Autonomy is earned, not switched on

Our agents do the work. A human approves every action that moves money. Here is how we built that, and why it is the only version of autonomy a real business should accept.

Every owner we talk to asks the same question within five minutes: if an AI runs part of my business, who is in control? It is the right question. Our answer, after running our own brands on agents, is simple. Agents should do almost all of the work. They should not be trusted with money until they have earned it, one action type at a time.

Agents propose, people approve

At Space AI, every agent works through one approval queue. Marketing, creative, support and engineering agents read live store, ad and support data, then file proposals. Nothing that changes the store or the ads happens until someone approves it (changelog).

That queue is not a formality. Each item is approved or rejected by a person accountable for the result. Over time it becomes the most useful dataset we own: a record of what good judgement looks like in our business.

Guardrails before trust

Approval alone is not enough. A tired founder clicking "yes" forty times a day is not oversight. So the system limits what an agent can even propose:

• Price bands. Each product has a floor and a ceiling. Any price change is checked against the band, the unit economics and the live price before it reaches a person.
• Small budget steps. Agents move ad budget one small, capped step at a time toward what is converting, never past the daily spend cap.
• A wallet that cannot overspend. Meta campaigns are funded from a prepaid balance, and a campaign cannot go live unless the balance covers it.
• Proposals checked before they are filed. Account numbers must be ours, links must load on our own domain and dates must be in the future. A proposal with made-up details never reaches anyone.
• Instructions that expire. Even our AI founder's instructions to the other agents lapse after 72 hours unless renewed.

Fewer, better decisions

The hard part was not building agents. It was making approvals worth a person's time. Our first version flooded the queue, and only about one in twenty-five resolved proposals was approved. We fixed that by showing at most the three approvals with the most money at stake, merging contradicting proposals into one card with options, and expiring stale cards after 72 hours.

The goal is a queue a founder can clear over morning chai, not a second inbox.

How autonomy is earned

Once an agent has a track record on one type of action, it can be given more room, but only for a set period and with a spend cap on every action. When the period ends, it goes back to asking. Extra trust is a trial that ends on its own, not a permanent switch.

Every decision also shows the data it was based on, and changes made by hand are logged separately. When something works, we know whether an agent or a person did it.

What this means for your business

If someone offers you a "fully autonomous" AI that runs your money with no one watching, walk away. The version that works looks like this:

1. Agents do the repetitive work: reading data, drafting, preparing every decision.
2. Rules decide what agents are allowed to propose.
3. A person approves anything that moves money, in minutes, not hours.
4. Autonomy grows only where the record earns it.

That is how we run our own brands, and it is how we build for clients. If you want to see where your business sits on that path, find your level.