What is a business-agent fleet — and why bring your own AI?

Everyone has a clever assistant now. Very few people have a business that runs better because of it. The difference is not the model — it is whether the work is held by something that knows the job and remembers your records.

The fleet vs a single assistant

A single assistant starts every conversation from nothing. You paste context, it produces something plausible, and the next day you paste it again. Nothing accumulates, and nothing is accountable.

A fleet is the opposite arrangement: agents that already know the shape of a job, working from the records you connected once. The books agent knows what a close is. The marketing agent knows your voice and last month’s numbers. You are not briefing them from scratch — you are asking them to carry on.

The departments

Four agents cover most of what a small company actually spends its week on:

• Books — reconciles your books to the penny, flags what doesn’t tie, files the month.
• Marketing — drafts posts and email in your voice, schedules them, waits for review.
• Product — keeps the roadmap honest: what shipped, what customers asked for, what is next.
• Projects — tracks what is moving, what is stuck, and what needs you today.

They roll up rather than sprawl: two ventures stay two separate books that report into one number, so you can look at either alone or both together without mixing them.

Why bring your own AI

You already pay for a good model. Paying a second time for someone else’s wrapper around it is a bad trade — and it locks your work to whichever model they resold you.

• No second model bill and no markup on tokens.
• You keep the assistant your team already knows.
• Change provider whenever you like; the work stays where it is.
• When a better model ships, you get it the day your provider does.

Model-agnostic is not a feature. It is the only honest way to sell a stack that sits behind someone else’s AI.

Your data stays yours

The agents are shared. The data is not. Your records live isolated on your own node, no other customer’s agents can read them, and nothing you connect is used to train a model. Access is granted per person and per area, checked on the server on every request, and fails closed — no grant means no data is sent and no screen is drawn.

And a person stays in front of the door. Agents draft, calculate and prepare; publishing, sending and spending wait for a human. Where an agent is unsure, it stops and asks.