Is a private AI assistant safe?

A private AI assistant can be reasonably safe if three things hold: it keeps your information separate from everyone else's, it cannot send, pay or change an account without your clear yes, and its business model does not depend on selling what it learns. No software is risk-free, so ask any vendor how each is enforced.

Disclosure: Housefly is made by Drifted Marketing, and this page is published by Drifted Marketing. Read it as a maker describing its own product, and check anything that matters to you on a call.

What does "safe" actually mean for an AI assistant?

For a business owner, the honest answer is three separate questions rather than one. Who can see what the assistant learns about me? What can it do without asking me? And who is paying for it, which tells you who it really works for? An assistant can be excellent on one of these and weak on another, so it is worth asking all three.

There is also a question no page can fully answer for you: how much risk is acceptable for your business. We would rather say that plainly than claim any assistant, ours included, is perfectly safe. No software is.

Who can see what the assistant learns about me?

Many assistants are one big brain that serves everyone. That is fine for weather questions. It is not fine for your client list, your prices or the way you like a proposal written.

The safer design keeps each person's information in its own store. Housefly does this: what it learns about you is kept in a sealed memory that belongs to you alone, made of your memory (the facts of your business, your clients and your commitments) and your taste (how you like things written, what you hate, what you always say no to). When Housefly writes to you, it is shown your box and nothing else, not another member's box and not Drifted's own notes about our agency. There is no shared brain between members.

A related question is who the assistant answers at all. With Housefly, only members on the list get a reply. A message from a number it does not know gets no reply, and nothing from it is kept.

What can it do without asking me?

This is where most of the real risk sits. The worst thing an assistant can do is act on a distracted "yeah". So the question to ask any vendor is: which actions run on their own, and which wait for me?

Housefly's line is drawn in one place. Reading your numbers, keeping your lists, preparing a brief and writing a draft happen on their own, because none of those leave your hands. Sending a message or email in your name, paying for anything, or changing a setting or record in one of your accounts always waits for your yes.

The order matters too. Housefly names the consequence first, in one sentence, and only then asks: "Ready: send the revised quote from your email to your client." You see what will happen, from which account and to whom, before you answer. And "not yet" is a full answer; it does not keep asking the same question in new words. The full detail is on nothing leaves without your yes.

Who is paying for the assistant?

When an assistant costs nothing, someone else is paying for what it learns about you. That is not a claim about any particular product; it is a reason to read the business model before you trust the product.

Housefly Concierge is paid for by its members: Solo at $3,000 a quarter or Operator at $6,000 a quarter, paid up front. There are no ads, and your data is not for sale. Because the only people paying are members, the only people it works for are members.

What questions should I ask any private AI assistant vendor?

• Is my information kept separate from other customers', and is it ever used to answer someone else?
• Exactly which actions can it take without my approval?
• When it needs my approval, does it tell me the consequence before it asks?
• If I say no, does it keep asking?
• How is the product paid for? Are there ads, or is data sold?
• Who answers messages from people who are not customers?
• Where is my information stored, and how do I get it removed?
• Is there a person I can talk to when something looks wrong?

Most of these have a direct answer for Housefly on this site. One does not: this site does not yet publish technical detail about where member data is stored or how it is removed. We would rather leave that unsaid than describe it loosely, so ask us on the 30-minute call and get the answer from a person.

What are the risks that remain?

Being straight about this is part of being safe. An assistant that reads your inbox and numbers is, by design, close to sensitive information, so who you let in matters. Drafts can be wrong, which is exactly why sending is a separate yes and why you should read a draft before approving it. And a text thread lives on your phone, so your phone's own security is part of the picture. None of these are reasons to avoid an assistant; they are reasons to choose one whose rules you understand.

How do I decide whether it is safe enough for me?

Ask the questions above, get the answers in plain words, and only go ahead if they hold up. With Housefly that starts with a 30-minute call. There is no payment link on the site, so you can ask anything before anything is paid. Concierge is opening to ten founding members; see founding seats and pricing.

Frequently asked questions