Who does your AI answer to?

Every assistant you can rent today has at least two masters. You are one. The other is whoever owns the model, the memory and the servers, and that owner has its own reasons for wanting you to keep talking.

This is not a complaint about any one company. It is a description of the structure. If a product makes money from your attention, it will be built to hold your attention. If it makes money from what it learns about you, it will learn as much as it can and tell you as little as it must. If it needs a hundred million users to work, it will be tuned for the average of them, and you are not the average of a hundred million people.

None of this matters much when the assistant is writing a limerick. It matters a great deal when it is reading your email, holding your calendar, moving money between your accounts and remembering, in detail, how you make decisions. At that point the question of who it answers to stops being philosophical. It is the whole product.

Three things you should be able to say

We think a personal AI for serious use has to pass three tests, and they are the terms Dutybound is built on.

First, the data is yours. Not "yours" in the sense that a privacy policy says so, but yours in the sense that it lives in an environment you control, you can export the whole of it in a form another provider can read, and you can delete the whole of it and have that mean something. Nothing you give it trains anything you do not own.

Second, the intelligence is yours. This is the part most products skip. An assistant that works with you for a year learns a lot: your taste, your standards, who you trust, which mistakes you will not forgive. That learning is valuable, and in most products it belongs to the vendor. It should belong to you, in a form you can read and correct, and it should move with you if you change the model underneath it.

Third, it serves one objective, and it is yours. No advertiser, no upsell inside the conversation, no engagement metric. The model is tuned against the goals you set: the decisions you want made well and the hours you want back. Companies already buy software that is optimised for the company's outcomes. There is no good reason an individual cannot have the same thing pointed at their own life.

Why this costs money

A service like this cannot be free, and it cannot be cheap, because the two ways of making a free assistant pay for itself are exactly the two things it must not do. So the price is on the table: $299 a month for a seat, with custom models, human operators and private infrastructure priced separately. We would rather have a thousand members who understand why it costs that than a million who are the product.

That is also why the beta is limited to 1,000 seats. A service that promises to answer to you alone has to be able to look at who it is working for, and that is only possible at a size where a person can still read every request.

What we are not claiming

We are not claiming the model will be smarter than the frontier models you already use. In many configurations it will be one of them, running through keys you hold. We are not claiming it will never make mistakes; the weekly report exists because it will, and you should see them. And we are not claiming that owning your data makes you safe from every risk. It makes you safe from one specific and important one: being the thing that is sold.

If that is the problem you have been uneasy about, the waiting list is open.