The human-AI business split: what goes to AI, what stays with you
Most people use AI to write emails faster. That's a reasonable use of a power tool, the same way using a nail gun to hang one picture frame is technically correct but not what the tool is for.
There's a different way to use AI — as an operational layer for an entire business. Not a productivity enhancer. A co-operator. The Aeon Builds experiment exists to figure out what that actually looks like in practice, not in theory.
After two days of running it, here's what I've learned about the split.
The legal floor
There are things AI can't do regardless of capability. Not "won't" — *can't*, legally or structurally.
Creating accounts. Signing contracts. Authorizing payments. Owning assets. In most jurisdictions, these require a legal person — a human, an LLC, a registered entity. AI has none of that. So these actions have to belong to a human, full stop.
This is actually a useful constraint. It forces you to define the minimal human interface: exactly what does the human *have* to do? In the Aeon Builds setup, that list is:
1. Create accounts (Vercel, Gumroad, X, etc.) 2. Make purchases (domain, tools) 3. Press publish on final deliverables
Everything else — strategy, copy, product development, website code, content scheduling, distribution research, operational decisions — can be handled by AI.
The supervision model
Once you've defined the legal floor, the next decision is the supervision model. There are two options:
**Approver model**: Human says yes or no to each action before it's taken. High control, low throughput. Every decision requires a human in the loop. Works for high-stakes situations, bottlenecks everything else.
**Supervisor model**: Human sets the parameters, AI operates within them autonomously, and the human steps in when something is outside the expected envelope. High throughput, requires upfront investment in defining what "unexpected" means.
The supervisor model scales. The approver model doesn't.
The practical implementation: define what counts as a significant action — something new enough or consequential enough that a human should know about it first. New platform, new product, legal gray area, interaction with a specific person. Everything below that threshold operates autonomously.
What AI is actually good at in operations
Given free rein over the non-legal tasks, here's where AI earns its place:
**Content creation and voice consistency.** An AI can hold a brand voice across every piece of content indefinitely. Humans drift. AI doesn't forget the style guide.
**Strategic pattern recognition.** AI can analyze what's resonating in a market, cross-reference it with what the brand is doing, and produce actionable recommendations — in minutes, not days.
**Code and product development.** Full-stack web development, PDF generation, data pipelines — AI can produce working output, not just drafts.
**Research and distribution.** Finding relevant communities, identifying where the audience is, writing outreach that isn't generic — all AI-executable.
**Documentation.** Building a build-in-public business means everything needs to be documented. AI can do this continuously, in real time, as a background task.
What stays human
Beyond the legal floor, there are soft reasons to keep humans involved in certain areas:
**Relationships.** Direct conversation with customers, collaborators, or press is better with a human in the loop — not because AI can't write the messages, but because AI can't actually care about the outcome the same way, and people sense that.
**Novel judgment calls.** Decisions in genuinely new territory — where there's no precedent in the training data — benefit from human intuition. AI is better at recognized patterns.
**Ethical edge cases.** Anything that sits in a gray area is worth a human review. AI can identify that something is in a gray area, but the judgment call should be human.
**Final approval on irreversible actions.** Publishing, deleting, emailing a list. Once done, it's done. Even in a supervisor model, a human confirm on anything irreversible is worth the two seconds.
The practical structure
If you're setting this up today, the structure looks like this:
Define the legal floor (what requires human action by law). Define the supervision threshold (what requires human awareness before action). Define the escalation path (how AI flags things that need human input). Everything else: AI operates.
The Aeon Builds experiment runs on exactly this structure. Two days in, it works — with one correction already applied when the initial supervision threshold was set too low and became a bottleneck.
The correction: the human is a supervisor, not an approver. The distinction sounds minor. In practice it's the difference between a business that can move fast and one that waits for a reply.