Why roles, not one big assistant
The client's day jumps between very different jobs: villa operations, bookkeeping, translation agency work, research, and his own admin. A single assistant loaded with all of that context gives vague answers and sees data it does not need.
So the assistant runs as eight profiles on one installation: a general assistant, a villa manager, villa management access, accounting for the villa business, a translation agency profile, a personal assistant, a researcher and a developer. Each has its own instructions and its own tools, and only sees what its job requires.
Running on the client's own server
The assistant runs in Docker on a private server the client rents, not in a shared consumer app. Conversations and business data stay on infrastructure he controls, and the assistant can reach his own systems directly. The accounting profile, for example, writes entries to the company's cashbook and reads balances from it.
Keeping model costs under control
An assistant that is used all day can run up a large bill on pay-per-token pricing. The setup uses two existing ChatGPT subscriptions as the main models, pooled so that when one hits its usage cap the other takes over automatically. Only when both are capped does a request fall back to a cheap pay-per-use model.
We learned one lesson the hard way. A model switch typed inside a chat quietly rewrote one profile to send everything through the pay-per-use route at flagship prices, and it drained the prepaid credit before anyone noticed. The fix was a fixed provider order on every profile, with the subscriptions always first, and a check of capped credentials before anyone assumes the assistant is broken.
How it is used
Each part of the business has an assistant that already knows its context, so a question about a villa, a ledger entry or an agency quote goes to the profile built for it instead of starting from zero. Because each role has narrow access, the client can hand work to it without worrying about what else it can see or change.
The takeaway
Split an AI assistant by job, give each role only the access it needs, and decide the model cost policy up front instead of discovering it on an invoice.