidates: a weekly calendar summary, draft replies that always pass through you, tracking an open topic, collecting background before a meeting.
The second list is what never gets delegated: payments and orders, sensitive messages to family or your boss, any decision that affects another person. The line is not the tool's capability but the kind of mistake possible: the harder an action is to undo and the more it touches someone else, the more it stays in your hands.
Each connection is a separate decision
Mail, calendar, files and contacts are not a bundle. Ask three questions about each connection: which delegated task needs it, what happens if this content leaks, and how do you disconnect it. Information you are obligated to protect (a client's, an employer's) does not enter a personal assistant without the owner's explicit approval. Start with one connection, such as calendar only, work for two weeks, then consider the next.
Some actions always need your yes
Define the list up front: sending a message to another person, purchases and bookings, changing or deleting information, sharing a file. The practical test: before letting the assistant act alone, imagine the worst plausible mistake. If it embarrasses you in front of someone or costs money, it stays behind approval.
Check outputs without spending the day on it
An assistant that knows your context produces more convincing output, so errors are easier to miss. Keep fixed checkpoints per output type. Meeting summary: decisions and owners. Email draft: recipient, tone, numbers. Week plan: dates and conflicts. The warning sign is when you stopped checking because "it is usually right".
Manage memory actively
Once a quarter, review what the assistant stored about you: delete what is stale or sensitive, update preferences. Stale memory is a source of errors, like recommendations based on a preference you changed. Ask before you start, not after: "If I want to leave in a year, how do I take what is mine and delete the rest?"
A week of caution saves a month of cleanup.
This English adaptation was prepared with AI assistance.
Yaniv Shenhav.
Top comments (0)